Spaces:
Runtime error
Runtime error
Songwei Ge
commited on
Commit
•
4c022fe
1
Parent(s):
14a857e
demo!
Browse files- app.py +223 -5
- models/attention.py +892 -0
- models/region_diffusion.py +307 -0
- models/unet_2d_blocks.py +1670 -0
- models/unet_2d_condition.py +411 -0
- sample.py +109 -0
- utils/.DS_Store +0 -0
- utils/attention_utils.py +201 -0
- utils/richtext_utils.py +234 -0
app.py
CHANGED
@@ -1,10 +1,228 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import gradio as gr
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
|
4 |
-
|
|
|
5 |
|
6 |
-
def greet(name):
|
7 |
-
return HTML, "Hello " + name + "!!"
|
8 |
|
9 |
-
|
10 |
-
|
|
|
1 |
+
import math
|
2 |
+
import random
|
3 |
+
import os
|
4 |
+
import json
|
5 |
+
import time
|
6 |
+
import argparse
|
7 |
+
import imageio
|
8 |
+
import torch
|
9 |
+
import numpy as np
|
10 |
+
from torchvision import transforms
|
11 |
+
|
12 |
+
from models.region_diffusion import RegionDiffusion
|
13 |
+
from utils.attention_utils import get_token_maps
|
14 |
+
from utils.richtext_utils import seed_everything, parse_json, get_region_diffusion_input,\
|
15 |
+
get_attention_control_input, get_gradient_guidance_input
|
16 |
+
|
17 |
+
|
18 |
import gradio as gr
|
19 |
+
from PIL import Image, ImageOps
|
20 |
+
|
21 |
+
|
22 |
+
help_text = """
|
23 |
+
Instructions placeholder.
|
24 |
+
"""
|
25 |
+
|
26 |
+
|
27 |
+
example_instructions = [
|
28 |
+
"Make it a picasso painting",
|
29 |
+
"as if it were by modigliani",
|
30 |
+
"convert to a bronze statue",
|
31 |
+
"Turn it into an anime.",
|
32 |
+
"have it look like a graphic novel",
|
33 |
+
"make him gain weight",
|
34 |
+
"what would he look like bald?",
|
35 |
+
"Have him smile",
|
36 |
+
"Put him in a cocktail party.",
|
37 |
+
"move him at the beach.",
|
38 |
+
"add dramatic lighting",
|
39 |
+
"Convert to black and white",
|
40 |
+
"What if it were snowing?",
|
41 |
+
"Give him a leather jacket",
|
42 |
+
"Turn him into a cyborg!",
|
43 |
+
"make him wear a beanie",
|
44 |
+
]
|
45 |
+
|
46 |
+
def main():
|
47 |
+
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
48 |
+
model = RegionDiffusion(device)
|
49 |
+
|
50 |
+
def generate(
|
51 |
+
text_input: str,
|
52 |
+
negative_text: str,
|
53 |
+
height: int,
|
54 |
+
width: int,
|
55 |
+
seed: int,
|
56 |
+
steps: int,
|
57 |
+
guidance_weight: float,
|
58 |
+
):
|
59 |
+
run_dir = 'results/'
|
60 |
+
# Load region diffusion model.
|
61 |
+
steps = 41 if not steps else steps
|
62 |
+
guidance_weight = 8.5 if not guidance_weight else guidance_weight
|
63 |
+
|
64 |
+
# parse json to span attributes
|
65 |
+
base_text_prompt, style_text_prompts, footnote_text_prompts, footnote_target_tokens,\
|
66 |
+
color_text_prompts, color_names, color_rgbs, size_text_prompts_and_sizes, use_grad_guidance = parse_json(
|
67 |
+
text_input)
|
68 |
+
|
69 |
+
# create control input for region diffusion
|
70 |
+
region_text_prompts, region_target_token_ids, base_tokens = get_region_diffusion_input(
|
71 |
+
model, base_text_prompt, style_text_prompts, footnote_text_prompts,
|
72 |
+
footnote_target_tokens, color_text_prompts, color_names)
|
73 |
+
|
74 |
+
# create control input for cross attention
|
75 |
+
text_format_dict = get_attention_control_input(
|
76 |
+
model, base_tokens, size_text_prompts_and_sizes)
|
77 |
+
|
78 |
+
# create control input for region guidance
|
79 |
+
text_format_dict, color_target_token_ids = get_gradient_guidance_input(
|
80 |
+
model, base_tokens, color_text_prompts, color_rgbs, text_format_dict)
|
81 |
+
|
82 |
+
seed_everything(seed)
|
83 |
+
|
84 |
+
# get token maps from plain text to image generation.
|
85 |
+
begin_time = time.time()
|
86 |
+
if model.attention_maps is None:
|
87 |
+
model.register_evaluation_hooks()
|
88 |
+
else:
|
89 |
+
model.reset_attention_maps()
|
90 |
+
plain_img = model.produce_attn_maps([base_text_prompt], [negative_text],
|
91 |
+
height=height, width=width, num_inference_steps=steps,
|
92 |
+
guidance_scale=guidance_weight)
|
93 |
+
print('time lapses to get attention maps: %.4f' % (time.time()-begin_time))
|
94 |
+
color_obj_masks = get_token_maps(
|
95 |
+
model.attention_maps, run_dir, width//8, height//8, color_target_token_ids, seed)
|
96 |
+
model.masks = get_token_maps(
|
97 |
+
model.attention_maps, run_dir, width//8, height//8, region_target_token_ids, seed, base_tokens)
|
98 |
+
color_obj_masks = [transforms.functional.resize(color_obj_mask, (height, width),
|
99 |
+
interpolation=transforms.InterpolationMode.BICUBIC,
|
100 |
+
antialias=True)
|
101 |
+
for color_obj_mask in color_obj_masks]
|
102 |
+
text_format_dict['color_obj_atten'] = color_obj_masks
|
103 |
+
model.remove_evaluation_hooks()
|
104 |
+
|
105 |
+
# generate image from rich text
|
106 |
+
begin_time = time.time()
|
107 |
+
seed_everything(seed)
|
108 |
+
rich_img = model.prompt_to_img(region_text_prompts, [negative_text],
|
109 |
+
height=height, width=width, num_inference_steps=steps,
|
110 |
+
guidance_scale=guidance_weight, use_grad_guidance=use_grad_guidance,
|
111 |
+
text_format_dict=text_format_dict)
|
112 |
+
print('time lapses to generate image from rich text: %.4f' %
|
113 |
+
(time.time()-begin_time))
|
114 |
+
return [plain_img[0], rich_img[0]]
|
115 |
+
|
116 |
+
with gr.Blocks() as demo:
|
117 |
+
gr.HTML("""<h1 style="font-weight: 900; margin-bottom: 7px;">Expressive Text-to-Image Generation with Rich Text</h1>
|
118 |
+
<p> Visit our <a href="https://rich-text-to-image.github.io/rich-text-to-json.html">rich-text-to-json interface</a> to generate rich-text JSON input.<p/>""")
|
119 |
+
with gr.Row():
|
120 |
+
with gr.Column():
|
121 |
+
text_input = gr.Textbox(
|
122 |
+
label='Rich-text JSON Input',
|
123 |
+
max_lines=1,
|
124 |
+
placeholder='Example: \'{"ops":[{"insert":"a Gothic "},{"attributes":{"color":"#b26b00"},"insert":"church"},{"insert":" in a the sunset with a beautiful landscape in the background.\n"}]}\'')
|
125 |
+
negative_prompt = gr.Textbox(
|
126 |
+
label='Negative Prompt',
|
127 |
+
max_lines=1,
|
128 |
+
placeholder='')
|
129 |
+
seed = gr.Slider(label='Seed',
|
130 |
+
minimum=0,
|
131 |
+
maximum=100000,
|
132 |
+
step=1,
|
133 |
+
value=6)
|
134 |
+
with gr.Accordion('Other Parameters', open=False):
|
135 |
+
steps = gr.Slider(label='Number of Steps',
|
136 |
+
minimum=0,
|
137 |
+
maximum=500,
|
138 |
+
step=1,
|
139 |
+
value=41)
|
140 |
+
guidance_weight = gr.Slider(label='CFG weight',
|
141 |
+
minimum=0,
|
142 |
+
maximum=50,
|
143 |
+
step=0.1,
|
144 |
+
value=8.5)
|
145 |
+
width = gr.Dropdown(choices=[512, 768, 896],
|
146 |
+
value=512,
|
147 |
+
label='Width',
|
148 |
+
visible=True)
|
149 |
+
height = gr.Dropdown(choices=[512, 768, 896],
|
150 |
+
value=512,
|
151 |
+
label='height',
|
152 |
+
visible=True)
|
153 |
+
|
154 |
+
with gr.Row():
|
155 |
+
with gr.Column(scale=1, min_width=100):
|
156 |
+
generate_button = gr.Button("Generate")
|
157 |
+
|
158 |
+
with gr.Column():
|
159 |
+
result = gr.Image(label='Result')
|
160 |
+
token_map = gr.Image(label='TokenMap')
|
161 |
+
|
162 |
+
with gr.Row():
|
163 |
+
examples = [
|
164 |
+
[
|
165 |
+
'{"ops":[{"insert":"a Gothic "},{"attributes":{"color":"#b26b00"},"insert":"church"},{"insert":" in a the sunset with a beautiful landscape in the background.\n"}]}',
|
166 |
+
'',
|
167 |
+
512,
|
168 |
+
512,
|
169 |
+
6,
|
170 |
+
],
|
171 |
+
[
|
172 |
+
'{"ops": [{"insert": "A pizza with "}, {"attributes": {"size": "50px"}, "insert": "pineapples"}, {"insert": ", pepperonis, and mushrooms on the top, 4k, photorealistic\n"}]}',
|
173 |
+
'blurry, art, painting, rendering, drawing, sketch, ugly, duplicate, morbid, mutilated, mutated, deformed, disfigured low quality, worst quality',
|
174 |
+
768,
|
175 |
+
896,
|
176 |
+
6,
|
177 |
+
],
|
178 |
+
[
|
179 |
+
'{"ops":[{"insert":"a "},{"attributes":{"font":"mirza"},"insert":"beautiful garden"},{"insert":" with a "},{"attributes":{"font":"roboto"},"insert":"snow mountain in the background"},{"insert":"\n"}]}',
|
180 |
+
'',
|
181 |
+
512,
|
182 |
+
512,
|
183 |
+
3,
|
184 |
+
],
|
185 |
+
[
|
186 |
+
'{"ops":[{"insert":"A close-up 4k dslr photo of a "},{"attributes":{"link":"A cat wearing sunglasses and a bandana around its neck."},"insert":"cat"},{"insert":" riding a scooter. Palm trees in the background.\n"}]}',
|
187 |
+
'',
|
188 |
+
512,
|
189 |
+
512,
|
190 |
+
6,
|
191 |
+
],
|
192 |
+
]
|
193 |
+
gr.Examples(examples=examples,
|
194 |
+
inputs=[
|
195 |
+
text_input,
|
196 |
+
negative_prompt,
|
197 |
+
height,
|
198 |
+
width,
|
199 |
+
seed,
|
200 |
+
],
|
201 |
+
outputs=[
|
202 |
+
result,
|
203 |
+
token_map,
|
204 |
+
],
|
205 |
+
fn=generate,
|
206 |
+
# cache_examples=True,
|
207 |
+
examples_per_page=20)
|
208 |
|
209 |
+
generate_button.click(
|
210 |
+
fn=generate,
|
211 |
+
inputs=[
|
212 |
+
text_input,
|
213 |
+
negative_prompt,
|
214 |
+
height,
|
215 |
+
width,
|
216 |
+
seed,
|
217 |
+
steps,
|
218 |
+
guidance_weight,
|
219 |
+
],
|
220 |
+
outputs=[result, token_map],
|
221 |
+
)
|
222 |
|
223 |
+
demo.queue(concurrency_count=1)
|
224 |
+
demo.launch(share=False)
|
225 |
|
|
|
|
|
226 |
|
227 |
+
if __name__ == "__main__":
|
228 |
+
main()
|
models/attention.py
ADDED
@@ -0,0 +1,892 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright 2022 The HuggingFace Team. All rights reserved.
|
2 |
+
#
|
3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4 |
+
# you may not use this file except in compliance with the License.
|
5 |
+
# You may obtain a copy of the License at
|
6 |
+
#
|
7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
8 |
+
#
|
9 |
+
# Unless required by applicable law or agreed to in writing, software
|
10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12 |
+
# See the License for the specific language governing permissions and
|
13 |
+
# limitations under the License.
|
14 |
+
import math
|
15 |
+
import warnings
|
16 |
+
from dataclasses import dataclass
|
17 |
+
from typing import Optional
|
18 |
+
|
19 |
+
import torch
|
20 |
+
import torch.nn.functional as F
|
21 |
+
from torch import nn
|
22 |
+
|
23 |
+
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
24 |
+
from diffusers.models.modeling_utils import ModelMixin
|
25 |
+
from diffusers.models.embeddings import ImagePositionalEmbeddings
|
26 |
+
from diffusers.utils import BaseOutput
|
27 |
+
from diffusers.utils.import_utils import is_xformers_available
|
28 |
+
|
29 |
+
|
30 |
+
@dataclass
|
31 |
+
class Transformer2DModelOutput(BaseOutput):
|
32 |
+
"""
|
33 |
+
Args:
|
34 |
+
sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
|
35 |
+
Hidden states conditioned on `encoder_hidden_states` input. If discrete, returns probability distributions
|
36 |
+
for the unnoised latent pixels.
|
37 |
+
"""
|
38 |
+
|
39 |
+
sample: torch.FloatTensor
|
40 |
+
|
41 |
+
|
42 |
+
if is_xformers_available():
|
43 |
+
import xformers
|
44 |
+
import xformers.ops
|
45 |
+
else:
|
46 |
+
xformers = None
|
47 |
+
|
48 |
+
|
49 |
+
class Transformer2DModel(ModelMixin, ConfigMixin):
|
50 |
+
"""
|
51 |
+
Transformer model for image-like data. Takes either discrete (classes of vector embeddings) or continuous (actual
|
52 |
+
embeddings) inputs.
|
53 |
+
|
54 |
+
When input is continuous: First, project the input (aka embedding) and reshape to b, t, d. Then apply standard
|
55 |
+
transformer action. Finally, reshape to image.
|
56 |
+
|
57 |
+
When input is discrete: First, input (classes of latent pixels) is converted to embeddings and has positional
|
58 |
+
embeddings applied, see `ImagePositionalEmbeddings`. Then apply standard transformer action. Finally, predict
|
59 |
+
classes of unnoised image.
|
60 |
+
|
61 |
+
Note that it is assumed one of the input classes is the masked latent pixel. The predicted classes of the unnoised
|
62 |
+
image do not contain a prediction for the masked pixel as the unnoised image cannot be masked.
|
63 |
+
|
64 |
+
Parameters:
|
65 |
+
num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
|
66 |
+
attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
|
67 |
+
in_channels (`int`, *optional*):
|
68 |
+
Pass if the input is continuous. The number of channels in the input and output.
|
69 |
+
num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
|
70 |
+
dropout (`float`, *optional*, defaults to 0.1): The dropout probability to use.
|
71 |
+
cross_attention_dim (`int`, *optional*): The number of context dimensions to use.
|
72 |
+
sample_size (`int`, *optional*): Pass if the input is discrete. The width of the latent images.
|
73 |
+
Note that this is fixed at training time as it is used for learning a number of position embeddings. See
|
74 |
+
`ImagePositionalEmbeddings`.
|
75 |
+
num_vector_embeds (`int`, *optional*):
|
76 |
+
Pass if the input is discrete. The number of classes of the vector embeddings of the latent pixels.
|
77 |
+
Includes the class for the masked latent pixel.
|
78 |
+
activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
|
79 |
+
num_embeds_ada_norm ( `int`, *optional*): Pass if at least one of the norm_layers is `AdaLayerNorm`.
|
80 |
+
The number of diffusion steps used during training. Note that this is fixed at training time as it is used
|
81 |
+
to learn a number of embeddings that are added to the hidden states. During inference, you can denoise for
|
82 |
+
up to but not more than steps than `num_embeds_ada_norm`.
|
83 |
+
attention_bias (`bool`, *optional*):
|
84 |
+
Configure if the TransformerBlocks' attention should contain a bias parameter.
|
85 |
+
"""
|
86 |
+
|
87 |
+
@register_to_config
|
88 |
+
def __init__(
|
89 |
+
self,
|
90 |
+
num_attention_heads: int = 16,
|
91 |
+
attention_head_dim: int = 88,
|
92 |
+
in_channels: Optional[int] = None,
|
93 |
+
num_layers: int = 1,
|
94 |
+
dropout: float = 0.0,
|
95 |
+
norm_num_groups: int = 32,
|
96 |
+
cross_attention_dim: Optional[int] = None,
|
97 |
+
attention_bias: bool = False,
|
98 |
+
sample_size: Optional[int] = None,
|
99 |
+
num_vector_embeds: Optional[int] = None,
|
100 |
+
activation_fn: str = "geglu",
|
101 |
+
num_embeds_ada_norm: Optional[int] = None,
|
102 |
+
use_linear_projection: bool = False,
|
103 |
+
only_cross_attention: bool = False,
|
104 |
+
):
|
105 |
+
super().__init__()
|
106 |
+
self.use_linear_projection = use_linear_projection
|
107 |
+
self.num_attention_heads = num_attention_heads
|
108 |
+
self.attention_head_dim = attention_head_dim
|
109 |
+
inner_dim = num_attention_heads * attention_head_dim
|
110 |
+
|
111 |
+
# 1. Transformer2DModel can process both standard continous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
|
112 |
+
# Define whether input is continuous or discrete depending on configuration
|
113 |
+
self.is_input_continuous = in_channels is not None
|
114 |
+
self.is_input_vectorized = num_vector_embeds is not None
|
115 |
+
|
116 |
+
if self.is_input_continuous and self.is_input_vectorized:
|
117 |
+
raise ValueError(
|
118 |
+
f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
|
119 |
+
" sure that either `in_channels` or `num_vector_embeds` is None."
|
120 |
+
)
|
121 |
+
elif not self.is_input_continuous and not self.is_input_vectorized:
|
122 |
+
raise ValueError(
|
123 |
+
f"Has to define either `in_channels`: {in_channels} or `num_vector_embeds`: {num_vector_embeds}. Make"
|
124 |
+
" sure that either `in_channels` or `num_vector_embeds` is not None."
|
125 |
+
)
|
126 |
+
|
127 |
+
# 2. Define input layers
|
128 |
+
if self.is_input_continuous:
|
129 |
+
self.in_channels = in_channels
|
130 |
+
|
131 |
+
self.norm = torch.nn.GroupNorm(
|
132 |
+
num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
|
133 |
+
if use_linear_projection:
|
134 |
+
self.proj_in = nn.Linear(in_channels, inner_dim)
|
135 |
+
else:
|
136 |
+
self.proj_in = nn.Conv2d(
|
137 |
+
in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
|
138 |
+
elif self.is_input_vectorized:
|
139 |
+
assert sample_size is not None, "Transformer2DModel over discrete input must provide sample_size"
|
140 |
+
assert num_vector_embeds is not None, "Transformer2DModel over discrete input must provide num_embed"
|
141 |
+
|
142 |
+
self.height = sample_size
|
143 |
+
self.width = sample_size
|
144 |
+
self.num_vector_embeds = num_vector_embeds
|
145 |
+
self.num_latent_pixels = self.height * self.width
|
146 |
+
|
147 |
+
self.latent_image_embedding = ImagePositionalEmbeddings(
|
148 |
+
num_embed=num_vector_embeds, embed_dim=inner_dim, height=self.height, width=self.width
|
149 |
+
)
|
150 |
+
|
151 |
+
# 3. Define transformers blocks
|
152 |
+
self.transformer_blocks = nn.ModuleList(
|
153 |
+
[
|
154 |
+
BasicTransformerBlock(
|
155 |
+
inner_dim,
|
156 |
+
num_attention_heads,
|
157 |
+
attention_head_dim,
|
158 |
+
dropout=dropout,
|
159 |
+
cross_attention_dim=cross_attention_dim,
|
160 |
+
activation_fn=activation_fn,
|
161 |
+
num_embeds_ada_norm=num_embeds_ada_norm,
|
162 |
+
attention_bias=attention_bias,
|
163 |
+
only_cross_attention=only_cross_attention,
|
164 |
+
)
|
165 |
+
for d in range(num_layers)
|
166 |
+
]
|
167 |
+
)
|
168 |
+
|
169 |
+
# 4. Define output layers
|
170 |
+
if self.is_input_continuous:
|
171 |
+
if use_linear_projection:
|
172 |
+
self.proj_out = nn.Linear(in_channels, inner_dim)
|
173 |
+
else:
|
174 |
+
self.proj_out = nn.Conv2d(
|
175 |
+
inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
|
176 |
+
elif self.is_input_vectorized:
|
177 |
+
self.norm_out = nn.LayerNorm(inner_dim)
|
178 |
+
self.out = nn.Linear(inner_dim, self.num_vector_embeds - 1)
|
179 |
+
|
180 |
+
def _set_attention_slice(self, slice_size):
|
181 |
+
for block in self.transformer_blocks:
|
182 |
+
block._set_attention_slice(slice_size)
|
183 |
+
|
184 |
+
def forward(self, hidden_states, encoder_hidden_states=None, timestep=None,
|
185 |
+
text_format_dict={}, return_dict: bool = True):
|
186 |
+
"""
|
187 |
+
Args:
|
188 |
+
hidden_states ( When discrete, `torch.LongTensor` of shape `(batch size, num latent pixels)`.
|
189 |
+
When continous, `torch.FloatTensor` of shape `(batch size, channel, height, width)`): Input
|
190 |
+
hidden_states
|
191 |
+
encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, context dim)`, *optional*):
|
192 |
+
Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
|
193 |
+
self-attention.
|
194 |
+
timestep ( `torch.long`, *optional*):
|
195 |
+
Optional timestep to be applied as an embedding in AdaLayerNorm's. Used to indicate denoising step.
|
196 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
197 |
+
Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple.
|
198 |
+
|
199 |
+
Returns:
|
200 |
+
[`~models.attention.Transformer2DModelOutput`] or `tuple`: [`~models.attention.Transformer2DModelOutput`]
|
201 |
+
if `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is the sample
|
202 |
+
tensor.
|
203 |
+
"""
|
204 |
+
# 1. Input
|
205 |
+
if self.is_input_continuous:
|
206 |
+
batch, channel, height, weight = hidden_states.shape
|
207 |
+
residual = hidden_states
|
208 |
+
|
209 |
+
hidden_states = self.norm(hidden_states)
|
210 |
+
if not self.use_linear_projection:
|
211 |
+
hidden_states = self.proj_in(hidden_states)
|
212 |
+
inner_dim = hidden_states.shape[1]
|
213 |
+
hidden_states = hidden_states.permute(
|
214 |
+
0, 2, 3, 1).reshape(batch, height * weight, inner_dim)
|
215 |
+
else:
|
216 |
+
inner_dim = hidden_states.shape[1]
|
217 |
+
hidden_states = hidden_states.permute(
|
218 |
+
0, 2, 3, 1).reshape(batch, height * weight, inner_dim)
|
219 |
+
hidden_states = self.proj_in(hidden_states)
|
220 |
+
elif self.is_input_vectorized:
|
221 |
+
hidden_states = self.latent_image_embedding(hidden_states)
|
222 |
+
|
223 |
+
# 2. Blocks
|
224 |
+
for block in self.transformer_blocks:
|
225 |
+
hidden_states = block(hidden_states, context=encoder_hidden_states, timestep=timestep,
|
226 |
+
text_format_dict=text_format_dict)
|
227 |
+
|
228 |
+
# 3. Output
|
229 |
+
if self.is_input_continuous:
|
230 |
+
if not self.use_linear_projection:
|
231 |
+
hidden_states = (
|
232 |
+
hidden_states.reshape(batch, height, weight, inner_dim).permute(
|
233 |
+
0, 3, 1, 2).contiguous()
|
234 |
+
)
|
235 |
+
hidden_states = self.proj_out(hidden_states)
|
236 |
+
else:
|
237 |
+
hidden_states = self.proj_out(hidden_states)
|
238 |
+
hidden_states = (
|
239 |
+
hidden_states.reshape(batch, height, weight, inner_dim).permute(
|
240 |
+
0, 3, 1, 2).contiguous()
|
241 |
+
)
|
242 |
+
|
243 |
+
output = hidden_states + residual
|
244 |
+
elif self.is_input_vectorized:
|
245 |
+
hidden_states = self.norm_out(hidden_states)
|
246 |
+
logits = self.out(hidden_states)
|
247 |
+
# (batch, self.num_vector_embeds - 1, self.num_latent_pixels)
|
248 |
+
logits = logits.permute(0, 2, 1)
|
249 |
+
|
250 |
+
# log(p(x_0))
|
251 |
+
output = F.log_softmax(logits.double(), dim=1).float()
|
252 |
+
|
253 |
+
if not return_dict:
|
254 |
+
return (output,)
|
255 |
+
|
256 |
+
return Transformer2DModelOutput(sample=output)
|
257 |
+
|
258 |
+
def _set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
|
259 |
+
for block in self.transformer_blocks:
|
260 |
+
block._set_use_memory_efficient_attention_xformers(
|
261 |
+
use_memory_efficient_attention_xformers)
|
262 |
+
|
263 |
+
|
264 |
+
class AttentionBlock(nn.Module):
|
265 |
+
"""
|
266 |
+
An attention block that allows spatial positions to attend to each other. Originally ported from here, but adapted
|
267 |
+
to the N-d case.
|
268 |
+
https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66.
|
269 |
+
Uses three q, k, v linear layers to compute attention.
|
270 |
+
|
271 |
+
Parameters:
|
272 |
+
channels (`int`): The number of channels in the input and output.
|
273 |
+
num_head_channels (`int`, *optional*):
|
274 |
+
The number of channels in each head. If None, then `num_heads` = 1.
|
275 |
+
norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for group norm.
|
276 |
+
rescale_output_factor (`float`, *optional*, defaults to 1.0): The factor to rescale the output by.
|
277 |
+
eps (`float`, *optional*, defaults to 1e-5): The epsilon value to use for group norm.
|
278 |
+
"""
|
279 |
+
|
280 |
+
def __init__(
|
281 |
+
self,
|
282 |
+
channels: int,
|
283 |
+
num_head_channels: Optional[int] = None,
|
284 |
+
norm_num_groups: int = 32,
|
285 |
+
rescale_output_factor: float = 1.0,
|
286 |
+
eps: float = 1e-5,
|
287 |
+
):
|
288 |
+
super().__init__()
|
289 |
+
self.channels = channels
|
290 |
+
|
291 |
+
self.num_heads = channels // num_head_channels if num_head_channels is not None else 1
|
292 |
+
self.num_head_size = num_head_channels
|
293 |
+
self.group_norm = nn.GroupNorm(
|
294 |
+
num_channels=channels, num_groups=norm_num_groups, eps=eps, affine=True)
|
295 |
+
|
296 |
+
# define q,k,v as linear layers
|
297 |
+
self.query = nn.Linear(channels, channels)
|
298 |
+
self.key = nn.Linear(channels, channels)
|
299 |
+
self.value = nn.Linear(channels, channels)
|
300 |
+
|
301 |
+
self.rescale_output_factor = rescale_output_factor
|
302 |
+
self.proj_attn = nn.Linear(channels, channels, 1)
|
303 |
+
|
304 |
+
def transpose_for_scores(self, projection: torch.Tensor) -> torch.Tensor:
|
305 |
+
new_projection_shape = projection.size()[:-1] + (self.num_heads, -1)
|
306 |
+
# move heads to 2nd position (B, T, H * D) -> (B, T, H, D) -> (B, H, T, D)
|
307 |
+
new_projection = projection.view(
|
308 |
+
new_projection_shape).permute(0, 2, 1, 3)
|
309 |
+
return new_projection
|
310 |
+
|
311 |
+
def forward(self, hidden_states):
|
312 |
+
residual = hidden_states
|
313 |
+
batch, channel, height, width = hidden_states.shape
|
314 |
+
|
315 |
+
# norm
|
316 |
+
hidden_states = self.group_norm(hidden_states)
|
317 |
+
|
318 |
+
hidden_states = hidden_states.view(
|
319 |
+
batch, channel, height * width).transpose(1, 2)
|
320 |
+
|
321 |
+
# proj to q, k, v
|
322 |
+
query_proj = self.query(hidden_states)
|
323 |
+
key_proj = self.key(hidden_states)
|
324 |
+
value_proj = self.value(hidden_states)
|
325 |
+
|
326 |
+
scale = 1 / math.sqrt(self.channels / self.num_heads)
|
327 |
+
|
328 |
+
# get scores
|
329 |
+
if self.num_heads > 1:
|
330 |
+
query_states = self.transpose_for_scores(query_proj)
|
331 |
+
key_states = self.transpose_for_scores(key_proj)
|
332 |
+
value_states = self.transpose_for_scores(value_proj)
|
333 |
+
|
334 |
+
# TODO: is there a way to perform batched matmul (e.g. baddbmm) on 4D tensors?
|
335 |
+
# or reformulate this into a 3D problem?
|
336 |
+
# TODO: measure whether on MPS device it would be faster to do this matmul via einsum
|
337 |
+
# as some matmuls can be 1.94x slower than an equivalent einsum on MPS
|
338 |
+
# https://gist.github.com/Birch-san/cba16789ec27bb20996a4b4831b13ce0
|
339 |
+
attention_scores = torch.matmul(
|
340 |
+
query_states, key_states.transpose(-1, -2)) * scale
|
341 |
+
else:
|
342 |
+
query_states, key_states, value_states = query_proj, key_proj, value_proj
|
343 |
+
|
344 |
+
attention_scores = torch.baddbmm(
|
345 |
+
torch.empty(
|
346 |
+
query_states.shape[0],
|
347 |
+
query_states.shape[1],
|
348 |
+
key_states.shape[1],
|
349 |
+
dtype=query_states.dtype,
|
350 |
+
device=query_states.device,
|
351 |
+
),
|
352 |
+
query_states,
|
353 |
+
key_states.transpose(-1, -2),
|
354 |
+
beta=0,
|
355 |
+
alpha=scale,
|
356 |
+
)
|
357 |
+
|
358 |
+
attention_probs = torch.softmax(
|
359 |
+
attention_scores.float(), dim=-1).type(attention_scores.dtype)
|
360 |
+
|
361 |
+
# compute attention output
|
362 |
+
if self.num_heads > 1:
|
363 |
+
# TODO: is there a way to perform batched matmul (e.g. bmm) on 4D tensors?
|
364 |
+
# or reformulate this into a 3D problem?
|
365 |
+
# TODO: measure whether on MPS device it would be faster to do this matmul via einsum
|
366 |
+
# as some matmuls can be 1.94x slower than an equivalent einsum on MPS
|
367 |
+
# https://gist.github.com/Birch-san/cba16789ec27bb20996a4b4831b13ce0
|
368 |
+
hidden_states = torch.matmul(attention_probs, value_states)
|
369 |
+
hidden_states = hidden_states.permute(0, 2, 1, 3).contiguous()
|
370 |
+
new_hidden_states_shape = hidden_states.size()[
|
371 |
+
:-2] + (self.channels,)
|
372 |
+
hidden_states = hidden_states.view(new_hidden_states_shape)
|
373 |
+
else:
|
374 |
+
hidden_states = torch.bmm(attention_probs, value_states)
|
375 |
+
|
376 |
+
# compute next hidden_states
|
377 |
+
hidden_states = self.proj_attn(hidden_states)
|
378 |
+
hidden_states = hidden_states.transpose(
|
379 |
+
-1, -2).reshape(batch, channel, height, width)
|
380 |
+
|
381 |
+
# res connect and rescale
|
382 |
+
hidden_states = (hidden_states + residual) / self.rescale_output_factor
|
383 |
+
return hidden_states
|
384 |
+
|
385 |
+
|
386 |
+
class BasicTransformerBlock(nn.Module):
|
387 |
+
r"""
|
388 |
+
A basic Transformer block.
|
389 |
+
|
390 |
+
Parameters:
|
391 |
+
dim (`int`): The number of channels in the input and output.
|
392 |
+
num_attention_heads (`int`): The number of heads to use for multi-head attention.
|
393 |
+
attention_head_dim (`int`): The number of channels in each head.
|
394 |
+
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
|
395 |
+
cross_attention_dim (`int`, *optional*): The size of the context vector for cross attention.
|
396 |
+
activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
|
397 |
+
num_embeds_ada_norm (:
|
398 |
+
obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
|
399 |
+
attention_bias (:
|
400 |
+
obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
|
401 |
+
"""
|
402 |
+
|
403 |
+
def __init__(
|
404 |
+
self,
|
405 |
+
dim: int,
|
406 |
+
num_attention_heads: int,
|
407 |
+
attention_head_dim: int,
|
408 |
+
dropout=0.0,
|
409 |
+
cross_attention_dim: Optional[int] = None,
|
410 |
+
activation_fn: str = "geglu",
|
411 |
+
num_embeds_ada_norm: Optional[int] = None,
|
412 |
+
attention_bias: bool = False,
|
413 |
+
only_cross_attention: bool = False,
|
414 |
+
):
|
415 |
+
super().__init__()
|
416 |
+
self.only_cross_attention = only_cross_attention
|
417 |
+
self.attn1 = CrossAttention(
|
418 |
+
query_dim=dim,
|
419 |
+
heads=num_attention_heads,
|
420 |
+
dim_head=attention_head_dim,
|
421 |
+
dropout=dropout,
|
422 |
+
bias=attention_bias,
|
423 |
+
cross_attention_dim=cross_attention_dim if only_cross_attention else None,
|
424 |
+
) # is a self-attention
|
425 |
+
self.ff = FeedForward(dim, dropout=dropout,
|
426 |
+
activation_fn=activation_fn)
|
427 |
+
self.attn2 = CrossAttention(
|
428 |
+
query_dim=dim,
|
429 |
+
cross_attention_dim=cross_attention_dim,
|
430 |
+
heads=num_attention_heads,
|
431 |
+
dim_head=attention_head_dim,
|
432 |
+
dropout=dropout,
|
433 |
+
bias=attention_bias,
|
434 |
+
) # is self-attn if context is none
|
435 |
+
|
436 |
+
# layer norms
|
437 |
+
self.use_ada_layer_norm = num_embeds_ada_norm is not None
|
438 |
+
if self.use_ada_layer_norm:
|
439 |
+
self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
|
440 |
+
self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm)
|
441 |
+
else:
|
442 |
+
self.norm1 = nn.LayerNorm(dim)
|
443 |
+
self.norm2 = nn.LayerNorm(dim)
|
444 |
+
self.norm3 = nn.LayerNorm(dim)
|
445 |
+
|
446 |
+
# if xformers is installed try to use memory_efficient_attention by default
|
447 |
+
if is_xformers_available():
|
448 |
+
try:
|
449 |
+
self._set_use_memory_efficient_attention_xformers(True)
|
450 |
+
except Exception as e:
|
451 |
+
warnings.warn(
|
452 |
+
"Could not enable memory efficient attention. Make sure xformers is installed"
|
453 |
+
f" correctly and a GPU is available: {e}"
|
454 |
+
)
|
455 |
+
|
456 |
+
def _set_attention_slice(self, slice_size):
|
457 |
+
self.attn1._slice_size = slice_size
|
458 |
+
self.attn2._slice_size = slice_size
|
459 |
+
|
460 |
+
def _set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
|
461 |
+
if not is_xformers_available():
|
462 |
+
print("Here is how to install it")
|
463 |
+
raise ModuleNotFoundError(
|
464 |
+
"Refer to https://github.com/facebookresearch/xformers for more information on how to install"
|
465 |
+
" xformers",
|
466 |
+
name="xformers",
|
467 |
+
)
|
468 |
+
elif not torch.cuda.is_available():
|
469 |
+
raise ValueError(
|
470 |
+
"torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is only"
|
471 |
+
" available for GPU "
|
472 |
+
)
|
473 |
+
else:
|
474 |
+
try:
|
475 |
+
# Make sure we can run the memory efficient attention
|
476 |
+
_ = xformers.ops.memory_efficient_attention(
|
477 |
+
torch.randn((1, 2, 40), device="cuda"),
|
478 |
+
torch.randn((1, 2, 40), device="cuda"),
|
479 |
+
torch.randn((1, 2, 40), device="cuda"),
|
480 |
+
)
|
481 |
+
except Exception as e:
|
482 |
+
raise e
|
483 |
+
self.attn1._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers
|
484 |
+
self.attn2._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers
|
485 |
+
|
486 |
+
def forward(self, hidden_states, context=None, timestep=None, text_format_dict={}):
|
487 |
+
# 1. Self-Attention
|
488 |
+
norm_hidden_states = (
|
489 |
+
self.norm1(hidden_states, timestep) if self.use_ada_layer_norm else self.norm1(
|
490 |
+
hidden_states)
|
491 |
+
)
|
492 |
+
|
493 |
+
if self.only_cross_attention:
|
494 |
+
attn_out, _ = self.attn1(
|
495 |
+
norm_hidden_states, context, text_format_dict=text_format_dict) + hidden_states
|
496 |
+
hidden_states = attn_out + hidden_states
|
497 |
+
else:
|
498 |
+
attn_out, _ = self.attn1(norm_hidden_states)
|
499 |
+
hidden_states = attn_out + hidden_states
|
500 |
+
|
501 |
+
# 2. Cross-Attention
|
502 |
+
norm_hidden_states = (
|
503 |
+
self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(
|
504 |
+
hidden_states)
|
505 |
+
)
|
506 |
+
attn_out, _ = self.attn2(
|
507 |
+
norm_hidden_states, context=context, text_format_dict=text_format_dict)
|
508 |
+
hidden_states = attn_out + hidden_states
|
509 |
+
|
510 |
+
# 3. Feed-forward
|
511 |
+
hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
|
512 |
+
|
513 |
+
return hidden_states
|
514 |
+
|
515 |
+
|
516 |
+
class CrossAttention(nn.Module):
|
517 |
+
r"""
|
518 |
+
A cross attention layer.
|
519 |
+
|
520 |
+
Parameters:
|
521 |
+
query_dim (`int`): The number of channels in the query.
|
522 |
+
cross_attention_dim (`int`, *optional*):
|
523 |
+
The number of channels in the context. If not given, defaults to `query_dim`.
|
524 |
+
heads (`int`, *optional*, defaults to 8): The number of heads to use for multi-head attention.
|
525 |
+
dim_head (`int`, *optional*, defaults to 64): The number of channels in each head.
|
526 |
+
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
|
527 |
+
bias (`bool`, *optional*, defaults to False):
|
528 |
+
Set to `True` for the query, key, and value linear layers to contain a bias parameter.
|
529 |
+
"""
|
530 |
+
|
531 |
+
def __init__(
|
532 |
+
self,
|
533 |
+
query_dim: int,
|
534 |
+
cross_attention_dim: Optional[int] = None,
|
535 |
+
heads: int = 8,
|
536 |
+
dim_head: int = 64,
|
537 |
+
dropout: float = 0.0,
|
538 |
+
bias=False,
|
539 |
+
):
|
540 |
+
super().__init__()
|
541 |
+
inner_dim = dim_head * heads
|
542 |
+
self.is_cross_attn = cross_attention_dim is not None
|
543 |
+
cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim
|
544 |
+
|
545 |
+
self.scale = dim_head**-0.5
|
546 |
+
self.heads = heads
|
547 |
+
# for slice_size > 0 the attention score computation
|
548 |
+
# is split across the batch axis to save memory
|
549 |
+
# You can set slice_size with `set_attention_slice`
|
550 |
+
self._slice_size = None
|
551 |
+
self._use_memory_efficient_attention_xformers = False
|
552 |
+
|
553 |
+
self.to_q = nn.Linear(query_dim, inner_dim, bias=bias)
|
554 |
+
self.to_k = nn.Linear(cross_attention_dim, inner_dim, bias=bias)
|
555 |
+
self.to_v = nn.Linear(cross_attention_dim, inner_dim, bias=bias)
|
556 |
+
|
557 |
+
self.to_out = nn.ModuleList([])
|
558 |
+
self.to_out.append(nn.Linear(inner_dim, query_dim))
|
559 |
+
self.to_out.append(nn.Dropout(dropout))
|
560 |
+
|
561 |
+
def reshape_heads_to_batch_dim(self, tensor):
|
562 |
+
batch_size, seq_len, dim = tensor.shape
|
563 |
+
head_size = self.heads
|
564 |
+
tensor = tensor.reshape(batch_size, seq_len,
|
565 |
+
head_size, dim // head_size)
|
566 |
+
tensor = tensor.permute(0, 2, 1, 3).reshape(
|
567 |
+
batch_size * head_size, seq_len, dim // head_size)
|
568 |
+
return tensor
|
569 |
+
|
570 |
+
def reshape_batch_dim_to_heads(self, tensor):
|
571 |
+
batch_size, seq_len, dim = tensor.shape
|
572 |
+
head_size = self.heads
|
573 |
+
tensor = tensor.reshape(batch_size // head_size,
|
574 |
+
head_size, seq_len, dim)
|
575 |
+
tensor = tensor.permute(0, 2, 1, 3).reshape(
|
576 |
+
batch_size // head_size, seq_len, dim * head_size)
|
577 |
+
return tensor
|
578 |
+
|
579 |
+
def reshape_batch_dim_to_heads_and_average(self, tensor):
|
580 |
+
batch_size, seq_len, seq_len2 = tensor.shape
|
581 |
+
head_size = self.heads
|
582 |
+
tensor = tensor.reshape(batch_size // head_size,
|
583 |
+
head_size, seq_len, seq_len2)
|
584 |
+
return tensor.mean(1)
|
585 |
+
|
586 |
+
def forward(self, hidden_states, context=None, mask=None, text_format_dict={}):
|
587 |
+
batch_size, sequence_length, _ = hidden_states.shape
|
588 |
+
|
589 |
+
query = self.to_q(hidden_states)
|
590 |
+
context = context if context is not None else hidden_states
|
591 |
+
key = self.to_k(context)
|
592 |
+
value = self.to_v(context)
|
593 |
+
|
594 |
+
dim = query.shape[-1]
|
595 |
+
|
596 |
+
query = self.reshape_heads_to_batch_dim(query)
|
597 |
+
key = self.reshape_heads_to_batch_dim(key)
|
598 |
+
value = self.reshape_heads_to_batch_dim(value)
|
599 |
+
|
600 |
+
# attention, what we cannot get enough of
|
601 |
+
if self._use_memory_efficient_attention_xformers:
|
602 |
+
hidden_states = self._memory_efficient_attention_xformers(
|
603 |
+
query, key, value)
|
604 |
+
# Some versions of xformers return output in fp32, cast it back to the dtype of the input
|
605 |
+
hidden_states = hidden_states.to(query.dtype)
|
606 |
+
else:
|
607 |
+
if self._slice_size is None or query.shape[0] // self._slice_size == 1:
|
608 |
+
# only this attention function is used
|
609 |
+
hidden_states, attn_probs = self._attention(
|
610 |
+
query, key, value, **text_format_dict)
|
611 |
+
|
612 |
+
# linear proj
|
613 |
+
hidden_states = self.to_out[0](hidden_states)
|
614 |
+
# dropout
|
615 |
+
hidden_states = self.to_out[1](hidden_states)
|
616 |
+
return hidden_states, attn_probs
|
617 |
+
|
618 |
+
def _qk(self, query, key):
|
619 |
+
return torch.baddbmm(
|
620 |
+
torch.empty(query.shape[0], query.shape[1], key.shape[1],
|
621 |
+
dtype=query.dtype, device=query.device),
|
622 |
+
query,
|
623 |
+
key.transpose(-1, -2),
|
624 |
+
beta=0,
|
625 |
+
alpha=self.scale,
|
626 |
+
)
|
627 |
+
|
628 |
+
def _attention(self, query, key, value, word_pos=None, font_size=None,
|
629 |
+
**kwargs):
|
630 |
+
attention_scores = self._qk(query, key)
|
631 |
+
|
632 |
+
# Font size:
|
633 |
+
if self.is_cross_attn and word_pos is not None and font_size is not None:
|
634 |
+
assert key.shape[1] == 77
|
635 |
+
attention_score_exp = attention_scores.exp()
|
636 |
+
font_size_abs, font_size_sign = font_size.abs(), font_size.sign()
|
637 |
+
attention_score_exp[:, :, word_pos] = attention_score_exp[:, :, word_pos].clone(
|
638 |
+
)*font_size_abs
|
639 |
+
attention_probs = attention_score_exp / \
|
640 |
+
attention_score_exp.sum(-1, True)
|
641 |
+
attention_probs[:, :, word_pos] *= font_size_sign
|
642 |
+
else:
|
643 |
+
attention_probs = attention_scores.softmax(dim=-1)
|
644 |
+
|
645 |
+
hidden_states = torch.bmm(attention_probs, value)
|
646 |
+
|
647 |
+
# reshape hidden_states
|
648 |
+
hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
|
649 |
+
attention_probs = self.reshape_batch_dim_to_heads_and_average(
|
650 |
+
attention_probs)
|
651 |
+
return hidden_states, attention_probs
|
652 |
+
|
653 |
+
def _memory_efficient_attention_xformers(self, query, key, value):
|
654 |
+
query = query.contiguous()
|
655 |
+
key = key.contiguous()
|
656 |
+
value = value.contiguous()
|
657 |
+
hidden_states = xformers.ops.memory_efficient_attention(
|
658 |
+
query, key, value, attn_bias=None)
|
659 |
+
hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
|
660 |
+
return hidden_states
|
661 |
+
|
662 |
+
|
663 |
+
class FeedForward(nn.Module):
|
664 |
+
r"""
|
665 |
+
A feed-forward layer.
|
666 |
+
|
667 |
+
Parameters:
|
668 |
+
dim (`int`): The number of channels in the input.
|
669 |
+
dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`.
|
670 |
+
mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension.
|
671 |
+
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
|
672 |
+
activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
|
673 |
+
"""
|
674 |
+
|
675 |
+
def __init__(
|
676 |
+
self,
|
677 |
+
dim: int,
|
678 |
+
dim_out: Optional[int] = None,
|
679 |
+
mult: int = 4,
|
680 |
+
dropout: float = 0.0,
|
681 |
+
activation_fn: str = "geglu",
|
682 |
+
):
|
683 |
+
super().__init__()
|
684 |
+
inner_dim = int(dim * mult)
|
685 |
+
dim_out = dim_out if dim_out is not None else dim
|
686 |
+
|
687 |
+
if activation_fn == "geglu":
|
688 |
+
geglu = GEGLU(dim, inner_dim)
|
689 |
+
elif activation_fn == "geglu-approximate":
|
690 |
+
geglu = ApproximateGELU(dim, inner_dim)
|
691 |
+
|
692 |
+
self.net = nn.ModuleList([])
|
693 |
+
# project in
|
694 |
+
self.net.append(geglu)
|
695 |
+
# project dropout
|
696 |
+
self.net.append(nn.Dropout(dropout))
|
697 |
+
# project out
|
698 |
+
self.net.append(nn.Linear(inner_dim, dim_out))
|
699 |
+
|
700 |
+
def forward(self, hidden_states):
|
701 |
+
for module in self.net:
|
702 |
+
hidden_states = module(hidden_states)
|
703 |
+
return hidden_states
|
704 |
+
|
705 |
+
|
706 |
+
# feedforward
|
707 |
+
class GEGLU(nn.Module):
|
708 |
+
r"""
|
709 |
+
A variant of the gated linear unit activation function from https://arxiv.org/abs/2002.05202.
|
710 |
+
|
711 |
+
Parameters:
|
712 |
+
dim_in (`int`): The number of channels in the input.
|
713 |
+
dim_out (`int`): The number of channels in the output.
|
714 |
+
"""
|
715 |
+
|
716 |
+
def __init__(self, dim_in: int, dim_out: int):
|
717 |
+
super().__init__()
|
718 |
+
self.proj = nn.Linear(dim_in, dim_out * 2)
|
719 |
+
|
720 |
+
def gelu(self, gate):
|
721 |
+
if gate.device.type != "mps":
|
722 |
+
return F.gelu(gate)
|
723 |
+
# mps: gelu is not implemented for float16
|
724 |
+
return F.gelu(gate.to(dtype=torch.float32)).to(dtype=gate.dtype)
|
725 |
+
|
726 |
+
def forward(self, hidden_states):
|
727 |
+
hidden_states, gate = self.proj(hidden_states).chunk(2, dim=-1)
|
728 |
+
return hidden_states * self.gelu(gate)
|
729 |
+
|
730 |
+
|
731 |
+
class ApproximateGELU(nn.Module):
|
732 |
+
"""
|
733 |
+
The approximate form of Gaussian Error Linear Unit (GELU)
|
734 |
+
|
735 |
+
For more details, see section 2: https://arxiv.org/abs/1606.08415
|
736 |
+
"""
|
737 |
+
|
738 |
+
def __init__(self, dim_in: int, dim_out: int):
|
739 |
+
super().__init__()
|
740 |
+
self.proj = nn.Linear(dim_in, dim_out)
|
741 |
+
|
742 |
+
def forward(self, x):
|
743 |
+
x = self.proj(x)
|
744 |
+
return x * torch.sigmoid(1.702 * x)
|
745 |
+
|
746 |
+
|
747 |
+
class AdaLayerNorm(nn.Module):
|
748 |
+
"""
|
749 |
+
Norm layer modified to incorporate timestep embeddings.
|
750 |
+
"""
|
751 |
+
|
752 |
+
def __init__(self, embedding_dim, num_embeddings):
|
753 |
+
super().__init__()
|
754 |
+
self.emb = nn.Embedding(num_embeddings, embedding_dim)
|
755 |
+
self.silu = nn.SiLU()
|
756 |
+
self.linear = nn.Linear(embedding_dim, embedding_dim * 2)
|
757 |
+
self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False)
|
758 |
+
|
759 |
+
def forward(self, x, timestep):
|
760 |
+
emb = self.linear(self.silu(self.emb(timestep)))
|
761 |
+
scale, shift = torch.chunk(emb, 2)
|
762 |
+
x = self.norm(x) * (1 + scale) + shift
|
763 |
+
return x
|
764 |
+
|
765 |
+
|
766 |
+
class DualTransformer2DModel(nn.Module):
|
767 |
+
"""
|
768 |
+
Dual transformer wrapper that combines two `Transformer2DModel`s for mixed inference.
|
769 |
+
|
770 |
+
Parameters:
|
771 |
+
num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
|
772 |
+
attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
|
773 |
+
in_channels (`int`, *optional*):
|
774 |
+
Pass if the input is continuous. The number of channels in the input and output.
|
775 |
+
num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
|
776 |
+
dropout (`float`, *optional*, defaults to 0.1): The dropout probability to use.
|
777 |
+
cross_attention_dim (`int`, *optional*): The number of context dimensions to use.
|
778 |
+
sample_size (`int`, *optional*): Pass if the input is discrete. The width of the latent images.
|
779 |
+
Note that this is fixed at training time as it is used for learning a number of position embeddings. See
|
780 |
+
`ImagePositionalEmbeddings`.
|
781 |
+
num_vector_embeds (`int`, *optional*):
|
782 |
+
Pass if the input is discrete. The number of classes of the vector embeddings of the latent pixels.
|
783 |
+
Includes the class for the masked latent pixel.
|
784 |
+
activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
|
785 |
+
num_embeds_ada_norm ( `int`, *optional*): Pass if at least one of the norm_layers is `AdaLayerNorm`.
|
786 |
+
The number of diffusion steps used during training. Note that this is fixed at training time as it is used
|
787 |
+
to learn a number of embeddings that are added to the hidden states. During inference, you can denoise for
|
788 |
+
up to but not more than steps than `num_embeds_ada_norm`.
|
789 |
+
attention_bias (`bool`, *optional*):
|
790 |
+
Configure if the TransformerBlocks' attention should contain a bias parameter.
|
791 |
+
"""
|
792 |
+
|
793 |
+
def __init__(
|
794 |
+
self,
|
795 |
+
num_attention_heads: int = 16,
|
796 |
+
attention_head_dim: int = 88,
|
797 |
+
in_channels: Optional[int] = None,
|
798 |
+
num_layers: int = 1,
|
799 |
+
dropout: float = 0.0,
|
800 |
+
norm_num_groups: int = 32,
|
801 |
+
cross_attention_dim: Optional[int] = None,
|
802 |
+
attention_bias: bool = False,
|
803 |
+
sample_size: Optional[int] = None,
|
804 |
+
num_vector_embeds: Optional[int] = None,
|
805 |
+
activation_fn: str = "geglu",
|
806 |
+
num_embeds_ada_norm: Optional[int] = None,
|
807 |
+
):
|
808 |
+
super().__init__()
|
809 |
+
self.transformers = nn.ModuleList(
|
810 |
+
[
|
811 |
+
Transformer2DModel(
|
812 |
+
num_attention_heads=num_attention_heads,
|
813 |
+
attention_head_dim=attention_head_dim,
|
814 |
+
in_channels=in_channels,
|
815 |
+
num_layers=num_layers,
|
816 |
+
dropout=dropout,
|
817 |
+
norm_num_groups=norm_num_groups,
|
818 |
+
cross_attention_dim=cross_attention_dim,
|
819 |
+
attention_bias=attention_bias,
|
820 |
+
sample_size=sample_size,
|
821 |
+
num_vector_embeds=num_vector_embeds,
|
822 |
+
activation_fn=activation_fn,
|
823 |
+
num_embeds_ada_norm=num_embeds_ada_norm,
|
824 |
+
)
|
825 |
+
for _ in range(2)
|
826 |
+
]
|
827 |
+
)
|
828 |
+
|
829 |
+
# Variables that can be set by a pipeline:
|
830 |
+
|
831 |
+
# The ratio of transformer1 to transformer2's output states to be combined during inference
|
832 |
+
self.mix_ratio = 0.5
|
833 |
+
|
834 |
+
# The shape of `encoder_hidden_states` is expected to be
|
835 |
+
# `(batch_size, condition_lengths[0]+condition_lengths[1], num_features)`
|
836 |
+
self.condition_lengths = [77, 257]
|
837 |
+
|
838 |
+
# Which transformer to use to encode which condition.
|
839 |
+
# E.g. `(1, 0)` means that we'll use `transformers[1](conditions[0])` and `transformers[0](conditions[1])`
|
840 |
+
self.transformer_index_for_condition = [1, 0]
|
841 |
+
|
842 |
+
def forward(self, hidden_states, encoder_hidden_states, timestep=None, return_dict: bool = True):
|
843 |
+
"""
|
844 |
+
Args:
|
845 |
+
hidden_states ( When discrete, `torch.LongTensor` of shape `(batch size, num latent pixels)`.
|
846 |
+
When continuous, `torch.FloatTensor` of shape `(batch size, channel, height, width)`): Input
|
847 |
+
hidden_states
|
848 |
+
encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, context dim)`, *optional*):
|
849 |
+
Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
|
850 |
+
self-attention.
|
851 |
+
timestep ( `torch.long`, *optional*):
|
852 |
+
Optional timestep to be applied as an embedding in AdaLayerNorm's. Used to indicate denoising step.
|
853 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
854 |
+
Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple.
|
855 |
+
|
856 |
+
Returns:
|
857 |
+
[`~models.attention.Transformer2DModelOutput`] or `tuple`: [`~models.attention.Transformer2DModelOutput`]
|
858 |
+
if `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is the sample
|
859 |
+
tensor.
|
860 |
+
"""
|
861 |
+
input_states = hidden_states
|
862 |
+
|
863 |
+
encoded_states = []
|
864 |
+
tokens_start = 0
|
865 |
+
for i in range(2):
|
866 |
+
# for each of the two transformers, pass the corresponding condition tokens
|
867 |
+
condition_state = encoder_hidden_states[:,
|
868 |
+
tokens_start: tokens_start + self.condition_lengths[i]]
|
869 |
+
transformer_index = self.transformer_index_for_condition[i]
|
870 |
+
encoded_state = self.transformers[transformer_index](input_states, condition_state, timestep, return_dict)[
|
871 |
+
0
|
872 |
+
]
|
873 |
+
encoded_states.append(encoded_state - input_states)
|
874 |
+
tokens_start += self.condition_lengths[i]
|
875 |
+
|
876 |
+
output_states = encoded_states[0] * self.mix_ratio + \
|
877 |
+
encoded_states[1] * (1 - self.mix_ratio)
|
878 |
+
output_states = output_states + input_states
|
879 |
+
|
880 |
+
if not return_dict:
|
881 |
+
return (output_states,)
|
882 |
+
|
883 |
+
return Transformer2DModelOutput(sample=output_states)
|
884 |
+
|
885 |
+
def _set_attention_slice(self, slice_size):
|
886 |
+
for transformer in self.transformers:
|
887 |
+
transformer._set_attention_slice(slice_size)
|
888 |
+
|
889 |
+
def _set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
|
890 |
+
for transformer in self.transformers:
|
891 |
+
transformer._set_use_memory_efficient_attention_xformers(
|
892 |
+
use_memory_efficient_attention_xformers)
|
models/region_diffusion.py
ADDED
@@ -0,0 +1,307 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import torch
|
3 |
+
import collections
|
4 |
+
import torch.nn as nn
|
5 |
+
from functools import partial
|
6 |
+
from transformers import CLIPTextModel, CLIPTokenizer, logging
|
7 |
+
from diffusers import AutoencoderKL, PNDMScheduler, EulerDiscreteScheduler, DPMSolverMultistepScheduler
|
8 |
+
from models.unet_2d_condition import UNet2DConditionModel
|
9 |
+
|
10 |
+
# suppress partial model loading warning
|
11 |
+
logging.set_verbosity_error()
|
12 |
+
|
13 |
+
|
14 |
+
class RegionDiffusion(nn.Module):
|
15 |
+
def __init__(self, device):
|
16 |
+
super().__init__()
|
17 |
+
|
18 |
+
try:
|
19 |
+
with open('./TOKEN', 'r') as f:
|
20 |
+
self.token = f.read().replace('\n', '') # remove the last \n!
|
21 |
+
print(f'[INFO] loaded hugging face access token from ./TOKEN!')
|
22 |
+
except FileNotFoundError as e:
|
23 |
+
self.token = True
|
24 |
+
print(f'[INFO] try to load hugging face access token from the default place, make sure you have run `huggingface-cli login`.')
|
25 |
+
|
26 |
+
self.device = device
|
27 |
+
self.num_train_timesteps = 1000
|
28 |
+
self.clip_gradient = False
|
29 |
+
|
30 |
+
print(f'[INFO] loading stable diffusion...')
|
31 |
+
local_pretrained_dir = f'pretrained-guidance/v1'
|
32 |
+
if not os.path.isdir(local_pretrained_dir):
|
33 |
+
save_pretrained = True
|
34 |
+
load_paths = 'runwayml/stable-diffusion-v1-5'
|
35 |
+
os.makedirs(local_pretrained_dir, exist_ok=True)
|
36 |
+
else:
|
37 |
+
save_pretrained = False
|
38 |
+
load_paths = local_pretrained_dir
|
39 |
+
|
40 |
+
# 1. Load the autoencoder model which will be used to decode the latents into image space.
|
41 |
+
self.vae = AutoencoderKL.from_pretrained(
|
42 |
+
load_paths, subfolder="vae", use_auth_token=self.token).to(self.device)
|
43 |
+
|
44 |
+
# 2. Load the tokenizer and text encoder to tokenize and encode the text.
|
45 |
+
self.tokenizer = CLIPTokenizer.from_pretrained(
|
46 |
+
load_paths, subfolder='tokenizer', use_auth_token=self.token)
|
47 |
+
self.text_encoder = CLIPTextModel.from_pretrained(
|
48 |
+
load_paths, subfolder='text_encoder', use_auth_token=self.token).to(self.device)
|
49 |
+
|
50 |
+
# 3. The UNet model for generating the latents.
|
51 |
+
self.unet = UNet2DConditionModel.from_pretrained(
|
52 |
+
load_paths, subfolder="unet", use_auth_token=self.token).to(self.device)
|
53 |
+
|
54 |
+
if save_pretrained:
|
55 |
+
self.vae.save_pretrained(os.path.join(local_pretrained_dir, 'vae'))
|
56 |
+
self.tokenizer.save_pretrained(
|
57 |
+
os.path.join(local_pretrained_dir, 'tokenizer'))
|
58 |
+
self.text_encoder.save_pretrained(
|
59 |
+
os.path.join(local_pretrained_dir, 'text_encoder'))
|
60 |
+
self.unet.save_pretrained(
|
61 |
+
os.path.join(local_pretrained_dir, 'unet'))
|
62 |
+
|
63 |
+
self.scheduler = PNDMScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear",
|
64 |
+
num_train_timesteps=self.num_train_timesteps, skip_prk_steps=True, steps_offset=1)
|
65 |
+
self.alphas_cumprod = self.scheduler.alphas_cumprod.to(self.device)
|
66 |
+
|
67 |
+
self.masks = []
|
68 |
+
self.attention_maps = None
|
69 |
+
self.color_loss = torch.nn.functional.mse_loss
|
70 |
+
|
71 |
+
print(f'[INFO] loaded stable diffusion!')
|
72 |
+
|
73 |
+
def get_text_embeds(self, prompt, negative_prompt):
|
74 |
+
# prompt, negative_prompt: [str]
|
75 |
+
|
76 |
+
# Tokenize text and get embeddings
|
77 |
+
text_input = self.tokenizer(
|
78 |
+
prompt, padding='max_length', max_length=self.tokenizer.model_max_length, truncation=True, return_tensors='pt')
|
79 |
+
|
80 |
+
with torch.no_grad():
|
81 |
+
text_embeddings = self.text_encoder(
|
82 |
+
text_input.input_ids.to(self.device))[0]
|
83 |
+
|
84 |
+
# Do the same for unconditional embeddings
|
85 |
+
uncond_input = self.tokenizer(negative_prompt, padding='max_length',
|
86 |
+
max_length=self.tokenizer.model_max_length, return_tensors='pt')
|
87 |
+
|
88 |
+
with torch.no_grad():
|
89 |
+
uncond_embeddings = self.text_encoder(
|
90 |
+
uncond_input.input_ids.to(self.device))[0]
|
91 |
+
|
92 |
+
# Cat for final embeddings
|
93 |
+
text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
|
94 |
+
return text_embeddings
|
95 |
+
|
96 |
+
def get_text_embeds_list(self, prompts):
|
97 |
+
# prompts: [list]
|
98 |
+
text_embeddings = []
|
99 |
+
for prompt in prompts:
|
100 |
+
# Tokenize text and get embeddings
|
101 |
+
text_input = self.tokenizer(
|
102 |
+
[prompt], padding='max_length', max_length=self.tokenizer.model_max_length, truncation=True, return_tensors='pt')
|
103 |
+
|
104 |
+
with torch.no_grad():
|
105 |
+
text_embeddings.append(self.text_encoder(
|
106 |
+
text_input.input_ids.to(self.device))[0])
|
107 |
+
|
108 |
+
return text_embeddings
|
109 |
+
|
110 |
+
def produce_latents(self, text_embeddings, height=512, width=512, num_inference_steps=50, guidance_scale=7.5,
|
111 |
+
latents=None, use_grad_guidance=False, text_format_dict={}):
|
112 |
+
|
113 |
+
if latents is None:
|
114 |
+
latents = torch.randn(
|
115 |
+
(1, self.unet.in_channels, height // 8, width // 8), device=self.device)
|
116 |
+
|
117 |
+
self.scheduler.set_timesteps(num_inference_steps)
|
118 |
+
n_styles = text_embeddings.shape[0]-1
|
119 |
+
assert n_styles == len(self.masks)
|
120 |
+
|
121 |
+
with torch.autocast('cuda'):
|
122 |
+
for i, t in enumerate(self.scheduler.timesteps):
|
123 |
+
|
124 |
+
# predict the noise residual
|
125 |
+
with torch.no_grad():
|
126 |
+
noise_pred_uncond = self.unet(latents, t, encoder_hidden_states=text_embeddings[:1],
|
127 |
+
text_format_dict={})['sample']
|
128 |
+
noise_pred_text = None
|
129 |
+
for style_i, mask in enumerate(self.masks):
|
130 |
+
if style_i < len(self.masks) - 1:
|
131 |
+
masked_latent = latents
|
132 |
+
noise_pred_text_cur = self.unet(masked_latent, t, encoder_hidden_states=text_embeddings[style_i+1:style_i+2],
|
133 |
+
text_format_dict={})['sample']
|
134 |
+
else:
|
135 |
+
noise_pred_text_cur = self.unet(latents, t, encoder_hidden_states=text_embeddings[style_i+1:style_i+2],
|
136 |
+
text_format_dict=text_format_dict)['sample']
|
137 |
+
if noise_pred_text is None:
|
138 |
+
noise_pred_text = noise_pred_text_cur * mask
|
139 |
+
else:
|
140 |
+
noise_pred_text = noise_pred_text + noise_pred_text_cur*mask
|
141 |
+
|
142 |
+
# perform classifier-free guidance
|
143 |
+
noise_pred = noise_pred_uncond + guidance_scale * \
|
144 |
+
(noise_pred_text - noise_pred_uncond)
|
145 |
+
|
146 |
+
# compute the previous noisy sample x_t -> x_t-1
|
147 |
+
latents = self.scheduler.step(noise_pred, t, latents)[
|
148 |
+
'prev_sample']
|
149 |
+
|
150 |
+
# apply gradient guidance
|
151 |
+
if use_grad_guidance and t < text_format_dict['guidance_start_step']:
|
152 |
+
with torch.enable_grad():
|
153 |
+
if not latents.requires_grad:
|
154 |
+
latents.requires_grad = True
|
155 |
+
latents_0 = self.predict_x0(latents, noise_pred, t)
|
156 |
+
latents_inp = 1 / 0.18215 * latents_0
|
157 |
+
imgs = self.vae.decode(latents_inp).sample
|
158 |
+
imgs = (imgs / 2 + 0.5).clamp(0, 1)
|
159 |
+
loss_total = 0.
|
160 |
+
for attn_map, rgb_val in zip(text_format_dict['color_obj_atten'], text_format_dict['target_RGB']):
|
161 |
+
avg_rgb = (
|
162 |
+
imgs*attn_map[:, 0]).sum(2).sum(2)/attn_map[:, 0].sum()
|
163 |
+
loss = self.color_loss(
|
164 |
+
avg_rgb, rgb_val[:, :, 0, 0])*100
|
165 |
+
# print(loss)
|
166 |
+
loss_total += loss
|
167 |
+
loss_total.backward()
|
168 |
+
latents = (
|
169 |
+
latents - latents.grad * text_format_dict['color_guidance_weight']).detach().clone()
|
170 |
+
|
171 |
+
return latents
|
172 |
+
|
173 |
+
def predict_x0(self, x_t, eps_t, t):
|
174 |
+
alpha_t = self.scheduler.alphas_cumprod[t]
|
175 |
+
return (x_t - eps_t * torch.sqrt(1-alpha_t)) / torch.sqrt(alpha_t)
|
176 |
+
|
177 |
+
def produce_attn_maps(self, prompts, negative_prompts='', height=512, width=512, num_inference_steps=50,
|
178 |
+
guidance_scale=7.5, latents=None):
|
179 |
+
|
180 |
+
if isinstance(prompts, str):
|
181 |
+
prompts = [prompts]
|
182 |
+
|
183 |
+
if isinstance(negative_prompts, str):
|
184 |
+
negative_prompts = [negative_prompts]
|
185 |
+
|
186 |
+
# Prompts -> text embeds
|
187 |
+
text_embeddings = self.get_text_embeds(
|
188 |
+
prompts, negative_prompts) # [2, 77, 768]
|
189 |
+
if latents is None:
|
190 |
+
latents = torch.randn(
|
191 |
+
(text_embeddings.shape[0] // 2, self.unet.in_channels, height // 8, width // 8), device=self.device)
|
192 |
+
|
193 |
+
self.scheduler.set_timesteps(num_inference_steps)
|
194 |
+
|
195 |
+
with torch.autocast('cuda'):
|
196 |
+
for i, t in enumerate(self.scheduler.timesteps):
|
197 |
+
# expand the latents if we are doing classifier-free guidance to avoid doing two forward passes.
|
198 |
+
latent_model_input = torch.cat([latents] * 2)
|
199 |
+
|
200 |
+
# predict the noise residual
|
201 |
+
with torch.no_grad():
|
202 |
+
noise_pred = self.unet(
|
203 |
+
latent_model_input, t, encoder_hidden_states=text_embeddings)['sample']
|
204 |
+
|
205 |
+
# perform guidance
|
206 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
207 |
+
noise_pred = noise_pred_uncond + guidance_scale * \
|
208 |
+
(noise_pred_text - noise_pred_uncond)
|
209 |
+
|
210 |
+
# compute the previous noisy sample x_t -> x_t-1
|
211 |
+
latents = self.scheduler.step(noise_pred, t, latents)[
|
212 |
+
'prev_sample']
|
213 |
+
|
214 |
+
# Img latents -> imgs
|
215 |
+
imgs = self.decode_latents(latents) # [1, 3, 512, 512]
|
216 |
+
|
217 |
+
# Img to Numpy
|
218 |
+
imgs = imgs.detach().cpu().permute(0, 2, 3, 1).numpy()
|
219 |
+
imgs = (imgs * 255).round().astype('uint8')
|
220 |
+
|
221 |
+
return imgs
|
222 |
+
|
223 |
+
def decode_latents(self, latents):
|
224 |
+
|
225 |
+
latents = 1 / 0.18215 * latents
|
226 |
+
|
227 |
+
with torch.no_grad():
|
228 |
+
imgs = self.vae.decode(latents).sample
|
229 |
+
|
230 |
+
imgs = (imgs / 2 + 0.5).clamp(0, 1)
|
231 |
+
|
232 |
+
return imgs
|
233 |
+
|
234 |
+
def prompt_to_img(self, prompts, negative_prompts='', height=512, width=512, num_inference_steps=50,
|
235 |
+
guidance_scale=7.5, latents=None, text_format_dict={}, use_grad_guidance=False):
|
236 |
+
|
237 |
+
if isinstance(prompts, str):
|
238 |
+
prompts = [prompts]
|
239 |
+
|
240 |
+
if isinstance(negative_prompts, str):
|
241 |
+
negative_prompts = [negative_prompts]
|
242 |
+
|
243 |
+
# Prompts -> text embeds
|
244 |
+
text_embeds = self.get_text_embeds(
|
245 |
+
prompts, negative_prompts) # [2, 77, 768]
|
246 |
+
|
247 |
+
if len(text_format_dict) > 0:
|
248 |
+
if 'font_styles' in text_format_dict and text_format_dict['font_styles'] is not None:
|
249 |
+
text_format_dict['font_styles_embs'] = self.get_text_embeds_list(
|
250 |
+
text_format_dict['font_styles']) # [2, 77, 768]
|
251 |
+
else:
|
252 |
+
text_format_dict['font_styles_embs'] = None
|
253 |
+
|
254 |
+
# else:
|
255 |
+
latents = self.produce_latents(text_embeds, height=height, width=width, latents=latents,
|
256 |
+
num_inference_steps=num_inference_steps, guidance_scale=guidance_scale,
|
257 |
+
use_grad_guidance=use_grad_guidance, text_format_dict=text_format_dict) # [1, 4, 64, 64]
|
258 |
+
|
259 |
+
# Img latents -> imgs
|
260 |
+
imgs = self.decode_latents(latents) # [1, 3, 512, 512]
|
261 |
+
|
262 |
+
# Img to Numpy
|
263 |
+
imgs = imgs.detach().cpu().permute(0, 2, 3, 1).numpy()
|
264 |
+
imgs = (imgs * 255).round().astype('uint8')
|
265 |
+
|
266 |
+
return imgs
|
267 |
+
|
268 |
+
def reset_attention_maps(self):
|
269 |
+
r"""Function to reset attention maps.
|
270 |
+
We reset attention maps because we append them while getting hooks
|
271 |
+
to visualize attention maps for every step.
|
272 |
+
"""
|
273 |
+
for key in self.attention_maps:
|
274 |
+
self.attention_maps[key] = []
|
275 |
+
|
276 |
+
def register_evaluation_hooks(self):
|
277 |
+
r"""Function for registering hooks during evaluation.
|
278 |
+
We mainly store activation maps averaged over queries.
|
279 |
+
"""
|
280 |
+
self.forward_hooks = []
|
281 |
+
|
282 |
+
def save_activations(activations, name, module, inp, out):
|
283 |
+
r"""
|
284 |
+
PyTorch Forward hook to save outputs at each forward pass.
|
285 |
+
"""
|
286 |
+
# out[0] - final output of attention layer
|
287 |
+
# out[1] - attention probability matrix
|
288 |
+
if 'attn2' in name:
|
289 |
+
assert out[1].shape[-1] == 77
|
290 |
+
activations[name].append(out[1].detach().cpu())
|
291 |
+
else:
|
292 |
+
assert out[1].shape[-1] != 77
|
293 |
+
attention_dict = collections.defaultdict(list)
|
294 |
+
for name, module in self.unet.named_modules():
|
295 |
+
leaf_name = name.split('.')[-1]
|
296 |
+
if 'attn' in leaf_name:
|
297 |
+
# Register hook to obtain outputs at every attention layer.
|
298 |
+
self.forward_hooks.append(module.register_forward_hook(
|
299 |
+
partial(save_activations, attention_dict, name)
|
300 |
+
))
|
301 |
+
# attention_dict is a dictionary containing attention maps for every attention layer
|
302 |
+
self.attention_maps = attention_dict
|
303 |
+
|
304 |
+
def remove_evaluation_hooks(self):
|
305 |
+
for hook in self.forward_hooks:
|
306 |
+
hook.remove()
|
307 |
+
self.attention_maps = None
|
models/unet_2d_blocks.py
ADDED
@@ -0,0 +1,1670 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright 2022 The HuggingFace Team. All rights reserved.
|
2 |
+
#
|
3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4 |
+
# you may not use this file except in compliance with the License.
|
5 |
+
# You may obtain a copy of the License at
|
6 |
+
#
|
7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
8 |
+
#
|
9 |
+
# Unless required by applicable law or agreed to in writing, software
|
10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12 |
+
# See the License for the specific language governing permissions and
|
13 |
+
# limitations under the License.
|
14 |
+
import numpy as np
|
15 |
+
import torch
|
16 |
+
from torch import nn
|
17 |
+
|
18 |
+
from .attention import AttentionBlock, DualTransformer2DModel, Transformer2DModel
|
19 |
+
from diffusers.models.resnet import Downsample2D, FirDownsample2D, FirUpsample2D, ResnetBlock2D, Upsample2D
|
20 |
+
|
21 |
+
|
22 |
+
def get_down_block(
|
23 |
+
down_block_type,
|
24 |
+
num_layers,
|
25 |
+
in_channels,
|
26 |
+
out_channels,
|
27 |
+
temb_channels,
|
28 |
+
add_downsample,
|
29 |
+
resnet_eps,
|
30 |
+
resnet_act_fn,
|
31 |
+
attn_num_head_channels,
|
32 |
+
resnet_groups=None,
|
33 |
+
cross_attention_dim=None,
|
34 |
+
downsample_padding=None,
|
35 |
+
dual_cross_attention=False,
|
36 |
+
use_linear_projection=False,
|
37 |
+
only_cross_attention=False,
|
38 |
+
):
|
39 |
+
down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
|
40 |
+
if down_block_type == "DownBlock2D":
|
41 |
+
return DownBlock2D(
|
42 |
+
num_layers=num_layers,
|
43 |
+
in_channels=in_channels,
|
44 |
+
out_channels=out_channels,
|
45 |
+
temb_channels=temb_channels,
|
46 |
+
add_downsample=add_downsample,
|
47 |
+
resnet_eps=resnet_eps,
|
48 |
+
resnet_act_fn=resnet_act_fn,
|
49 |
+
resnet_groups=resnet_groups,
|
50 |
+
downsample_padding=downsample_padding,
|
51 |
+
)
|
52 |
+
elif down_block_type == "AttnDownBlock2D":
|
53 |
+
return AttnDownBlock2D(
|
54 |
+
num_layers=num_layers,
|
55 |
+
in_channels=in_channels,
|
56 |
+
out_channels=out_channels,
|
57 |
+
temb_channels=temb_channels,
|
58 |
+
add_downsample=add_downsample,
|
59 |
+
resnet_eps=resnet_eps,
|
60 |
+
resnet_act_fn=resnet_act_fn,
|
61 |
+
resnet_groups=resnet_groups,
|
62 |
+
downsample_padding=downsample_padding,
|
63 |
+
attn_num_head_channels=attn_num_head_channels,
|
64 |
+
)
|
65 |
+
elif down_block_type == "CrossAttnDownBlock2D":
|
66 |
+
if cross_attention_dim is None:
|
67 |
+
raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D")
|
68 |
+
return CrossAttnDownBlock2D(
|
69 |
+
num_layers=num_layers,
|
70 |
+
in_channels=in_channels,
|
71 |
+
out_channels=out_channels,
|
72 |
+
temb_channels=temb_channels,
|
73 |
+
add_downsample=add_downsample,
|
74 |
+
resnet_eps=resnet_eps,
|
75 |
+
resnet_act_fn=resnet_act_fn,
|
76 |
+
resnet_groups=resnet_groups,
|
77 |
+
downsample_padding=downsample_padding,
|
78 |
+
cross_attention_dim=cross_attention_dim,
|
79 |
+
attn_num_head_channels=attn_num_head_channels,
|
80 |
+
dual_cross_attention=dual_cross_attention,
|
81 |
+
use_linear_projection=use_linear_projection,
|
82 |
+
only_cross_attention=only_cross_attention,
|
83 |
+
)
|
84 |
+
elif down_block_type == "SkipDownBlock2D":
|
85 |
+
return SkipDownBlock2D(
|
86 |
+
num_layers=num_layers,
|
87 |
+
in_channels=in_channels,
|
88 |
+
out_channels=out_channels,
|
89 |
+
temb_channels=temb_channels,
|
90 |
+
add_downsample=add_downsample,
|
91 |
+
resnet_eps=resnet_eps,
|
92 |
+
resnet_act_fn=resnet_act_fn,
|
93 |
+
downsample_padding=downsample_padding,
|
94 |
+
)
|
95 |
+
elif down_block_type == "AttnSkipDownBlock2D":
|
96 |
+
return AttnSkipDownBlock2D(
|
97 |
+
num_layers=num_layers,
|
98 |
+
in_channels=in_channels,
|
99 |
+
out_channels=out_channels,
|
100 |
+
temb_channels=temb_channels,
|
101 |
+
add_downsample=add_downsample,
|
102 |
+
resnet_eps=resnet_eps,
|
103 |
+
resnet_act_fn=resnet_act_fn,
|
104 |
+
downsample_padding=downsample_padding,
|
105 |
+
attn_num_head_channels=attn_num_head_channels,
|
106 |
+
)
|
107 |
+
elif down_block_type == "DownEncoderBlock2D":
|
108 |
+
return DownEncoderBlock2D(
|
109 |
+
num_layers=num_layers,
|
110 |
+
in_channels=in_channels,
|
111 |
+
out_channels=out_channels,
|
112 |
+
add_downsample=add_downsample,
|
113 |
+
resnet_eps=resnet_eps,
|
114 |
+
resnet_act_fn=resnet_act_fn,
|
115 |
+
resnet_groups=resnet_groups,
|
116 |
+
downsample_padding=downsample_padding,
|
117 |
+
)
|
118 |
+
elif down_block_type == "AttnDownEncoderBlock2D":
|
119 |
+
return AttnDownEncoderBlock2D(
|
120 |
+
num_layers=num_layers,
|
121 |
+
in_channels=in_channels,
|
122 |
+
out_channels=out_channels,
|
123 |
+
add_downsample=add_downsample,
|
124 |
+
resnet_eps=resnet_eps,
|
125 |
+
resnet_act_fn=resnet_act_fn,
|
126 |
+
resnet_groups=resnet_groups,
|
127 |
+
downsample_padding=downsample_padding,
|
128 |
+
attn_num_head_channels=attn_num_head_channels,
|
129 |
+
)
|
130 |
+
raise ValueError(f"{down_block_type} does not exist.")
|
131 |
+
|
132 |
+
|
133 |
+
def get_up_block(
|
134 |
+
up_block_type,
|
135 |
+
num_layers,
|
136 |
+
in_channels,
|
137 |
+
out_channels,
|
138 |
+
prev_output_channel,
|
139 |
+
temb_channels,
|
140 |
+
add_upsample,
|
141 |
+
resnet_eps,
|
142 |
+
resnet_act_fn,
|
143 |
+
attn_num_head_channels,
|
144 |
+
resnet_groups=None,
|
145 |
+
cross_attention_dim=None,
|
146 |
+
dual_cross_attention=False,
|
147 |
+
use_linear_projection=False,
|
148 |
+
only_cross_attention=False,
|
149 |
+
):
|
150 |
+
up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
|
151 |
+
if up_block_type == "UpBlock2D":
|
152 |
+
return UpBlock2D(
|
153 |
+
num_layers=num_layers,
|
154 |
+
in_channels=in_channels,
|
155 |
+
out_channels=out_channels,
|
156 |
+
prev_output_channel=prev_output_channel,
|
157 |
+
temb_channels=temb_channels,
|
158 |
+
add_upsample=add_upsample,
|
159 |
+
resnet_eps=resnet_eps,
|
160 |
+
resnet_act_fn=resnet_act_fn,
|
161 |
+
resnet_groups=resnet_groups,
|
162 |
+
)
|
163 |
+
elif up_block_type == "CrossAttnUpBlock2D":
|
164 |
+
if cross_attention_dim is None:
|
165 |
+
raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock2D")
|
166 |
+
return CrossAttnUpBlock2D(
|
167 |
+
num_layers=num_layers,
|
168 |
+
in_channels=in_channels,
|
169 |
+
out_channels=out_channels,
|
170 |
+
prev_output_channel=prev_output_channel,
|
171 |
+
temb_channels=temb_channels,
|
172 |
+
add_upsample=add_upsample,
|
173 |
+
resnet_eps=resnet_eps,
|
174 |
+
resnet_act_fn=resnet_act_fn,
|
175 |
+
resnet_groups=resnet_groups,
|
176 |
+
cross_attention_dim=cross_attention_dim,
|
177 |
+
attn_num_head_channels=attn_num_head_channels,
|
178 |
+
dual_cross_attention=dual_cross_attention,
|
179 |
+
use_linear_projection=use_linear_projection,
|
180 |
+
only_cross_attention=only_cross_attention,
|
181 |
+
)
|
182 |
+
elif up_block_type == "AttnUpBlock2D":
|
183 |
+
return AttnUpBlock2D(
|
184 |
+
num_layers=num_layers,
|
185 |
+
in_channels=in_channels,
|
186 |
+
out_channels=out_channels,
|
187 |
+
prev_output_channel=prev_output_channel,
|
188 |
+
temb_channels=temb_channels,
|
189 |
+
add_upsample=add_upsample,
|
190 |
+
resnet_eps=resnet_eps,
|
191 |
+
resnet_act_fn=resnet_act_fn,
|
192 |
+
resnet_groups=resnet_groups,
|
193 |
+
attn_num_head_channels=attn_num_head_channels,
|
194 |
+
)
|
195 |
+
elif up_block_type == "SkipUpBlock2D":
|
196 |
+
return SkipUpBlock2D(
|
197 |
+
num_layers=num_layers,
|
198 |
+
in_channels=in_channels,
|
199 |
+
out_channels=out_channels,
|
200 |
+
prev_output_channel=prev_output_channel,
|
201 |
+
temb_channels=temb_channels,
|
202 |
+
add_upsample=add_upsample,
|
203 |
+
resnet_eps=resnet_eps,
|
204 |
+
resnet_act_fn=resnet_act_fn,
|
205 |
+
)
|
206 |
+
elif up_block_type == "AttnSkipUpBlock2D":
|
207 |
+
return AttnSkipUpBlock2D(
|
208 |
+
num_layers=num_layers,
|
209 |
+
in_channels=in_channels,
|
210 |
+
out_channels=out_channels,
|
211 |
+
prev_output_channel=prev_output_channel,
|
212 |
+
temb_channels=temb_channels,
|
213 |
+
add_upsample=add_upsample,
|
214 |
+
resnet_eps=resnet_eps,
|
215 |
+
resnet_act_fn=resnet_act_fn,
|
216 |
+
attn_num_head_channels=attn_num_head_channels,
|
217 |
+
)
|
218 |
+
elif up_block_type == "UpDecoderBlock2D":
|
219 |
+
return UpDecoderBlock2D(
|
220 |
+
num_layers=num_layers,
|
221 |
+
in_channels=in_channels,
|
222 |
+
out_channels=out_channels,
|
223 |
+
add_upsample=add_upsample,
|
224 |
+
resnet_eps=resnet_eps,
|
225 |
+
resnet_act_fn=resnet_act_fn,
|
226 |
+
resnet_groups=resnet_groups,
|
227 |
+
)
|
228 |
+
elif up_block_type == "AttnUpDecoderBlock2D":
|
229 |
+
return AttnUpDecoderBlock2D(
|
230 |
+
num_layers=num_layers,
|
231 |
+
in_channels=in_channels,
|
232 |
+
out_channels=out_channels,
|
233 |
+
add_upsample=add_upsample,
|
234 |
+
resnet_eps=resnet_eps,
|
235 |
+
resnet_act_fn=resnet_act_fn,
|
236 |
+
resnet_groups=resnet_groups,
|
237 |
+
attn_num_head_channels=attn_num_head_channels,
|
238 |
+
)
|
239 |
+
raise ValueError(f"{up_block_type} does not exist.")
|
240 |
+
|
241 |
+
|
242 |
+
class UNetMidBlock2D(nn.Module):
|
243 |
+
def __init__(
|
244 |
+
self,
|
245 |
+
in_channels: int,
|
246 |
+
temb_channels: int,
|
247 |
+
dropout: float = 0.0,
|
248 |
+
num_layers: int = 1,
|
249 |
+
resnet_eps: float = 1e-6,
|
250 |
+
resnet_time_scale_shift: str = "default",
|
251 |
+
resnet_act_fn: str = "swish",
|
252 |
+
resnet_groups: int = 32,
|
253 |
+
resnet_pre_norm: bool = True,
|
254 |
+
attn_num_head_channels=1,
|
255 |
+
attention_type="default",
|
256 |
+
output_scale_factor=1.0,
|
257 |
+
):
|
258 |
+
super().__init__()
|
259 |
+
|
260 |
+
self.attention_type = attention_type
|
261 |
+
resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
|
262 |
+
|
263 |
+
# there is always at least one resnet
|
264 |
+
resnets = [
|
265 |
+
ResnetBlock2D(
|
266 |
+
in_channels=in_channels,
|
267 |
+
out_channels=in_channels,
|
268 |
+
temb_channels=temb_channels,
|
269 |
+
eps=resnet_eps,
|
270 |
+
groups=resnet_groups,
|
271 |
+
dropout=dropout,
|
272 |
+
time_embedding_norm=resnet_time_scale_shift,
|
273 |
+
non_linearity=resnet_act_fn,
|
274 |
+
output_scale_factor=output_scale_factor,
|
275 |
+
pre_norm=resnet_pre_norm,
|
276 |
+
)
|
277 |
+
]
|
278 |
+
attentions = []
|
279 |
+
|
280 |
+
for _ in range(num_layers):
|
281 |
+
attentions.append(
|
282 |
+
AttentionBlock(
|
283 |
+
in_channels,
|
284 |
+
num_head_channels=attn_num_head_channels,
|
285 |
+
rescale_output_factor=output_scale_factor,
|
286 |
+
eps=resnet_eps,
|
287 |
+
norm_num_groups=resnet_groups,
|
288 |
+
)
|
289 |
+
)
|
290 |
+
resnets.append(
|
291 |
+
ResnetBlock2D(
|
292 |
+
in_channels=in_channels,
|
293 |
+
out_channels=in_channels,
|
294 |
+
temb_channels=temb_channels,
|
295 |
+
eps=resnet_eps,
|
296 |
+
groups=resnet_groups,
|
297 |
+
dropout=dropout,
|
298 |
+
time_embedding_norm=resnet_time_scale_shift,
|
299 |
+
non_linearity=resnet_act_fn,
|
300 |
+
output_scale_factor=output_scale_factor,
|
301 |
+
pre_norm=resnet_pre_norm,
|
302 |
+
)
|
303 |
+
)
|
304 |
+
|
305 |
+
self.attentions = nn.ModuleList(attentions)
|
306 |
+
self.resnets = nn.ModuleList(resnets)
|
307 |
+
|
308 |
+
def forward(self, hidden_states, temb=None, encoder_states=None):
|
309 |
+
hidden_states = self.resnets[0](hidden_states, temb)
|
310 |
+
for attn, resnet in zip(self.attentions, self.resnets[1:]):
|
311 |
+
if self.attention_type == "default":
|
312 |
+
hidden_states = attn(hidden_states)
|
313 |
+
else:
|
314 |
+
hidden_states = attn(hidden_states, encoder_states)
|
315 |
+
hidden_states = resnet(hidden_states, temb)
|
316 |
+
|
317 |
+
return hidden_states
|
318 |
+
|
319 |
+
|
320 |
+
class UNetMidBlock2DCrossAttn(nn.Module):
|
321 |
+
def __init__(
|
322 |
+
self,
|
323 |
+
in_channels: int,
|
324 |
+
temb_channels: int,
|
325 |
+
dropout: float = 0.0,
|
326 |
+
num_layers: int = 1,
|
327 |
+
resnet_eps: float = 1e-6,
|
328 |
+
resnet_time_scale_shift: str = "default",
|
329 |
+
resnet_act_fn: str = "swish",
|
330 |
+
resnet_groups: int = 32,
|
331 |
+
resnet_pre_norm: bool = True,
|
332 |
+
attn_num_head_channels=1,
|
333 |
+
attention_type="default",
|
334 |
+
output_scale_factor=1.0,
|
335 |
+
cross_attention_dim=1280,
|
336 |
+
dual_cross_attention=False,
|
337 |
+
use_linear_projection=False,
|
338 |
+
):
|
339 |
+
super().__init__()
|
340 |
+
|
341 |
+
self.attention_type = attention_type
|
342 |
+
self.attn_num_head_channels = attn_num_head_channels
|
343 |
+
resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
|
344 |
+
|
345 |
+
# there is always at least one resnet
|
346 |
+
resnets = [
|
347 |
+
ResnetBlock2D(
|
348 |
+
in_channels=in_channels,
|
349 |
+
out_channels=in_channels,
|
350 |
+
temb_channels=temb_channels,
|
351 |
+
eps=resnet_eps,
|
352 |
+
groups=resnet_groups,
|
353 |
+
dropout=dropout,
|
354 |
+
time_embedding_norm=resnet_time_scale_shift,
|
355 |
+
non_linearity=resnet_act_fn,
|
356 |
+
output_scale_factor=output_scale_factor,
|
357 |
+
pre_norm=resnet_pre_norm,
|
358 |
+
)
|
359 |
+
]
|
360 |
+
attentions = []
|
361 |
+
|
362 |
+
for _ in range(num_layers):
|
363 |
+
if not dual_cross_attention:
|
364 |
+
attentions.append(
|
365 |
+
Transformer2DModel(
|
366 |
+
attn_num_head_channels,
|
367 |
+
in_channels // attn_num_head_channels,
|
368 |
+
in_channels=in_channels,
|
369 |
+
num_layers=1,
|
370 |
+
cross_attention_dim=cross_attention_dim,
|
371 |
+
norm_num_groups=resnet_groups,
|
372 |
+
use_linear_projection=use_linear_projection,
|
373 |
+
)
|
374 |
+
)
|
375 |
+
else:
|
376 |
+
attentions.append(
|
377 |
+
DualTransformer2DModel(
|
378 |
+
attn_num_head_channels,
|
379 |
+
in_channels // attn_num_head_channels,
|
380 |
+
in_channels=in_channels,
|
381 |
+
num_layers=1,
|
382 |
+
cross_attention_dim=cross_attention_dim,
|
383 |
+
norm_num_groups=resnet_groups,
|
384 |
+
)
|
385 |
+
)
|
386 |
+
resnets.append(
|
387 |
+
ResnetBlock2D(
|
388 |
+
in_channels=in_channels,
|
389 |
+
out_channels=in_channels,
|
390 |
+
temb_channels=temb_channels,
|
391 |
+
eps=resnet_eps,
|
392 |
+
groups=resnet_groups,
|
393 |
+
dropout=dropout,
|
394 |
+
time_embedding_norm=resnet_time_scale_shift,
|
395 |
+
non_linearity=resnet_act_fn,
|
396 |
+
output_scale_factor=output_scale_factor,
|
397 |
+
pre_norm=resnet_pre_norm,
|
398 |
+
)
|
399 |
+
)
|
400 |
+
|
401 |
+
self.attentions = nn.ModuleList(attentions)
|
402 |
+
self.resnets = nn.ModuleList(resnets)
|
403 |
+
|
404 |
+
def set_attention_slice(self, slice_size):
|
405 |
+
head_dims = self.attn_num_head_channels
|
406 |
+
head_dims = [head_dims] if isinstance(head_dims, int) else head_dims
|
407 |
+
if slice_size is not None and any(dim % slice_size != 0 for dim in head_dims):
|
408 |
+
raise ValueError(
|
409 |
+
f"Make sure slice_size {slice_size} is a common divisor of "
|
410 |
+
f"the number of heads used in cross_attention: {head_dims}"
|
411 |
+
)
|
412 |
+
if slice_size is not None and slice_size > min(head_dims):
|
413 |
+
raise ValueError(
|
414 |
+
f"slice_size {slice_size} has to be smaller or equal to "
|
415 |
+
f"the lowest number of heads used in cross_attention: min({head_dims}) = {min(head_dims)}"
|
416 |
+
)
|
417 |
+
|
418 |
+
for attn in self.attentions:
|
419 |
+
attn._set_attention_slice(slice_size)
|
420 |
+
|
421 |
+
def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
|
422 |
+
for attn in self.attentions:
|
423 |
+
attn._set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers)
|
424 |
+
|
425 |
+
def forward(self, hidden_states, temb=None, encoder_hidden_states=None,
|
426 |
+
text_format_dict={}):
|
427 |
+
hidden_states = self.resnets[0](hidden_states, temb)
|
428 |
+
for attn, resnet in zip(self.attentions, self.resnets[1:]):
|
429 |
+
hidden_states = attn(hidden_states, encoder_hidden_states,
|
430 |
+
text_format_dict).sample
|
431 |
+
hidden_states = resnet(hidden_states, temb)
|
432 |
+
|
433 |
+
return hidden_states
|
434 |
+
|
435 |
+
|
436 |
+
class AttnDownBlock2D(nn.Module):
|
437 |
+
def __init__(
|
438 |
+
self,
|
439 |
+
in_channels: int,
|
440 |
+
out_channels: int,
|
441 |
+
temb_channels: int,
|
442 |
+
dropout: float = 0.0,
|
443 |
+
num_layers: int = 1,
|
444 |
+
resnet_eps: float = 1e-6,
|
445 |
+
resnet_time_scale_shift: str = "default",
|
446 |
+
resnet_act_fn: str = "swish",
|
447 |
+
resnet_groups: int = 32,
|
448 |
+
resnet_pre_norm: bool = True,
|
449 |
+
attn_num_head_channels=1,
|
450 |
+
attention_type="default",
|
451 |
+
output_scale_factor=1.0,
|
452 |
+
downsample_padding=1,
|
453 |
+
add_downsample=True,
|
454 |
+
):
|
455 |
+
super().__init__()
|
456 |
+
resnets = []
|
457 |
+
attentions = []
|
458 |
+
|
459 |
+
self.attention_type = attention_type
|
460 |
+
|
461 |
+
for i in range(num_layers):
|
462 |
+
in_channels = in_channels if i == 0 else out_channels
|
463 |
+
resnets.append(
|
464 |
+
ResnetBlock2D(
|
465 |
+
in_channels=in_channels,
|
466 |
+
out_channels=out_channels,
|
467 |
+
temb_channels=temb_channels,
|
468 |
+
eps=resnet_eps,
|
469 |
+
groups=resnet_groups,
|
470 |
+
dropout=dropout,
|
471 |
+
time_embedding_norm=resnet_time_scale_shift,
|
472 |
+
non_linearity=resnet_act_fn,
|
473 |
+
output_scale_factor=output_scale_factor,
|
474 |
+
pre_norm=resnet_pre_norm,
|
475 |
+
)
|
476 |
+
)
|
477 |
+
attentions.append(
|
478 |
+
AttentionBlock(
|
479 |
+
out_channels,
|
480 |
+
num_head_channels=attn_num_head_channels,
|
481 |
+
rescale_output_factor=output_scale_factor,
|
482 |
+
eps=resnet_eps,
|
483 |
+
norm_num_groups=resnet_groups,
|
484 |
+
)
|
485 |
+
)
|
486 |
+
|
487 |
+
self.attentions = nn.ModuleList(attentions)
|
488 |
+
self.resnets = nn.ModuleList(resnets)
|
489 |
+
|
490 |
+
if add_downsample:
|
491 |
+
self.downsamplers = nn.ModuleList(
|
492 |
+
[
|
493 |
+
Downsample2D(
|
494 |
+
out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
|
495 |
+
)
|
496 |
+
]
|
497 |
+
)
|
498 |
+
else:
|
499 |
+
self.downsamplers = None
|
500 |
+
|
501 |
+
def forward(self, hidden_states, temb=None):
|
502 |
+
output_states = ()
|
503 |
+
|
504 |
+
for resnet, attn in zip(self.resnets, self.attentions):
|
505 |
+
hidden_states = resnet(hidden_states, temb)
|
506 |
+
hidden_states = attn(hidden_states)
|
507 |
+
output_states += (hidden_states,)
|
508 |
+
|
509 |
+
if self.downsamplers is not None:
|
510 |
+
for downsampler in self.downsamplers:
|
511 |
+
hidden_states = downsampler(hidden_states)
|
512 |
+
|
513 |
+
output_states += (hidden_states,)
|
514 |
+
|
515 |
+
return hidden_states, output_states
|
516 |
+
|
517 |
+
|
518 |
+
class CrossAttnDownBlock2D(nn.Module):
|
519 |
+
def __init__(
|
520 |
+
self,
|
521 |
+
in_channels: int,
|
522 |
+
out_channels: int,
|
523 |
+
temb_channels: int,
|
524 |
+
dropout: float = 0.0,
|
525 |
+
num_layers: int = 1,
|
526 |
+
resnet_eps: float = 1e-6,
|
527 |
+
resnet_time_scale_shift: str = "default",
|
528 |
+
resnet_act_fn: str = "swish",
|
529 |
+
resnet_groups: int = 32,
|
530 |
+
resnet_pre_norm: bool = True,
|
531 |
+
attn_num_head_channels=1,
|
532 |
+
cross_attention_dim=1280,
|
533 |
+
attention_type="default",
|
534 |
+
output_scale_factor=1.0,
|
535 |
+
downsample_padding=1,
|
536 |
+
add_downsample=True,
|
537 |
+
dual_cross_attention=False,
|
538 |
+
use_linear_projection=False,
|
539 |
+
only_cross_attention=False,
|
540 |
+
):
|
541 |
+
super().__init__()
|
542 |
+
resnets = []
|
543 |
+
attentions = []
|
544 |
+
|
545 |
+
self.attention_type = attention_type
|
546 |
+
self.attn_num_head_channels = attn_num_head_channels
|
547 |
+
|
548 |
+
for i in range(num_layers):
|
549 |
+
in_channels = in_channels if i == 0 else out_channels
|
550 |
+
resnets.append(
|
551 |
+
ResnetBlock2D(
|
552 |
+
in_channels=in_channels,
|
553 |
+
out_channels=out_channels,
|
554 |
+
temb_channels=temb_channels,
|
555 |
+
eps=resnet_eps,
|
556 |
+
groups=resnet_groups,
|
557 |
+
dropout=dropout,
|
558 |
+
time_embedding_norm=resnet_time_scale_shift,
|
559 |
+
non_linearity=resnet_act_fn,
|
560 |
+
output_scale_factor=output_scale_factor,
|
561 |
+
pre_norm=resnet_pre_norm,
|
562 |
+
)
|
563 |
+
)
|
564 |
+
if not dual_cross_attention:
|
565 |
+
attentions.append(
|
566 |
+
Transformer2DModel(
|
567 |
+
attn_num_head_channels,
|
568 |
+
out_channels // attn_num_head_channels,
|
569 |
+
in_channels=out_channels,
|
570 |
+
num_layers=1,
|
571 |
+
cross_attention_dim=cross_attention_dim,
|
572 |
+
norm_num_groups=resnet_groups,
|
573 |
+
use_linear_projection=use_linear_projection,
|
574 |
+
only_cross_attention=only_cross_attention,
|
575 |
+
)
|
576 |
+
)
|
577 |
+
else:
|
578 |
+
attentions.append(
|
579 |
+
DualTransformer2DModel(
|
580 |
+
attn_num_head_channels,
|
581 |
+
out_channels // attn_num_head_channels,
|
582 |
+
in_channels=out_channels,
|
583 |
+
num_layers=1,
|
584 |
+
cross_attention_dim=cross_attention_dim,
|
585 |
+
norm_num_groups=resnet_groups,
|
586 |
+
)
|
587 |
+
)
|
588 |
+
self.attentions = nn.ModuleList(attentions)
|
589 |
+
self.resnets = nn.ModuleList(resnets)
|
590 |
+
|
591 |
+
if add_downsample:
|
592 |
+
self.downsamplers = nn.ModuleList(
|
593 |
+
[
|
594 |
+
Downsample2D(
|
595 |
+
out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
|
596 |
+
)
|
597 |
+
]
|
598 |
+
)
|
599 |
+
else:
|
600 |
+
self.downsamplers = None
|
601 |
+
|
602 |
+
self.gradient_checkpointing = False
|
603 |
+
|
604 |
+
def set_attention_slice(self, slice_size):
|
605 |
+
head_dims = self.attn_num_head_channels
|
606 |
+
head_dims = [head_dims] if isinstance(head_dims, int) else head_dims
|
607 |
+
if slice_size is not None and any(dim % slice_size != 0 for dim in head_dims):
|
608 |
+
raise ValueError(
|
609 |
+
f"Make sure slice_size {slice_size} is a common divisor of "
|
610 |
+
f"the number of heads used in cross_attention: {head_dims}"
|
611 |
+
)
|
612 |
+
if slice_size is not None and slice_size > min(head_dims):
|
613 |
+
raise ValueError(
|
614 |
+
f"slice_size {slice_size} has to be smaller or equal to "
|
615 |
+
f"the lowest number of heads used in cross_attention: min({head_dims}) = {min(head_dims)}"
|
616 |
+
)
|
617 |
+
|
618 |
+
for attn in self.attentions:
|
619 |
+
attn._set_attention_slice(slice_size)
|
620 |
+
|
621 |
+
def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
|
622 |
+
for attn in self.attentions:
|
623 |
+
attn._set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers)
|
624 |
+
|
625 |
+
def forward(self, hidden_states, temb=None, encoder_hidden_states=None,
|
626 |
+
text_format_dict={}):
|
627 |
+
output_states = ()
|
628 |
+
|
629 |
+
for resnet, attn in zip(self.resnets, self.attentions):
|
630 |
+
if self.training and self.gradient_checkpointing:
|
631 |
+
|
632 |
+
def create_custom_forward(module, return_dict=None):
|
633 |
+
def custom_forward(*inputs):
|
634 |
+
if return_dict is not None:
|
635 |
+
return module(*inputs, return_dict=return_dict)
|
636 |
+
else:
|
637 |
+
return module(*inputs)
|
638 |
+
|
639 |
+
return custom_forward
|
640 |
+
|
641 |
+
hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
|
642 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
643 |
+
create_custom_forward(attn, return_dict=False), hidden_states, encoder_hidden_states,
|
644 |
+
text_format_dict
|
645 |
+
)[0]
|
646 |
+
else:
|
647 |
+
hidden_states = resnet(hidden_states, temb)
|
648 |
+
hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states,
|
649 |
+
text_format_dict=text_format_dict).sample
|
650 |
+
|
651 |
+
output_states += (hidden_states,)
|
652 |
+
|
653 |
+
if self.downsamplers is not None:
|
654 |
+
for downsampler in self.downsamplers:
|
655 |
+
hidden_states = downsampler(hidden_states)
|
656 |
+
|
657 |
+
output_states += (hidden_states,)
|
658 |
+
|
659 |
+
return hidden_states, output_states
|
660 |
+
|
661 |
+
|
662 |
+
class DownBlock2D(nn.Module):
|
663 |
+
def __init__(
|
664 |
+
self,
|
665 |
+
in_channels: int,
|
666 |
+
out_channels: int,
|
667 |
+
temb_channels: int,
|
668 |
+
dropout: float = 0.0,
|
669 |
+
num_layers: int = 1,
|
670 |
+
resnet_eps: float = 1e-6,
|
671 |
+
resnet_time_scale_shift: str = "default",
|
672 |
+
resnet_act_fn: str = "swish",
|
673 |
+
resnet_groups: int = 32,
|
674 |
+
resnet_pre_norm: bool = True,
|
675 |
+
output_scale_factor=1.0,
|
676 |
+
add_downsample=True,
|
677 |
+
downsample_padding=1,
|
678 |
+
):
|
679 |
+
super().__init__()
|
680 |
+
resnets = []
|
681 |
+
|
682 |
+
for i in range(num_layers):
|
683 |
+
in_channels = in_channels if i == 0 else out_channels
|
684 |
+
resnets.append(
|
685 |
+
ResnetBlock2D(
|
686 |
+
in_channels=in_channels,
|
687 |
+
out_channels=out_channels,
|
688 |
+
temb_channels=temb_channels,
|
689 |
+
eps=resnet_eps,
|
690 |
+
groups=resnet_groups,
|
691 |
+
dropout=dropout,
|
692 |
+
time_embedding_norm=resnet_time_scale_shift,
|
693 |
+
non_linearity=resnet_act_fn,
|
694 |
+
output_scale_factor=output_scale_factor,
|
695 |
+
pre_norm=resnet_pre_norm,
|
696 |
+
)
|
697 |
+
)
|
698 |
+
|
699 |
+
self.resnets = nn.ModuleList(resnets)
|
700 |
+
|
701 |
+
if add_downsample:
|
702 |
+
self.downsamplers = nn.ModuleList(
|
703 |
+
[
|
704 |
+
Downsample2D(
|
705 |
+
out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
|
706 |
+
)
|
707 |
+
]
|
708 |
+
)
|
709 |
+
else:
|
710 |
+
self.downsamplers = None
|
711 |
+
|
712 |
+
self.gradient_checkpointing = False
|
713 |
+
|
714 |
+
def forward(self, hidden_states, temb=None):
|
715 |
+
output_states = ()
|
716 |
+
|
717 |
+
for resnet in self.resnets:
|
718 |
+
if self.training and self.gradient_checkpointing:
|
719 |
+
|
720 |
+
def create_custom_forward(module):
|
721 |
+
def custom_forward(*inputs):
|
722 |
+
return module(*inputs)
|
723 |
+
|
724 |
+
return custom_forward
|
725 |
+
|
726 |
+
hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
|
727 |
+
else:
|
728 |
+
hidden_states = resnet(hidden_states, temb)
|
729 |
+
|
730 |
+
output_states += (hidden_states,)
|
731 |
+
|
732 |
+
if self.downsamplers is not None:
|
733 |
+
for downsampler in self.downsamplers:
|
734 |
+
hidden_states = downsampler(hidden_states)
|
735 |
+
|
736 |
+
output_states += (hidden_states,)
|
737 |
+
|
738 |
+
return hidden_states, output_states
|
739 |
+
|
740 |
+
|
741 |
+
class DownEncoderBlock2D(nn.Module):
|
742 |
+
def __init__(
|
743 |
+
self,
|
744 |
+
in_channels: int,
|
745 |
+
out_channels: int,
|
746 |
+
dropout: float = 0.0,
|
747 |
+
num_layers: int = 1,
|
748 |
+
resnet_eps: float = 1e-6,
|
749 |
+
resnet_time_scale_shift: str = "default",
|
750 |
+
resnet_act_fn: str = "swish",
|
751 |
+
resnet_groups: int = 32,
|
752 |
+
resnet_pre_norm: bool = True,
|
753 |
+
output_scale_factor=1.0,
|
754 |
+
add_downsample=True,
|
755 |
+
downsample_padding=1,
|
756 |
+
):
|
757 |
+
super().__init__()
|
758 |
+
resnets = []
|
759 |
+
|
760 |
+
for i in range(num_layers):
|
761 |
+
in_channels = in_channels if i == 0 else out_channels
|
762 |
+
resnets.append(
|
763 |
+
ResnetBlock2D(
|
764 |
+
in_channels=in_channels,
|
765 |
+
out_channels=out_channels,
|
766 |
+
temb_channels=None,
|
767 |
+
eps=resnet_eps,
|
768 |
+
groups=resnet_groups,
|
769 |
+
dropout=dropout,
|
770 |
+
time_embedding_norm=resnet_time_scale_shift,
|
771 |
+
non_linearity=resnet_act_fn,
|
772 |
+
output_scale_factor=output_scale_factor,
|
773 |
+
pre_norm=resnet_pre_norm,
|
774 |
+
)
|
775 |
+
)
|
776 |
+
|
777 |
+
self.resnets = nn.ModuleList(resnets)
|
778 |
+
|
779 |
+
if add_downsample:
|
780 |
+
self.downsamplers = nn.ModuleList(
|
781 |
+
[
|
782 |
+
Downsample2D(
|
783 |
+
out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
|
784 |
+
)
|
785 |
+
]
|
786 |
+
)
|
787 |
+
else:
|
788 |
+
self.downsamplers = None
|
789 |
+
|
790 |
+
def forward(self, hidden_states):
|
791 |
+
for resnet in self.resnets:
|
792 |
+
hidden_states = resnet(hidden_states, temb=None)
|
793 |
+
|
794 |
+
if self.downsamplers is not None:
|
795 |
+
for downsampler in self.downsamplers:
|
796 |
+
hidden_states = downsampler(hidden_states)
|
797 |
+
|
798 |
+
return hidden_states
|
799 |
+
|
800 |
+
|
801 |
+
class AttnDownEncoderBlock2D(nn.Module):
|
802 |
+
def __init__(
|
803 |
+
self,
|
804 |
+
in_channels: int,
|
805 |
+
out_channels: int,
|
806 |
+
dropout: float = 0.0,
|
807 |
+
num_layers: int = 1,
|
808 |
+
resnet_eps: float = 1e-6,
|
809 |
+
resnet_time_scale_shift: str = "default",
|
810 |
+
resnet_act_fn: str = "swish",
|
811 |
+
resnet_groups: int = 32,
|
812 |
+
resnet_pre_norm: bool = True,
|
813 |
+
attn_num_head_channels=1,
|
814 |
+
output_scale_factor=1.0,
|
815 |
+
add_downsample=True,
|
816 |
+
downsample_padding=1,
|
817 |
+
):
|
818 |
+
super().__init__()
|
819 |
+
resnets = []
|
820 |
+
attentions = []
|
821 |
+
|
822 |
+
for i in range(num_layers):
|
823 |
+
in_channels = in_channels if i == 0 else out_channels
|
824 |
+
resnets.append(
|
825 |
+
ResnetBlock2D(
|
826 |
+
in_channels=in_channels,
|
827 |
+
out_channels=out_channels,
|
828 |
+
temb_channels=None,
|
829 |
+
eps=resnet_eps,
|
830 |
+
groups=resnet_groups,
|
831 |
+
dropout=dropout,
|
832 |
+
time_embedding_norm=resnet_time_scale_shift,
|
833 |
+
non_linearity=resnet_act_fn,
|
834 |
+
output_scale_factor=output_scale_factor,
|
835 |
+
pre_norm=resnet_pre_norm,
|
836 |
+
)
|
837 |
+
)
|
838 |
+
attentions.append(
|
839 |
+
AttentionBlock(
|
840 |
+
out_channels,
|
841 |
+
num_head_channels=attn_num_head_channels,
|
842 |
+
rescale_output_factor=output_scale_factor,
|
843 |
+
eps=resnet_eps,
|
844 |
+
norm_num_groups=resnet_groups,
|
845 |
+
)
|
846 |
+
)
|
847 |
+
|
848 |
+
self.attentions = nn.ModuleList(attentions)
|
849 |
+
self.resnets = nn.ModuleList(resnets)
|
850 |
+
|
851 |
+
if add_downsample:
|
852 |
+
self.downsamplers = nn.ModuleList(
|
853 |
+
[
|
854 |
+
Downsample2D(
|
855 |
+
out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
|
856 |
+
)
|
857 |
+
]
|
858 |
+
)
|
859 |
+
else:
|
860 |
+
self.downsamplers = None
|
861 |
+
|
862 |
+
def forward(self, hidden_states):
|
863 |
+
for resnet, attn in zip(self.resnets, self.attentions):
|
864 |
+
hidden_states = resnet(hidden_states, temb=None)
|
865 |
+
hidden_states = attn(hidden_states)
|
866 |
+
|
867 |
+
if self.downsamplers is not None:
|
868 |
+
for downsampler in self.downsamplers:
|
869 |
+
hidden_states = downsampler(hidden_states)
|
870 |
+
|
871 |
+
return hidden_states
|
872 |
+
|
873 |
+
|
874 |
+
class AttnSkipDownBlock2D(nn.Module):
|
875 |
+
def __init__(
|
876 |
+
self,
|
877 |
+
in_channels: int,
|
878 |
+
out_channels: int,
|
879 |
+
temb_channels: int,
|
880 |
+
dropout: float = 0.0,
|
881 |
+
num_layers: int = 1,
|
882 |
+
resnet_eps: float = 1e-6,
|
883 |
+
resnet_time_scale_shift: str = "default",
|
884 |
+
resnet_act_fn: str = "swish",
|
885 |
+
resnet_pre_norm: bool = True,
|
886 |
+
attn_num_head_channels=1,
|
887 |
+
attention_type="default",
|
888 |
+
output_scale_factor=np.sqrt(2.0),
|
889 |
+
downsample_padding=1,
|
890 |
+
add_downsample=True,
|
891 |
+
):
|
892 |
+
super().__init__()
|
893 |
+
self.attentions = nn.ModuleList([])
|
894 |
+
self.resnets = nn.ModuleList([])
|
895 |
+
|
896 |
+
self.attention_type = attention_type
|
897 |
+
|
898 |
+
for i in range(num_layers):
|
899 |
+
in_channels = in_channels if i == 0 else out_channels
|
900 |
+
self.resnets.append(
|
901 |
+
ResnetBlock2D(
|
902 |
+
in_channels=in_channels,
|
903 |
+
out_channels=out_channels,
|
904 |
+
temb_channels=temb_channels,
|
905 |
+
eps=resnet_eps,
|
906 |
+
groups=min(in_channels // 4, 32),
|
907 |
+
groups_out=min(out_channels // 4, 32),
|
908 |
+
dropout=dropout,
|
909 |
+
time_embedding_norm=resnet_time_scale_shift,
|
910 |
+
non_linearity=resnet_act_fn,
|
911 |
+
output_scale_factor=output_scale_factor,
|
912 |
+
pre_norm=resnet_pre_norm,
|
913 |
+
)
|
914 |
+
)
|
915 |
+
self.attentions.append(
|
916 |
+
AttentionBlock(
|
917 |
+
out_channels,
|
918 |
+
num_head_channels=attn_num_head_channels,
|
919 |
+
rescale_output_factor=output_scale_factor,
|
920 |
+
eps=resnet_eps,
|
921 |
+
)
|
922 |
+
)
|
923 |
+
|
924 |
+
if add_downsample:
|
925 |
+
self.resnet_down = ResnetBlock2D(
|
926 |
+
in_channels=out_channels,
|
927 |
+
out_channels=out_channels,
|
928 |
+
temb_channels=temb_channels,
|
929 |
+
eps=resnet_eps,
|
930 |
+
groups=min(out_channels // 4, 32),
|
931 |
+
dropout=dropout,
|
932 |
+
time_embedding_norm=resnet_time_scale_shift,
|
933 |
+
non_linearity=resnet_act_fn,
|
934 |
+
output_scale_factor=output_scale_factor,
|
935 |
+
pre_norm=resnet_pre_norm,
|
936 |
+
use_in_shortcut=True,
|
937 |
+
down=True,
|
938 |
+
kernel="fir",
|
939 |
+
)
|
940 |
+
self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)])
|
941 |
+
self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1))
|
942 |
+
else:
|
943 |
+
self.resnet_down = None
|
944 |
+
self.downsamplers = None
|
945 |
+
self.skip_conv = None
|
946 |
+
|
947 |
+
def forward(self, hidden_states, temb=None, skip_sample=None):
|
948 |
+
output_states = ()
|
949 |
+
|
950 |
+
for resnet, attn in zip(self.resnets, self.attentions):
|
951 |
+
hidden_states = resnet(hidden_states, temb)
|
952 |
+
hidden_states = attn(hidden_states)
|
953 |
+
output_states += (hidden_states,)
|
954 |
+
|
955 |
+
if self.downsamplers is not None:
|
956 |
+
hidden_states = self.resnet_down(hidden_states, temb)
|
957 |
+
for downsampler in self.downsamplers:
|
958 |
+
skip_sample = downsampler(skip_sample)
|
959 |
+
|
960 |
+
hidden_states = self.skip_conv(skip_sample) + hidden_states
|
961 |
+
|
962 |
+
output_states += (hidden_states,)
|
963 |
+
|
964 |
+
return hidden_states, output_states, skip_sample
|
965 |
+
|
966 |
+
|
967 |
+
class SkipDownBlock2D(nn.Module):
|
968 |
+
def __init__(
|
969 |
+
self,
|
970 |
+
in_channels: int,
|
971 |
+
out_channels: int,
|
972 |
+
temb_channels: int,
|
973 |
+
dropout: float = 0.0,
|
974 |
+
num_layers: int = 1,
|
975 |
+
resnet_eps: float = 1e-6,
|
976 |
+
resnet_time_scale_shift: str = "default",
|
977 |
+
resnet_act_fn: str = "swish",
|
978 |
+
resnet_pre_norm: bool = True,
|
979 |
+
output_scale_factor=np.sqrt(2.0),
|
980 |
+
add_downsample=True,
|
981 |
+
downsample_padding=1,
|
982 |
+
):
|
983 |
+
super().__init__()
|
984 |
+
self.resnets = nn.ModuleList([])
|
985 |
+
|
986 |
+
for i in range(num_layers):
|
987 |
+
in_channels = in_channels if i == 0 else out_channels
|
988 |
+
self.resnets.append(
|
989 |
+
ResnetBlock2D(
|
990 |
+
in_channels=in_channels,
|
991 |
+
out_channels=out_channels,
|
992 |
+
temb_channels=temb_channels,
|
993 |
+
eps=resnet_eps,
|
994 |
+
groups=min(in_channels // 4, 32),
|
995 |
+
groups_out=min(out_channels // 4, 32),
|
996 |
+
dropout=dropout,
|
997 |
+
time_embedding_norm=resnet_time_scale_shift,
|
998 |
+
non_linearity=resnet_act_fn,
|
999 |
+
output_scale_factor=output_scale_factor,
|
1000 |
+
pre_norm=resnet_pre_norm,
|
1001 |
+
)
|
1002 |
+
)
|
1003 |
+
|
1004 |
+
if add_downsample:
|
1005 |
+
self.resnet_down = ResnetBlock2D(
|
1006 |
+
in_channels=out_channels,
|
1007 |
+
out_channels=out_channels,
|
1008 |
+
temb_channels=temb_channels,
|
1009 |
+
eps=resnet_eps,
|
1010 |
+
groups=min(out_channels // 4, 32),
|
1011 |
+
dropout=dropout,
|
1012 |
+
time_embedding_norm=resnet_time_scale_shift,
|
1013 |
+
non_linearity=resnet_act_fn,
|
1014 |
+
output_scale_factor=output_scale_factor,
|
1015 |
+
pre_norm=resnet_pre_norm,
|
1016 |
+
use_in_shortcut=True,
|
1017 |
+
down=True,
|
1018 |
+
kernel="fir",
|
1019 |
+
)
|
1020 |
+
self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)])
|
1021 |
+
self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1))
|
1022 |
+
else:
|
1023 |
+
self.resnet_down = None
|
1024 |
+
self.downsamplers = None
|
1025 |
+
self.skip_conv = None
|
1026 |
+
|
1027 |
+
def forward(self, hidden_states, temb=None, skip_sample=None):
|
1028 |
+
output_states = ()
|
1029 |
+
|
1030 |
+
for resnet in self.resnets:
|
1031 |
+
hidden_states = resnet(hidden_states, temb)
|
1032 |
+
output_states += (hidden_states,)
|
1033 |
+
|
1034 |
+
if self.downsamplers is not None:
|
1035 |
+
hidden_states = self.resnet_down(hidden_states, temb)
|
1036 |
+
for downsampler in self.downsamplers:
|
1037 |
+
skip_sample = downsampler(skip_sample)
|
1038 |
+
|
1039 |
+
hidden_states = self.skip_conv(skip_sample) + hidden_states
|
1040 |
+
|
1041 |
+
output_states += (hidden_states,)
|
1042 |
+
|
1043 |
+
return hidden_states, output_states, skip_sample
|
1044 |
+
|
1045 |
+
|
1046 |
+
class AttnUpBlock2D(nn.Module):
|
1047 |
+
def __init__(
|
1048 |
+
self,
|
1049 |
+
in_channels: int,
|
1050 |
+
prev_output_channel: int,
|
1051 |
+
out_channels: int,
|
1052 |
+
temb_channels: int,
|
1053 |
+
dropout: float = 0.0,
|
1054 |
+
num_layers: int = 1,
|
1055 |
+
resnet_eps: float = 1e-6,
|
1056 |
+
resnet_time_scale_shift: str = "default",
|
1057 |
+
resnet_act_fn: str = "swish",
|
1058 |
+
resnet_groups: int = 32,
|
1059 |
+
resnet_pre_norm: bool = True,
|
1060 |
+
attention_type="default",
|
1061 |
+
attn_num_head_channels=1,
|
1062 |
+
output_scale_factor=1.0,
|
1063 |
+
add_upsample=True,
|
1064 |
+
):
|
1065 |
+
super().__init__()
|
1066 |
+
resnets = []
|
1067 |
+
attentions = []
|
1068 |
+
|
1069 |
+
self.attention_type = attention_type
|
1070 |
+
|
1071 |
+
for i in range(num_layers):
|
1072 |
+
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
|
1073 |
+
resnet_in_channels = prev_output_channel if i == 0 else out_channels
|
1074 |
+
|
1075 |
+
resnets.append(
|
1076 |
+
ResnetBlock2D(
|
1077 |
+
in_channels=resnet_in_channels + res_skip_channels,
|
1078 |
+
out_channels=out_channels,
|
1079 |
+
temb_channels=temb_channels,
|
1080 |
+
eps=resnet_eps,
|
1081 |
+
groups=resnet_groups,
|
1082 |
+
dropout=dropout,
|
1083 |
+
time_embedding_norm=resnet_time_scale_shift,
|
1084 |
+
non_linearity=resnet_act_fn,
|
1085 |
+
output_scale_factor=output_scale_factor,
|
1086 |
+
pre_norm=resnet_pre_norm,
|
1087 |
+
)
|
1088 |
+
)
|
1089 |
+
attentions.append(
|
1090 |
+
AttentionBlock(
|
1091 |
+
out_channels,
|
1092 |
+
num_head_channels=attn_num_head_channels,
|
1093 |
+
rescale_output_factor=output_scale_factor,
|
1094 |
+
eps=resnet_eps,
|
1095 |
+
norm_num_groups=resnet_groups,
|
1096 |
+
)
|
1097 |
+
)
|
1098 |
+
|
1099 |
+
self.attentions = nn.ModuleList(attentions)
|
1100 |
+
self.resnets = nn.ModuleList(resnets)
|
1101 |
+
|
1102 |
+
if add_upsample:
|
1103 |
+
self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
|
1104 |
+
else:
|
1105 |
+
self.upsamplers = None
|
1106 |
+
|
1107 |
+
def forward(self, hidden_states, res_hidden_states_tuple, temb=None):
|
1108 |
+
for resnet, attn in zip(self.resnets, self.attentions):
|
1109 |
+
# pop res hidden states
|
1110 |
+
res_hidden_states = res_hidden_states_tuple[-1]
|
1111 |
+
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
|
1112 |
+
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
|
1113 |
+
|
1114 |
+
hidden_states = resnet(hidden_states, temb)
|
1115 |
+
hidden_states = attn(hidden_states)
|
1116 |
+
|
1117 |
+
if self.upsamplers is not None:
|
1118 |
+
for upsampler in self.upsamplers:
|
1119 |
+
hidden_states = upsampler(hidden_states)
|
1120 |
+
|
1121 |
+
return hidden_states
|
1122 |
+
|
1123 |
+
|
1124 |
+
class CrossAttnUpBlock2D(nn.Module):
|
1125 |
+
def __init__(
|
1126 |
+
self,
|
1127 |
+
in_channels: int,
|
1128 |
+
out_channels: int,
|
1129 |
+
prev_output_channel: int,
|
1130 |
+
temb_channels: int,
|
1131 |
+
dropout: float = 0.0,
|
1132 |
+
num_layers: int = 1,
|
1133 |
+
resnet_eps: float = 1e-6,
|
1134 |
+
resnet_time_scale_shift: str = "default",
|
1135 |
+
resnet_act_fn: str = "swish",
|
1136 |
+
resnet_groups: int = 32,
|
1137 |
+
resnet_pre_norm: bool = True,
|
1138 |
+
attn_num_head_channels=1,
|
1139 |
+
cross_attention_dim=1280,
|
1140 |
+
attention_type="default",
|
1141 |
+
output_scale_factor=1.0,
|
1142 |
+
add_upsample=True,
|
1143 |
+
dual_cross_attention=False,
|
1144 |
+
use_linear_projection=False,
|
1145 |
+
only_cross_attention=False,
|
1146 |
+
):
|
1147 |
+
super().__init__()
|
1148 |
+
resnets = []
|
1149 |
+
attentions = []
|
1150 |
+
|
1151 |
+
self.attention_type = attention_type
|
1152 |
+
self.attn_num_head_channels = attn_num_head_channels
|
1153 |
+
|
1154 |
+
for i in range(num_layers):
|
1155 |
+
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
|
1156 |
+
resnet_in_channels = prev_output_channel if i == 0 else out_channels
|
1157 |
+
|
1158 |
+
resnets.append(
|
1159 |
+
ResnetBlock2D(
|
1160 |
+
in_channels=resnet_in_channels + res_skip_channels,
|
1161 |
+
out_channels=out_channels,
|
1162 |
+
temb_channels=temb_channels,
|
1163 |
+
eps=resnet_eps,
|
1164 |
+
groups=resnet_groups,
|
1165 |
+
dropout=dropout,
|
1166 |
+
time_embedding_norm=resnet_time_scale_shift,
|
1167 |
+
non_linearity=resnet_act_fn,
|
1168 |
+
output_scale_factor=output_scale_factor,
|
1169 |
+
pre_norm=resnet_pre_norm,
|
1170 |
+
)
|
1171 |
+
)
|
1172 |
+
if not dual_cross_attention:
|
1173 |
+
attentions.append(
|
1174 |
+
Transformer2DModel(
|
1175 |
+
attn_num_head_channels,
|
1176 |
+
out_channels // attn_num_head_channels,
|
1177 |
+
in_channels=out_channels,
|
1178 |
+
num_layers=1,
|
1179 |
+
cross_attention_dim=cross_attention_dim,
|
1180 |
+
norm_num_groups=resnet_groups,
|
1181 |
+
use_linear_projection=use_linear_projection,
|
1182 |
+
only_cross_attention=only_cross_attention,
|
1183 |
+
)
|
1184 |
+
)
|
1185 |
+
else:
|
1186 |
+
attentions.append(
|
1187 |
+
DualTransformer2DModel(
|
1188 |
+
attn_num_head_channels,
|
1189 |
+
out_channels // attn_num_head_channels,
|
1190 |
+
in_channels=out_channels,
|
1191 |
+
num_layers=1,
|
1192 |
+
cross_attention_dim=cross_attention_dim,
|
1193 |
+
norm_num_groups=resnet_groups,
|
1194 |
+
)
|
1195 |
+
)
|
1196 |
+
self.attentions = nn.ModuleList(attentions)
|
1197 |
+
self.resnets = nn.ModuleList(resnets)
|
1198 |
+
|
1199 |
+
if add_upsample:
|
1200 |
+
self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
|
1201 |
+
else:
|
1202 |
+
self.upsamplers = None
|
1203 |
+
|
1204 |
+
self.gradient_checkpointing = False
|
1205 |
+
|
1206 |
+
def set_attention_slice(self, slice_size):
|
1207 |
+
head_dims = self.attn_num_head_channels
|
1208 |
+
head_dims = [head_dims] if isinstance(head_dims, int) else head_dims
|
1209 |
+
if slice_size is not None and any(dim % slice_size != 0 for dim in head_dims):
|
1210 |
+
raise ValueError(
|
1211 |
+
f"Make sure slice_size {slice_size} is a common divisor of "
|
1212 |
+
f"the number of heads used in cross_attention: {head_dims}"
|
1213 |
+
)
|
1214 |
+
if slice_size is not None and slice_size > min(head_dims):
|
1215 |
+
raise ValueError(
|
1216 |
+
f"slice_size {slice_size} has to be smaller or equal to "
|
1217 |
+
f"the lowest number of heads used in cross_attention: min({head_dims}) = {min(head_dims)}"
|
1218 |
+
)
|
1219 |
+
|
1220 |
+
for attn in self.attentions:
|
1221 |
+
attn._set_attention_slice(slice_size)
|
1222 |
+
|
1223 |
+
self.gradient_checkpointing = False
|
1224 |
+
|
1225 |
+
def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
|
1226 |
+
for attn in self.attentions:
|
1227 |
+
attn._set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers)
|
1228 |
+
|
1229 |
+
def forward(
|
1230 |
+
self,
|
1231 |
+
hidden_states,
|
1232 |
+
res_hidden_states_tuple,
|
1233 |
+
temb=None,
|
1234 |
+
encoder_hidden_states=None,
|
1235 |
+
upsample_size=None,
|
1236 |
+
text_format_dict={}
|
1237 |
+
):
|
1238 |
+
for resnet, attn in zip(self.resnets, self.attentions):
|
1239 |
+
# pop res hidden states
|
1240 |
+
res_hidden_states = res_hidden_states_tuple[-1]
|
1241 |
+
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
|
1242 |
+
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
|
1243 |
+
|
1244 |
+
if self.training and self.gradient_checkpointing:
|
1245 |
+
|
1246 |
+
def create_custom_forward(module, return_dict=None):
|
1247 |
+
def custom_forward(*inputs):
|
1248 |
+
if return_dict is not None:
|
1249 |
+
return module(*inputs, return_dict=return_dict)
|
1250 |
+
else:
|
1251 |
+
return module(*inputs)
|
1252 |
+
|
1253 |
+
return custom_forward
|
1254 |
+
|
1255 |
+
hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
|
1256 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
1257 |
+
create_custom_forward(attn, return_dict=False), hidden_states, encoder_hidden_states,
|
1258 |
+
text_format_dict
|
1259 |
+
)[0]
|
1260 |
+
else:
|
1261 |
+
hidden_states = resnet(hidden_states, temb)
|
1262 |
+
hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states,
|
1263 |
+
text_format_dict=text_format_dict).sample
|
1264 |
+
|
1265 |
+
if self.upsamplers is not None:
|
1266 |
+
for upsampler in self.upsamplers:
|
1267 |
+
hidden_states = upsampler(hidden_states, upsample_size)
|
1268 |
+
|
1269 |
+
return hidden_states
|
1270 |
+
|
1271 |
+
|
1272 |
+
class UpBlock2D(nn.Module):
|
1273 |
+
def __init__(
|
1274 |
+
self,
|
1275 |
+
in_channels: int,
|
1276 |
+
prev_output_channel: int,
|
1277 |
+
out_channels: int,
|
1278 |
+
temb_channels: int,
|
1279 |
+
dropout: float = 0.0,
|
1280 |
+
num_layers: int = 1,
|
1281 |
+
resnet_eps: float = 1e-6,
|
1282 |
+
resnet_time_scale_shift: str = "default",
|
1283 |
+
resnet_act_fn: str = "swish",
|
1284 |
+
resnet_groups: int = 32,
|
1285 |
+
resnet_pre_norm: bool = True,
|
1286 |
+
output_scale_factor=1.0,
|
1287 |
+
add_upsample=True,
|
1288 |
+
):
|
1289 |
+
super().__init__()
|
1290 |
+
resnets = []
|
1291 |
+
|
1292 |
+
for i in range(num_layers):
|
1293 |
+
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
|
1294 |
+
resnet_in_channels = prev_output_channel if i == 0 else out_channels
|
1295 |
+
|
1296 |
+
resnets.append(
|
1297 |
+
ResnetBlock2D(
|
1298 |
+
in_channels=resnet_in_channels + res_skip_channels,
|
1299 |
+
out_channels=out_channels,
|
1300 |
+
temb_channels=temb_channels,
|
1301 |
+
eps=resnet_eps,
|
1302 |
+
groups=resnet_groups,
|
1303 |
+
dropout=dropout,
|
1304 |
+
time_embedding_norm=resnet_time_scale_shift,
|
1305 |
+
non_linearity=resnet_act_fn,
|
1306 |
+
output_scale_factor=output_scale_factor,
|
1307 |
+
pre_norm=resnet_pre_norm,
|
1308 |
+
)
|
1309 |
+
)
|
1310 |
+
|
1311 |
+
self.resnets = nn.ModuleList(resnets)
|
1312 |
+
|
1313 |
+
if add_upsample:
|
1314 |
+
self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
|
1315 |
+
else:
|
1316 |
+
self.upsamplers = None
|
1317 |
+
|
1318 |
+
self.gradient_checkpointing = False
|
1319 |
+
|
1320 |
+
def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None):
|
1321 |
+
for resnet in self.resnets:
|
1322 |
+
# pop res hidden states
|
1323 |
+
res_hidden_states = res_hidden_states_tuple[-1]
|
1324 |
+
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
|
1325 |
+
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
|
1326 |
+
|
1327 |
+
if self.training and self.gradient_checkpointing:
|
1328 |
+
|
1329 |
+
def create_custom_forward(module):
|
1330 |
+
def custom_forward(*inputs):
|
1331 |
+
return module(*inputs)
|
1332 |
+
|
1333 |
+
return custom_forward
|
1334 |
+
|
1335 |
+
hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
|
1336 |
+
else:
|
1337 |
+
hidden_states = resnet(hidden_states, temb)
|
1338 |
+
|
1339 |
+
if self.upsamplers is not None:
|
1340 |
+
for upsampler in self.upsamplers:
|
1341 |
+
hidden_states = upsampler(hidden_states, upsample_size)
|
1342 |
+
|
1343 |
+
return hidden_states
|
1344 |
+
|
1345 |
+
|
1346 |
+
class UpDecoderBlock2D(nn.Module):
|
1347 |
+
def __init__(
|
1348 |
+
self,
|
1349 |
+
in_channels: int,
|
1350 |
+
out_channels: int,
|
1351 |
+
dropout: float = 0.0,
|
1352 |
+
num_layers: int = 1,
|
1353 |
+
resnet_eps: float = 1e-6,
|
1354 |
+
resnet_time_scale_shift: str = "default",
|
1355 |
+
resnet_act_fn: str = "swish",
|
1356 |
+
resnet_groups: int = 32,
|
1357 |
+
resnet_pre_norm: bool = True,
|
1358 |
+
output_scale_factor=1.0,
|
1359 |
+
add_upsample=True,
|
1360 |
+
):
|
1361 |
+
super().__init__()
|
1362 |
+
resnets = []
|
1363 |
+
|
1364 |
+
for i in range(num_layers):
|
1365 |
+
input_channels = in_channels if i == 0 else out_channels
|
1366 |
+
|
1367 |
+
resnets.append(
|
1368 |
+
ResnetBlock2D(
|
1369 |
+
in_channels=input_channels,
|
1370 |
+
out_channels=out_channels,
|
1371 |
+
temb_channels=None,
|
1372 |
+
eps=resnet_eps,
|
1373 |
+
groups=resnet_groups,
|
1374 |
+
dropout=dropout,
|
1375 |
+
time_embedding_norm=resnet_time_scale_shift,
|
1376 |
+
non_linearity=resnet_act_fn,
|
1377 |
+
output_scale_factor=output_scale_factor,
|
1378 |
+
pre_norm=resnet_pre_norm,
|
1379 |
+
)
|
1380 |
+
)
|
1381 |
+
|
1382 |
+
self.resnets = nn.ModuleList(resnets)
|
1383 |
+
|
1384 |
+
if add_upsample:
|
1385 |
+
self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
|
1386 |
+
else:
|
1387 |
+
self.upsamplers = None
|
1388 |
+
|
1389 |
+
def forward(self, hidden_states):
|
1390 |
+
for resnet in self.resnets:
|
1391 |
+
hidden_states = resnet(hidden_states, temb=None)
|
1392 |
+
|
1393 |
+
if self.upsamplers is not None:
|
1394 |
+
for upsampler in self.upsamplers:
|
1395 |
+
hidden_states = upsampler(hidden_states)
|
1396 |
+
|
1397 |
+
return hidden_states
|
1398 |
+
|
1399 |
+
|
1400 |
+
class AttnUpDecoderBlock2D(nn.Module):
|
1401 |
+
def __init__(
|
1402 |
+
self,
|
1403 |
+
in_channels: int,
|
1404 |
+
out_channels: int,
|
1405 |
+
dropout: float = 0.0,
|
1406 |
+
num_layers: int = 1,
|
1407 |
+
resnet_eps: float = 1e-6,
|
1408 |
+
resnet_time_scale_shift: str = "default",
|
1409 |
+
resnet_act_fn: str = "swish",
|
1410 |
+
resnet_groups: int = 32,
|
1411 |
+
resnet_pre_norm: bool = True,
|
1412 |
+
attn_num_head_channels=1,
|
1413 |
+
output_scale_factor=1.0,
|
1414 |
+
add_upsample=True,
|
1415 |
+
):
|
1416 |
+
super().__init__()
|
1417 |
+
resnets = []
|
1418 |
+
attentions = []
|
1419 |
+
|
1420 |
+
for i in range(num_layers):
|
1421 |
+
input_channels = in_channels if i == 0 else out_channels
|
1422 |
+
|
1423 |
+
resnets.append(
|
1424 |
+
ResnetBlock2D(
|
1425 |
+
in_channels=input_channels,
|
1426 |
+
out_channels=out_channels,
|
1427 |
+
temb_channels=None,
|
1428 |
+
eps=resnet_eps,
|
1429 |
+
groups=resnet_groups,
|
1430 |
+
dropout=dropout,
|
1431 |
+
time_embedding_norm=resnet_time_scale_shift,
|
1432 |
+
non_linearity=resnet_act_fn,
|
1433 |
+
output_scale_factor=output_scale_factor,
|
1434 |
+
pre_norm=resnet_pre_norm,
|
1435 |
+
)
|
1436 |
+
)
|
1437 |
+
attentions.append(
|
1438 |
+
AttentionBlock(
|
1439 |
+
out_channels,
|
1440 |
+
num_head_channels=attn_num_head_channels,
|
1441 |
+
rescale_output_factor=output_scale_factor,
|
1442 |
+
eps=resnet_eps,
|
1443 |
+
norm_num_groups=resnet_groups,
|
1444 |
+
)
|
1445 |
+
)
|
1446 |
+
|
1447 |
+
self.attentions = nn.ModuleList(attentions)
|
1448 |
+
self.resnets = nn.ModuleList(resnets)
|
1449 |
+
|
1450 |
+
if add_upsample:
|
1451 |
+
self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
|
1452 |
+
else:
|
1453 |
+
self.upsamplers = None
|
1454 |
+
|
1455 |
+
def forward(self, hidden_states):
|
1456 |
+
for resnet, attn in zip(self.resnets, self.attentions):
|
1457 |
+
hidden_states = resnet(hidden_states, temb=None)
|
1458 |
+
hidden_states = attn(hidden_states)
|
1459 |
+
|
1460 |
+
if self.upsamplers is not None:
|
1461 |
+
for upsampler in self.upsamplers:
|
1462 |
+
hidden_states = upsampler(hidden_states)
|
1463 |
+
|
1464 |
+
return hidden_states
|
1465 |
+
|
1466 |
+
|
1467 |
+
class AttnSkipUpBlock2D(nn.Module):
|
1468 |
+
def __init__(
|
1469 |
+
self,
|
1470 |
+
in_channels: int,
|
1471 |
+
prev_output_channel: int,
|
1472 |
+
out_channels: int,
|
1473 |
+
temb_channels: int,
|
1474 |
+
dropout: float = 0.0,
|
1475 |
+
num_layers: int = 1,
|
1476 |
+
resnet_eps: float = 1e-6,
|
1477 |
+
resnet_time_scale_shift: str = "default",
|
1478 |
+
resnet_act_fn: str = "swish",
|
1479 |
+
resnet_pre_norm: bool = True,
|
1480 |
+
attn_num_head_channels=1,
|
1481 |
+
attention_type="default",
|
1482 |
+
output_scale_factor=np.sqrt(2.0),
|
1483 |
+
upsample_padding=1,
|
1484 |
+
add_upsample=True,
|
1485 |
+
):
|
1486 |
+
super().__init__()
|
1487 |
+
self.attentions = nn.ModuleList([])
|
1488 |
+
self.resnets = nn.ModuleList([])
|
1489 |
+
|
1490 |
+
self.attention_type = attention_type
|
1491 |
+
|
1492 |
+
for i in range(num_layers):
|
1493 |
+
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
|
1494 |
+
resnet_in_channels = prev_output_channel if i == 0 else out_channels
|
1495 |
+
|
1496 |
+
self.resnets.append(
|
1497 |
+
ResnetBlock2D(
|
1498 |
+
in_channels=resnet_in_channels + res_skip_channels,
|
1499 |
+
out_channels=out_channels,
|
1500 |
+
temb_channels=temb_channels,
|
1501 |
+
eps=resnet_eps,
|
1502 |
+
groups=min(resnet_in_channels + res_skip_channels // 4, 32),
|
1503 |
+
groups_out=min(out_channels // 4, 32),
|
1504 |
+
dropout=dropout,
|
1505 |
+
time_embedding_norm=resnet_time_scale_shift,
|
1506 |
+
non_linearity=resnet_act_fn,
|
1507 |
+
output_scale_factor=output_scale_factor,
|
1508 |
+
pre_norm=resnet_pre_norm,
|
1509 |
+
)
|
1510 |
+
)
|
1511 |
+
|
1512 |
+
self.attentions.append(
|
1513 |
+
AttentionBlock(
|
1514 |
+
out_channels,
|
1515 |
+
num_head_channels=attn_num_head_channels,
|
1516 |
+
rescale_output_factor=output_scale_factor,
|
1517 |
+
eps=resnet_eps,
|
1518 |
+
)
|
1519 |
+
)
|
1520 |
+
|
1521 |
+
self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels)
|
1522 |
+
if add_upsample:
|
1523 |
+
self.resnet_up = ResnetBlock2D(
|
1524 |
+
in_channels=out_channels,
|
1525 |
+
out_channels=out_channels,
|
1526 |
+
temb_channels=temb_channels,
|
1527 |
+
eps=resnet_eps,
|
1528 |
+
groups=min(out_channels // 4, 32),
|
1529 |
+
groups_out=min(out_channels // 4, 32),
|
1530 |
+
dropout=dropout,
|
1531 |
+
time_embedding_norm=resnet_time_scale_shift,
|
1532 |
+
non_linearity=resnet_act_fn,
|
1533 |
+
output_scale_factor=output_scale_factor,
|
1534 |
+
pre_norm=resnet_pre_norm,
|
1535 |
+
use_in_shortcut=True,
|
1536 |
+
up=True,
|
1537 |
+
kernel="fir",
|
1538 |
+
)
|
1539 |
+
self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
|
1540 |
+
self.skip_norm = torch.nn.GroupNorm(
|
1541 |
+
num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True
|
1542 |
+
)
|
1543 |
+
self.act = nn.SiLU()
|
1544 |
+
else:
|
1545 |
+
self.resnet_up = None
|
1546 |
+
self.skip_conv = None
|
1547 |
+
self.skip_norm = None
|
1548 |
+
self.act = None
|
1549 |
+
|
1550 |
+
def forward(self, hidden_states, res_hidden_states_tuple, temb=None, skip_sample=None):
|
1551 |
+
for resnet in self.resnets:
|
1552 |
+
# pop res hidden states
|
1553 |
+
res_hidden_states = res_hidden_states_tuple[-1]
|
1554 |
+
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
|
1555 |
+
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
|
1556 |
+
|
1557 |
+
hidden_states = resnet(hidden_states, temb)
|
1558 |
+
|
1559 |
+
hidden_states = self.attentions[0](hidden_states)
|
1560 |
+
|
1561 |
+
if skip_sample is not None:
|
1562 |
+
skip_sample = self.upsampler(skip_sample)
|
1563 |
+
else:
|
1564 |
+
skip_sample = 0
|
1565 |
+
|
1566 |
+
if self.resnet_up is not None:
|
1567 |
+
skip_sample_states = self.skip_norm(hidden_states)
|
1568 |
+
skip_sample_states = self.act(skip_sample_states)
|
1569 |
+
skip_sample_states = self.skip_conv(skip_sample_states)
|
1570 |
+
|
1571 |
+
skip_sample = skip_sample + skip_sample_states
|
1572 |
+
|
1573 |
+
hidden_states = self.resnet_up(hidden_states, temb)
|
1574 |
+
|
1575 |
+
return hidden_states, skip_sample
|
1576 |
+
|
1577 |
+
|
1578 |
+
class SkipUpBlock2D(nn.Module):
|
1579 |
+
def __init__(
|
1580 |
+
self,
|
1581 |
+
in_channels: int,
|
1582 |
+
prev_output_channel: int,
|
1583 |
+
out_channels: int,
|
1584 |
+
temb_channels: int,
|
1585 |
+
dropout: float = 0.0,
|
1586 |
+
num_layers: int = 1,
|
1587 |
+
resnet_eps: float = 1e-6,
|
1588 |
+
resnet_time_scale_shift: str = "default",
|
1589 |
+
resnet_act_fn: str = "swish",
|
1590 |
+
resnet_pre_norm: bool = True,
|
1591 |
+
output_scale_factor=np.sqrt(2.0),
|
1592 |
+
add_upsample=True,
|
1593 |
+
upsample_padding=1,
|
1594 |
+
):
|
1595 |
+
super().__init__()
|
1596 |
+
self.resnets = nn.ModuleList([])
|
1597 |
+
|
1598 |
+
for i in range(num_layers):
|
1599 |
+
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
|
1600 |
+
resnet_in_channels = prev_output_channel if i == 0 else out_channels
|
1601 |
+
|
1602 |
+
self.resnets.append(
|
1603 |
+
ResnetBlock2D(
|
1604 |
+
in_channels=resnet_in_channels + res_skip_channels,
|
1605 |
+
out_channels=out_channels,
|
1606 |
+
temb_channels=temb_channels,
|
1607 |
+
eps=resnet_eps,
|
1608 |
+
groups=min((resnet_in_channels + res_skip_channels) // 4, 32),
|
1609 |
+
groups_out=min(out_channels // 4, 32),
|
1610 |
+
dropout=dropout,
|
1611 |
+
time_embedding_norm=resnet_time_scale_shift,
|
1612 |
+
non_linearity=resnet_act_fn,
|
1613 |
+
output_scale_factor=output_scale_factor,
|
1614 |
+
pre_norm=resnet_pre_norm,
|
1615 |
+
)
|
1616 |
+
)
|
1617 |
+
|
1618 |
+
self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels)
|
1619 |
+
if add_upsample:
|
1620 |
+
self.resnet_up = ResnetBlock2D(
|
1621 |
+
in_channels=out_channels,
|
1622 |
+
out_channels=out_channels,
|
1623 |
+
temb_channels=temb_channels,
|
1624 |
+
eps=resnet_eps,
|
1625 |
+
groups=min(out_channels // 4, 32),
|
1626 |
+
groups_out=min(out_channels // 4, 32),
|
1627 |
+
dropout=dropout,
|
1628 |
+
time_embedding_norm=resnet_time_scale_shift,
|
1629 |
+
non_linearity=resnet_act_fn,
|
1630 |
+
output_scale_factor=output_scale_factor,
|
1631 |
+
pre_norm=resnet_pre_norm,
|
1632 |
+
use_in_shortcut=True,
|
1633 |
+
up=True,
|
1634 |
+
kernel="fir",
|
1635 |
+
)
|
1636 |
+
self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
|
1637 |
+
self.skip_norm = torch.nn.GroupNorm(
|
1638 |
+
num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True
|
1639 |
+
)
|
1640 |
+
self.act = nn.SiLU()
|
1641 |
+
else:
|
1642 |
+
self.resnet_up = None
|
1643 |
+
self.skip_conv = None
|
1644 |
+
self.skip_norm = None
|
1645 |
+
self.act = None
|
1646 |
+
|
1647 |
+
def forward(self, hidden_states, res_hidden_states_tuple, temb=None, skip_sample=None):
|
1648 |
+
for resnet in self.resnets:
|
1649 |
+
# pop res hidden states
|
1650 |
+
res_hidden_states = res_hidden_states_tuple[-1]
|
1651 |
+
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
|
1652 |
+
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
|
1653 |
+
|
1654 |
+
hidden_states = resnet(hidden_states, temb)
|
1655 |
+
|
1656 |
+
if skip_sample is not None:
|
1657 |
+
skip_sample = self.upsampler(skip_sample)
|
1658 |
+
else:
|
1659 |
+
skip_sample = 0
|
1660 |
+
|
1661 |
+
if self.resnet_up is not None:
|
1662 |
+
skip_sample_states = self.skip_norm(hidden_states)
|
1663 |
+
skip_sample_states = self.act(skip_sample_states)
|
1664 |
+
skip_sample_states = self.skip_conv(skip_sample_states)
|
1665 |
+
|
1666 |
+
skip_sample = skip_sample + skip_sample_states
|
1667 |
+
|
1668 |
+
hidden_states = self.resnet_up(hidden_states, temb)
|
1669 |
+
|
1670 |
+
return hidden_states, skip_sample
|
models/unet_2d_condition.py
ADDED
@@ -0,0 +1,411 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright 2022 The HuggingFace Team. All rights reserved.
|
2 |
+
#
|
3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4 |
+
# you may not use this file except in compliance with the License.
|
5 |
+
# You may obtain a copy of the License at
|
6 |
+
#
|
7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
8 |
+
#
|
9 |
+
# Unless required by applicable law or agreed to in writing, software
|
10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12 |
+
# See the License for the specific language governing permissions and
|
13 |
+
# limitations under the License.
|
14 |
+
from dataclasses import dataclass
|
15 |
+
from typing import Optional, Tuple, Union
|
16 |
+
|
17 |
+
import torch
|
18 |
+
import torch.nn as nn
|
19 |
+
import torch.utils.checkpoint
|
20 |
+
|
21 |
+
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
22 |
+
from diffusers.models.modeling_utils import ModelMixin
|
23 |
+
from diffusers.utils import BaseOutput, logging
|
24 |
+
from diffusers.models.embeddings import TimestepEmbedding, Timesteps
|
25 |
+
from .unet_2d_blocks import (
|
26 |
+
CrossAttnDownBlock2D,
|
27 |
+
CrossAttnUpBlock2D,
|
28 |
+
DownBlock2D,
|
29 |
+
UNetMidBlock2DCrossAttn,
|
30 |
+
UpBlock2D,
|
31 |
+
get_down_block,
|
32 |
+
get_up_block,
|
33 |
+
)
|
34 |
+
|
35 |
+
|
36 |
+
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
37 |
+
|
38 |
+
|
39 |
+
@dataclass
|
40 |
+
class UNet2DConditionOutput(BaseOutput):
|
41 |
+
"""
|
42 |
+
Args:
|
43 |
+
sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
|
44 |
+
Hidden states conditioned on `encoder_hidden_states` input. Output of last layer of model.
|
45 |
+
"""
|
46 |
+
|
47 |
+
sample: torch.FloatTensor
|
48 |
+
|
49 |
+
|
50 |
+
class UNet2DConditionModel(ModelMixin, ConfigMixin):
|
51 |
+
r"""
|
52 |
+
UNet2DConditionModel is a conditional 2D UNet model that takes in a noisy sample, conditional state, and a timestep
|
53 |
+
and returns sample shaped output.
|
54 |
+
|
55 |
+
This model inherits from [`ModelMixin`]. Check the superclass documentation for the generic methods the library
|
56 |
+
implements for all the models (such as downloading or saving, etc.)
|
57 |
+
|
58 |
+
Parameters:
|
59 |
+
sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
|
60 |
+
Height and width of input/output sample.
|
61 |
+
in_channels (`int`, *optional*, defaults to 4): The number of channels in the input sample.
|
62 |
+
out_channels (`int`, *optional*, defaults to 4): The number of channels in the output.
|
63 |
+
center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
|
64 |
+
flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
|
65 |
+
Whether to flip the sin to cos in the time embedding.
|
66 |
+
freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
|
67 |
+
down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
|
68 |
+
The tuple of downsample blocks to use.
|
69 |
+
up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D",)`):
|
70 |
+
The tuple of upsample blocks to use.
|
71 |
+
block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
|
72 |
+
The tuple of output channels for each block.
|
73 |
+
layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
|
74 |
+
downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
|
75 |
+
mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
|
76 |
+
act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
|
77 |
+
norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
|
78 |
+
norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
|
79 |
+
cross_attention_dim (`int`, *optional*, defaults to 1280): The dimension of the cross attention features.
|
80 |
+
attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
|
81 |
+
"""
|
82 |
+
|
83 |
+
_supports_gradient_checkpointing = True
|
84 |
+
|
85 |
+
@register_to_config
|
86 |
+
def __init__(
|
87 |
+
self,
|
88 |
+
sample_size: Optional[int] = None,
|
89 |
+
in_channels: int = 4,
|
90 |
+
out_channels: int = 4,
|
91 |
+
center_input_sample: bool = False,
|
92 |
+
flip_sin_to_cos: bool = True,
|
93 |
+
freq_shift: int = 0,
|
94 |
+
down_block_types: Tuple[str] = (
|
95 |
+
"CrossAttnDownBlock2D",
|
96 |
+
"CrossAttnDownBlock2D",
|
97 |
+
"CrossAttnDownBlock2D",
|
98 |
+
"DownBlock2D",
|
99 |
+
),
|
100 |
+
up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
|
101 |
+
only_cross_attention: Union[bool, Tuple[bool]] = False,
|
102 |
+
block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
|
103 |
+
layers_per_block: int = 2,
|
104 |
+
downsample_padding: int = 1,
|
105 |
+
mid_block_scale_factor: float = 1,
|
106 |
+
act_fn: str = "silu",
|
107 |
+
norm_num_groups: int = 32,
|
108 |
+
norm_eps: float = 1e-5,
|
109 |
+
cross_attention_dim: int = 1280,
|
110 |
+
attention_head_dim: Union[int, Tuple[int]] = 8,
|
111 |
+
dual_cross_attention: bool = False,
|
112 |
+
use_linear_projection: bool = False,
|
113 |
+
num_class_embeds: Optional[int] = None,
|
114 |
+
):
|
115 |
+
super().__init__()
|
116 |
+
|
117 |
+
self.sample_size = sample_size
|
118 |
+
time_embed_dim = block_out_channels[0] * 4
|
119 |
+
# import ipdb;ipdb.set_trace()
|
120 |
+
|
121 |
+
# input
|
122 |
+
self.conv_in = nn.Conv2d(in_channels, block_out_channels[0], kernel_size=3, padding=(1, 1))
|
123 |
+
|
124 |
+
# time
|
125 |
+
self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
|
126 |
+
timestep_input_dim = block_out_channels[0]
|
127 |
+
|
128 |
+
self.time_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
|
129 |
+
|
130 |
+
# class embedding
|
131 |
+
if num_class_embeds is not None:
|
132 |
+
self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
|
133 |
+
|
134 |
+
self.down_blocks = nn.ModuleList([])
|
135 |
+
self.mid_block = None
|
136 |
+
self.up_blocks = nn.ModuleList([])
|
137 |
+
|
138 |
+
if isinstance(only_cross_attention, bool):
|
139 |
+
only_cross_attention = [only_cross_attention] * len(down_block_types)
|
140 |
+
|
141 |
+
if isinstance(attention_head_dim, int):
|
142 |
+
attention_head_dim = (attention_head_dim,) * len(down_block_types)
|
143 |
+
|
144 |
+
# down
|
145 |
+
output_channel = block_out_channels[0]
|
146 |
+
for i, down_block_type in enumerate(down_block_types):
|
147 |
+
input_channel = output_channel
|
148 |
+
output_channel = block_out_channels[i]
|
149 |
+
is_final_block = i == len(block_out_channels) - 1
|
150 |
+
|
151 |
+
down_block = get_down_block(
|
152 |
+
down_block_type,
|
153 |
+
num_layers=layers_per_block,
|
154 |
+
in_channels=input_channel,
|
155 |
+
out_channels=output_channel,
|
156 |
+
temb_channels=time_embed_dim,
|
157 |
+
add_downsample=not is_final_block,
|
158 |
+
resnet_eps=norm_eps,
|
159 |
+
resnet_act_fn=act_fn,
|
160 |
+
resnet_groups=norm_num_groups,
|
161 |
+
cross_attention_dim=cross_attention_dim,
|
162 |
+
attn_num_head_channels=attention_head_dim[i],
|
163 |
+
downsample_padding=downsample_padding,
|
164 |
+
dual_cross_attention=dual_cross_attention,
|
165 |
+
use_linear_projection=use_linear_projection,
|
166 |
+
only_cross_attention=only_cross_attention[i],
|
167 |
+
)
|
168 |
+
self.down_blocks.append(down_block)
|
169 |
+
|
170 |
+
# mid
|
171 |
+
self.mid_block = UNetMidBlock2DCrossAttn(
|
172 |
+
in_channels=block_out_channels[-1],
|
173 |
+
temb_channels=time_embed_dim,
|
174 |
+
resnet_eps=norm_eps,
|
175 |
+
resnet_act_fn=act_fn,
|
176 |
+
output_scale_factor=mid_block_scale_factor,
|
177 |
+
resnet_time_scale_shift="default",
|
178 |
+
cross_attention_dim=cross_attention_dim,
|
179 |
+
attn_num_head_channels=attention_head_dim[-1],
|
180 |
+
resnet_groups=norm_num_groups,
|
181 |
+
dual_cross_attention=dual_cross_attention,
|
182 |
+
use_linear_projection=use_linear_projection,
|
183 |
+
)
|
184 |
+
|
185 |
+
# count how many layers upsample the images
|
186 |
+
self.num_upsamplers = 0
|
187 |
+
|
188 |
+
# up
|
189 |
+
reversed_block_out_channels = list(reversed(block_out_channels))
|
190 |
+
reversed_attention_head_dim = list(reversed(attention_head_dim))
|
191 |
+
only_cross_attention = list(reversed(only_cross_attention))
|
192 |
+
output_channel = reversed_block_out_channels[0]
|
193 |
+
for i, up_block_type in enumerate(up_block_types):
|
194 |
+
is_final_block = i == len(block_out_channels) - 1
|
195 |
+
|
196 |
+
prev_output_channel = output_channel
|
197 |
+
output_channel = reversed_block_out_channels[i]
|
198 |
+
input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
|
199 |
+
|
200 |
+
# add upsample block for all BUT final layer
|
201 |
+
if not is_final_block:
|
202 |
+
add_upsample = True
|
203 |
+
self.num_upsamplers += 1
|
204 |
+
else:
|
205 |
+
add_upsample = False
|
206 |
+
|
207 |
+
up_block = get_up_block(
|
208 |
+
up_block_type,
|
209 |
+
num_layers=layers_per_block + 1,
|
210 |
+
in_channels=input_channel,
|
211 |
+
out_channels=output_channel,
|
212 |
+
prev_output_channel=prev_output_channel,
|
213 |
+
temb_channels=time_embed_dim,
|
214 |
+
add_upsample=add_upsample,
|
215 |
+
resnet_eps=norm_eps,
|
216 |
+
resnet_act_fn=act_fn,
|
217 |
+
resnet_groups=norm_num_groups,
|
218 |
+
cross_attention_dim=cross_attention_dim,
|
219 |
+
attn_num_head_channels=reversed_attention_head_dim[i],
|
220 |
+
dual_cross_attention=dual_cross_attention,
|
221 |
+
use_linear_projection=use_linear_projection,
|
222 |
+
only_cross_attention=only_cross_attention[i],
|
223 |
+
)
|
224 |
+
self.up_blocks.append(up_block)
|
225 |
+
prev_output_channel = output_channel
|
226 |
+
|
227 |
+
# out
|
228 |
+
self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps)
|
229 |
+
self.conv_act = nn.SiLU()
|
230 |
+
self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, kernel_size=3, padding=1)
|
231 |
+
|
232 |
+
def set_attention_slice(self, slice_size):
|
233 |
+
head_dims = self.config.attention_head_dim
|
234 |
+
head_dims = [head_dims] if isinstance(head_dims, int) else head_dims
|
235 |
+
if slice_size is not None and any(dim % slice_size != 0 for dim in head_dims):
|
236 |
+
raise ValueError(
|
237 |
+
f"Make sure slice_size {slice_size} is a common divisor of "
|
238 |
+
f"the number of heads used in cross_attention: {head_dims}"
|
239 |
+
)
|
240 |
+
if slice_size is not None and slice_size > min(head_dims):
|
241 |
+
raise ValueError(
|
242 |
+
f"slice_size {slice_size} has to be smaller or equal to "
|
243 |
+
f"the lowest number of heads used in cross_attention: min({head_dims}) = {min(head_dims)}"
|
244 |
+
)
|
245 |
+
|
246 |
+
for block in self.down_blocks:
|
247 |
+
if hasattr(block, "attentions") and block.attentions is not None:
|
248 |
+
block.set_attention_slice(slice_size)
|
249 |
+
|
250 |
+
self.mid_block.set_attention_slice(slice_size)
|
251 |
+
|
252 |
+
for block in self.up_blocks:
|
253 |
+
if hasattr(block, "attentions") and block.attentions is not None:
|
254 |
+
block.set_attention_slice(slice_size)
|
255 |
+
|
256 |
+
def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
|
257 |
+
for block in self.down_blocks:
|
258 |
+
if hasattr(block, "attentions") and block.attentions is not None:
|
259 |
+
block.set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers)
|
260 |
+
|
261 |
+
self.mid_block.set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers)
|
262 |
+
|
263 |
+
for block in self.up_blocks:
|
264 |
+
if hasattr(block, "attentions") and block.attentions is not None:
|
265 |
+
block.set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers)
|
266 |
+
|
267 |
+
def _set_gradient_checkpointing(self, module, value=False):
|
268 |
+
if isinstance(module, (CrossAttnDownBlock2D, DownBlock2D, CrossAttnUpBlock2D, UpBlock2D)):
|
269 |
+
module.gradient_checkpointing = value
|
270 |
+
|
271 |
+
def forward(
|
272 |
+
self,
|
273 |
+
sample: torch.FloatTensor,
|
274 |
+
timestep: Union[torch.Tensor, float, int],
|
275 |
+
encoder_hidden_states: torch.Tensor,
|
276 |
+
class_labels: Optional[torch.Tensor] = None,
|
277 |
+
text_format_dict = {},
|
278 |
+
return_dict: bool = True,
|
279 |
+
) -> Union[UNet2DConditionOutput, Tuple]:
|
280 |
+
r"""
|
281 |
+
Args:
|
282 |
+
sample (`torch.FloatTensor`): (batch, channel, height, width) noisy inputs tensor
|
283 |
+
timestep (`torch.FloatTensor` or `float` or `int`): (batch) timesteps
|
284 |
+
encoder_hidden_states (`torch.FloatTensor`): (batch, channel, height, width) encoder hidden states
|
285 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
286 |
+
Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple.
|
287 |
+
|
288 |
+
Returns:
|
289 |
+
[`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
|
290 |
+
[`~models.unet_2d_condition.UNet2DConditionOutput`] if `return_dict` is True, otherwise a `tuple`. When
|
291 |
+
returning a tuple, the first element is the sample tensor.
|
292 |
+
"""
|
293 |
+
# By default samples have to be AT least a multiple of the overall upsampling factor.
|
294 |
+
# The overall upsampling factor is equal to 2 ** (# num of upsampling layears).
|
295 |
+
# However, the upsampling interpolation output size can be forced to fit any upsampling size
|
296 |
+
# on the fly if necessary.
|
297 |
+
default_overall_up_factor = 2**self.num_upsamplers
|
298 |
+
|
299 |
+
# upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
|
300 |
+
forward_upsample_size = False
|
301 |
+
upsample_size = None
|
302 |
+
|
303 |
+
if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
|
304 |
+
logger.info("Forward upsample size to force interpolation output size.")
|
305 |
+
forward_upsample_size = True
|
306 |
+
|
307 |
+
# 0. center input if necessary
|
308 |
+
if self.config.center_input_sample:
|
309 |
+
sample = 2 * sample - 1.0
|
310 |
+
|
311 |
+
# 1. time
|
312 |
+
timesteps = timestep
|
313 |
+
if not torch.is_tensor(timesteps):
|
314 |
+
# TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
|
315 |
+
timesteps = torch.tensor([timesteps], dtype=torch.long, device=sample.device)
|
316 |
+
elif torch.is_tensor(timesteps) and len(timesteps.shape) == 0:
|
317 |
+
timesteps = timesteps[None].to(sample.device)
|
318 |
+
|
319 |
+
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
|
320 |
+
timesteps = timesteps.expand(sample.shape[0])
|
321 |
+
|
322 |
+
t_emb = self.time_proj(timesteps)
|
323 |
+
|
324 |
+
# timesteps does not contain any weights and will always return f32 tensors
|
325 |
+
# but time_embedding might actually be running in fp16. so we need to cast here.
|
326 |
+
# there might be better ways to encapsulate this.
|
327 |
+
t_emb = t_emb.to(dtype=self.dtype)
|
328 |
+
emb = self.time_embedding(t_emb)
|
329 |
+
|
330 |
+
if self.config.num_class_embeds is not None:
|
331 |
+
if class_labels is None:
|
332 |
+
raise ValueError("class_labels should be provided when num_class_embeds > 0")
|
333 |
+
class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
|
334 |
+
emb = emb + class_emb
|
335 |
+
|
336 |
+
# 2. pre-process
|
337 |
+
sample = self.conv_in(sample)
|
338 |
+
|
339 |
+
# 3. down
|
340 |
+
down_block_res_samples = (sample,)
|
341 |
+
for downsample_block in self.down_blocks:
|
342 |
+
if hasattr(downsample_block, "attentions") and downsample_block.attentions is not None:
|
343 |
+
if isinstance(downsample_block, CrossAttnDownBlock2D):
|
344 |
+
sample, res_samples = downsample_block(
|
345 |
+
hidden_states=sample,
|
346 |
+
temb=emb,
|
347 |
+
encoder_hidden_states=encoder_hidden_states,
|
348 |
+
text_format_dict=text_format_dict
|
349 |
+
)
|
350 |
+
else:
|
351 |
+
sample, res_samples = downsample_block(
|
352 |
+
hidden_states=sample,
|
353 |
+
temb=emb,
|
354 |
+
encoder_hidden_states=encoder_hidden_states,
|
355 |
+
)
|
356 |
+
else:
|
357 |
+
if isinstance(downsample_block, CrossAttnDownBlock2D):
|
358 |
+
import ipdb;ipdb.set_trace()
|
359 |
+
sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
|
360 |
+
down_block_res_samples += res_samples
|
361 |
+
|
362 |
+
# 4. mid
|
363 |
+
sample = self.mid_block(sample, emb, encoder_hidden_states=encoder_hidden_states,
|
364 |
+
text_format_dict=text_format_dict)
|
365 |
+
|
366 |
+
# 5. up
|
367 |
+
for i, upsample_block in enumerate(self.up_blocks):
|
368 |
+
is_final_block = i == len(self.up_blocks) - 1
|
369 |
+
|
370 |
+
res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
|
371 |
+
down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
|
372 |
+
|
373 |
+
# if we have not reached the final block and need to forward the
|
374 |
+
# upsample size, we do it here
|
375 |
+
if not is_final_block and forward_upsample_size:
|
376 |
+
upsample_size = down_block_res_samples[-1].shape[2:]
|
377 |
+
|
378 |
+
if hasattr(upsample_block, "attentions") and upsample_block.attentions is not None:
|
379 |
+
if isinstance(upsample_block, CrossAttnUpBlock2D):
|
380 |
+
sample = upsample_block(
|
381 |
+
hidden_states=sample,
|
382 |
+
temb=emb,
|
383 |
+
res_hidden_states_tuple=res_samples,
|
384 |
+
encoder_hidden_states=encoder_hidden_states,
|
385 |
+
upsample_size=upsample_size,
|
386 |
+
text_format_dict=text_format_dict
|
387 |
+
)
|
388 |
+
else:
|
389 |
+
sample = upsample_block(
|
390 |
+
hidden_states=sample,
|
391 |
+
temb=emb,
|
392 |
+
res_hidden_states_tuple=res_samples,
|
393 |
+
encoder_hidden_states=encoder_hidden_states,
|
394 |
+
upsample_size=upsample_size,
|
395 |
+
)
|
396 |
+
else:
|
397 |
+
if isinstance(upsample_block, CrossAttnUpBlock2D):
|
398 |
+
upsample_block.attentions
|
399 |
+
import ipdb;ipdb.set_trace()
|
400 |
+
sample = upsample_block(
|
401 |
+
hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples, upsample_size=upsample_size
|
402 |
+
)
|
403 |
+
# 6. post-process
|
404 |
+
sample = self.conv_norm_out(sample)
|
405 |
+
sample = self.conv_act(sample)
|
406 |
+
sample = self.conv_out(sample)
|
407 |
+
|
408 |
+
if not return_dict:
|
409 |
+
return (sample,)
|
410 |
+
|
411 |
+
return UNet2DConditionOutput(sample=sample)
|
sample.py
ADDED
@@ -0,0 +1,109 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import json
|
3 |
+
import time
|
4 |
+
import argparse
|
5 |
+
import imageio
|
6 |
+
import torch
|
7 |
+
import numpy as np
|
8 |
+
from torchvision import transforms
|
9 |
+
|
10 |
+
from models.region_diffusion import RegionDiffusion
|
11 |
+
from utils.attention_utils import get_token_maps
|
12 |
+
from utils.richtext_utils import seed_everything, parse_json, get_region_diffusion_input,\
|
13 |
+
get_attention_control_input, get_gradient_guidance_input
|
14 |
+
|
15 |
+
|
16 |
+
def main(args, param):
|
17 |
+
|
18 |
+
# Create the folder to store outputs.
|
19 |
+
run_dir = args.run_dir
|
20 |
+
os.makedirs(args.run_dir, exist_ok=True)
|
21 |
+
|
22 |
+
# Load region diffusion model.
|
23 |
+
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
24 |
+
model = RegionDiffusion(device)
|
25 |
+
|
26 |
+
# parse json to span attributes
|
27 |
+
base_text_prompt, style_text_prompts, footnote_text_prompts, footnote_target_tokens,\
|
28 |
+
color_text_prompts, color_names, color_rgbs, size_text_prompts_and_sizes, use_grad_guidance = parse_json(
|
29 |
+
param['text_input'])
|
30 |
+
|
31 |
+
# create control input for region diffusion
|
32 |
+
region_text_prompts, region_target_token_ids, base_tokens = get_region_diffusion_input(
|
33 |
+
model, base_text_prompt, style_text_prompts, footnote_text_prompts,
|
34 |
+
footnote_target_tokens, color_text_prompts, color_names)
|
35 |
+
|
36 |
+
# create control input for cross attention
|
37 |
+
text_format_dict = get_attention_control_input(
|
38 |
+
model, base_tokens, size_text_prompts_and_sizes)
|
39 |
+
|
40 |
+
# create control input for region guidance
|
41 |
+
text_format_dict, color_target_token_ids = get_gradient_guidance_input(
|
42 |
+
model, base_tokens, color_text_prompts, color_rgbs, text_format_dict)
|
43 |
+
|
44 |
+
height = param['height']
|
45 |
+
width = param['width']
|
46 |
+
seed = param['noise_index']
|
47 |
+
negative_text = param['negative_prompt']
|
48 |
+
seed_everything(seed)
|
49 |
+
|
50 |
+
# get token maps from plain text to image generation.
|
51 |
+
begin_time = time.time()
|
52 |
+
if model.attention_maps is None:
|
53 |
+
model.register_evaluation_hooks()
|
54 |
+
else:
|
55 |
+
model.reset_attention_maps()
|
56 |
+
plain_img = model.produce_attn_maps([base_text_prompt], [negative_text],
|
57 |
+
height=height, width=width, num_inference_steps=param['steps'],
|
58 |
+
guidance_scale=param['guidance_weight'])
|
59 |
+
fn_base = os.path.join(run_dir, 'seed%d_plain.png' % (seed))
|
60 |
+
imageio.imwrite(fn_base, plain_img[0])
|
61 |
+
print('time lapses to get attention maps: %.4f' % (time.time()-begin_time))
|
62 |
+
color_obj_masks = get_token_maps(
|
63 |
+
model.attention_maps, run_dir, width//8, height//8, color_target_token_ids, seed)
|
64 |
+
model.masks = get_token_maps(
|
65 |
+
model.attention_maps, run_dir, width//8, height//8, region_target_token_ids, seed, base_tokens)
|
66 |
+
color_obj_masks = [transforms.functional.resize(color_obj_mask, (height, width),
|
67 |
+
interpolation=transforms.InterpolationMode.BICUBIC,
|
68 |
+
antialias=True)
|
69 |
+
for color_obj_mask in color_obj_masks]
|
70 |
+
text_format_dict['color_obj_atten'] = color_obj_masks
|
71 |
+
model.remove_evaluation_hooks()
|
72 |
+
|
73 |
+
# generate image from rich text
|
74 |
+
begin_time = time.time()
|
75 |
+
seed_everything(seed)
|
76 |
+
rich_img = model.prompt_to_img(region_text_prompts, [negative_text],
|
77 |
+
height=height, width=width, num_inference_steps=param['steps'],
|
78 |
+
guidance_scale=param['guidance_weight'], use_grad_guidance=use_grad_guidance,
|
79 |
+
text_format_dict=text_format_dict)
|
80 |
+
print('time lapses to generate image from rich text: %.4f' %
|
81 |
+
(time.time()-begin_time))
|
82 |
+
fn_style = os.path.join(run_dir, 'seed%d_rich.png' % (seed))
|
83 |
+
imageio.imwrite(fn_style, rich_img[0])
|
84 |
+
# imageio.imwrite(fn_cat, np.concatenate([img[0], rich_img[0]], 1))
|
85 |
+
|
86 |
+
|
87 |
+
if __name__ == '__main__':
|
88 |
+
parser = argparse.ArgumentParser()
|
89 |
+
parser.add_argument('--run_dir', type=str, default='results/release/debug')
|
90 |
+
parser.add_argument('--height', type=int, default=512)
|
91 |
+
parser.add_argument('--width', type=int, default=512)
|
92 |
+
parser.add_argument('--seed', type=int, default=6)
|
93 |
+
parser.add_argument('--sample_steps', type=int, default=41)
|
94 |
+
parser.add_argument('--rich_text_json', type=str,
|
95 |
+
default='{"ops":[{"insert":"A close-up 4k dslr photo of a "},{"attributes":{"link":"A cat wearing sunglasses and a bandana around its neck."},"insert":"cat"},{"insert":" riding a scooter. There are palm trees in the background."}]}')
|
96 |
+
parser.add_argument('--negative_prompt', type=str, default='')
|
97 |
+
parser.add_argument('--guidance_weight', type=float, default=8.5)
|
98 |
+
args = parser.parse_args()
|
99 |
+
param = {
|
100 |
+
'text_input': json.loads(args.rich_text_json),
|
101 |
+
'height': args.height,
|
102 |
+
'width': args.width,
|
103 |
+
'guidance_weight': args.guidance_weight,
|
104 |
+
'steps': args.sample_steps,
|
105 |
+
'noise_index': args.seed,
|
106 |
+
'negative_prompt': args.negative_prompt,
|
107 |
+
}
|
108 |
+
|
109 |
+
main(args, param)
|
utils/.DS_Store
ADDED
Binary file (6.15 kB). View file
|
|
utils/attention_utils.py
ADDED
@@ -0,0 +1,201 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
import os
|
3 |
+
import matplotlib as mpl
|
4 |
+
import matplotlib.pyplot as plt
|
5 |
+
import seaborn as sns
|
6 |
+
import torch
|
7 |
+
import torchvision
|
8 |
+
|
9 |
+
from pathlib import Path
|
10 |
+
import skimage
|
11 |
+
from skimage.morphology import erosion, square
|
12 |
+
|
13 |
+
|
14 |
+
def split_attention_maps_over_steps(attention_maps):
|
15 |
+
r"""Function for splitting attention maps over steps.
|
16 |
+
Args:
|
17 |
+
attention_maps (dict): Dictionary of attention maps.
|
18 |
+
sampler_order (int): Order of the sampler.
|
19 |
+
"""
|
20 |
+
# This function splits attention maps into unconditional and conditional score and over steps
|
21 |
+
|
22 |
+
attention_maps_cond = dict() # Maps corresponding to conditional score
|
23 |
+
attention_maps_uncond = dict() # Maps corresponding to unconditional score
|
24 |
+
|
25 |
+
for layer in attention_maps.keys():
|
26 |
+
|
27 |
+
for step_num in range(len(attention_maps[layer])):
|
28 |
+
if step_num not in attention_maps_cond:
|
29 |
+
attention_maps_cond[step_num] = dict()
|
30 |
+
attention_maps_uncond[step_num] = dict()
|
31 |
+
|
32 |
+
attention_maps_uncond[step_num].update(
|
33 |
+
{layer: attention_maps[layer][step_num][:1]})
|
34 |
+
attention_maps_cond[step_num].update(
|
35 |
+
{layer: attention_maps[layer][step_num][1:2]})
|
36 |
+
|
37 |
+
return attention_maps_cond, attention_maps_uncond
|
38 |
+
|
39 |
+
|
40 |
+
def plot_attention_maps(atten_map_list, obj_tokens, save_dir, seed, tokens_vis=None):
|
41 |
+
atten_names = ['presoftmax', 'postsoftmax', 'postsoftmax_erosion']
|
42 |
+
for i, (attn_map, obj_token) in enumerate(zip(atten_map_list, obj_tokens)):
|
43 |
+
n_obj = len(attn_map)
|
44 |
+
plt.figure()
|
45 |
+
plt.clf()
|
46 |
+
|
47 |
+
fig, axs = plt.subplots(
|
48 |
+
ncols=n_obj+1, gridspec_kw=dict(width_ratios=[1 for _ in range(n_obj)]+[0.1]))
|
49 |
+
|
50 |
+
fig.set_figheight(3)
|
51 |
+
fig.set_figwidth(3*n_obj+0.1)
|
52 |
+
|
53 |
+
cmap = plt.get_cmap('OrRd')
|
54 |
+
|
55 |
+
vmax = 0
|
56 |
+
vmin = 1
|
57 |
+
for tid in range(n_obj):
|
58 |
+
attention_map_cur = attn_map[tid]
|
59 |
+
vmax = max(vmax, float(attention_map_cur.max()))
|
60 |
+
vmin = min(vmin, float(attention_map_cur.min()))
|
61 |
+
|
62 |
+
for tid in range(n_obj):
|
63 |
+
sns.heatmap(
|
64 |
+
attn_map[tid][0], annot=False, cbar=False, ax=axs[tid],
|
65 |
+
cmap=cmap, vmin=vmin, vmax=vmax
|
66 |
+
)
|
67 |
+
axs[tid].set_axis_off()
|
68 |
+
if tokens_vis is not None:
|
69 |
+
if tid == n_obj-1:
|
70 |
+
axs_xlabel = 'other tokens'
|
71 |
+
else:
|
72 |
+
axs_xlabel = ''
|
73 |
+
for token_id in obj_tokens[tid]:
|
74 |
+
axs_xlabel += tokens_vis[token_id.item() -
|
75 |
+
1][:-len('</w>')]
|
76 |
+
axs[tid].set_title(axs_xlabel)
|
77 |
+
|
78 |
+
norm = mpl.colors.Normalize(vmin=vmin, vmax=vmax)
|
79 |
+
sm = plt.cm.ScalarMappable(cmap=cmap, norm=norm)
|
80 |
+
fig.colorbar(sm, cax=axs[-1])
|
81 |
+
|
82 |
+
fig.tight_layout()
|
83 |
+
plt.savefig(os.path.join(
|
84 |
+
save_dir, 'token_mapes_seed%d_%s.png' % (seed, atten_names[i])), dpi=100)
|
85 |
+
plt.close('all')
|
86 |
+
|
87 |
+
|
88 |
+
def get_token_maps(attention_maps, save_dir, width, height, obj_tokens, seed=0, tokens_vis=None,
|
89 |
+
preprocess=False):
|
90 |
+
r"""Function to visualize attention maps.
|
91 |
+
Args:
|
92 |
+
save_dir (str): Path to save attention maps
|
93 |
+
batch_size (int): Batch size
|
94 |
+
sampler_order (int): Sampler order
|
95 |
+
"""
|
96 |
+
|
97 |
+
# Split attention maps over steps
|
98 |
+
attention_maps_cond, _ = split_attention_maps_over_steps(
|
99 |
+
attention_maps
|
100 |
+
)
|
101 |
+
|
102 |
+
selected_layers = [
|
103 |
+
# 'down_blocks.0.attentions.0.transformer_blocks.0.attn2',
|
104 |
+
# 'down_blocks.0.attentions.1.transformer_blocks.0.attn2',
|
105 |
+
'down_blocks.1.attentions.0.transformer_blocks.0.attn2',
|
106 |
+
# 'down_blocks.1.attentions.1.transformer_blocks.0.attn2',
|
107 |
+
'down_blocks.2.attentions.0.transformer_blocks.0.attn2',
|
108 |
+
'down_blocks.2.attentions.1.transformer_blocks.0.attn2',
|
109 |
+
'mid_block.attentions.0.transformer_blocks.0.attn2',
|
110 |
+
'up_blocks.1.attentions.0.transformer_blocks.0.attn2',
|
111 |
+
'up_blocks.1.attentions.1.transformer_blocks.0.attn2',
|
112 |
+
'up_blocks.1.attentions.2.transformer_blocks.0.attn2',
|
113 |
+
# 'up_blocks.2.attentions.0.transformer_blocks.0.attn2',
|
114 |
+
'up_blocks.2.attentions.1.transformer_blocks.0.attn2',
|
115 |
+
# 'up_blocks.2.attentions.2.transformer_blocks.0.attn2',
|
116 |
+
# 'up_blocks.3.attentions.0.transformer_blocks.0.attn2',
|
117 |
+
# 'up_blocks.3.attentions.1.transformer_blocks.0.attn2',
|
118 |
+
# 'up_blocks.3.attentions.2.transformer_blocks.0.attn2'
|
119 |
+
]
|
120 |
+
|
121 |
+
nsteps = len(attention_maps_cond)
|
122 |
+
hw_ori = width * height
|
123 |
+
|
124 |
+
attention_maps = []
|
125 |
+
for obj_token in obj_tokens:
|
126 |
+
attention_maps.append([])
|
127 |
+
|
128 |
+
for step_num in range(nsteps):
|
129 |
+
attention_maps_cur = attention_maps_cond[step_num]
|
130 |
+
|
131 |
+
for layer in attention_maps_cur.keys():
|
132 |
+
if step_num < 10 or layer not in selected_layers:
|
133 |
+
continue
|
134 |
+
|
135 |
+
attention_ind = attention_maps_cur[layer].cpu()
|
136 |
+
|
137 |
+
# Attention maps are of shape [batch_size, nkeys, 77]
|
138 |
+
# since they are averaged out while collecting from hooks to save memory.
|
139 |
+
# Now split the heads from batch dimension
|
140 |
+
bs, hw, nclip = attention_ind.shape
|
141 |
+
down_ratio = np.sqrt(hw_ori // hw)
|
142 |
+
width_cur = int(width // down_ratio)
|
143 |
+
height_cur = int(height // down_ratio)
|
144 |
+
attention_ind = attention_ind.reshape(
|
145 |
+
bs, height_cur, width_cur, nclip)
|
146 |
+
for obj_id, obj_token in enumerate(obj_tokens):
|
147 |
+
if obj_token[0] == -1:
|
148 |
+
attention_map_prev = torch.stack(
|
149 |
+
[attention_maps[i][-1] for i in range(obj_id)]).sum(0)
|
150 |
+
attention_maps[obj_id].append(
|
151 |
+
attention_map_prev.max()-attention_map_prev)
|
152 |
+
else:
|
153 |
+
obj_attention_map = attention_ind[:, :, :, obj_token].max(-1, True)[
|
154 |
+
0].permute([3, 0, 1, 2])
|
155 |
+
obj_attention_map = torchvision.transforms.functional.resize(obj_attention_map, (height, width),
|
156 |
+
interpolation=torchvision.transforms.InterpolationMode.BICUBIC, antialias=True)
|
157 |
+
attention_maps[obj_id].append(obj_attention_map)
|
158 |
+
|
159 |
+
# average attention maps over steps
|
160 |
+
attention_maps_averaged = []
|
161 |
+
for obj_id, obj_token in enumerate(obj_tokens):
|
162 |
+
if obj_id == len(obj_tokens) - 1:
|
163 |
+
attention_maps_averaged.append(
|
164 |
+
torch.cat(attention_maps[obj_id]).mean(0))
|
165 |
+
else:
|
166 |
+
attention_maps_averaged.append(
|
167 |
+
torch.cat(attention_maps[obj_id]).mean(0))
|
168 |
+
|
169 |
+
# normalize attention maps into [0, 1]
|
170 |
+
attention_maps_averaged_normalized = []
|
171 |
+
attention_maps_averaged_sum = torch.cat(attention_maps_averaged).sum(0)
|
172 |
+
for obj_id, obj_token in enumerate(obj_tokens):
|
173 |
+
attention_maps_averaged_normalized.append(
|
174 |
+
attention_maps_averaged[obj_id]/attention_maps_averaged_sum)
|
175 |
+
|
176 |
+
# softmax
|
177 |
+
attention_maps_averaged_normalized = (
|
178 |
+
torch.cat(attention_maps_averaged)/0.001).softmax(0)
|
179 |
+
attention_maps_averaged_normalized = [
|
180 |
+
attention_maps_averaged_normalized[i:i+1] for i in range(attention_maps_averaged_normalized.shape[0])]
|
181 |
+
|
182 |
+
if preprocess:
|
183 |
+
# it is possible to preprocess the attention maps here
|
184 |
+
selem = square(5)
|
185 |
+
attention_maps_averaged_eroded = [erosion(skimage.img_as_float(
|
186 |
+
map[0].numpy()*255), selem) for map in attention_maps_averaged_normalized[:2]]
|
187 |
+
attention_maps_averaged_eroded = [(torch.from_numpy(map).unsqueeze(
|
188 |
+
0)/255. > 0.8).float() for map in attention_maps_averaged_eroded]
|
189 |
+
attention_maps_averaged_eroded.append(
|
190 |
+
1 - torch.cat(attention_maps_averaged_eroded).sum(0, True))
|
191 |
+
plot_attention_maps([attention_maps_averaged, attention_maps_averaged_normalized,
|
192 |
+
attention_maps_averaged_eroded], obj_tokens, save_dir, seed, tokens_vis)
|
193 |
+
attention_maps_averaged_eroded = [attn_mask.unsqueeze(1).repeat(
|
194 |
+
[1, 4, 1, 1]).cuda() for attn_mask in attention_maps_averaged_eroded]
|
195 |
+
return attention_maps_averaged_eroded
|
196 |
+
else:
|
197 |
+
plot_attention_maps([attention_maps_averaged, attention_maps_averaged_normalized],
|
198 |
+
obj_tokens, save_dir, seed, tokens_vis)
|
199 |
+
attention_maps_averaged_normalized = [attn_mask.unsqueeze(1).repeat(
|
200 |
+
[1, 4, 1, 1]).cuda() for attn_mask in attention_maps_averaged_normalized]
|
201 |
+
return attention_maps_averaged_normalized
|
utils/richtext_utils.py
ADDED
@@ -0,0 +1,234 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import json
|
3 |
+
import torch
|
4 |
+
import random
|
5 |
+
import numpy as np
|
6 |
+
|
7 |
+
COLORS = {
|
8 |
+
'brown': [165, 42, 42],
|
9 |
+
'red': [255, 0, 0],
|
10 |
+
'pink': [253, 108, 158],
|
11 |
+
'orange': [255, 165, 0],
|
12 |
+
'yellow': [255, 255, 0],
|
13 |
+
'purple': [128, 0, 128],
|
14 |
+
'green': [0, 128, 0],
|
15 |
+
'blue': [0, 0, 255],
|
16 |
+
'white': [255, 255, 255],
|
17 |
+
'gray': [128, 128, 128],
|
18 |
+
'black': [0, 0, 0],
|
19 |
+
}
|
20 |
+
|
21 |
+
|
22 |
+
def seed_everything(seed):
|
23 |
+
random.seed(seed)
|
24 |
+
os.environ['PYTHONHASHSEED'] = str(seed)
|
25 |
+
np.random.seed(seed)
|
26 |
+
torch.manual_seed(seed)
|
27 |
+
torch.cuda.manual_seed(seed)
|
28 |
+
|
29 |
+
|
30 |
+
def hex_to_rgb(hex_string, return_nearest_color=False):
|
31 |
+
r"""
|
32 |
+
Covert Hex triplet to RGB triplet.
|
33 |
+
"""
|
34 |
+
# Remove '#' symbol if present
|
35 |
+
hex_string = hex_string.lstrip('#')
|
36 |
+
# Convert hex values to integers
|
37 |
+
red = int(hex_string[0:2], 16)
|
38 |
+
green = int(hex_string[2:4], 16)
|
39 |
+
blue = int(hex_string[4:6], 16)
|
40 |
+
rgb = torch.FloatTensor((red, green, blue))[None, :, None, None]/255.
|
41 |
+
if return_nearest_color:
|
42 |
+
nearest_color = find_nearest_color(rgb)
|
43 |
+
return rgb.cuda(), nearest_color
|
44 |
+
return rgb.cuda()
|
45 |
+
|
46 |
+
|
47 |
+
def find_nearest_color(rgb):
|
48 |
+
r"""
|
49 |
+
Find the nearest neighbor color given the RGB value.
|
50 |
+
"""
|
51 |
+
if isinstance(rgb, list) or isinstance(rgb, tuple):
|
52 |
+
rgb = torch.FloatTensor(rgb)[None, :, None, None]/255.
|
53 |
+
color_distance = torch.FloatTensor([np.linalg.norm(
|
54 |
+
rgb - torch.FloatTensor(COLORS[color])[None, :, None, None]/255.) for color in COLORS.keys()])
|
55 |
+
nearest_color = list(COLORS.keys())[torch.argmin(color_distance).item()]
|
56 |
+
return nearest_color
|
57 |
+
|
58 |
+
|
59 |
+
def font2style(font):
|
60 |
+
r"""
|
61 |
+
Convert the font name to the style name.
|
62 |
+
"""
|
63 |
+
return {'mirza': 'Claud Monet, impressionism, oil on canvas',
|
64 |
+
'roboto': 'Ukiyoe',
|
65 |
+
'cursive': 'Cyber Punk, futuristic, blade runner, william gibson, trending on artstation hq',
|
66 |
+
'sofia': 'Pop Art, masterpiece, andy warhol',
|
67 |
+
'slabo': 'Vincent Van Gogh',
|
68 |
+
'inconsolata': 'Pixel Art, 8 bits, 16 bits',
|
69 |
+
'ubuntu': 'Rembrandt',
|
70 |
+
'Monoton': 'neon art, colorful light, highly details, octane render',
|
71 |
+
'Akronim': 'Abstract Cubism, Pablo Picasso', }[font]
|
72 |
+
|
73 |
+
|
74 |
+
def parse_json(json_str):
|
75 |
+
r"""
|
76 |
+
Convert the JSON string to attributes.
|
77 |
+
"""
|
78 |
+
# initialze region-base attributes.
|
79 |
+
base_text_prompt = ''
|
80 |
+
style_text_prompts = []
|
81 |
+
footnote_text_prompts = []
|
82 |
+
footnote_target_tokens = []
|
83 |
+
color_text_prompts = []
|
84 |
+
color_rgbs = []
|
85 |
+
color_names = []
|
86 |
+
size_text_prompts_and_sizes = []
|
87 |
+
|
88 |
+
# parse the attributes from JSON.
|
89 |
+
prev_style = None
|
90 |
+
prev_color_rgb = None
|
91 |
+
use_grad_guidance = False
|
92 |
+
for span in json_str['ops']:
|
93 |
+
text_prompt = span['insert'].rstrip('\n')
|
94 |
+
base_text_prompt += span['insert'].rstrip('\n')
|
95 |
+
if text_prompt == ' ':
|
96 |
+
continue
|
97 |
+
if 'attributes' in span:
|
98 |
+
if 'font' in span['attributes']:
|
99 |
+
style = font2style(span['attributes']['font'])
|
100 |
+
if prev_style == style:
|
101 |
+
prev_text_prompt = style_text_prompts[-1].split('in the style of')[
|
102 |
+
0]
|
103 |
+
style_text_prompts[-1] = prev_text_prompt + \
|
104 |
+
' ' + text_prompt + f' in the style of {style}'
|
105 |
+
else:
|
106 |
+
style_text_prompts.append(
|
107 |
+
text_prompt + f' in the style of {style}')
|
108 |
+
prev_style = style
|
109 |
+
else:
|
110 |
+
prev_style = None
|
111 |
+
if 'link' in span['attributes']:
|
112 |
+
footnote_text_prompts.append(span['attributes']['link'])
|
113 |
+
footnote_target_tokens.append(text_prompt)
|
114 |
+
font_size = 1
|
115 |
+
if 'size' in span['attributes'] and 'strike' not in span['attributes']:
|
116 |
+
font_size = float(span['attributes']['size'][:-2])/3.
|
117 |
+
elif 'size' in span['attributes'] and 'strike' in span['attributes']:
|
118 |
+
font_size = -float(span['attributes']['size'][:-2])/3.
|
119 |
+
elif 'size' not in span['attributes'] and 'strike' not in span['attributes']:
|
120 |
+
font_size = 1
|
121 |
+
if 'color' in span['attributes']:
|
122 |
+
use_grad_guidance = True
|
123 |
+
color_rgb, nearest_color = hex_to_rgb(
|
124 |
+
span['attributes']['color'], True)
|
125 |
+
if prev_color_rgb == color_rgb:
|
126 |
+
prev_text_prompt = color_text_prompts[-1]
|
127 |
+
color_text_prompts[-1] = prev_text_prompt + \
|
128 |
+
' ' + text_prompt
|
129 |
+
else:
|
130 |
+
color_rgbs.append(color_rgb)
|
131 |
+
color_names.append(nearest_color)
|
132 |
+
color_text_prompts.append(text_prompt)
|
133 |
+
if font_size != 1:
|
134 |
+
size_text_prompts_and_sizes.append([text_prompt, font_size])
|
135 |
+
return base_text_prompt, style_text_prompts, footnote_text_prompts, footnote_target_tokens,\
|
136 |
+
color_text_prompts, color_names, color_rgbs, size_text_prompts_and_sizes, use_grad_guidance
|
137 |
+
|
138 |
+
|
139 |
+
def get_region_diffusion_input(model, base_text_prompt, style_text_prompts, footnote_text_prompts,
|
140 |
+
footnote_target_tokens, color_text_prompts, color_names):
|
141 |
+
r"""
|
142 |
+
Algorithm 1 in the paper.
|
143 |
+
"""
|
144 |
+
region_text_prompts = []
|
145 |
+
region_target_token_ids = []
|
146 |
+
base_tokens = model.tokenizer._tokenize(base_text_prompt)
|
147 |
+
# process the style text prompt
|
148 |
+
for text_prompt in style_text_prompts:
|
149 |
+
region_text_prompts.append(text_prompt)
|
150 |
+
region_target_token_ids.append([])
|
151 |
+
style_tokens = model.tokenizer._tokenize(
|
152 |
+
text_prompt.split('in the style of')[0])
|
153 |
+
for style_token in style_tokens:
|
154 |
+
region_target_token_ids[-1].append(
|
155 |
+
base_tokens.index(style_token)+1)
|
156 |
+
|
157 |
+
# process the complementary text prompt
|
158 |
+
for footnote_text_prompt, text_prompt in zip(footnote_text_prompts, footnote_target_tokens):
|
159 |
+
region_target_token_ids.append([])
|
160 |
+
region_text_prompts.append(footnote_text_prompt)
|
161 |
+
style_tokens = model.tokenizer._tokenize(text_prompt)
|
162 |
+
for style_token in style_tokens:
|
163 |
+
region_target_token_ids[-1].append(
|
164 |
+
base_tokens.index(style_token)+1)
|
165 |
+
|
166 |
+
# process the color text prompt
|
167 |
+
for color_text_prompt, color_name in zip(color_text_prompts, color_names):
|
168 |
+
region_target_token_ids.append([])
|
169 |
+
region_text_prompts.append(color_name+' '+color_text_prompt)
|
170 |
+
style_tokens = model.tokenizer._tokenize(color_text_prompt)
|
171 |
+
for style_token in style_tokens:
|
172 |
+
region_target_token_ids[-1].append(
|
173 |
+
base_tokens.index(style_token)+1)
|
174 |
+
|
175 |
+
# process the remaining tokens without any attributes
|
176 |
+
region_text_prompts.append(base_text_prompt)
|
177 |
+
region_target_token_ids_all = [
|
178 |
+
id for ids in region_target_token_ids for id in ids]
|
179 |
+
target_token_ids_rest = [id for id in range(
|
180 |
+
1, len(base_tokens)+1) if id not in region_target_token_ids_all]
|
181 |
+
region_target_token_ids.append(target_token_ids_rest)
|
182 |
+
|
183 |
+
region_target_token_ids = [torch.LongTensor(
|
184 |
+
obj_token_id) for obj_token_id in region_target_token_ids]
|
185 |
+
return region_text_prompts, region_target_token_ids, base_tokens
|
186 |
+
|
187 |
+
|
188 |
+
def get_attention_control_input(model, base_tokens, size_text_prompts_and_sizes):
|
189 |
+
r"""
|
190 |
+
Control the token impact using font sizes.
|
191 |
+
"""
|
192 |
+
word_pos = []
|
193 |
+
font_sizes = []
|
194 |
+
for text_prompt, font_size in size_text_prompts_and_sizes:
|
195 |
+
size_tokens = model.tokenizer._tokenize(text_prompt)
|
196 |
+
for size_token in size_tokens:
|
197 |
+
word_pos.append(base_tokens.index(size_token)+1)
|
198 |
+
font_sizes.append(font_size)
|
199 |
+
if len(word_pos) > 0:
|
200 |
+
word_pos = torch.LongTensor(word_pos).cuda()
|
201 |
+
font_sizes = torch.FloatTensor(font_sizes).cuda()
|
202 |
+
else:
|
203 |
+
word_pos = None
|
204 |
+
font_sizes = None
|
205 |
+
text_format_dict = {
|
206 |
+
'word_pos': word_pos,
|
207 |
+
'font_size': font_sizes,
|
208 |
+
}
|
209 |
+
return text_format_dict
|
210 |
+
|
211 |
+
|
212 |
+
def get_gradient_guidance_input(model, base_tokens, color_text_prompts, color_rgbs, text_format_dict,
|
213 |
+
guidance_start_step=999, color_guidance_weight=1):
|
214 |
+
r"""
|
215 |
+
Control the token impact using font sizes.
|
216 |
+
"""
|
217 |
+
color_target_token_ids = []
|
218 |
+
for text_prompt in color_text_prompts:
|
219 |
+
color_target_token_ids.append([])
|
220 |
+
color_tokens = model.tokenizer._tokenize(text_prompt)
|
221 |
+
for color_token in color_tokens:
|
222 |
+
color_target_token_ids[-1].append(base_tokens.index(color_token)+1)
|
223 |
+
color_target_token_ids_all = [
|
224 |
+
id for ids in color_target_token_ids for id in ids]
|
225 |
+
color_target_token_ids_rest = [id for id in range(
|
226 |
+
1, len(base_tokens)+1) if id not in color_target_token_ids_all]
|
227 |
+
color_target_token_ids.append(color_target_token_ids_rest)
|
228 |
+
color_target_token_ids = [torch.LongTensor(
|
229 |
+
obj_token_id) for obj_token_id in color_target_token_ids]
|
230 |
+
|
231 |
+
text_format_dict['target_RGB'] = color_rgbs
|
232 |
+
text_format_dict['guidance_start_step'] = guidance_start_step
|
233 |
+
text_format_dict['color_guidance_weight'] = color_guidance_weight
|
234 |
+
return text_format_dict, color_target_token_ids
|