Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -6,13 +6,16 @@ import os
|
|
6 |
|
7 |
import gradio as gr
|
8 |
import requests
|
9 |
-
from openai import OpenAI
|
10 |
from func_timeout import FunctionTimedOut, func_timeout
|
11 |
from tqdm import tqdm
|
12 |
|
13 |
HUGGINGFACE=True
|
14 |
-
MOCK =
|
15 |
TEST_FOLDER = "c4f5"
|
|
|
|
|
|
|
16 |
|
17 |
if HUGGINGFACE:
|
18 |
MODEL_NAME="xu3kev/deepseekcoder-7b-logo-pbe"
|
@@ -36,7 +39,7 @@ MOCK_RESPONSE = [
|
|
36 |
forward(2*i)
|
37 |
left(90.0)
|
38 |
"""
|
39 |
-
] *
|
40 |
|
41 |
LOGO_HEADER = """from myturtle_cv import Turtle
|
42 |
from myturtle import HALF_INF, INF, EPS_DIST, EPS_ANGLE
|
@@ -218,13 +221,14 @@ def generate_grid_images(gif_results):
|
|
218 |
plt.close(fig)
|
219 |
return image_array
|
220 |
|
|
|
221 |
@spaces.GPU
|
222 |
def llm_call(question_prompt, model_name,
|
223 |
temperature=1, max_tokens=320,
|
224 |
top_p=1, n_samples=64, stop=None):
|
225 |
if HUGGINGFACE:
|
226 |
model_inputs = hug_tokenizer([question_prompt], return_tensors="pt").to('cuda')
|
227 |
-
generated_ids = hug_model.generate(**model_inputs, max_length=1400, temperature=1, num_return_sequences=
|
228 |
responses = hug_tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
|
229 |
codes = []
|
230 |
for response in responses:
|
@@ -334,8 +338,9 @@ def run(img_str):
|
|
334 |
|
335 |
|
336 |
|
337 |
-
for code in tqdm(codes):
|
338 |
-
|
|
|
339 |
|
340 |
from concurrent.futures import ProcessPoolExecutor
|
341 |
from concurrent.futures import as_completed
|
@@ -381,10 +386,26 @@ def create_tmp_folder():
|
|
381 |
return folder_name
|
382 |
|
383 |
|
|
|
384 |
def img_to_code_img(sketchpad_img):
|
385 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
386 |
image_array = np.array(img)
|
387 |
-
image_array =
|
|
|
388 |
|
389 |
# height, width = image_array.shape
|
390 |
# output_size = 512
|
@@ -417,13 +438,16 @@ def img_to_code_img(sketchpad_img):
|
|
417 |
# return generated_grid_img[0]
|
418 |
|
419 |
folder = create_tmp_folder()
|
420 |
-
|
|
|
421 |
for i in range(len(gif_results)):
|
422 |
if gif_results[i]:
|
423 |
with open(f"{folder}/img{i}.gif", "wb") as f:
|
424 |
f.write(gif_results[i])
|
425 |
-
|
426 |
-
|
|
|
|
|
427 |
|
428 |
|
429 |
def main():
|
@@ -434,22 +458,46 @@ def main():
|
|
434 |
from gradio import Brush
|
435 |
theme = gr.themes.Default().set(
|
436 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
437 |
with gr.Blocks(theme=theme) as demo:
|
438 |
gr.Markdown('# Visual Program Synthesis with LLM')
|
439 |
gr.Markdown("""LOGO/Turtle graphics Programming-by-Example problems aims to synthesize a program that generates the given target image, where the program uses drawing library similar to Python Turtle.""")
|
440 |
gr.Markdown("""Here we can draw a target image using the sketchpad, and see what kinds of graphics program LLM generates. To allow the LLM to visually perceive the input image, we convert the image to ASCII strings.""")
|
441 |
gr.Markdown("Please check out our [paper](https://arxiv.org/abs/2406.08316) for more details!")
|
442 |
-
gr.Markdown("##
|
443 |
with gr.Row():
|
444 |
with gr.Column(scale=1):
|
445 |
-
|
446 |
-
|
|
|
447 |
with gr.Column(scale=4):
|
448 |
output_gallery = gr.Gallery(
|
449 |
-
|
450 |
-
|
451 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
452 |
|
|
|
|
|
453 |
submit_button.click(img_to_code_img, inputs=canvas, outputs=output_gallery)
|
454 |
demo.load(
|
455 |
None,
|
|
|
6 |
|
7 |
import gradio as gr
|
8 |
import requests
|
9 |
+
# from openai import OpenAI
|
10 |
from func_timeout import FunctionTimedOut, func_timeout
|
11 |
from tqdm import tqdm
|
12 |
|
13 |
HUGGINGFACE=True
|
14 |
+
MOCK = not HUGGINGFACE
|
15 |
TEST_FOLDER = "c4f5"
|
16 |
+
NUM_RETURN_SEQ = 10
|
17 |
+
|
18 |
+
DROPDOWN = None
|
19 |
|
20 |
if HUGGINGFACE:
|
21 |
MODEL_NAME="xu3kev/deepseekcoder-7b-logo-pbe"
|
|
|
39 |
forward(2*i)
|
40 |
left(90.0)
|
41 |
"""
|
42 |
+
] * 10
|
43 |
|
44 |
LOGO_HEADER = """from myturtle_cv import Turtle
|
45 |
from myturtle import HALF_INF, INF, EPS_DIST, EPS_ANGLE
|
|
|
221 |
plt.close(fig)
|
222 |
return image_array
|
223 |
|
224 |
+
|
225 |
@spaces.GPU
|
226 |
def llm_call(question_prompt, model_name,
|
227 |
temperature=1, max_tokens=320,
|
228 |
top_p=1, n_samples=64, stop=None):
|
229 |
if HUGGINGFACE:
|
230 |
model_inputs = hug_tokenizer([question_prompt], return_tensors="pt").to('cuda')
|
231 |
+
generated_ids = hug_model.generate(**model_inputs, max_length=1400, temperature=1, num_return_sequences=NUM_RETURN_SEQ, do_sample=True)
|
232 |
responses = hug_tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
|
233 |
codes = []
|
234 |
for response in responses:
|
|
|
338 |
|
339 |
|
340 |
|
341 |
+
# for code in tqdm(codes):
|
342 |
+
# pass
|
343 |
+
print(f"Running {len(codes)} codes")
|
344 |
|
345 |
from concurrent.futures import ProcessPoolExecutor
|
346 |
from concurrent.futures import as_completed
|
|
|
386 |
return folder_name
|
387 |
|
388 |
|
389 |
+
CODES = []
|
390 |
def img_to_code_img(sketchpad_img):
|
391 |
+
|
392 |
+
from PIL import Image
|
393 |
+
# with open("debug_background.png", "wb") as f:
|
394 |
+
# # convert numpy to png
|
395 |
+
# numpy_array = sketchpad_img['background']
|
396 |
+
# img = Image.fromarray(numpy_array)
|
397 |
+
# img.save(f)
|
398 |
+
# with open("debug_composite.png", "wb") as f:
|
399 |
+
# # convert numpy to png
|
400 |
+
# numpy_array = sketchpad_img['composite']
|
401 |
+
# img = Image.fromarray(numpy_array)
|
402 |
+
# img.save(f)
|
403 |
+
|
404 |
+
# img = sketchpad_img['layers'][0]
|
405 |
+
img = sketchpad_img['composite']
|
406 |
image_array = np.array(img)
|
407 |
+
image_array = image_array[:,:,0]
|
408 |
+
# image_array = 255 - image_array[:,:,3]
|
409 |
|
410 |
# height, width = image_array.shape
|
411 |
# output_size = 512
|
|
|
438 |
# return generated_grid_img[0]
|
439 |
|
440 |
folder = create_tmp_folder()
|
441 |
+
global CODES
|
442 |
+
CODES = []
|
443 |
for i in range(len(gif_results)):
|
444 |
if gif_results[i]:
|
445 |
with open(f"{folder}/img{i}.gif", "wb") as f:
|
446 |
f.write(gif_results[i])
|
447 |
+
CODES.append(f"```python\n{codes[i]}\n```")
|
448 |
+
else:
|
449 |
+
CODES.append("#### Execution Error/Timeout; Skip")
|
450 |
+
return [f"{folder}/img{i}.gif" for i in range(len(gif_results))]
|
451 |
|
452 |
|
453 |
def main():
|
|
|
458 |
from gradio import Brush
|
459 |
theme = gr.themes.Default().set(
|
460 |
)
|
461 |
+
import os
|
462 |
+
# get all png files under demo_example
|
463 |
+
|
464 |
+
example_input_images = []
|
465 |
+
for root, dirs, files in os.walk("demo_example"):
|
466 |
+
for file in files:
|
467 |
+
if file.endswith(".png"):
|
468 |
+
example_input_images.append(os.path.join(root, file))
|
469 |
+
|
470 |
+
canvas = gr.Sketchpad(canvas_size=(512,512), brush=Brush(colors=["black"], default_size=2, color_mode='fixed'))
|
471 |
with gr.Blocks(theme=theme) as demo:
|
472 |
gr.Markdown('# Visual Program Synthesis with LLM')
|
473 |
gr.Markdown("""LOGO/Turtle graphics Programming-by-Example problems aims to synthesize a program that generates the given target image, where the program uses drawing library similar to Python Turtle.""")
|
474 |
gr.Markdown("""Here we can draw a target image using the sketchpad, and see what kinds of graphics program LLM generates. To allow the LLM to visually perceive the input image, we convert the image to ASCII strings.""")
|
475 |
gr.Markdown("Please check out our [paper](https://arxiv.org/abs/2406.08316) for more details!")
|
476 |
+
gr.Markdown("## Select an example logo input or draw your own logo!")
|
477 |
with gr.Row():
|
478 |
with gr.Column(scale=1):
|
479 |
+
gr.Examples(example_input_images, inputs=canvas)
|
480 |
+
canvas.render()
|
481 |
+
submit_button = gr.Button("Generate Programs")
|
482 |
with gr.Column(scale=4):
|
483 |
output_gallery = gr.Gallery(
|
484 |
+
label="Generated Images", show_label=True, elem_id="gallery"
|
485 |
+
, columns=[5], rows=[2], object_fit="contain", height="auto")
|
486 |
+
with gr.Group():
|
487 |
+
dropdown = gr.Dropdown([f"sample {i+1}" for i in range(NUM_RETURN_SEQ)], label='show generated program samples')
|
488 |
+
code_block = gr.Markdown('')
|
489 |
+
def update_code(sample_idx):
|
490 |
+
int_idx = int(sample_idx.split(" ")[1]) - 1
|
491 |
+
if int_idx < len(CODES):
|
492 |
+
return CODES[int_idx]
|
493 |
+
else:
|
494 |
+
return "### Please submit an image to generate programs."
|
495 |
+
#return gr.Markdown('333')
|
496 |
+
dropdown.input(update_code, dropdown, code_block)
|
497 |
+
# output_image = gr.Image(label="output")
|
498 |
|
499 |
+
global DROPDOWN
|
500 |
+
DROPDOWN = dropdown
|
501 |
submit_button.click(img_to_code_img, inputs=canvas, outputs=output_gallery)
|
502 |
demo.load(
|
503 |
None,
|