Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -7,64 +7,49 @@ Original file is located at
|
|
7 |
https://colab.research.google.com/drive/1vzsQ17-W1Vy6yJ60XUwFy0QRkOR_SIg7
|
8 |
"""
|
9 |
|
10 |
-
from transformers import
|
11 |
-
from qwen_vl_utils import process_vision_info
|
12 |
import torch
|
13 |
import gradio as gr
|
14 |
from PIL import Image
|
15 |
-
|
|
|
|
|
16 |
|
17 |
processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-2B-Instruct")
|
|
|
18 |
|
19 |
-
# Initialize the model with float16 precision and handle fallback to CPU
|
20 |
-
# Simplified model loading function for CPU
|
21 |
def load_model():
|
22 |
-
return
|
23 |
"Qwen/Qwen2-VL-2B-Instruct",
|
24 |
-
torch_dtype=torch.float32,
|
25 |
-
low_cpu_mem_usage=True
|
|
|
26 |
)
|
27 |
|
28 |
-
# Load the model
|
29 |
vlm = load_model()
|
30 |
|
31 |
-
# OCR function to extract text from an image
|
32 |
def ocr_image(image, query="Extract text from the image", keyword=""):
|
|
|
|
|
33 |
messages = [
|
34 |
{
|
35 |
"role": "user",
|
36 |
"content": [
|
37 |
{
|
38 |
"type": "image",
|
39 |
-
"image":
|
40 |
},
|
41 |
{"type": "text", "text": query},
|
42 |
],
|
43 |
}
|
44 |
]
|
45 |
|
46 |
-
|
47 |
-
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
48 |
-
image_inputs, video_inputs = process_vision_info(messages)
|
49 |
-
inputs = processor(
|
50 |
-
text=[text],
|
51 |
-
images=image_inputs,
|
52 |
-
videos=video_inputs,
|
53 |
-
padding=True,
|
54 |
-
return_tensors="pt",
|
55 |
-
)
|
56 |
inputs = inputs.to("cpu")
|
57 |
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
|
62 |
-
]
|
63 |
-
|
64 |
-
output_text = processor.batch_decode(
|
65 |
-
generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
|
66 |
-
)[0]
|
67 |
-
|
68 |
if keyword:
|
69 |
keyword_lower = keyword.lower()
|
70 |
if keyword_lower in output_text.lower():
|
@@ -75,14 +60,14 @@ def ocr_image(image, query="Extract text from the image", keyword=""):
|
|
75 |
else:
|
76 |
return output_text
|
77 |
|
78 |
-
# Gradio interface
|
79 |
def process_image(image, keyword=""):
|
|
|
80 |
max_size = 1024
|
81 |
if max(image.size) > max_size:
|
82 |
image.thumbnail((max_size, max_size))
|
83 |
return ocr_image(image, keyword=keyword)
|
84 |
|
85 |
-
#
|
86 |
interface = gr.Interface(
|
87 |
fn=process_image,
|
88 |
inputs=[
|
|
|
7 |
https://colab.research.google.com/drive/1vzsQ17-W1Vy6yJ60XUwFy0QRkOR_SIg7
|
8 |
"""
|
9 |
|
10 |
+
from transformers import AutoProcessor
|
|
|
11 |
import torch
|
12 |
import gradio as gr
|
13 |
from PIL import Image
|
14 |
+
# Hypothetical imports
|
15 |
+
from byaldi import ByaldiProcessor
|
16 |
+
from colpali import ColPaliQwen2VLModel
|
17 |
|
18 |
processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-2B-Instruct")
|
19 |
+
byaldi_processor = ByaldiProcessor()
|
20 |
|
|
|
|
|
21 |
def load_model():
|
22 |
+
return ColPaliQwen2VLModel.from_pretrained(
|
23 |
"Qwen/Qwen2-VL-2B-Instruct",
|
24 |
+
torch_dtype=torch.float32,
|
25 |
+
low_cpu_mem_usage=True,
|
26 |
+
device_map="auto"
|
27 |
)
|
28 |
|
|
|
29 |
vlm = load_model()
|
30 |
|
|
|
31 |
def ocr_image(image, query="Extract text from the image", keyword=""):
|
32 |
+
processed_image = byaldi_processor.process_image(image)
|
33 |
+
|
34 |
messages = [
|
35 |
{
|
36 |
"role": "user",
|
37 |
"content": [
|
38 |
{
|
39 |
"type": "image",
|
40 |
+
"image": processed_image,
|
41 |
},
|
42 |
{"type": "text", "text": query},
|
43 |
],
|
44 |
}
|
45 |
]
|
46 |
|
47 |
+
inputs = processor(messages, return_tensors="pt")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
48 |
inputs = inputs.to("cpu")
|
49 |
|
50 |
+
output = vlm.generate(**inputs, max_new_tokens=512)
|
51 |
+
output_text = processor.decode(output[0], skip_special_tokens=True)
|
52 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
53 |
if keyword:
|
54 |
keyword_lower = keyword.lower()
|
55 |
if keyword_lower in output_text.lower():
|
|
|
60 |
else:
|
61 |
return output_text
|
62 |
|
|
|
63 |
def process_image(image, keyword=""):
|
64 |
+
# Resize image if it's too large
|
65 |
max_size = 1024
|
66 |
if max(image.size) > max_size:
|
67 |
image.thumbnail((max_size, max_size))
|
68 |
return ocr_image(image, keyword=keyword)
|
69 |
|
70 |
+
# Gradio interface:
|
71 |
interface = gr.Interface(
|
72 |
fn=process_image,
|
73 |
inputs=[
|