Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -1,10 +1,4 @@
|
|
1 |
import gradio as gr
|
2 |
-
from transformers import (
|
3 |
-
Qwen2VLForConditionalGeneration,
|
4 |
-
AutoProcessor,
|
5 |
-
TextIteratorStreamer,
|
6 |
-
AutoModelForImageTextToText,
|
7 |
-
)
|
8 |
from transformers.image_utils import load_image
|
9 |
from threading import Thread
|
10 |
import time
|
@@ -13,6 +7,14 @@ import spaces
|
|
13 |
from PIL import Image
|
14 |
import requests
|
15 |
from io import BytesIO
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
|
17 |
# Helper function to return a progress bar HTML snippet.
|
18 |
def progress_bar_html(label: str) -> str:
|
@@ -20,7 +22,7 @@ def progress_bar_html(label: str) -> str:
|
|
20 |
<div style="display: flex; align-items: center;">
|
21 |
<span style="margin-right: 10px; font-size: 14px;">{label}</span>
|
22 |
<div style="width: 110px; height: 5px; background-color: #FFB6C1; border-radius: 2px; overflow: hidden;">
|
23 |
-
<div style="width: 100%; height: 100%; background-color: #FF69B4
|
24 |
</div>
|
25 |
</div>
|
26 |
<style>
|
@@ -31,7 +33,29 @@ def progress_bar_html(label: str) -> str:
|
|
31 |
</style>
|
32 |
'''
|
33 |
|
34 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
35 |
qwen_processor = AutoProcessor.from_pretrained(QV_MODEL_ID, trust_remote_code=True)
|
36 |
qwen_model = Qwen2VLForConditionalGeneration.from_pretrained(
|
37 |
QV_MODEL_ID,
|
@@ -39,25 +63,77 @@ qwen_model = Qwen2VLForConditionalGeneration.from_pretrained(
|
|
39 |
torch_dtype=torch.float16
|
40 |
).to("cuda").eval()
|
41 |
|
|
|
42 |
AYA_MODEL_ID = "CohereForAI/aya-vision-8b"
|
43 |
aya_processor = AutoProcessor.from_pretrained(AYA_MODEL_ID)
|
44 |
aya_model = AutoModelForImageTextToText.from_pretrained(
|
45 |
AYA_MODEL_ID, device_map="auto", torch_dtype=torch.float16
|
46 |
)
|
47 |
|
|
|
|
|
|
|
48 |
@spaces.GPU
|
49 |
def model_inference(input_dict, history):
|
50 |
text = input_dict["text"].strip()
|
51 |
files = input_dict.get("files", [])
|
52 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
53 |
if text.lower().startswith("@aya-vision"):
|
54 |
-
# Remove the command prefix and trim the prompt.
|
55 |
text_prompt = text[len("@aya-vision"):].strip()
|
56 |
if not files:
|
57 |
yield "Error: Please provide an image for the @aya-vision feature."
|
58 |
return
|
59 |
else:
|
60 |
-
#
|
61 |
image = load_image(files[0])
|
62 |
yield progress_bar_html("Processing with Aya-Vision-8b")
|
63 |
messages = [{
|
@@ -75,7 +151,6 @@ def model_inference(input_dict, history):
|
|
75 |
return_dict=True,
|
76 |
return_tensors="pt"
|
77 |
).to(aya_model.device)
|
78 |
-
# Set up a streamer for Aya-Vision output
|
79 |
streamer = TextIteratorStreamer(aya_processor, skip_prompt=True, skip_special_tokens=True)
|
80 |
generation_kwargs = dict(
|
81 |
inputs,
|
@@ -94,7 +169,7 @@ def model_inference(input_dict, history):
|
|
94 |
yield buffer
|
95 |
return
|
96 |
|
97 |
-
#
|
98 |
if len(files) > 1:
|
99 |
images = [load_image(image) for image in files]
|
100 |
elif len(files) == 1:
|
@@ -102,7 +177,6 @@ def model_inference(input_dict, history):
|
|
102 |
else:
|
103 |
images = []
|
104 |
|
105 |
-
# Validate input: require both text and (optionally) image(s).
|
106 |
if text == "" and not images:
|
107 |
yield "Error: Please input a query and optionally image(s)."
|
108 |
return
|
@@ -110,7 +184,6 @@ def model_inference(input_dict, history):
|
|
110 |
yield "Error: Please input a text query along with the image(s)."
|
111 |
return
|
112 |
|
113 |
-
# Prepare messages for the Qwen2-VL model.
|
114 |
messages = [{
|
115 |
"role": "user",
|
116 |
"content": [
|
@@ -129,11 +202,9 @@ def model_inference(input_dict, history):
|
|
129 |
padding=True,
|
130 |
).to("cuda")
|
131 |
|
132 |
-
# Set up a streamer for real-time output.
|
133 |
streamer = TextIteratorStreamer(qwen_processor, skip_prompt=True, skip_special_tokens=True)
|
134 |
generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=1024)
|
135 |
|
136 |
-
# Start generation in a separate thread.
|
137 |
thread = Thread(target=qwen_model.generate, kwargs=generation_kwargs)
|
138 |
thread.start()
|
139 |
|
@@ -145,7 +216,11 @@ def model_inference(input_dict, history):
|
|
145 |
time.sleep(0.01)
|
146 |
yield buffer
|
147 |
|
|
|
|
|
|
|
148 |
examples = [
|
|
|
149 |
[{"text": "@aya-vision Summarize the letter", "files": ["examples/1.png"]}],
|
150 |
[{"text": "@aya-vision Extract JSON from the image", "files": ["example_images/document.jpg"]}],
|
151 |
[{"text": "Extract as JSON table from the table", "files": ["examples/4.jpg"]}],
|
@@ -160,13 +235,13 @@ examples = [
|
|
160 |
|
161 |
demo = gr.ChatInterface(
|
162 |
fn=model_inference,
|
163 |
-
description="# **Multimodal OCR
|
164 |
examples=examples,
|
165 |
textbox=gr.MultimodalTextbox(
|
166 |
label="Query Input",
|
167 |
-
file_types=["image"],
|
168 |
file_count="multiple",
|
169 |
-
placeholder="
|
170 |
),
|
171 |
stop_btn="Stop Generation",
|
172 |
multimodal=True,
|
|
|
1 |
import gradio as gr
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
from transformers.image_utils import load_image
|
3 |
from threading import Thread
|
4 |
import time
|
|
|
7 |
from PIL import Image
|
8 |
import requests
|
9 |
from io import BytesIO
|
10 |
+
import cv2
|
11 |
+
import numpy as np
|
12 |
+
from transformers import (
|
13 |
+
Qwen2VLForConditionalGeneration,
|
14 |
+
AutoProcessor,
|
15 |
+
TextIteratorStreamer,
|
16 |
+
AutoModelForImageTextToText,
|
17 |
+
)
|
18 |
|
19 |
# Helper function to return a progress bar HTML snippet.
|
20 |
def progress_bar_html(label: str) -> str:
|
|
|
22 |
<div style="display: flex; align-items: center;">
|
23 |
<span style="margin-right: 10px; font-size: 14px;">{label}</span>
|
24 |
<div style="width: 110px; height: 5px; background-color: #FFB6C1; border-radius: 2px; overflow: hidden;">
|
25 |
+
<div style="width: 100%; height: 100%; background-color: #FF69B4; animation: loading 1.5s linear infinite;"></div>
|
26 |
</div>
|
27 |
</div>
|
28 |
<style>
|
|
|
33 |
</style>
|
34 |
'''
|
35 |
|
36 |
+
# Helper function to downsample a video into 10 evenly spaced frames.
|
37 |
+
def downsample_video(video_path):
|
38 |
+
vidcap = cv2.VideoCapture(video_path)
|
39 |
+
total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))
|
40 |
+
fps = vidcap.get(cv2.CAP_PROP_FPS)
|
41 |
+
frames = []
|
42 |
+
# Sample 10 evenly spaced frames.
|
43 |
+
frame_indices = np.linspace(0, total_frames - 1, 10, dtype=int)
|
44 |
+
for i in frame_indices:
|
45 |
+
vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)
|
46 |
+
success, image = vidcap.read()
|
47 |
+
if success:
|
48 |
+
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
|
49 |
+
pil_image = Image.fromarray(image)
|
50 |
+
timestamp = round(i / fps, 2)
|
51 |
+
frames.append((pil_image, timestamp))
|
52 |
+
vidcap.release()
|
53 |
+
return frames
|
54 |
+
|
55 |
+
# Model and processor setups
|
56 |
+
|
57 |
+
# Setup for Qwen2VL OCR branch (default).
|
58 |
+
QV_MODEL_ID = "prithivMLmods/Qwen2-VL-OCR-2B-Instruct" # or use "prithivMLmods/Qwen2-VL-OCR2-2B-Instruct"
|
59 |
qwen_processor = AutoProcessor.from_pretrained(QV_MODEL_ID, trust_remote_code=True)
|
60 |
qwen_model = Qwen2VLForConditionalGeneration.from_pretrained(
|
61 |
QV_MODEL_ID,
|
|
|
63 |
torch_dtype=torch.float16
|
64 |
).to("cuda").eval()
|
65 |
|
66 |
+
# Setup for Aya-Vision branch.
|
67 |
AYA_MODEL_ID = "CohereForAI/aya-vision-8b"
|
68 |
aya_processor = AutoProcessor.from_pretrained(AYA_MODEL_ID)
|
69 |
aya_model = AutoModelForImageTextToText.from_pretrained(
|
70 |
AYA_MODEL_ID, device_map="auto", torch_dtype=torch.float16
|
71 |
)
|
72 |
|
73 |
+
# ---------------------------
|
74 |
+
# Main Inference Function
|
75 |
+
# ---------------------------
|
76 |
@spaces.GPU
|
77 |
def model_inference(input_dict, history):
|
78 |
text = input_dict["text"].strip()
|
79 |
files = input_dict.get("files", [])
|
80 |
|
81 |
+
# Branch for video inference with Aya-Vision using @video-infer.
|
82 |
+
if text.lower().startswith("@video-infer"):
|
83 |
+
prompt = text[len("@video-infer"):].strip()
|
84 |
+
if not files:
|
85 |
+
yield "Error: Please provide a video for the @video-infer feature."
|
86 |
+
return
|
87 |
+
video_path = files[0]
|
88 |
+
frames = downsample_video(video_path)
|
89 |
+
if not frames:
|
90 |
+
yield "Error: Could not extract frames from the video."
|
91 |
+
return
|
92 |
+
# Build messages: start with the prompt then add each frame with its timestamp.
|
93 |
+
content_list = []
|
94 |
+
content_list.append({"type": "text", "text": prompt})
|
95 |
+
for frame, timestamp in frames:
|
96 |
+
content_list.append({"type": "text", "text": f"Frame {timestamp}:"})
|
97 |
+
content_list.append({"type": "image", "image": frame})
|
98 |
+
messages = [{
|
99 |
+
"role": "user",
|
100 |
+
"content": content_list,
|
101 |
+
}]
|
102 |
+
inputs = aya_processor.apply_chat_template(
|
103 |
+
messages,
|
104 |
+
padding=True,
|
105 |
+
add_generation_prompt=True,
|
106 |
+
tokenize=True,
|
107 |
+
return_dict=True,
|
108 |
+
return_tensors="pt"
|
109 |
+
).to(aya_model.device)
|
110 |
+
streamer = TextIteratorStreamer(aya_processor, skip_prompt=True, skip_special_tokens=True)
|
111 |
+
generation_kwargs = dict(
|
112 |
+
inputs,
|
113 |
+
streamer=streamer,
|
114 |
+
max_new_tokens=1024,
|
115 |
+
do_sample=True,
|
116 |
+
temperature=0.3
|
117 |
+
)
|
118 |
+
thread = Thread(target=aya_model.generate, kwargs=generation_kwargs)
|
119 |
+
thread.start()
|
120 |
+
buffer = ""
|
121 |
+
yield progress_bar_html("Processing video with Aya-Vision-8b")
|
122 |
+
for new_text in streamer:
|
123 |
+
buffer += new_text
|
124 |
+
buffer = buffer.replace("<|im_end|>", "")
|
125 |
+
time.sleep(0.01)
|
126 |
+
yield buffer
|
127 |
+
return
|
128 |
+
|
129 |
+
# Branch for single image inference with Aya-Vision using @aya-vision.
|
130 |
if text.lower().startswith("@aya-vision"):
|
|
|
131 |
text_prompt = text[len("@aya-vision"):].strip()
|
132 |
if not files:
|
133 |
yield "Error: Please provide an image for the @aya-vision feature."
|
134 |
return
|
135 |
else:
|
136 |
+
# Use the first provided image.
|
137 |
image = load_image(files[0])
|
138 |
yield progress_bar_html("Processing with Aya-Vision-8b")
|
139 |
messages = [{
|
|
|
151 |
return_dict=True,
|
152 |
return_tensors="pt"
|
153 |
).to(aya_model.device)
|
|
|
154 |
streamer = TextIteratorStreamer(aya_processor, skip_prompt=True, skip_special_tokens=True)
|
155 |
generation_kwargs = dict(
|
156 |
inputs,
|
|
|
169 |
yield buffer
|
170 |
return
|
171 |
|
172 |
+
# Default branch: Use Qwen2VL OCR for text (with optional images).
|
173 |
if len(files) > 1:
|
174 |
images = [load_image(image) for image in files]
|
175 |
elif len(files) == 1:
|
|
|
177 |
else:
|
178 |
images = []
|
179 |
|
|
|
180 |
if text == "" and not images:
|
181 |
yield "Error: Please input a query and optionally image(s)."
|
182 |
return
|
|
|
184 |
yield "Error: Please input a text query along with the image(s)."
|
185 |
return
|
186 |
|
|
|
187 |
messages = [{
|
188 |
"role": "user",
|
189 |
"content": [
|
|
|
202 |
padding=True,
|
203 |
).to("cuda")
|
204 |
|
|
|
205 |
streamer = TextIteratorStreamer(qwen_processor, skip_prompt=True, skip_special_tokens=True)
|
206 |
generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=1024)
|
207 |
|
|
|
208 |
thread = Thread(target=qwen_model.generate, kwargs=generation_kwargs)
|
209 |
thread.start()
|
210 |
|
|
|
216 |
time.sleep(0.01)
|
217 |
yield buffer
|
218 |
|
219 |
+
|
220 |
+
# Gradio Interface Setup
|
221 |
+
|
222 |
examples = [
|
223 |
+
[{"text": "@video-infer Summarize the video content", "files": ["examples/videoplayback.mp4"]}],
|
224 |
[{"text": "@aya-vision Summarize the letter", "files": ["examples/1.png"]}],
|
225 |
[{"text": "@aya-vision Extract JSON from the image", "files": ["example_images/document.jpg"]}],
|
226 |
[{"text": "Extract as JSON table from the table", "files": ["examples/4.jpg"]}],
|
|
|
235 |
|
236 |
demo = gr.ChatInterface(
|
237 |
fn=model_inference,
|
238 |
+
description="# **Multimodal OCR and Video Inference with Aya-Vision (@aya-vision for image, @video-infer for video) and Qwen2VL OCR (default)**",
|
239 |
examples=examples,
|
240 |
textbox=gr.MultimodalTextbox(
|
241 |
label="Query Input",
|
242 |
+
file_types=["image", "video"],
|
243 |
file_count="multiple",
|
244 |
+
placeholder="Tag @aya-vision for Aya-Vision image infer, @video-infer for Aya-Vision video infer, default runs Qwen2VL OCR"
|
245 |
),
|
246 |
stop_btn="Stop Generation",
|
247 |
multimodal=True,
|