Spaces:
Running
on
Zero
Running
on
Zero
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,127 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
from pdf2image import convert_from_path
|
3 |
+
|
4 |
+
import torch
|
5 |
+
from PIL import Image
|
6 |
+
from torch.utils.data import DataLoader
|
7 |
+
from tqdm import tqdm
|
8 |
+
from transformers import AutoProcessor
|
9 |
+
|
10 |
+
from custom_colbert.models.paligemma_colbert_architecture import ColPali
|
11 |
+
from custom_colbert.trainer.retrieval_evaluator import CustomEvaluator
|
12 |
+
|
13 |
+
|
14 |
+
def process_images(processor, images, max_length: int = 50):
|
15 |
+
texts_doc = ["Describe the image."] * len(images)
|
16 |
+
images = [image.convert("RGB") for image in images]
|
17 |
+
|
18 |
+
batch_doc = processor(
|
19 |
+
text=texts_doc,
|
20 |
+
images=images,
|
21 |
+
return_tensors="pt",
|
22 |
+
padding="longest",
|
23 |
+
max_length=max_length + processor.image_seq_length,
|
24 |
+
)
|
25 |
+
return batch_doc
|
26 |
+
|
27 |
+
|
28 |
+
def process_queries(processor, queries, mock_image, max_length: int = 50):
|
29 |
+
texts_query = []
|
30 |
+
for query in queries:
|
31 |
+
query = f"Question: {query}<unused0><unused0><unused0><unused0><unused0>"
|
32 |
+
texts_query.append(query)
|
33 |
+
|
34 |
+
batch_query = processor(
|
35 |
+
images=[mock_image.convert("RGB")] * len(texts_query),
|
36 |
+
# NOTE: the image is not used in batch_query but it is required for calling the processor
|
37 |
+
text=texts_query,
|
38 |
+
return_tensors="pt",
|
39 |
+
padding="longest",
|
40 |
+
max_length=max_length + processor.image_seq_length,
|
41 |
+
)
|
42 |
+
del batch_query["pixel_values"]
|
43 |
+
|
44 |
+
batch_query["input_ids"] = batch_query["input_ids"][..., processor.image_seq_length :]
|
45 |
+
batch_query["attention_mask"] = batch_query["attention_mask"][..., processor.image_seq_length :]
|
46 |
+
return batch_query
|
47 |
+
|
48 |
+
|
49 |
+
def search(query: str, ds, images) -> str:
|
50 |
+
qs = []
|
51 |
+
with torch.no_grad():
|
52 |
+
batch_query = process_queries(processor, [query], mock_image)
|
53 |
+
batch_query = {k: v.to(device) for k, v in batch_query.items()}
|
54 |
+
embeddings_query = model(**batch_query)
|
55 |
+
qs.extend(list(torch.unbind(embeddings_query.to("cpu"))))
|
56 |
+
|
57 |
+
# run evaluation
|
58 |
+
retriever_evaluator = CustomEvaluator(is_multi_vector=True)
|
59 |
+
scores = retriever_evaluator.evaluate(qs, ds)
|
60 |
+
|
61 |
+
return f"The most relevant page is {scores.argmax(axis=1)}", images[scores.argmax(axis=1)]
|
62 |
+
# return f"Query: {query}, most relevant page: 1, {len(ds)}", images[1]
|
63 |
+
|
64 |
+
|
65 |
+
def index(file):
|
66 |
+
"""Example script to run inference with ColPali"""
|
67 |
+
images = []
|
68 |
+
for f in file:
|
69 |
+
images.extend(convert_from_path(f))
|
70 |
+
|
71 |
+
# run inference - docs
|
72 |
+
dataloader = DataLoader(
|
73 |
+
images,
|
74 |
+
batch_size=4,
|
75 |
+
shuffle=False,
|
76 |
+
collate_fn=lambda x: process_images(processor, x),
|
77 |
+
)
|
78 |
+
ds = ["test", "double test"]
|
79 |
+
for batch_doc in tqdm(dataloader):
|
80 |
+
with torch.no_grad():
|
81 |
+
batch_doc = {k: v.to(device) for k, v in batch_doc.items()}
|
82 |
+
embeddings_doc = model(**batch_doc)
|
83 |
+
ds.extend(list(torch.unbind(embeddings_doc.to("cpu"))))
|
84 |
+
return f"Uploaded and converted {len(images)} pages", ds, images
|
85 |
+
|
86 |
+
|
87 |
+
COLORS = ['#4285f4', '#db4437', '#f4b400', '#0f9d58', '#e48ef1']
|
88 |
+
# Load model
|
89 |
+
model_name = "coldoc/colpali-3b-mix-448"
|
90 |
+
model = ColPali.from_pretrained("google/paligemma-3b-mix-448", torch_dtype=torch.bfloat16, device_map="cuda").eval()
|
91 |
+
model.load_adapter(model_name)
|
92 |
+
processor = AutoProcessor.from_pretrained(model_name)
|
93 |
+
device = model.device
|
94 |
+
mock_image = Image.new("RGB", (448, 448), (255, 255, 255))
|
95 |
+
|
96 |
+
with gr.Blocks() as demo:
|
97 |
+
gr.Markdown("# PDF to π€ Dataset")
|
98 |
+
gr.Markdown("## 1οΈβ£ Upload PDFs")
|
99 |
+
file = gr.File(file_types=["pdf"], file_count="multiple")
|
100 |
+
|
101 |
+
gr.Markdown("## 2οΈβ£ Convert the PDFs and upload")
|
102 |
+
convert_button = gr.Button("π Convert and upload")
|
103 |
+
message = gr.Textbox("Files not yet uploaded")
|
104 |
+
embeds = gr.State()
|
105 |
+
imgs = gr.State()
|
106 |
+
|
107 |
+
# Define the actions
|
108 |
+
convert_button.click(
|
109 |
+
index,
|
110 |
+
inputs=[file],
|
111 |
+
outputs=[message, embeds, imgs]
|
112 |
+
)
|
113 |
+
|
114 |
+
gr.Markdown("## 3οΈβ£ Search")
|
115 |
+
query = gr.Textbox(placeholder="Enter your query here")
|
116 |
+
search_button = gr.Button("π Search")
|
117 |
+
message2 = gr.Textbox("Query not yet set")
|
118 |
+
output_img = gr.Image()
|
119 |
+
|
120 |
+
search_button.click(
|
121 |
+
search, inputs=[query, embeds, imgs],
|
122 |
+
outputs=[message2, output_img]
|
123 |
+
)
|
124 |
+
|
125 |
+
|
126 |
+
if __name__ == "__main__":
|
127 |
+
demo.queue(max_size=10).launch(debug=True)
|