liltom-eth
commited on
Commit
·
f00035f
1
Parent(s):
f255825
Upload inference.py with huggingface_hub
Browse files- inference.py +93 -0
inference.py
ADDED
@@ -0,0 +1,93 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import requests
|
2 |
+
from PIL import Image
|
3 |
+
from io import BytesIO
|
4 |
+
import torch
|
5 |
+
from transformers import AutoTokenizer
|
6 |
+
|
7 |
+
from llava.model import LlavaLlamaForCausalLM
|
8 |
+
from llava.utils import disable_torch_init
|
9 |
+
from llava.mm_utils import tokenizer_image_token, KeywordsStoppingCriteria
|
10 |
+
|
11 |
+
from llava.conversation import conv_templates, SeparatorStyle
|
12 |
+
from llava.constants import (
|
13 |
+
IMAGE_TOKEN_INDEX,
|
14 |
+
DEFAULT_IMAGE_TOKEN,
|
15 |
+
DEFAULT_IM_START_TOKEN,
|
16 |
+
DEFAULT_IM_END_TOKEN,
|
17 |
+
)
|
18 |
+
|
19 |
+
|
20 |
+
def model_fn(model_dir):
|
21 |
+
kwargs = {"device_map": "auto"}
|
22 |
+
kwargs["torch_dtype"] = torch.float16
|
23 |
+
model = LlavaLlamaForCausalLM.from_pretrained(
|
24 |
+
model_dir, low_cpu_mem_usage=True, **kwargs
|
25 |
+
)
|
26 |
+
tokenizer = AutoTokenizer.from_pretrained(model_dir, use_fast=False)
|
27 |
+
|
28 |
+
vision_tower = model.get_vision_tower()
|
29 |
+
if not vision_tower.is_loaded:
|
30 |
+
vision_tower.load_model()
|
31 |
+
vision_tower.to(device="cuda", dtype=torch.float16)
|
32 |
+
image_processor = vision_tower.image_processor
|
33 |
+
return model, tokenizer, image_processor
|
34 |
+
|
35 |
+
|
36 |
+
def predict_fn(data, model_and_tokenizer):
|
37 |
+
# unpack model and tokenizer
|
38 |
+
model, tokenizer, image_processor = model_and_tokenizer
|
39 |
+
|
40 |
+
# get prompt & parameters
|
41 |
+
image_file = data.pop("image", data)
|
42 |
+
raw_prompt = data.pop("question", data)
|
43 |
+
|
44 |
+
max_new_tokens = data.pop("max_new_tokens", 1024)
|
45 |
+
temperature = data.pop("temperature", 0.2)
|
46 |
+
conv_mode = data.pop("conv_mode", "llava_v1")
|
47 |
+
|
48 |
+
# conv_mode = "llava_v1"
|
49 |
+
conv = conv_templates[conv_mode].copy()
|
50 |
+
roles = conv.roles
|
51 |
+
inp = f"{roles[0]}: {raw_prompt}"
|
52 |
+
inp = (
|
53 |
+
DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + "\n" + inp
|
54 |
+
)
|
55 |
+
conv.append_message(conv.roles[0], inp)
|
56 |
+
conv.append_message(conv.roles[1], None)
|
57 |
+
prompt = conv.get_prompt()
|
58 |
+
stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
|
59 |
+
|
60 |
+
if image_file.startswith("http") or image_file.startswith("https"):
|
61 |
+
response = requests.get(image_file)
|
62 |
+
image = Image.open(BytesIO(response.content)).convert("RGB")
|
63 |
+
else:
|
64 |
+
image = Image.open(image_file).convert("RGB")
|
65 |
+
|
66 |
+
disable_torch_init()
|
67 |
+
image_tensor = (
|
68 |
+
image_processor.preprocess(image, return_tensors="pt")["pixel_values"]
|
69 |
+
.half()
|
70 |
+
.cuda()
|
71 |
+
)
|
72 |
+
|
73 |
+
keywords = [stop_str]
|
74 |
+
input_ids = (
|
75 |
+
tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt")
|
76 |
+
.unsqueeze(0)
|
77 |
+
.cuda()
|
78 |
+
)
|
79 |
+
stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids)
|
80 |
+
with torch.inference_mode():
|
81 |
+
output_ids = model.generate(
|
82 |
+
input_ids,
|
83 |
+
images=image_tensor,
|
84 |
+
do_sample=True,
|
85 |
+
temperature=temperature,
|
86 |
+
max_new_tokens=max_new_tokens,
|
87 |
+
use_cache=True,
|
88 |
+
stopping_criteria=[stopping_criteria],
|
89 |
+
)
|
90 |
+
outputs = tokenizer.decode(
|
91 |
+
output_ids[0, input_ids.shape[1] :], skip_special_tokens=True
|
92 |
+
).strip()
|
93 |
+
return outputs
|