Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -1,21 +1,23 @@
|
|
1 |
import gradio as gr
|
2 |
-
import cv2
|
3 |
-
import numpy as np
|
4 |
-
import time
|
5 |
-
import torch
|
6 |
-
import spaces
|
7 |
-
from threading import Thread
|
8 |
-
from PIL import Image
|
9 |
from transformers import (
|
10 |
AutoProcessor,
|
11 |
Qwen2_5_VLForConditionalGeneration,
|
12 |
TextIteratorStreamer,
|
13 |
-
AutoTokenizer,
|
14 |
AutoModelForCausalLM,
|
|
|
15 |
)
|
16 |
from transformers.image_utils import load_image
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
17 |
|
|
|
18 |
# Progress Bar Helper
|
|
|
19 |
def progress_bar_html(label: str) -> str:
|
20 |
"""
|
21 |
Returns an HTML snippet for a thin progress bar with a label.
|
@@ -36,7 +38,9 @@ def progress_bar_html(label: str) -> str:
|
|
36 |
</style>
|
37 |
'''
|
38 |
|
39 |
-
#
|
|
|
|
|
40 |
def downsample_video(video_path):
|
41 |
"""
|
42 |
Downsamples the video to 10 evenly spaced frames.
|
@@ -62,7 +66,9 @@ def downsample_video(video_path):
|
|
62 |
vidcap.release()
|
63 |
return frames
|
64 |
|
65 |
-
#
|
|
|
|
|
66 |
MODEL_ID_VL = "Qwen/Qwen2.5-VL-7B-Instruct" # Alternatively: "Qwen/Qwen2.5-VL-3B-Instruct"
|
67 |
processor = AutoProcessor.from_pretrained(MODEL_ID_VL, trust_remote_code=True)
|
68 |
vl_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
@@ -71,8 +77,10 @@ vl_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
|
71 |
torch_dtype=torch.bfloat16
|
72 |
).to("cuda").eval()
|
73 |
|
74 |
-
#
|
75 |
-
|
|
|
|
|
76 |
tg_tokenizer = AutoTokenizer.from_pretrained(TG_MODEL_ID)
|
77 |
tg_model = AutoModelForCausalLM.from_pretrained(
|
78 |
TG_MODEL_ID,
|
@@ -81,38 +89,37 @@ tg_model = AutoModelForCausalLM.from_pretrained(
|
|
81 |
)
|
82 |
tg_model.eval()
|
83 |
|
|
|
|
|
|
|
84 |
@spaces.GPU
|
85 |
def model_inference(input_dict, history):
|
86 |
text = input_dict["text"]
|
87 |
-
files = input_dict
|
88 |
|
89 |
-
# Video inference branch
|
90 |
if text.strip().lower().startswith("@video-infer"):
|
91 |
-
# Remove the tag from the query.
|
92 |
text = text[len("@video-infer"):].strip()
|
93 |
if not files:
|
94 |
-
gr.Error("Please upload a video file along with your @video-infer query.")
|
95 |
return
|
96 |
-
# Assume the first file is a video.
|
97 |
video_path = files[0]
|
98 |
frames = downsample_video(video_path)
|
99 |
if not frames:
|
100 |
-
gr.Error("Could not process video.")
|
101 |
return
|
102 |
-
# Build messages
|
103 |
messages = [
|
104 |
{
|
105 |
"role": "user",
|
106 |
"content": [{"type": "text", "text": text}]
|
107 |
}
|
108 |
]
|
109 |
-
# Append each frame with a timestamp label.
|
110 |
for image, timestamp in frames:
|
111 |
messages[0]["content"].append({"type": "text", "text": f"Frame {timestamp}:"})
|
112 |
messages[0]["content"].append({"type": "image", "image": image})
|
113 |
-
# Collect
|
114 |
video_images = [image for image, _ in frames]
|
115 |
-
# Prepare the prompt.
|
116 |
prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
117 |
inputs = processor(
|
118 |
text=[prompt],
|
@@ -120,7 +127,6 @@ def model_inference(input_dict, history):
|
|
120 |
return_tensors="pt",
|
121 |
padding=True,
|
122 |
).to("cuda")
|
123 |
-
# Set up streaming generation.
|
124 |
streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
|
125 |
generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=1024)
|
126 |
thread = Thread(target=vl_model.generate, kwargs=generation_kwargs)
|
@@ -133,12 +139,20 @@ def model_inference(input_dict, history):
|
|
133 |
yield buffer
|
134 |
return
|
135 |
|
136 |
-
#
|
137 |
if files:
|
|
|
138 |
if len(files) > 1:
|
139 |
images = [load_image(image) for image in files]
|
140 |
elif len(files) == 1:
|
141 |
images = [load_image(files[0])]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
142 |
messages = [
|
143 |
{
|
144 |
"role": "user",
|
@@ -167,34 +181,37 @@ def model_inference(input_dict, history):
|
|
167 |
yield buffer
|
168 |
return
|
169 |
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
streamer = TextIteratorStreamer(tg_tokenizer, skip_prompt=True, skip_special_tokens=True)
|
174 |
-
generation_kwargs = {
|
175 |
-
"input_ids": input_ids,
|
176 |
-
"streamer": streamer,
|
177 |
-
"max_new_tokens": 1024,
|
178 |
-
"do_sample": True,
|
179 |
-
"temperature": 0.7,
|
180 |
-
"top_p": 0.9,
|
181 |
-
}
|
182 |
-
thread = Thread(target=tg_model.generate, kwargs=generation_kwargs)
|
183 |
-
thread.start()
|
184 |
-
buffer = ""
|
185 |
-
yield progress_bar_html("Processing text with Ganymede Model")
|
186 |
-
for new_text in streamer:
|
187 |
-
buffer += new_text
|
188 |
-
time.sleep(0.01)
|
189 |
-
yield buffer
|
190 |
return
|
191 |
|
192 |
-
|
193 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
194 |
|
195 |
-
#
|
|
|
|
|
196 |
examples = [
|
197 |
-
[{"text": "
|
198 |
[{"text": "Tell me a story about a brave knight."}],
|
199 |
[{"text": "@video-infer Explain the content of the Advertisement", "files": ["example_images/videoplayback.mp4"]}],
|
200 |
[{"text": "@video-infer Explain the content of the video in detail", "files": ["example_images/breakfast.mp4"]}],
|
|
|
1 |
import gradio as gr
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
from transformers import (
|
3 |
AutoProcessor,
|
4 |
Qwen2_5_VLForConditionalGeneration,
|
5 |
TextIteratorStreamer,
|
|
|
6 |
AutoModelForCausalLM,
|
7 |
+
AutoTokenizer,
|
8 |
)
|
9 |
from transformers.image_utils import load_image
|
10 |
+
from threading import Thread
|
11 |
+
import time
|
12 |
+
import torch
|
13 |
+
import spaces
|
14 |
+
import cv2
|
15 |
+
import numpy as np
|
16 |
+
from PIL import Image
|
17 |
|
18 |
+
# -----------------------
|
19 |
# Progress Bar Helper
|
20 |
+
# -----------------------
|
21 |
def progress_bar_html(label: str) -> str:
|
22 |
"""
|
23 |
Returns an HTML snippet for a thin progress bar with a label.
|
|
|
38 |
</style>
|
39 |
'''
|
40 |
|
41 |
+
# -----------------------
|
42 |
+
# Video Processing Helper
|
43 |
+
# -----------------------
|
44 |
def downsample_video(video_path):
|
45 |
"""
|
46 |
Downsamples the video to 10 evenly spaced frames.
|
|
|
66 |
vidcap.release()
|
67 |
return frames
|
68 |
|
69 |
+
# -----------------------
|
70 |
+
# Qwen2.5-VL Model (Multimodal)
|
71 |
+
# -----------------------
|
72 |
MODEL_ID_VL = "Qwen/Qwen2.5-VL-7B-Instruct" # Alternatively: "Qwen/Qwen2.5-VL-3B-Instruct"
|
73 |
processor = AutoProcessor.from_pretrained(MODEL_ID_VL, trust_remote_code=True)
|
74 |
vl_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
|
|
77 |
torch_dtype=torch.bfloat16
|
78 |
).to("cuda").eval()
|
79 |
|
80 |
+
# -----------------------
|
81 |
+
# Text Generation Setup (DeepHermes)
|
82 |
+
# -----------------------
|
83 |
+
TG_MODEL_ID = "prithivMLmods/DeepHermes-3-Llama-3-3B-Preview-abliterated"
|
84 |
tg_tokenizer = AutoTokenizer.from_pretrained(TG_MODEL_ID)
|
85 |
tg_model = AutoModelForCausalLM.from_pretrained(
|
86 |
TG_MODEL_ID,
|
|
|
89 |
)
|
90 |
tg_model.eval()
|
91 |
|
92 |
+
# -----------------------
|
93 |
+
# Main Inference Function
|
94 |
+
# -----------------------
|
95 |
@spaces.GPU
|
96 |
def model_inference(input_dict, history):
|
97 |
text = input_dict["text"]
|
98 |
+
files = input_dict["files"]
|
99 |
|
100 |
+
# Video inference branch
|
101 |
if text.strip().lower().startswith("@video-infer"):
|
|
|
102 |
text = text[len("@video-infer"):].strip()
|
103 |
if not files:
|
104 |
+
yield gr.Error("Please upload a video file along with your @video-infer query.")
|
105 |
return
|
|
|
106 |
video_path = files[0]
|
107 |
frames = downsample_video(video_path)
|
108 |
if not frames:
|
109 |
+
yield gr.Error("Could not process video.")
|
110 |
return
|
111 |
+
# Build messages starting with the text prompt and then add each frame with its timestamp.
|
112 |
messages = [
|
113 |
{
|
114 |
"role": "user",
|
115 |
"content": [{"type": "text", "text": text}]
|
116 |
}
|
117 |
]
|
|
|
118 |
for image, timestamp in frames:
|
119 |
messages[0]["content"].append({"type": "text", "text": f"Frame {timestamp}:"})
|
120 |
messages[0]["content"].append({"type": "image", "image": image})
|
121 |
+
# Collect images from the frames.
|
122 |
video_images = [image for image, _ in frames]
|
|
|
123 |
prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
124 |
inputs = processor(
|
125 |
text=[prompt],
|
|
|
127 |
return_tensors="pt",
|
128 |
padding=True,
|
129 |
).to("cuda")
|
|
|
130 |
streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
|
131 |
generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=1024)
|
132 |
thread = Thread(target=vl_model.generate, kwargs=generation_kwargs)
|
|
|
139 |
yield buffer
|
140 |
return
|
141 |
|
142 |
+
# Multimodal branch if images are provided (non-video)
|
143 |
if files:
|
144 |
+
# If more than one file is provided, load them as images.
|
145 |
if len(files) > 1:
|
146 |
images = [load_image(image) for image in files]
|
147 |
elif len(files) == 1:
|
148 |
images = [load_image(files[0])]
|
149 |
+
else:
|
150 |
+
images = []
|
151 |
+
|
152 |
+
if text == "":
|
153 |
+
yield gr.Error("Please input a text query along with the image(s).")
|
154 |
+
return
|
155 |
+
|
156 |
messages = [
|
157 |
{
|
158 |
"role": "user",
|
|
|
181 |
yield buffer
|
182 |
return
|
183 |
|
184 |
+
# Text-only branch using DeepHermes text generation.
|
185 |
+
if text.strip() == "":
|
186 |
+
yield gr.Error("Please input a query.")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
187 |
return
|
188 |
|
189 |
+
input_ids = tg_tokenizer(text, return_tensors="pt").to(tg_model.device)
|
190 |
+
streamer = TextIteratorStreamer(tg_tokenizer, skip_prompt=True, skip_special_tokens=True)
|
191 |
+
generation_kwargs = {
|
192 |
+
"input_ids": input_ids,
|
193 |
+
"streamer": streamer,
|
194 |
+
"max_new_tokens": 2048,
|
195 |
+
"do_sample": True,
|
196 |
+
"top_p": 0.9,
|
197 |
+
"top_k": 50,
|
198 |
+
"temperature": 0.6,
|
199 |
+
"repetition_penalty": 1.2,
|
200 |
+
}
|
201 |
+
thread = Thread(target=tg_model.generate, kwargs=generation_kwargs)
|
202 |
+
thread.start()
|
203 |
+
buffer = ""
|
204 |
+
yield progress_bar_html("Processing text with DeepHermes Model")
|
205 |
+
for new_text in streamer:
|
206 |
+
buffer += new_text
|
207 |
+
time.sleep(0.01)
|
208 |
+
yield buffer
|
209 |
|
210 |
+
# -----------------------
|
211 |
+
# Gradio Chat Interface
|
212 |
+
# -----------------------
|
213 |
examples = [
|
214 |
+
[{"text": "Describe the Image?", "files": ["example_images/document.jpg"]}],
|
215 |
[{"text": "Tell me a story about a brave knight."}],
|
216 |
[{"text": "@video-infer Explain the content of the Advertisement", "files": ["example_images/videoplayback.mp4"]}],
|
217 |
[{"text": "@video-infer Explain the content of the video in detail", "files": ["example_images/breakfast.mp4"]}],
|