Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,64 +1,107 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
import gradio as gr
|
| 2 |
from transformers import AutoProcessor, AutoModelForImageTextToText
|
| 3 |
from PIL import Image
|
| 4 |
-
import
|
| 5 |
|
| 6 |
-
# Load
|
| 7 |
processor = AutoProcessor.from_pretrained("ds4sd/SmolDocling-256M-preview")
|
| 8 |
model = AutoModelForImageTextToText.from_pretrained("ds4sd/SmolDocling-256M-preview")
|
| 9 |
|
| 10 |
-
def
|
| 11 |
-
"""Run SmolDocling on an image and return numeric values inside <fcel> tags."""
|
| 12 |
-
# Prepare prompt for the model
|
| 13 |
messages = [
|
| 14 |
{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": prompt_text}]}
|
| 15 |
]
|
| 16 |
prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
|
| 17 |
inputs = processor(text=prompt, images=[image], return_tensors="pt")
|
| 18 |
-
|
| 19 |
-
# Generate output
|
| 20 |
-
outputs = model.generate(**inputs, max_new_tokens=2048)
|
| 21 |
prompt_length = inputs.input_ids.shape[1]
|
| 22 |
generated = outputs[:, prompt_length:]
|
| 23 |
result = processor.batch_decode(generated, skip_special_tokens=False)[0]
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
# Extract only <fcel> values
|
| 27 |
-
values = re.findall(r"<fcel>([\d.]+)", clean_text)
|
| 28 |
-
values = [float(v) for v in values] # convert to floats
|
| 29 |
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
# Calculate accuracy
|
| 38 |
-
if len(values1) == len(values2) and values1 == values2:
|
| 39 |
-
accuracy = 100.0
|
| 40 |
-
else:
|
| 41 |
-
matches = sum(1 for a, b in zip(values1, values2) if a == b)
|
| 42 |
-
total = max(len(values1), len(values2))
|
| 43 |
-
accuracy = (matches / total) * 100 if total > 0 else 0
|
| 44 |
-
|
| 45 |
-
return {
|
| 46 |
-
"Extracted Values 1": values1,
|
| 47 |
-
"Extracted Values 2": values2,
|
| 48 |
-
"Accuracy (%)": accuracy
|
| 49 |
-
}
|
| 50 |
|
| 51 |
# Gradio UI
|
| 52 |
demo = gr.Interface(
|
| 53 |
-
fn=
|
| 54 |
inputs=[
|
| 55 |
-
gr.Image(type="pil", label="Upload
|
| 56 |
-
gr.
|
| 57 |
-
gr.Textbox(lines=1, placeholder="Enter prompt (e.g. Extract table as OTSL)", label="Prompt")
|
| 58 |
],
|
| 59 |
-
outputs=
|
| 60 |
-
title="
|
| 61 |
-
description="
|
| 62 |
)
|
| 63 |
|
| 64 |
demo.launch()
|
|
|
|
|
|
| 1 |
+
# import gradio as gr
|
| 2 |
+
# from transformers import AutoProcessor, AutoModelForImageTextToText
|
| 3 |
+
# from PIL import Image
|
| 4 |
+
# import re
|
| 5 |
+
|
| 6 |
+
# # Load SmolDocling model & processor once
|
| 7 |
+
# processor = AutoProcessor.from_pretrained("ds4sd/SmolDocling-256M-preview")
|
| 8 |
+
# model = AutoModelForImageTextToText.from_pretrained("ds4sd/SmolDocling-256M-preview")
|
| 9 |
+
|
| 10 |
+
# def extract_fcel_values_from_image(image, prompt_text):
|
| 11 |
+
# """Run SmolDocling on an image and return numeric values inside <fcel> tags."""
|
| 12 |
+
# # Prepare prompt for the model
|
| 13 |
+
# messages = [
|
| 14 |
+
# {"role": "user", "content": [{"type": "image"}, {"type": "text", "text": prompt_text}]}
|
| 15 |
+
# ]
|
| 16 |
+
# prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
|
| 17 |
+
# inputs = processor(text=prompt, images=[image], return_tensors="pt")
|
| 18 |
+
|
| 19 |
+
# # Generate output
|
| 20 |
+
# outputs = model.generate(**inputs, max_new_tokens=2048)
|
| 21 |
+
# prompt_length = inputs.input_ids.shape[1]
|
| 22 |
+
# generated = outputs[:, prompt_length:]
|
| 23 |
+
# result = processor.batch_decode(generated, skip_special_tokens=False)[0]
|
| 24 |
+
# clean_text = result.replace("<end_of_utterance>", "").strip()
|
| 25 |
+
|
| 26 |
+
# # Extract only <fcel> values
|
| 27 |
+
# values = re.findall(r"<fcel>([\d.]+)", clean_text)
|
| 28 |
+
# values = [float(v) for v in values] # convert to floats
|
| 29 |
+
|
| 30 |
+
# return values, clean_text
|
| 31 |
+
|
| 32 |
+
# def compare_images(image1, image2, prompt_text):
|
| 33 |
+
# # Extract fcel values from both images
|
| 34 |
+
# values1, raw1 = extract_fcel_values_from_image(image1, prompt_text)
|
| 35 |
+
# values2, raw2 = extract_fcel_values_from_image(image2, prompt_text)
|
| 36 |
+
|
| 37 |
+
# # Calculate accuracy
|
| 38 |
+
# if len(values1) == len(values2) and values1 == values2:
|
| 39 |
+
# accuracy = 100.0
|
| 40 |
+
# else:
|
| 41 |
+
# matches = sum(1 for a, b in zip(values1, values2) if a == b)
|
| 42 |
+
# total = max(len(values1), len(values2))
|
| 43 |
+
# accuracy = (matches / total) * 100 if total > 0 else 0
|
| 44 |
+
|
| 45 |
+
# return {
|
| 46 |
+
# "Extracted Values 1": values1,
|
| 47 |
+
# "Extracted Values 2": values2,
|
| 48 |
+
# "Accuracy (%)": accuracy
|
| 49 |
+
# }
|
| 50 |
+
|
| 51 |
+
# # Gradio UI
|
| 52 |
+
# demo = gr.Interface(
|
| 53 |
+
# fn=compare_images,
|
| 54 |
+
# inputs=[
|
| 55 |
+
# gr.Image(type="pil", label="Upload First Table Image"),
|
| 56 |
+
# gr.Image(type="pil", label="Upload Second Table Image"),
|
| 57 |
+
# gr.Textbox(lines=1, placeholder="Enter prompt (e.g. Extract table as OTSL)", label="Prompt")
|
| 58 |
+
# ],
|
| 59 |
+
# outputs="json",
|
| 60 |
+
# title="Table Data Accuracy Checker (SmolDocling)",
|
| 61 |
+
# description="Uploads two table images, extracts only <fcel> values from OTSL output, and compares them for accuracy."
|
| 62 |
+
# )
|
| 63 |
+
|
| 64 |
+
# demo.launch()
|
| 65 |
+
|
| 66 |
import gradio as gr
|
| 67 |
from transformers import AutoProcessor, AutoModelForImageTextToText
|
| 68 |
from PIL import Image
|
| 69 |
+
import json
|
| 70 |
|
| 71 |
+
# Load model & processor once at startup
|
| 72 |
processor = AutoProcessor.from_pretrained("ds4sd/SmolDocling-256M-preview")
|
| 73 |
model = AutoModelForImageTextToText.from_pretrained("ds4sd/SmolDocling-256M-preview")
|
| 74 |
|
| 75 |
+
def smoldocling_readimage(image, prompt_text):
|
|
|
|
|
|
|
| 76 |
messages = [
|
| 77 |
{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": prompt_text}]}
|
| 78 |
]
|
| 79 |
prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
|
| 80 |
inputs = processor(text=prompt, images=[image], return_tensors="pt")
|
| 81 |
+
outputs = model.generate(**inputs, max_new_tokens=1024)
|
|
|
|
|
|
|
| 82 |
prompt_length = inputs.input_ids.shape[1]
|
| 83 |
generated = outputs[:, prompt_length:]
|
| 84 |
result = processor.batch_decode(generated, skip_special_tokens=False)[0]
|
| 85 |
+
clean_result = result.replace("<end_of_utterance>", "").strip()
|
|
|
|
|
|
|
|
|
|
|
|
|
| 86 |
|
| 87 |
+
# Try to parse as JSON
|
| 88 |
+
try:
|
| 89 |
+
json_result = json.loads(clean_result)
|
| 90 |
+
return json_result
|
| 91 |
+
except json.JSONDecodeError:
|
| 92 |
+
return {"error": "Output is not valid JSON", "raw_output": clean_result}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 93 |
|
| 94 |
# Gradio UI
|
| 95 |
demo = gr.Interface(
|
| 96 |
+
fn=smoldocling_readimage,
|
| 97 |
inputs=[
|
| 98 |
+
gr.Image(type="pil", label="Upload Image"),
|
| 99 |
+
gr.Textbox(lines=1, placeholder="Enter prompt (e.g. Convert to docling)", label="Prompt"),
|
|
|
|
| 100 |
],
|
| 101 |
+
outputs=gr.JSON(),
|
| 102 |
+
title="SmolDocling Web App",
|
| 103 |
+
description="Upload a document image and convert it to structured docling format."
|
| 104 |
)
|
| 105 |
|
| 106 |
demo.launch()
|
| 107 |
+
|