Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -62,7 +62,9 @@
|
|
| 62 |
# )
|
| 63 |
|
| 64 |
# demo.launch()
|
|
|
|
| 65 |
import re
|
|
|
|
| 66 |
import gradio as gr
|
| 67 |
from transformers import AutoProcessor, AutoModelForImageTextToText
|
| 68 |
from PIL import Image
|
|
@@ -82,12 +84,11 @@ def extract_values(docling_text):
|
|
| 82 |
continue
|
| 83 |
# Extract numbers inside <fcel> tags
|
| 84 |
values = re.findall(r"<fcel>(.*?)<fcel>", row)
|
| 85 |
-
# Convert to float list
|
| 86 |
float_values = [float(v) for v in values]
|
| 87 |
result.append(float_values)
|
| 88 |
return result
|
| 89 |
|
| 90 |
-
def
|
| 91 |
messages = [
|
| 92 |
{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": prompt_text}]}
|
| 93 |
]
|
|
@@ -97,20 +98,50 @@ def smoldocling_readimage(image, prompt_text):
|
|
| 97 |
prompt_length = inputs.input_ids.shape[1]
|
| 98 |
generated = outputs[:, prompt_length:]
|
| 99 |
raw_result = processor.batch_decode(generated, skip_special_tokens=False)[0]
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 103 |
|
| 104 |
-
# Gradio UI
|
| 105 |
demo = gr.Interface(
|
| 106 |
-
fn=
|
| 107 |
inputs=[
|
| 108 |
-
gr.Image(type="pil", label="Upload Image"),
|
|
|
|
| 109 |
gr.Textbox(lines=1, placeholder="Enter prompt (e.g. Convert to docling)", label="Prompt"),
|
| 110 |
],
|
| 111 |
outputs="text",
|
| 112 |
-
title="SmolDocling
|
| 113 |
-
description="Upload
|
| 114 |
)
|
| 115 |
|
| 116 |
demo.launch()
|
|
|
|
| 62 |
# )
|
| 63 |
|
| 64 |
# demo.launch()
|
| 65 |
+
|
| 66 |
import re
|
| 67 |
+
import numpy as np
|
| 68 |
import gradio as gr
|
| 69 |
from transformers import AutoProcessor, AutoModelForImageTextToText
|
| 70 |
from PIL import Image
|
|
|
|
| 84 |
continue
|
| 85 |
# Extract numbers inside <fcel> tags
|
| 86 |
values = re.findall(r"<fcel>(.*?)<fcel>", row)
|
|
|
|
| 87 |
float_values = [float(v) for v in values]
|
| 88 |
result.append(float_values)
|
| 89 |
return result
|
| 90 |
|
| 91 |
+
def get_array_from_image(image, prompt_text):
|
| 92 |
messages = [
|
| 93 |
{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": prompt_text}]}
|
| 94 |
]
|
|
|
|
| 98 |
prompt_length = inputs.input_ids.shape[1]
|
| 99 |
generated = outputs[:, prompt_length:]
|
| 100 |
raw_result = processor.batch_decode(generated, skip_special_tokens=False)[0]
|
| 101 |
+
return extract_values(raw_result)
|
| 102 |
+
|
| 103 |
+
def compare_arrays(arr1, arr2):
|
| 104 |
+
# Flatten both arrays (assumes 2D list)
|
| 105 |
+
flat1 = np.array(arr1).flatten()
|
| 106 |
+
flat2 = np.array(arr2).flatten()
|
| 107 |
+
|
| 108 |
+
# If shapes differ, compare only overlapping parts
|
| 109 |
+
min_len = min(len(flat1), len(flat2))
|
| 110 |
+
if min_len == 0:
|
| 111 |
+
return 0.0 # no data to compare
|
| 112 |
+
|
| 113 |
+
flat1 = flat1[:min_len]
|
| 114 |
+
flat2 = flat2[:min_len]
|
| 115 |
+
|
| 116 |
+
# Calculate similarity as 1 - normalized mean absolute error
|
| 117 |
+
mae = np.mean(np.abs(flat1 - flat2))
|
| 118 |
+
max_val = max(np.max(flat1), np.max(flat2), 1e-6) # avoid zero division
|
| 119 |
+
similarity = 1 - (mae / max_val)
|
| 120 |
+
similarity_percent = max(0, similarity) * 100 # clamp to >=0
|
| 121 |
+
|
| 122 |
+
return round(similarity_percent, 2)
|
| 123 |
+
|
| 124 |
+
def process_two_images(image1, image2, prompt_text):
|
| 125 |
+
arr1 = get_array_from_image(image1, prompt_text)
|
| 126 |
+
arr2 = get_array_from_image(image2, prompt_text)
|
| 127 |
+
similarity = compare_arrays(arr1, arr2)
|
| 128 |
+
|
| 129 |
+
return (
|
| 130 |
+
f"Extracted values from Image 1:\n{arr1}\n\n"
|
| 131 |
+
f"Extracted values from Image 2:\n{arr2}\n\n"
|
| 132 |
+
f"Similarity Accuracy: {similarity} %"
|
| 133 |
+
)
|
| 134 |
|
|
|
|
| 135 |
demo = gr.Interface(
|
| 136 |
+
fn=process_two_images,
|
| 137 |
inputs=[
|
| 138 |
+
gr.Image(type="pil", label="Upload Image 1"),
|
| 139 |
+
gr.Image(type="pil", label="Upload Image 2"),
|
| 140 |
gr.Textbox(lines=1, placeholder="Enter prompt (e.g. Convert to docling)", label="Prompt"),
|
| 141 |
],
|
| 142 |
outputs="text",
|
| 143 |
+
title="SmolDocling Image Comparison",
|
| 144 |
+
description="Upload two document images, extract numeric arrays, and compare their similarity."
|
| 145 |
)
|
| 146 |
|
| 147 |
demo.launch()
|