upload notebooks
Browse files- Aya-Vision-8B/aya_vision_8b.ipynb +189 -0
- Florence-2-Base/Florence_2_base.ipynb +94 -0
- Gemma3-VL/Gemma3_4B_it.ipynb +369 -0
- MiMo-VL-7B-RL/MiMo_VL_7B_RL.ipynb +242 -0
- MiMo-VL-7B-SFT/MiMo_VL_7B_SFT.ipynb +242 -0
- Qwen-2VL-MessyOCR/Qwen2_VL_OCR_2B_Instruct_prithivmlmods.ipynb +247 -0
- Qwen2-VL/Qwen2_VL_2B_Instruct.ipynb +242 -0
- Qwen2-VL/Qwen2_VL_7B_Instruct.ipynb +242 -0
- Qwen2.5-VL/Qwen2_5VL_3B.ipynb +242 -0
- Qwen2.5-VL/Qwen2_5VL_7B.ipynb +242 -0
- RolmOCR-Qwen2.5-VL/reducto_RolmOCR_Qwen2_5VL_7B.ipynb +242 -0
- olmOCR-Qwen2-VL/olmOCR_7B_0225.ipynb +242 -0
- typhoon-ocr-7b-Qwen2.5VL/typhoon_ocr_7b.ipynb +242 -0
Aya-Vision-8B/aya_vision_8b.ipynb
ADDED
@@ -0,0 +1,189 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cells": [
|
3 |
+
{
|
4 |
+
"cell_type": "code",
|
5 |
+
"execution_count": 1,
|
6 |
+
"metadata": {
|
7 |
+
"id": "m7rU-pjX3Y1O"
|
8 |
+
},
|
9 |
+
"outputs": [],
|
10 |
+
"source": [
|
11 |
+
"%%capture\n",
|
12 |
+
"!pip install gradio transformers accelerate numpy\n",
|
13 |
+
"!pip install torch torchvision av hf_xet spaces\n",
|
14 |
+
"!pip install pillow huggingface_hub opencv-python"
|
15 |
+
]
|
16 |
+
},
|
17 |
+
{
|
18 |
+
"cell_type": "code",
|
19 |
+
"execution_count": null,
|
20 |
+
"metadata": {
|
21 |
+
"id": "dZUVag_jJMck"
|
22 |
+
},
|
23 |
+
"outputs": [],
|
24 |
+
"source": [
|
25 |
+
"from huggingface_hub import notebook_login, HfApi\n",
|
26 |
+
"notebook_login()"
|
27 |
+
]
|
28 |
+
},
|
29 |
+
{
|
30 |
+
"cell_type": "code",
|
31 |
+
"execution_count": null,
|
32 |
+
"metadata": {
|
33 |
+
"id": "kW4MjaOs3c9E"
|
34 |
+
},
|
35 |
+
"outputs": [],
|
36 |
+
"source": [
|
37 |
+
"import gradio as gr\n",
|
38 |
+
"from transformers import AutoProcessor, TextIteratorStreamer, AutoModelForImageTextToText\n",
|
39 |
+
"from transformers.image_utils import load_image\n",
|
40 |
+
"from threading import Thread\n",
|
41 |
+
"import time\n",
|
42 |
+
"import torch\n",
|
43 |
+
"import spaces\n",
|
44 |
+
"import cv2\n",
|
45 |
+
"import numpy as np\n",
|
46 |
+
"from PIL import Image\n",
|
47 |
+
"\n",
|
48 |
+
"# Helper: progress bar HTML\n",
|
49 |
+
"def progress_bar_html(label: str) -> str:\n",
|
50 |
+
" return f'''\n",
|
51 |
+
"<div style=\"display: flex; align-items: center;\">\n",
|
52 |
+
" <span style=\"margin-right: 10px; font-size: 14px;\">{label}</span>\n",
|
53 |
+
" <div style=\"width: 110px; height: 5px; background-color: #FFB6C1; border-radius: 2px; overflow: hidden;\">\n",
|
54 |
+
" <div style=\"width: 100%; height: 100%; background-color: #FF69B4; animation: loading 1.5s linear infinite;\"></div>\n",
|
55 |
+
" </div>\n",
|
56 |
+
"</div>\n",
|
57 |
+
"<style>\n",
|
58 |
+
"@keyframes loading {{\n",
|
59 |
+
" 0% {{ transform: translateX(-100%); }}\n",
|
60 |
+
" 100% {{ transform: translateX(100%); }}\n",
|
61 |
+
"}}\n",
|
62 |
+
"</style>\n",
|
63 |
+
" '''\n",
|
64 |
+
"\n",
|
65 |
+
"# Aya Vision 8B setup\n",
|
66 |
+
"AYA_MODEL_ID = \"CohereForAI/aya-vision-8b\"\n",
|
67 |
+
"aya_processor = AutoProcessor.from_pretrained(AYA_MODEL_ID)\n",
|
68 |
+
"aya_model = AutoModelForImageTextToText.from_pretrained(\n",
|
69 |
+
" AYA_MODEL_ID,\n",
|
70 |
+
" device_map=\"auto\",\n",
|
71 |
+
" torch_dtype=torch.float16\n",
|
72 |
+
")\n",
|
73 |
+
"\n",
|
74 |
+
"def downsample_video(video_path, num_frames=10):\n",
|
75 |
+
" \"\"\"\n",
|
76 |
+
" Extract evenly spaced frames and timestamps from a video file.\n",
|
77 |
+
" Returns list of (PIL.Image, timestamp_sec).\n",
|
78 |
+
" \"\"\"\n",
|
79 |
+
" vidcap = cv2.VideoCapture(video_path)\n",
|
80 |
+
" total = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))\n",
|
81 |
+
" fps = vidcap.get(cv2.CAP_PROP_FPS) or 30\n",
|
82 |
+
" indices = np.linspace(0, total-1, num_frames, dtype=int)\n",
|
83 |
+
" frames = []\n",
|
84 |
+
" for idx in indices:\n",
|
85 |
+
" vidcap.set(cv2.CAP_PROP_POS_FRAMES, int(idx))\n",
|
86 |
+
" ret, frame = vidcap.read()\n",
|
87 |
+
" if not ret:\n",
|
88 |
+
" continue\n",
|
89 |
+
" frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)\n",
|
90 |
+
" pil = Image.fromarray(frame)\n",
|
91 |
+
" timestamp = round(idx / fps, 2)\n",
|
92 |
+
" frames.append((pil, timestamp))\n",
|
93 |
+
" vidcap.release()\n",
|
94 |
+
" return frames\n",
|
95 |
+
"\n",
|
96 |
+
"@spaces.GPU\n",
|
97 |
+
"def process_image(prompt: str, image: Image.Image):\n",
|
98 |
+
" if image is None:\n",
|
99 |
+
" yield \"Error: Please upload an image.\"\n",
|
100 |
+
" return\n",
|
101 |
+
" if not prompt.strip():\n",
|
102 |
+
" yield \"Error: Please provide a prompt with the image.\"\n",
|
103 |
+
" return\n",
|
104 |
+
" yield progress_bar_html(\"Processing Image with Aya Vision 8B\")\n",
|
105 |
+
" messages = [{\"role\": \"user\", \"content\": [\n",
|
106 |
+
" {\"type\": \"image\", \"image\": image},\n",
|
107 |
+
" {\"type\": \"text\", \"text\": prompt.strip()}\n",
|
108 |
+
" ]}]\n",
|
109 |
+
" inputs = aya_processor.apply_chat_template(\n",
|
110 |
+
" messages, padding=True, add_generation_prompt=True,\n",
|
111 |
+
" tokenize=True, return_dict=True, return_tensors=\"pt\"\n",
|
112 |
+
" ).to(aya_model.device)\n",
|
113 |
+
" streamer = TextIteratorStreamer(aya_processor, skip_prompt=True, skip_special_tokens=True)\n",
|
114 |
+
" thread = Thread(target=aya_model.generate, kwargs={**inputs, \"streamer\": streamer, \"max_new_tokens\": 1024, \"do_sample\": True, \"temperature\": 0.3})\n",
|
115 |
+
" thread.start()\n",
|
116 |
+
" buff = \"\"\n",
|
117 |
+
" for chunk in streamer:\n",
|
118 |
+
" buff += chunk.replace(\"<|im_end|>\", \"\")\n",
|
119 |
+
" time.sleep(0.01)\n",
|
120 |
+
" yield buff\n",
|
121 |
+
"\n",
|
122 |
+
"@spaces.GPU\n",
|
123 |
+
"def process_video(prompt: str, video_file: str):\n",
|
124 |
+
" if video_file is None:\n",
|
125 |
+
" yield \"Error: Please upload a video.\"\n",
|
126 |
+
" return\n",
|
127 |
+
" if not prompt.strip():\n",
|
128 |
+
" yield \"Error: Please provide a prompt with the video.\"\n",
|
129 |
+
" return\n",
|
130 |
+
" yield progress_bar_html(\"Processing Video with Aya Vision 8B\")\n",
|
131 |
+
" frames = downsample_video(video_file)\n",
|
132 |
+
" # Build chat messages with each frame and timestamp\n",
|
133 |
+
" content = [{\"type\": \"text\", \"text\": prompt.strip()}]\n",
|
134 |
+
" for img, ts in frames:\n",
|
135 |
+
" content.append({\"type\": \"text\", \"text\": f\"Frame at {ts}s:\"})\n",
|
136 |
+
" content.append({\"type\": \"image\", \"image\": img})\n",
|
137 |
+
" messages = [{\"role\": \"user\", \"content\": content}]\n",
|
138 |
+
" inputs = aya_processor.apply_chat_template(\n",
|
139 |
+
" messages, tokenize=True, add_generation_prompt=True,\n",
|
140 |
+
" return_dict=True, return_tensors=\"pt\"\n",
|
141 |
+
" ).to(aya_model.device)\n",
|
142 |
+
" streamer = TextIteratorStreamer(aya_processor, skip_prompt=True, skip_special_tokens=True)\n",
|
143 |
+
" thread = Thread(target=aya_model.generate, kwargs={**inputs, \"streamer\": streamer, \"max_new_tokens\": 1024, \"do_sample\": True, \"temperature\": 0.3})\n",
|
144 |
+
" thread.start()\n",
|
145 |
+
" buff = \"\"\n",
|
146 |
+
" for chunk in streamer:\n",
|
147 |
+
" buff += chunk.replace(\"<|im_end|>\", \"\")\n",
|
148 |
+
" time.sleep(0.01)\n",
|
149 |
+
" yield buff\n",
|
150 |
+
"\n",
|
151 |
+
"# Build Gradio UI\n",
|
152 |
+
"demo = gr.Blocks()\n",
|
153 |
+
"with demo:\n",
|
154 |
+
" gr.Markdown(\"# **Aya Vision 8B Multimodal: Image & Video**\")\n",
|
155 |
+
" with gr.Tabs():\n",
|
156 |
+
" with gr.TabItem(\"Image Inference\"):\n",
|
157 |
+
" txt_i = gr.Textbox(label=\"Prompt\", placeholder=\"Enter prompt...\")\n",
|
158 |
+
" img_u = gr.Image(type=\"filepath\", label=\"Image\")\n",
|
159 |
+
" btn_i = gr.Button(\"Run Image\")\n",
|
160 |
+
" out_i = gr.Textbox(label=\"Output\", interactive=False)\n",
|
161 |
+
" btn_i.click(fn=process_image, inputs=[txt_i, img_u], outputs=out_i)\n",
|
162 |
+
" with gr.TabItem(\"Video Inference\"):\n",
|
163 |
+
" txt_v = gr.Textbox(label=\"Prompt\", placeholder=\"Enter prompt...\")\n",
|
164 |
+
" vid_u = gr.Video(label=\"Video\")\n",
|
165 |
+
" btn_v = gr.Button(\"Run Video\")\n",
|
166 |
+
" out_v = gr.Textbox(label=\"Output\", interactive=False)\n",
|
167 |
+
" btn_v.click(fn=process_video, inputs=[txt_v, vid_u], outputs=out_v)\n",
|
168 |
+
"\n",
|
169 |
+
"demo.launch(debug=True, share=True)"
|
170 |
+
]
|
171 |
+
}
|
172 |
+
],
|
173 |
+
"metadata": {
|
174 |
+
"accelerator": "GPU",
|
175 |
+
"colab": {
|
176 |
+
"gpuType": "T4",
|
177 |
+
"provenance": []
|
178 |
+
},
|
179 |
+
"kernelspec": {
|
180 |
+
"display_name": "Python 3",
|
181 |
+
"name": "python3"
|
182 |
+
},
|
183 |
+
"language_info": {
|
184 |
+
"name": "python"
|
185 |
+
}
|
186 |
+
},
|
187 |
+
"nbformat": 4,
|
188 |
+
"nbformat_minor": 0
|
189 |
+
}
|
Florence-2-Base/Florence_2_base.ipynb
ADDED
@@ -0,0 +1,94 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"nbformat": 4,
|
3 |
+
"nbformat_minor": 0,
|
4 |
+
"metadata": {
|
5 |
+
"colab": {
|
6 |
+
"provenance": [],
|
7 |
+
"gpuType": "T4"
|
8 |
+
},
|
9 |
+
"kernelspec": {
|
10 |
+
"name": "python3",
|
11 |
+
"display_name": "Python 3"
|
12 |
+
},
|
13 |
+
"language_info": {
|
14 |
+
"name": "python"
|
15 |
+
},
|
16 |
+
"accelerator": "GPU"
|
17 |
+
},
|
18 |
+
"cells": [
|
19 |
+
{
|
20 |
+
"cell_type": "code",
|
21 |
+
"execution_count": 1,
|
22 |
+
"metadata": {
|
23 |
+
"id": "m7rU-pjX3Y1O"
|
24 |
+
},
|
25 |
+
"outputs": [],
|
26 |
+
"source": [
|
27 |
+
"%%capture\n",
|
28 |
+
"!pip install gradio transformers==4.30.2 pillow\n",
|
29 |
+
"!pip install torch torchvision hf_xet timm==1.0.10\n",
|
30 |
+
"!pip install flash-attn --no-build-isolation"
|
31 |
+
]
|
32 |
+
},
|
33 |
+
{
|
34 |
+
"cell_type": "code",
|
35 |
+
"source": [
|
36 |
+
"import gradio as gr\n",
|
37 |
+
"import torch\n",
|
38 |
+
"from PIL import Image\n",
|
39 |
+
"from transformers import AutoProcessor, AutoModelForCausalLM\n",
|
40 |
+
"\n",
|
41 |
+
"device = \"cuda\" if torch.cuda.is_available() else \"cpu\"\n",
|
42 |
+
"vision_language_model = AutoModelForCausalLM.from_pretrained('microsoft/Florence-2-base', trust_remote_code=True).to(device).eval()\n",
|
43 |
+
"vision_language_processor = AutoProcessor.from_pretrained('microsoft/Florence-2-base', trust_remote_code=True)\n",
|
44 |
+
"\n",
|
45 |
+
"def describe_image(uploaded_image):\n",
|
46 |
+
" \"\"\"\n",
|
47 |
+
" Generates a detailed description of the input image.\n",
|
48 |
+
"\n",
|
49 |
+
" Args:\n",
|
50 |
+
" uploaded_image (PIL.Image.Image or numpy.ndarray): The image to describe.\n",
|
51 |
+
"\n",
|
52 |
+
" Returns:\n",
|
53 |
+
" str: A detailed textual description of the image.\n",
|
54 |
+
" \"\"\"\n",
|
55 |
+
" if not isinstance(uploaded_image, Image.Image):\n",
|
56 |
+
" uploaded_image = Image.fromarray(uploaded_image)\n",
|
57 |
+
"\n",
|
58 |
+
" inputs = vision_language_processor(text=\"<MORE_DETAILED_CAPTION>\", images=uploaded_image, return_tensors=\"pt\").to(device)\n",
|
59 |
+
" with torch.no_grad():\n",
|
60 |
+
" generated_ids = vision_language_model.generate(\n",
|
61 |
+
" input_ids=inputs[\"input_ids\"],\n",
|
62 |
+
" pixel_values=inputs[\"pixel_values\"],\n",
|
63 |
+
" max_new_tokens=1024,\n",
|
64 |
+
" early_stopping=False,\n",
|
65 |
+
" do_sample=False,\n",
|
66 |
+
" num_beams=3,\n",
|
67 |
+
" )\n",
|
68 |
+
" generated_text = vision_language_processor.batch_decode(generated_ids, skip_special_tokens=False)[0]\n",
|
69 |
+
" processed_description = vision_language_processor.post_process_generation(\n",
|
70 |
+
" generated_text,\n",
|
71 |
+
" task=\"<MORE_DETAILED_CAPTION>\",\n",
|
72 |
+
" image_size=(uploaded_image.width, uploaded_image.height)\n",
|
73 |
+
" )\n",
|
74 |
+
" image_description = processed_description[\"<MORE_DETAILED_CAPTION>\"]\n",
|
75 |
+
" print(\"\\nImage description generated!:\", image_description)\n",
|
76 |
+
" return image_description\n",
|
77 |
+
"\n",
|
78 |
+
"image_description_interface = gr.Interface(\n",
|
79 |
+
" fn=describe_image,\n",
|
80 |
+
" inputs=gr.Image(label=\"Upload Image\"),\n",
|
81 |
+
" outputs=gr.Textbox(label=\"Generated Caption\", lines=4, show_copy_button=True),\n",
|
82 |
+
" live=False,\n",
|
83 |
+
")\n",
|
84 |
+
"\n",
|
85 |
+
"image_description_interface.launch(debug=True, ssr_mode=False)"
|
86 |
+
],
|
87 |
+
"metadata": {
|
88 |
+
"id": "kW4MjaOs3c9E"
|
89 |
+
},
|
90 |
+
"execution_count": null,
|
91 |
+
"outputs": []
|
92 |
+
}
|
93 |
+
]
|
94 |
+
}
|
Gemma3-VL/Gemma3_4B_it.ipynb
ADDED
@@ -0,0 +1,369 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"nbformat": 4,
|
3 |
+
"nbformat_minor": 0,
|
4 |
+
"metadata": {
|
5 |
+
"colab": {
|
6 |
+
"provenance": [],
|
7 |
+
"gpuType": "T4"
|
8 |
+
},
|
9 |
+
"kernelspec": {
|
10 |
+
"name": "python3",
|
11 |
+
"display_name": "Python 3"
|
12 |
+
},
|
13 |
+
"language_info": {
|
14 |
+
"name": "python"
|
15 |
+
},
|
16 |
+
"accelerator": "GPU"
|
17 |
+
},
|
18 |
+
"cells": [
|
19 |
+
{
|
20 |
+
"cell_type": "code",
|
21 |
+
"execution_count": 1,
|
22 |
+
"metadata": {
|
23 |
+
"id": "m7rU-pjX3Y1O"
|
24 |
+
},
|
25 |
+
"outputs": [],
|
26 |
+
"source": [
|
27 |
+
"%%capture\n",
|
28 |
+
"!pip install gradio transformers accelerate numpy requests\n",
|
29 |
+
"!pip install torch torchvision av hf_xet qwen-vl-utils\n",
|
30 |
+
"!pip install pillow huggingface_hub opencv-python spaces"
|
31 |
+
]
|
32 |
+
},
|
33 |
+
{
|
34 |
+
"cell_type": "code",
|
35 |
+
"source": [
|
36 |
+
"from huggingface_hub import notebook_login, HfApi\n",
|
37 |
+
"notebook_login()"
|
38 |
+
],
|
39 |
+
"metadata": {
|
40 |
+
"id": "dZUVag_jJMck"
|
41 |
+
},
|
42 |
+
"execution_count": null,
|
43 |
+
"outputs": []
|
44 |
+
},
|
45 |
+
{
|
46 |
+
"cell_type": "code",
|
47 |
+
"source": [
|
48 |
+
"import os\n",
|
49 |
+
"import random\n",
|
50 |
+
"import uuid\n",
|
51 |
+
"import json\n",
|
52 |
+
"import time\n",
|
53 |
+
"import asyncio\n",
|
54 |
+
"import re\n",
|
55 |
+
"from threading import Thread\n",
|
56 |
+
"\n",
|
57 |
+
"import gradio as gr\n",
|
58 |
+
"import spaces\n",
|
59 |
+
"import torch\n",
|
60 |
+
"import numpy as np\n",
|
61 |
+
"from PIL import Image\n",
|
62 |
+
"import cv2\n",
|
63 |
+
"\n",
|
64 |
+
"from transformers import (\n",
|
65 |
+
" AutoProcessor,\n",
|
66 |
+
" Gemma3ForConditionalGeneration,\n",
|
67 |
+
" Qwen2VLForConditionalGeneration,\n",
|
68 |
+
" TextIteratorStreamer,\n",
|
69 |
+
")\n",
|
70 |
+
"from transformers.image_utils import load_image\n",
|
71 |
+
"\n",
|
72 |
+
"# Constants\n",
|
73 |
+
"MAX_MAX_NEW_TOKENS = 2048\n",
|
74 |
+
"DEFAULT_MAX_NEW_TOKENS = 1024\n",
|
75 |
+
"MAX_INPUT_TOKEN_LENGTH = int(os.getenv(\"MAX_INPUT_TOKEN_LENGTH\", \"4096\"))\n",
|
76 |
+
"MAX_SEED = np.iinfo(np.int32).max\n",
|
77 |
+
"\n",
|
78 |
+
"device = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n",
|
79 |
+
"\n",
|
80 |
+
"# Helper function to return a progress bar HTML snippet.\n",
|
81 |
+
"def progress_bar_html(label: str) -> str:\n",
|
82 |
+
" return f'''\n",
|
83 |
+
"<div style=\"display: flex; align-items: center;\">\n",
|
84 |
+
" <span style=\"margin-right: 10px; font-size: 14px;\">{label}</span>\n",
|
85 |
+
" <div style=\"width: 110px; height: 5px; background-color: #F0FFF0; border-radius: 2px; overflow: hidden;\">\n",
|
86 |
+
" <div style=\"width: 100%; height: 100%; background-color: #00FF00; animation: loading 1.5s linear infinite;\"></div>\n",
|
87 |
+
" </div>\n",
|
88 |
+
"</div>\n",
|
89 |
+
"<style>\n",
|
90 |
+
"@keyframes loading {{\n",
|
91 |
+
" 0% {{ transform: translateX(-100%); }}\n",
|
92 |
+
" 100% {{ transform: translateX(100%); }}\n",
|
93 |
+
"}}\n",
|
94 |
+
"</style>\n",
|
95 |
+
" '''\n",
|
96 |
+
"\n",
|
97 |
+
"# Qwen2-VL (for optional image inference)\n",
|
98 |
+
"\n",
|
99 |
+
"MODEL_ID_VL = \"prithivMLmods/Qwen2-VL-OCR-2B-Instruct\"\n",
|
100 |
+
"processor = AutoProcessor.from_pretrained(MODEL_ID_VL, trust_remote_code=True)\n",
|
101 |
+
"model_m = Qwen2VLForConditionalGeneration.from_pretrained(\n",
|
102 |
+
" MODEL_ID_VL,\n",
|
103 |
+
" trust_remote_code=True,\n",
|
104 |
+
" torch_dtype=torch.float16\n",
|
105 |
+
").to(\"cuda\").eval()\n",
|
106 |
+
"\n",
|
107 |
+
"def clean_chat_history(chat_history):\n",
|
108 |
+
" cleaned = []\n",
|
109 |
+
" for msg in chat_history:\n",
|
110 |
+
" if isinstance(msg, dict) and isinstance(msg.get(\"content\"), str):\n",
|
111 |
+
" cleaned.append(msg)\n",
|
112 |
+
" return cleaned\n",
|
113 |
+
"\n",
|
114 |
+
"bad_words = json.loads(os.getenv('BAD_WORDS', \"[]\"))\n",
|
115 |
+
"bad_words_negative = json.loads(os.getenv('BAD_WORDS_NEGATIVE', \"[]\"))\n",
|
116 |
+
"default_negative = os.getenv(\"default_negative\", \"\")\n",
|
117 |
+
"\n",
|
118 |
+
"def check_text(prompt, negative=\"\"):\n",
|
119 |
+
" for i in bad_words:\n",
|
120 |
+
" if i in prompt:\n",
|
121 |
+
" return True\n",
|
122 |
+
" for i in bad_words_negative:\n",
|
123 |
+
" if i in negative:\n",
|
124 |
+
" return True\n",
|
125 |
+
" return False\n",
|
126 |
+
"\n",
|
127 |
+
"def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:\n",
|
128 |
+
" if randomize_seed:\n",
|
129 |
+
" seed = random.randint(0, MAX_SEED)\n",
|
130 |
+
" return seed\n",
|
131 |
+
"\n",
|
132 |
+
"CACHE_EXAMPLES = torch.cuda.is_available() and os.getenv(\"CACHE_EXAMPLES\", \"0\") == \"1\"\n",
|
133 |
+
"MAX_IMAGE_SIZE = int(os.getenv(\"MAX_IMAGE_SIZE\", \"2048\"))\n",
|
134 |
+
"USE_TORCH_COMPILE = os.getenv(\"USE_TORCH_COMPILE\", \"0\") == \"1\"\n",
|
135 |
+
"ENABLE_CPU_OFFLOAD = os.getenv(\"ENABLE_CPU_OFFLOAD\", \"0\") == \"1\"\n",
|
136 |
+
"\n",
|
137 |
+
"dtype = torch.float16 if device.type == \"cuda\" else torch.float32\n",
|
138 |
+
"\n",
|
139 |
+
"\n",
|
140 |
+
"# Gemma3 Model (default for text, image, & video inference)\n",
|
141 |
+
"\n",
|
142 |
+
"gemma3_model_id = \"google/gemma-3-4b-it\" # alternative: google/gemma-3-12b-it\n",
|
143 |
+
"gemma3_model = Gemma3ForConditionalGeneration.from_pretrained(\n",
|
144 |
+
" gemma3_model_id, device_map=\"auto\"\n",
|
145 |
+
").eval()\n",
|
146 |
+
"gemma3_processor = AutoProcessor.from_pretrained(gemma3_model_id)\n",
|
147 |
+
"\n",
|
148 |
+
"# VIDEO PROCESSING HELPER\n",
|
149 |
+
"\n",
|
150 |
+
"def downsample_video(video_path):\n",
|
151 |
+
" vidcap = cv2.VideoCapture(video_path)\n",
|
152 |
+
" total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))\n",
|
153 |
+
" fps = vidcap.get(cv2.CAP_PROP_FPS)\n",
|
154 |
+
" frames = []\n",
|
155 |
+
" # Sample 10 evenly spaced frames.\n",
|
156 |
+
" frame_indices = np.linspace(0, total_frames - 1, 10, dtype=int)\n",
|
157 |
+
" for i in frame_indices:\n",
|
158 |
+
" vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)\n",
|
159 |
+
" success, image = vidcap.read()\n",
|
160 |
+
" if success:\n",
|
161 |
+
" # Convert from BGR to RGB and then to PIL Image.\n",
|
162 |
+
" image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)\n",
|
163 |
+
" pil_image = Image.fromarray(image)\n",
|
164 |
+
" timestamp = round(i / fps, 2)\n",
|
165 |
+
" frames.append((pil_image, timestamp))\n",
|
166 |
+
" vidcap.release()\n",
|
167 |
+
" return frames\n",
|
168 |
+
"\n",
|
169 |
+
"# MAIN GENERATION FUNCTION\n",
|
170 |
+
"\n",
|
171 |
+
"@spaces.GPU\n",
|
172 |
+
"def generate(\n",
|
173 |
+
" input_dict: dict,\n",
|
174 |
+
" chat_history: list[dict],\n",
|
175 |
+
" max_new_tokens: int = DEFAULT_MAX_NEW_TOKENS,\n",
|
176 |
+
" temperature: float = 0.6,\n",
|
177 |
+
" top_p: float = 0.9,\n",
|
178 |
+
" top_k: int = 50,\n",
|
179 |
+
" repetition_penalty: float = 1.2,\n",
|
180 |
+
"):\n",
|
181 |
+
" text = input_dict[\"text\"]\n",
|
182 |
+
" files = input_dict.get(\"files\", [])\n",
|
183 |
+
" lower_text = text.lower().strip()\n",
|
184 |
+
"\n",
|
185 |
+
" # ----- Qwen2-VL branch (triggered with @qwen2-vl) -----\n",
|
186 |
+
" if lower_text.startswith(\"@qwen2-vl\"):\n",
|
187 |
+
" prompt_clean = re.sub(r\"@qwen2-vl\", \"\", text, flags=re.IGNORECASE).strip().strip('\"')\n",
|
188 |
+
" if files:\n",
|
189 |
+
" images = [load_image(f) for f in files]\n",
|
190 |
+
" messages = [{\n",
|
191 |
+
" \"role\": \"user\",\n",
|
192 |
+
" \"content\": [\n",
|
193 |
+
" *[{\"type\": \"image\", \"image\": image} for image in images],\n",
|
194 |
+
" {\"type\": \"text\", \"text\": prompt_clean},\n",
|
195 |
+
" ]\n",
|
196 |
+
" }]\n",
|
197 |
+
" prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)\n",
|
198 |
+
" inputs = processor(text=[prompt], images=images, return_tensors=\"pt\", padding=True).to(\"cuda\")\n",
|
199 |
+
" else:\n",
|
200 |
+
" messages = [\n",
|
201 |
+
" {\"role\": \"system\", \"content\": [{\"type\": \"text\", \"text\": \"You are a helpful assistant.\"}]},\n",
|
202 |
+
" {\"role\": \"user\", \"content\": [{\"type\": \"text\", \"text\": prompt_clean}]}\n",
|
203 |
+
" ]\n",
|
204 |
+
" inputs = processor.apply_chat_template(\n",
|
205 |
+
" messages, add_generation_prompt=True, tokenize=True,\n",
|
206 |
+
" return_dict=True, return_tensors=\"pt\"\n",
|
207 |
+
" ).to(\"cuda\", dtype=torch.float16)\n",
|
208 |
+
" streamer = TextIteratorStreamer(processor.tokenizer, timeout=20.0, skip_prompt=True, skip_special_tokens=True)\n",
|
209 |
+
" generation_kwargs = {\n",
|
210 |
+
" **inputs,\n",
|
211 |
+
" \"streamer\": streamer,\n",
|
212 |
+
" \"max_new_tokens\": max_new_tokens,\n",
|
213 |
+
" \"do_sample\": True,\n",
|
214 |
+
" \"temperature\": temperature,\n",
|
215 |
+
" \"top_p\": top_p,\n",
|
216 |
+
" \"top_k\": top_k,\n",
|
217 |
+
" \"repetition_penalty\": repetition_penalty,\n",
|
218 |
+
" }\n",
|
219 |
+
" thread = Thread(target=model_m.generate, kwargs=generation_kwargs)\n",
|
220 |
+
" thread.start()\n",
|
221 |
+
" buffer = \"\"\n",
|
222 |
+
" yield progress_bar_html(\"Processing with Qwen2VL\")\n",
|
223 |
+
" for new_text in streamer:\n",
|
224 |
+
" buffer += new_text\n",
|
225 |
+
" buffer = buffer.replace(\"<|im_end|>\", \"\")\n",
|
226 |
+
" time.sleep(0.01)\n",
|
227 |
+
" yield buffer\n",
|
228 |
+
" return\n",
|
229 |
+
"\n",
|
230 |
+
" # ----- Default branch: Gemma3 (for text, image, & video inference) -----\n",
|
231 |
+
" if files:\n",
|
232 |
+
" # Check if any provided file is a video based on extension.\n",
|
233 |
+
" video_extensions = (\".mp4\", \".mov\", \".avi\", \".mkv\", \".webm\")\n",
|
234 |
+
" if any(str(f).lower().endswith(video_extensions) for f in files):\n",
|
235 |
+
" # Video inference branch.\n",
|
236 |
+
" prompt_clean = re.sub(r\"@video-infer\", \"\", text, flags=re.IGNORECASE).strip().strip('\"')\n",
|
237 |
+
" video_path = files[0]\n",
|
238 |
+
" frames = downsample_video(video_path)\n",
|
239 |
+
" messages = [\n",
|
240 |
+
" {\"role\": \"system\", \"content\": [{\"type\": \"text\", \"text\": \"You are a helpful assistant.\"}]},\n",
|
241 |
+
" {\"role\": \"user\", \"content\": [{\"type\": \"text\", \"text\": prompt_clean}]}\n",
|
242 |
+
" ]\n",
|
243 |
+
" # Append each frame (with its timestamp) to the conversation.\n",
|
244 |
+
" for frame in frames:\n",
|
245 |
+
" image, timestamp = frame\n",
|
246 |
+
" image_path = f\"video_frame_{uuid.uuid4().hex}.png\"\n",
|
247 |
+
" image.save(image_path)\n",
|
248 |
+
" messages[1][\"content\"].append({\"type\": \"text\", \"text\": f\"Frame {timestamp}:\"})\n",
|
249 |
+
" messages[1][\"content\"].append({\"type\": \"image\", \"url\": image_path})\n",
|
250 |
+
" inputs = gemma3_processor.apply_chat_template(\n",
|
251 |
+
" messages, add_generation_prompt=True, tokenize=True,\n",
|
252 |
+
" return_dict=True, return_tensors=\"pt\"\n",
|
253 |
+
" ).to(gemma3_model.device, dtype=torch.bfloat16)\n",
|
254 |
+
" streamer = TextIteratorStreamer(gemma3_processor.tokenizer, timeout=20.0, skip_prompt=True, skip_special_tokens=True)\n",
|
255 |
+
" generation_kwargs = {\n",
|
256 |
+
" **inputs,\n",
|
257 |
+
" \"streamer\": streamer,\n",
|
258 |
+
" \"max_new_tokens\": max_new_tokens,\n",
|
259 |
+
" \"do_sample\": True,\n",
|
260 |
+
" \"temperature\": temperature,\n",
|
261 |
+
" \"top_p\": top_p,\n",
|
262 |
+
" \"top_k\": top_k,\n",
|
263 |
+
" \"repetition_penalty\": repetition_penalty,\n",
|
264 |
+
" }\n",
|
265 |
+
" thread = Thread(target=gemma3_model.generate, kwargs=generation_kwargs)\n",
|
266 |
+
" thread.start()\n",
|
267 |
+
" buffer = \"\"\n",
|
268 |
+
" yield progress_bar_html(\"Processing video with Gemma3\")\n",
|
269 |
+
" for new_text in streamer:\n",
|
270 |
+
" buffer += new_text\n",
|
271 |
+
" time.sleep(0.01)\n",
|
272 |
+
" yield buffer\n",
|
273 |
+
" return\n",
|
274 |
+
" else:\n",
|
275 |
+
" # Image inference branch.\n",
|
276 |
+
" prompt_clean = re.sub(r\"@gemma3\", \"\", text, flags=re.IGNORECASE).strip().strip('\"')\n",
|
277 |
+
" images = [load_image(f) for f in files]\n",
|
278 |
+
" messages = [{\n",
|
279 |
+
" \"role\": \"user\",\n",
|
280 |
+
" \"content\": [\n",
|
281 |
+
" *[{\"type\": \"image\", \"image\": image} for image in images],\n",
|
282 |
+
" {\"type\": \"text\", \"text\": prompt_clean},\n",
|
283 |
+
" ]\n",
|
284 |
+
" }]\n",
|
285 |
+
" inputs = gemma3_processor.apply_chat_template(\n",
|
286 |
+
" messages, tokenize=True, add_generation_prompt=True,\n",
|
287 |
+
" return_dict=True, return_tensors=\"pt\"\n",
|
288 |
+
" ).to(gemma3_model.device, dtype=torch.bfloat16)\n",
|
289 |
+
" streamer = TextIteratorStreamer(gemma3_processor.tokenizer, timeout=20.0, skip_prompt=True, skip_special_tokens=True)\n",
|
290 |
+
" generation_kwargs = {\n",
|
291 |
+
" **inputs,\n",
|
292 |
+
" \"streamer\": streamer,\n",
|
293 |
+
" \"max_new_tokens\": max_new_tokens,\n",
|
294 |
+
" \"do_sample\": True,\n",
|
295 |
+
" \"temperature\": temperature,\n",
|
296 |
+
" \"top_p\": top_p,\n",
|
297 |
+
" \"top_k\": top_k,\n",
|
298 |
+
" \"repetition_penalty\": repetition_penalty,\n",
|
299 |
+
" }\n",
|
300 |
+
" thread = Thread(target=gemma3_model.generate, kwargs=generation_kwargs)\n",
|
301 |
+
" thread.start()\n",
|
302 |
+
" buffer = \"\"\n",
|
303 |
+
" yield progress_bar_html(\"Processing with Gemma3\")\n",
|
304 |
+
" for new_text in streamer:\n",
|
305 |
+
" buffer += new_text\n",
|
306 |
+
" time.sleep(0.01)\n",
|
307 |
+
" yield buffer\n",
|
308 |
+
" return\n",
|
309 |
+
" else:\n",
|
310 |
+
" # Text-only inference branch.\n",
|
311 |
+
" messages = [\n",
|
312 |
+
" {\"role\": \"system\", \"content\": [{\"type\": \"text\", \"text\": \"You are a helpful assistant.\"}]},\n",
|
313 |
+
" {\"role\": \"user\", \"content\": [{\"type\": \"text\", \"text\": text}]}\n",
|
314 |
+
" ]\n",
|
315 |
+
" inputs = gemma3_processor.apply_chat_template(\n",
|
316 |
+
" messages, add_generation_prompt=True, tokenize=True,\n",
|
317 |
+
" return_dict=True, return_tensors=\"pt\"\n",
|
318 |
+
" ).to(gemma3_model.device, dtype=torch.bfloat16)\n",
|
319 |
+
" streamer = TextIteratorStreamer(gemma3_processor.tokenizer, timeout=20.0, skip_prompt=True, skip_special_tokens=True)\n",
|
320 |
+
" generation_kwargs = {\n",
|
321 |
+
" **inputs,\n",
|
322 |
+
" \"streamer\": streamer,\n",
|
323 |
+
" \"max_new_tokens\": max_new_tokens,\n",
|
324 |
+
" \"do_sample\": True,\n",
|
325 |
+
" \"temperature\": temperature,\n",
|
326 |
+
" \"top_p\": top_p,\n",
|
327 |
+
" \"top_k\": top_k,\n",
|
328 |
+
" \"repetition_penalty\": repetition_penalty,\n",
|
329 |
+
" }\n",
|
330 |
+
" thread = Thread(target=gemma3_model.generate, kwargs=generation_kwargs)\n",
|
331 |
+
" thread.start()\n",
|
332 |
+
" outputs = []\n",
|
333 |
+
" for new_text in streamer:\n",
|
334 |
+
" outputs.append(new_text)\n",
|
335 |
+
" yield \"\".join(outputs)\n",
|
336 |
+
" final_response = \"\".join(outputs)\n",
|
337 |
+
" yield final_response\n",
|
338 |
+
"\n",
|
339 |
+
"\n",
|
340 |
+
"# Gradio Interface\n",
|
341 |
+
"\n",
|
342 |
+
"demo = gr.ChatInterface(\n",
|
343 |
+
" fn=generate,\n",
|
344 |
+
" additional_inputs=[\n",
|
345 |
+
" gr.Slider(label=\"Max new tokens\", minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1, value=DEFAULT_MAX_NEW_TOKENS),\n",
|
346 |
+
" gr.Slider(label=\"Temperature\", minimum=0.1, maximum=4.0, step=0.1, value=0.6),\n",
|
347 |
+
" gr.Slider(label=\"Top-p (nucleus sampling)\", minimum=0.05, maximum=1.0, step=0.05, value=0.9),\n",
|
348 |
+
" gr.Slider(label=\"Top-k\", minimum=1, maximum=1000, step=1, value=50),\n",
|
349 |
+
" gr.Slider(label=\"Repetition penalty\", minimum=1.0, maximum=2.0, step=0.05, value=1.2),\n",
|
350 |
+
" ],\n",
|
351 |
+
" type=\"messages\",\n",
|
352 |
+
" description=\"# **Gemma 3 Multimodal** \\n`Use @qwen2-vl to switch to Qwen2-VL OCR for image inference and @video-infer for video input`\",\n",
|
353 |
+
" fill_height=True,\n",
|
354 |
+
" textbox=gr.MultimodalTextbox(label=\"Query Input\", file_types=[\"image\", \"video\"], file_count=\"multiple\", placeholder=\"Tag with @qwen2-vl for Qwen2-VL inference if needed.\"),\n",
|
355 |
+
" stop_btn=\"Stop Generation\",\n",
|
356 |
+
" multimodal=True,\n",
|
357 |
+
")\n",
|
358 |
+
"\n",
|
359 |
+
"if __name__ == \"__main__\":\n",
|
360 |
+
" demo.queue(max_size=20).launch(share=True)"
|
361 |
+
],
|
362 |
+
"metadata": {
|
363 |
+
"id": "kW4MjaOs3c9E"
|
364 |
+
},
|
365 |
+
"execution_count": null,
|
366 |
+
"outputs": []
|
367 |
+
}
|
368 |
+
]
|
369 |
+
}
|
MiMo-VL-7B-RL/MiMo_VL_7B_RL.ipynb
ADDED
@@ -0,0 +1,242 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cells": [
|
3 |
+
{
|
4 |
+
"cell_type": "code",
|
5 |
+
"execution_count": null,
|
6 |
+
"metadata": {
|
7 |
+
"id": "xL8y37Y6bORU"
|
8 |
+
},
|
9 |
+
"outputs": [],
|
10 |
+
"source": [
|
11 |
+
"%%capture\n",
|
12 |
+
"!pip install gradio spaces transformers accelerate numpy requests\n",
|
13 |
+
"!pip install torch torchvision qwen-vl-utils av hf_xet\n",
|
14 |
+
"!pip install pillow huggingface_hub opencv-python"
|
15 |
+
]
|
16 |
+
},
|
17 |
+
{
|
18 |
+
"cell_type": "code",
|
19 |
+
"execution_count": null,
|
20 |
+
"metadata": {
|
21 |
+
"id": "Y-NTbL1tdL9X"
|
22 |
+
},
|
23 |
+
"outputs": [],
|
24 |
+
"source": [
|
25 |
+
"import os\n",
|
26 |
+
"import random\n",
|
27 |
+
"import uuid\n",
|
28 |
+
"import json\n",
|
29 |
+
"import time\n",
|
30 |
+
"import asyncio\n",
|
31 |
+
"from threading import Thread\n",
|
32 |
+
"\n",
|
33 |
+
"import gradio as gr\n",
|
34 |
+
"import spaces\n",
|
35 |
+
"import torch\n",
|
36 |
+
"import numpy as np\n",
|
37 |
+
"from PIL import Image\n",
|
38 |
+
"import cv2\n",
|
39 |
+
"\n",
|
40 |
+
"from transformers import (\n",
|
41 |
+
" Qwen2_5_VLForConditionalGeneration,\n",
|
42 |
+
" AutoProcessor,\n",
|
43 |
+
" TextIteratorStreamer,\n",
|
44 |
+
")\n",
|
45 |
+
"from transformers.image_utils import load_image\n",
|
46 |
+
"\n",
|
47 |
+
"# Constants for text generation\n",
|
48 |
+
"MAX_MAX_NEW_TOKENS = 2048\n",
|
49 |
+
"DEFAULT_MAX_NEW_TOKENS = 1024\n",
|
50 |
+
"# Increase or disable input truncation to avoid token mismatches\n",
|
51 |
+
"MAX_INPUT_TOKEN_LENGTH = int(os.getenv(\"MAX_INPUT_TOKEN_LENGTH\", \"4096\"))\n",
|
52 |
+
"\n",
|
53 |
+
"device = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n",
|
54 |
+
"\n",
|
55 |
+
"MODEL_ID = \"XiaomiMiMo/MiMo-VL-7B-RL\"\n",
|
56 |
+
"processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)\n",
|
57 |
+
"model_m = Qwen2_5_VLForConditionalGeneration.from_pretrained(\n",
|
58 |
+
" MODEL_ID,\n",
|
59 |
+
" trust_remote_code=True,\n",
|
60 |
+
" torch_dtype=torch.float16\n",
|
61 |
+
").to(\"cuda\").eval()\n",
|
62 |
+
"\n",
|
63 |
+
"def downsample_video(video_path):\n",
|
64 |
+
" \"\"\"\n",
|
65 |
+
" Downsamples the video to evenly spaced frames.\n",
|
66 |
+
" Each frame is returned as a PIL image along with its timestamp.\n",
|
67 |
+
" \"\"\"\n",
|
68 |
+
" vidcap = cv2.VideoCapture(video_path)\n",
|
69 |
+
" total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))\n",
|
70 |
+
" fps = vidcap.get(cv2.CAP_PROP_FPS)\n",
|
71 |
+
" frames = []\n",
|
72 |
+
" # Sample 10 evenly spaced frames.\n",
|
73 |
+
" frame_indices = np.linspace(0, total_frames - 1, 10, dtype=int)\n",
|
74 |
+
" for i in frame_indices:\n",
|
75 |
+
" vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)\n",
|
76 |
+
" success, image = vidcap.read()\n",
|
77 |
+
" if success:\n",
|
78 |
+
" image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) # Convert BGR to RGB\n",
|
79 |
+
" pil_image = Image.fromarray(image)\n",
|
80 |
+
" timestamp = round(i / fps, 2)\n",
|
81 |
+
" frames.append((pil_image, timestamp))\n",
|
82 |
+
" vidcap.release()\n",
|
83 |
+
" return frames\n",
|
84 |
+
"\n",
|
85 |
+
"@spaces.GPU\n",
|
86 |
+
"def generate_image(text: str, image: Image.Image,\n",
|
87 |
+
" max_new_tokens: int = 1024,\n",
|
88 |
+
" temperature: float = 0.6,\n",
|
89 |
+
" top_p: float = 0.9,\n",
|
90 |
+
" top_k: int = 50,\n",
|
91 |
+
" repetition_penalty: float = 1.2):\n",
|
92 |
+
"\n",
|
93 |
+
" if image is None:\n",
|
94 |
+
" yield \"Please upload an image.\"\n",
|
95 |
+
" return\n",
|
96 |
+
"\n",
|
97 |
+
" messages = [{\n",
|
98 |
+
" \"role\": \"user\",\n",
|
99 |
+
" \"content\": [\n",
|
100 |
+
" {\"type\": \"image\", \"image\": image},\n",
|
101 |
+
" {\"type\": \"text\", \"text\": text},\n",
|
102 |
+
" ]\n",
|
103 |
+
" }]\n",
|
104 |
+
" prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)\n",
|
105 |
+
" inputs = processor(\n",
|
106 |
+
" text=[prompt_full],\n",
|
107 |
+
" images=[image],\n",
|
108 |
+
" return_tensors=\"pt\",\n",
|
109 |
+
" padding=True,\n",
|
110 |
+
" truncation=False,\n",
|
111 |
+
" max_length=MAX_INPUT_TOKEN_LENGTH\n",
|
112 |
+
" ).to(\"cuda\")\n",
|
113 |
+
" streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)\n",
|
114 |
+
" generation_kwargs = {**inputs, \"streamer\": streamer, \"max_new_tokens\": max_new_tokens}\n",
|
115 |
+
" thread = Thread(target=model_m.generate, kwargs=generation_kwargs)\n",
|
116 |
+
" thread.start()\n",
|
117 |
+
" buffer = \"\"\n",
|
118 |
+
" for new_text in streamer:\n",
|
119 |
+
" buffer += new_text\n",
|
120 |
+
" buffer = buffer.replace(\"<|im_end|>\", \"\")\n",
|
121 |
+
" time.sleep(0.01)\n",
|
122 |
+
" yield buffer\n",
|
123 |
+
"\n",
|
124 |
+
"@spaces.GPU\n",
|
125 |
+
"def generate_video(text: str, video_path: str,\n",
|
126 |
+
" max_new_tokens: int = 1024,\n",
|
127 |
+
" temperature: float = 0.6,\n",
|
128 |
+
" top_p: float = 0.9,\n",
|
129 |
+
" top_k: int = 50,\n",
|
130 |
+
" repetition_penalty: float = 1.2):\n",
|
131 |
+
"\n",
|
132 |
+
" if video_path is None:\n",
|
133 |
+
" yield \"Please upload a video.\"\n",
|
134 |
+
" return\n",
|
135 |
+
"\n",
|
136 |
+
" frames = downsample_video(video_path)\n",
|
137 |
+
" messages = [\n",
|
138 |
+
" {\"role\": \"system\", \"content\": [{\"type\": \"text\", \"text\": \"You are a helpful assistant.\"}]},\n",
|
139 |
+
" {\"role\": \"user\", \"content\": [{\"type\": \"text\", \"text\": text}]}\n",
|
140 |
+
" ]\n",
|
141 |
+
" # Append each frame with its timestamp.\n",
|
142 |
+
" for frame in frames:\n",
|
143 |
+
" image, timestamp = frame\n",
|
144 |
+
" messages[1][\"content\"].append({\"type\": \"text\", \"text\": f\"Frame {timestamp}:\"})\n",
|
145 |
+
" messages[1][\"content\"].append({\"type\": \"image\", \"image\": image})\n",
|
146 |
+
" inputs = processor.apply_chat_template(\n",
|
147 |
+
" messages,\n",
|
148 |
+
" tokenize=True,\n",
|
149 |
+
" add_generation_prompt=True,\n",
|
150 |
+
" return_dict=True,\n",
|
151 |
+
" return_tensors=\"pt\",\n",
|
152 |
+
" truncation=False,\n",
|
153 |
+
" max_length=MAX_INPUT_TOKEN_LENGTH\n",
|
154 |
+
" ).to(\"cuda\")\n",
|
155 |
+
" streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)\n",
|
156 |
+
" generation_kwargs = {\n",
|
157 |
+
" **inputs,\n",
|
158 |
+
" \"streamer\": streamer,\n",
|
159 |
+
" \"max_new_tokens\": max_new_tokens,\n",
|
160 |
+
" \"do_sample\": True,\n",
|
161 |
+
" \"temperature\": temperature,\n",
|
162 |
+
" \"top_p\": top_p,\n",
|
163 |
+
" \"top_k\": top_k,\n",
|
164 |
+
" \"repetition_penalty\": repetition_penalty,\n",
|
165 |
+
" }\n",
|
166 |
+
" thread = Thread(target=model_m.generate, kwargs=generation_kwargs)\n",
|
167 |
+
" thread.start()\n",
|
168 |
+
" buffer = \"\"\n",
|
169 |
+
" for new_text in streamer:\n",
|
170 |
+
" buffer += new_text\n",
|
171 |
+
" buffer = buffer.replace(\"<|im_end|>\", \"\")\n",
|
172 |
+
" time.sleep(0.01)\n",
|
173 |
+
" yield buffer\n",
|
174 |
+
"\n",
|
175 |
+
"css = \"\"\"\n",
|
176 |
+
".submit-btn {\n",
|
177 |
+
" background-color: #2980b9 !important;\n",
|
178 |
+
" color: white !important;\n",
|
179 |
+
"}\n",
|
180 |
+
".submit-btn:hover {\n",
|
181 |
+
" background-color: #3498db !important;\n",
|
182 |
+
"}\n",
|
183 |
+
"\"\"\"\n",
|
184 |
+
"\n",
|
185 |
+
"# Create the Gradio Interface\n",
|
186 |
+
"with gr.Blocks(css=css, theme=\"bethecloud/storj_theme\") as demo:\n",
|
187 |
+
" gr.Markdown(\"# **XiaomiMiMo/MiMo-VL-7B-RL**\")\n",
|
188 |
+
" with gr.Row():\n",
|
189 |
+
" with gr.Column():\n",
|
190 |
+
" with gr.Tabs():\n",
|
191 |
+
" with gr.TabItem(\"Image Inference\"):\n",
|
192 |
+
" image_query = gr.Textbox(label=\"Query Input\", placeholder=\"Enter your query here...\")\n",
|
193 |
+
" image_upload = gr.Image(type=\"pil\", label=\"Image\")\n",
|
194 |
+
" image_submit = gr.Button(\"Submit\", elem_classes=\"submit-btn\")\n",
|
195 |
+
"\n",
|
196 |
+
" with gr.TabItem(\"Video Inference\"):\n",
|
197 |
+
" video_query = gr.Textbox(label=\"Query Input\", placeholder=\"Enter your query here...\")\n",
|
198 |
+
" video_upload = gr.Video(label=\"Video\")\n",
|
199 |
+
" video_submit = gr.Button(\"Submit\", elem_classes=\"submit-btn\")\n",
|
200 |
+
"\n",
|
201 |
+
" with gr.Accordion(\"Advanced options\", open=False):\n",
|
202 |
+
" max_new_tokens = gr.Slider(label=\"Max new tokens\", minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1, value=DEFAULT_MAX_NEW_TOKENS)\n",
|
203 |
+
" temperature = gr.Slider(label=\"Temperature\", minimum=0.1, maximum=4.0, step=0.1, value=0.6)\n",
|
204 |
+
" top_p = gr.Slider(label=\"Top-p (nucleus sampling)\", minimum=0.05, maximum=1.0, step=0.05, value=0.9)\n",
|
205 |
+
" top_k = gr.Slider(label=\"Top-k\", minimum=1, maximum=1000, step=1, value=50)\n",
|
206 |
+
" repetition_penalty = gr.Slider(label=\"Repetition penalty\", minimum=1.0, maximum=2.0, step=0.05, value=1.2)\n",
|
207 |
+
" with gr.Column():\n",
|
208 |
+
" output = gr.Textbox(label=\"Output\", interactive=False)\n",
|
209 |
+
"\n",
|
210 |
+
" image_submit.click(\n",
|
211 |
+
" fn=generate_image,\n",
|
212 |
+
" inputs=[image_query, image_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],\n",
|
213 |
+
" outputs=output\n",
|
214 |
+
" )\n",
|
215 |
+
" video_submit.click(\n",
|
216 |
+
" fn=generate_video,\n",
|
217 |
+
" inputs=[video_query, video_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],\n",
|
218 |
+
" outputs=output\n",
|
219 |
+
" )\n",
|
220 |
+
"\n",
|
221 |
+
"if __name__ == \"__main__\":\n",
|
222 |
+
" demo.queue(max_size=30).launch(share=True, ssr_mode=False, show_error=True)"
|
223 |
+
]
|
224 |
+
}
|
225 |
+
],
|
226 |
+
"metadata": {
|
227 |
+
"accelerator": "GPU",
|
228 |
+
"colab": {
|
229 |
+
"gpuType": "T4",
|
230 |
+
"provenance": []
|
231 |
+
},
|
232 |
+
"kernelspec": {
|
233 |
+
"display_name": "Python 3",
|
234 |
+
"name": "python3"
|
235 |
+
},
|
236 |
+
"language_info": {
|
237 |
+
"name": "python"
|
238 |
+
}
|
239 |
+
},
|
240 |
+
"nbformat": 4,
|
241 |
+
"nbformat_minor": 0
|
242 |
+
}
|
MiMo-VL-7B-SFT/MiMo_VL_7B_SFT.ipynb
ADDED
@@ -0,0 +1,242 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cells": [
|
3 |
+
{
|
4 |
+
"cell_type": "code",
|
5 |
+
"execution_count": null,
|
6 |
+
"metadata": {
|
7 |
+
"id": "xL8y37Y6bORU"
|
8 |
+
},
|
9 |
+
"outputs": [],
|
10 |
+
"source": [
|
11 |
+
"%%capture\n",
|
12 |
+
"!pip install gradio spaces transformers accelerate numpy requests\n",
|
13 |
+
"!pip install torch torchvision qwen-vl-utils av hf_xet\n",
|
14 |
+
"!pip install pillow huggingface_hub opencv-python"
|
15 |
+
]
|
16 |
+
},
|
17 |
+
{
|
18 |
+
"cell_type": "code",
|
19 |
+
"execution_count": null,
|
20 |
+
"metadata": {
|
21 |
+
"id": "Y-NTbL1tdL9X"
|
22 |
+
},
|
23 |
+
"outputs": [],
|
24 |
+
"source": [
|
25 |
+
"import os\n",
|
26 |
+
"import random\n",
|
27 |
+
"import uuid\n",
|
28 |
+
"import json\n",
|
29 |
+
"import time\n",
|
30 |
+
"import asyncio\n",
|
31 |
+
"from threading import Thread\n",
|
32 |
+
"\n",
|
33 |
+
"import gradio as gr\n",
|
34 |
+
"import spaces\n",
|
35 |
+
"import torch\n",
|
36 |
+
"import numpy as np\n",
|
37 |
+
"from PIL import Image\n",
|
38 |
+
"import cv2\n",
|
39 |
+
"\n",
|
40 |
+
"from transformers import (\n",
|
41 |
+
" Qwen2_5_VLForConditionalGeneration,\n",
|
42 |
+
" AutoProcessor,\n",
|
43 |
+
" TextIteratorStreamer,\n",
|
44 |
+
")\n",
|
45 |
+
"from transformers.image_utils import load_image\n",
|
46 |
+
"\n",
|
47 |
+
"# Constants for text generation\n",
|
48 |
+
"MAX_MAX_NEW_TOKENS = 2048\n",
|
49 |
+
"DEFAULT_MAX_NEW_TOKENS = 1024\n",
|
50 |
+
"# Increase or disable input truncation to avoid token mismatches\n",
|
51 |
+
"MAX_INPUT_TOKEN_LENGTH = int(os.getenv(\"MAX_INPUT_TOKEN_LENGTH\", \"4096\"))\n",
|
52 |
+
"\n",
|
53 |
+
"device = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n",
|
54 |
+
"\n",
|
55 |
+
"MODEL_ID = \"XiaomiMiMo/MiMo-VL-7B-SFT\"\n",
|
56 |
+
"processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)\n",
|
57 |
+
"model_m = Qwen2_5_VLForConditionalGeneration.from_pretrained(\n",
|
58 |
+
" MODEL_ID,\n",
|
59 |
+
" trust_remote_code=True,\n",
|
60 |
+
" torch_dtype=torch.float16\n",
|
61 |
+
").to(\"cuda\").eval()\n",
|
62 |
+
"\n",
|
63 |
+
"def downsample_video(video_path):\n",
|
64 |
+
" \"\"\"\n",
|
65 |
+
" Downsamples the video to evenly spaced frames.\n",
|
66 |
+
" Each frame is returned as a PIL image along with its timestamp.\n",
|
67 |
+
" \"\"\"\n",
|
68 |
+
" vidcap = cv2.VideoCapture(video_path)\n",
|
69 |
+
" total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))\n",
|
70 |
+
" fps = vidcap.get(cv2.CAP_PROP_FPS)\n",
|
71 |
+
" frames = []\n",
|
72 |
+
" # Sample 10 evenly spaced frames.\n",
|
73 |
+
" frame_indices = np.linspace(0, total_frames - 1, 10, dtype=int)\n",
|
74 |
+
" for i in frame_indices:\n",
|
75 |
+
" vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)\n",
|
76 |
+
" success, image = vidcap.read()\n",
|
77 |
+
" if success:\n",
|
78 |
+
" image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) # Convert BGR to RGB\n",
|
79 |
+
" pil_image = Image.fromarray(image)\n",
|
80 |
+
" timestamp = round(i / fps, 2)\n",
|
81 |
+
" frames.append((pil_image, timestamp))\n",
|
82 |
+
" vidcap.release()\n",
|
83 |
+
" return frames\n",
|
84 |
+
"\n",
|
85 |
+
"@spaces.GPU\n",
|
86 |
+
"def generate_image(text: str, image: Image.Image,\n",
|
87 |
+
" max_new_tokens: int = 1024,\n",
|
88 |
+
" temperature: float = 0.6,\n",
|
89 |
+
" top_p: float = 0.9,\n",
|
90 |
+
" top_k: int = 50,\n",
|
91 |
+
" repetition_penalty: float = 1.2):\n",
|
92 |
+
"\n",
|
93 |
+
" if image is None:\n",
|
94 |
+
" yield \"Please upload an image.\"\n",
|
95 |
+
" return\n",
|
96 |
+
"\n",
|
97 |
+
" messages = [{\n",
|
98 |
+
" \"role\": \"user\",\n",
|
99 |
+
" \"content\": [\n",
|
100 |
+
" {\"type\": \"image\", \"image\": image},\n",
|
101 |
+
" {\"type\": \"text\", \"text\": text},\n",
|
102 |
+
" ]\n",
|
103 |
+
" }]\n",
|
104 |
+
" prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)\n",
|
105 |
+
" inputs = processor(\n",
|
106 |
+
" text=[prompt_full],\n",
|
107 |
+
" images=[image],\n",
|
108 |
+
" return_tensors=\"pt\",\n",
|
109 |
+
" padding=True,\n",
|
110 |
+
" truncation=False,\n",
|
111 |
+
" max_length=MAX_INPUT_TOKEN_LENGTH\n",
|
112 |
+
" ).to(\"cuda\")\n",
|
113 |
+
" streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)\n",
|
114 |
+
" generation_kwargs = {**inputs, \"streamer\": streamer, \"max_new_tokens\": max_new_tokens}\n",
|
115 |
+
" thread = Thread(target=model_m.generate, kwargs=generation_kwargs)\n",
|
116 |
+
" thread.start()\n",
|
117 |
+
" buffer = \"\"\n",
|
118 |
+
" for new_text in streamer:\n",
|
119 |
+
" buffer += new_text\n",
|
120 |
+
" buffer = buffer.replace(\"<|im_end|>\", \"\")\n",
|
121 |
+
" time.sleep(0.01)\n",
|
122 |
+
" yield buffer\n",
|
123 |
+
"\n",
|
124 |
+
"@spaces.GPU\n",
|
125 |
+
"def generate_video(text: str, video_path: str,\n",
|
126 |
+
" max_new_tokens: int = 1024,\n",
|
127 |
+
" temperature: float = 0.6,\n",
|
128 |
+
" top_p: float = 0.9,\n",
|
129 |
+
" top_k: int = 50,\n",
|
130 |
+
" repetition_penalty: float = 1.2):\n",
|
131 |
+
"\n",
|
132 |
+
" if video_path is None:\n",
|
133 |
+
" yield \"Please upload a video.\"\n",
|
134 |
+
" return\n",
|
135 |
+
"\n",
|
136 |
+
" frames = downsample_video(video_path)\n",
|
137 |
+
" messages = [\n",
|
138 |
+
" {\"role\": \"system\", \"content\": [{\"type\": \"text\", \"text\": \"You are a helpful assistant.\"}]},\n",
|
139 |
+
" {\"role\": \"user\", \"content\": [{\"type\": \"text\", \"text\": text}]}\n",
|
140 |
+
" ]\n",
|
141 |
+
" # Append each frame with its timestamp.\n",
|
142 |
+
" for frame in frames:\n",
|
143 |
+
" image, timestamp = frame\n",
|
144 |
+
" messages[1][\"content\"].append({\"type\": \"text\", \"text\": f\"Frame {timestamp}:\"})\n",
|
145 |
+
" messages[1][\"content\"].append({\"type\": \"image\", \"image\": image})\n",
|
146 |
+
" inputs = processor.apply_chat_template(\n",
|
147 |
+
" messages,\n",
|
148 |
+
" tokenize=True,\n",
|
149 |
+
" add_generation_prompt=True,\n",
|
150 |
+
" return_dict=True,\n",
|
151 |
+
" return_tensors=\"pt\",\n",
|
152 |
+
" truncation=False,\n",
|
153 |
+
" max_length=MAX_INPUT_TOKEN_LENGTH\n",
|
154 |
+
" ).to(\"cuda\")\n",
|
155 |
+
" streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)\n",
|
156 |
+
" generation_kwargs = {\n",
|
157 |
+
" **inputs,\n",
|
158 |
+
" \"streamer\": streamer,\n",
|
159 |
+
" \"max_new_tokens\": max_new_tokens,\n",
|
160 |
+
" \"do_sample\": True,\n",
|
161 |
+
" \"temperature\": temperature,\n",
|
162 |
+
" \"top_p\": top_p,\n",
|
163 |
+
" \"top_k\": top_k,\n",
|
164 |
+
" \"repetition_penalty\": repetition_penalty,\n",
|
165 |
+
" }\n",
|
166 |
+
" thread = Thread(target=model_m.generate, kwargs=generation_kwargs)\n",
|
167 |
+
" thread.start()\n",
|
168 |
+
" buffer = \"\"\n",
|
169 |
+
" for new_text in streamer:\n",
|
170 |
+
" buffer += new_text\n",
|
171 |
+
" buffer = buffer.replace(\"<|im_end|>\", \"\")\n",
|
172 |
+
" time.sleep(0.01)\n",
|
173 |
+
" yield buffer\n",
|
174 |
+
"\n",
|
175 |
+
"css = \"\"\"\n",
|
176 |
+
".submit-btn {\n",
|
177 |
+
" background-color: #2980b9 !important;\n",
|
178 |
+
" color: white !important;\n",
|
179 |
+
"}\n",
|
180 |
+
".submit-btn:hover {\n",
|
181 |
+
" background-color: #3498db !important;\n",
|
182 |
+
"}\n",
|
183 |
+
"\"\"\"\n",
|
184 |
+
"\n",
|
185 |
+
"# Create the Gradio Interface\n",
|
186 |
+
"with gr.Blocks(css=css, theme=\"bethecloud/storj_theme\") as demo:\n",
|
187 |
+
" gr.Markdown(\"# **XiaomiMiMo/MiMo-VL-7B-SFT**\")\n",
|
188 |
+
" with gr.Row():\n",
|
189 |
+
" with gr.Column():\n",
|
190 |
+
" with gr.Tabs():\n",
|
191 |
+
" with gr.TabItem(\"Image Inference\"):\n",
|
192 |
+
" image_query = gr.Textbox(label=\"Query Input\", placeholder=\"Enter your query here...\")\n",
|
193 |
+
" image_upload = gr.Image(type=\"pil\", label=\"Image\")\n",
|
194 |
+
" image_submit = gr.Button(\"Submit\", elem_classes=\"submit-btn\")\n",
|
195 |
+
"\n",
|
196 |
+
" with gr.TabItem(\"Video Inference\"):\n",
|
197 |
+
" video_query = gr.Textbox(label=\"Query Input\", placeholder=\"Enter your query here...\")\n",
|
198 |
+
" video_upload = gr.Video(label=\"Video\")\n",
|
199 |
+
" video_submit = gr.Button(\"Submit\", elem_classes=\"submit-btn\")\n",
|
200 |
+
"\n",
|
201 |
+
" with gr.Accordion(\"Advanced options\", open=False):\n",
|
202 |
+
" max_new_tokens = gr.Slider(label=\"Max new tokens\", minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1, value=DEFAULT_MAX_NEW_TOKENS)\n",
|
203 |
+
" temperature = gr.Slider(label=\"Temperature\", minimum=0.1, maximum=4.0, step=0.1, value=0.6)\n",
|
204 |
+
" top_p = gr.Slider(label=\"Top-p (nucleus sampling)\", minimum=0.05, maximum=1.0, step=0.05, value=0.9)\n",
|
205 |
+
" top_k = gr.Slider(label=\"Top-k\", minimum=1, maximum=1000, step=1, value=50)\n",
|
206 |
+
" repetition_penalty = gr.Slider(label=\"Repetition penalty\", minimum=1.0, maximum=2.0, step=0.05, value=1.2)\n",
|
207 |
+
" with gr.Column():\n",
|
208 |
+
" output = gr.Textbox(label=\"Output\", interactive=False)\n",
|
209 |
+
"\n",
|
210 |
+
" image_submit.click(\n",
|
211 |
+
" fn=generate_image,\n",
|
212 |
+
" inputs=[image_query, image_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],\n",
|
213 |
+
" outputs=output\n",
|
214 |
+
" )\n",
|
215 |
+
" video_submit.click(\n",
|
216 |
+
" fn=generate_video,\n",
|
217 |
+
" inputs=[video_query, video_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],\n",
|
218 |
+
" outputs=output\n",
|
219 |
+
" )\n",
|
220 |
+
"\n",
|
221 |
+
"if __name__ == \"__main__\":\n",
|
222 |
+
" demo.queue(max_size=30).launch(share=True, ssr_mode=False, show_error=True)"
|
223 |
+
]
|
224 |
+
}
|
225 |
+
],
|
226 |
+
"metadata": {
|
227 |
+
"accelerator": "GPU",
|
228 |
+
"colab": {
|
229 |
+
"gpuType": "T4",
|
230 |
+
"provenance": []
|
231 |
+
},
|
232 |
+
"kernelspec": {
|
233 |
+
"display_name": "Python 3",
|
234 |
+
"name": "python3"
|
235 |
+
},
|
236 |
+
"language_info": {
|
237 |
+
"name": "python"
|
238 |
+
}
|
239 |
+
},
|
240 |
+
"nbformat": 4,
|
241 |
+
"nbformat_minor": 0
|
242 |
+
}
|
Qwen-2VL-MessyOCR/Qwen2_VL_OCR_2B_Instruct_prithivmlmods.ipynb
ADDED
@@ -0,0 +1,247 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"nbformat": 4,
|
3 |
+
"nbformat_minor": 0,
|
4 |
+
"metadata": {
|
5 |
+
"colab": {
|
6 |
+
"provenance": [],
|
7 |
+
"gpuType": "T4"
|
8 |
+
},
|
9 |
+
"kernelspec": {
|
10 |
+
"name": "python3",
|
11 |
+
"display_name": "Python 3"
|
12 |
+
},
|
13 |
+
"language_info": {
|
14 |
+
"name": "python"
|
15 |
+
},
|
16 |
+
"accelerator": "GPU"
|
17 |
+
},
|
18 |
+
"cells": [
|
19 |
+
{
|
20 |
+
"cell_type": "code",
|
21 |
+
"execution_count": 1,
|
22 |
+
"metadata": {
|
23 |
+
"id": "xL8y37Y6bORU"
|
24 |
+
},
|
25 |
+
"outputs": [],
|
26 |
+
"source": [
|
27 |
+
"%%capture\n",
|
28 |
+
"!pip install gradio spaces transformers accelerate numpy requests\n",
|
29 |
+
"!pip install torch torchvision qwen-vl-utils av hf_xet\n",
|
30 |
+
"!pip install pillow huggingface_hub opencv-python"
|
31 |
+
]
|
32 |
+
},
|
33 |
+
{
|
34 |
+
"cell_type": "code",
|
35 |
+
"source": [
|
36 |
+
"import os\n",
|
37 |
+
"import time\n",
|
38 |
+
"import numpy as np\n",
|
39 |
+
"from threading import Thread\n",
|
40 |
+
"\n",
|
41 |
+
"import gradio as gr\n",
|
42 |
+
"import spaces\n",
|
43 |
+
"import torch\n",
|
44 |
+
"from PIL import Image\n",
|
45 |
+
"import cv2\n",
|
46 |
+
"\n",
|
47 |
+
"from transformers import (\n",
|
48 |
+
" Qwen2VLForConditionalGeneration,\n",
|
49 |
+
" AutoProcessor,\n",
|
50 |
+
" TextIteratorStreamer,\n",
|
51 |
+
")\n",
|
52 |
+
"\n",
|
53 |
+
"# Constants for text generation\n",
|
54 |
+
"MAX_MAX_NEW_TOKENS = 2048\n",
|
55 |
+
"DEFAULT_MAX_NEW_TOKENS = 1024\n",
|
56 |
+
"# Increase or disable input truncation to avoid token mismatches\n",
|
57 |
+
"MAX_INPUT_TOKEN_LENGTH = int(os.getenv(\"MAX_INPUT_TOKEN_LENGTH\", \"8192\"))\n",
|
58 |
+
"\n",
|
59 |
+
"device = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n",
|
60 |
+
"\n",
|
61 |
+
"MODEL_ID = \"prithivMLmods/Qwen2-VL-OCR-2B-Instruct\"\n",
|
62 |
+
"processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)\n",
|
63 |
+
"model_m = Qwen2VLForConditionalGeneration.from_pretrained(\n",
|
64 |
+
" MODEL_ID,\n",
|
65 |
+
" trust_remote_code=True,\n",
|
66 |
+
" torch_dtype=torch.float16\n",
|
67 |
+
").to(device).eval()\n",
|
68 |
+
"\n",
|
69 |
+
"def downsample_video(video_path):\n",
|
70 |
+
" \"\"\"\n",
|
71 |
+
" Downsamples the video to evenly spaced frames.\n",
|
72 |
+
" Each frame is returned as a PIL image along with its timestamp.\n",
|
73 |
+
" \"\"\"\n",
|
74 |
+
" vidcap = cv2.VideoCapture(video_path)\n",
|
75 |
+
" total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))\n",
|
76 |
+
" fps = vidcap.get(cv2.CAP_PROP_FPS)\n",
|
77 |
+
" frames = []\n",
|
78 |
+
" frame_indices = np.linspace(0, total_frames - 1, 10, dtype=int)\n",
|
79 |
+
" for i in frame_indices:\n",
|
80 |
+
" vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)\n",
|
81 |
+
" success, image = vidcap.read()\n",
|
82 |
+
" if success:\n",
|
83 |
+
" image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)\n",
|
84 |
+
" pil_image = Image.fromarray(image)\n",
|
85 |
+
" timestamp = round(i / fps, 2)\n",
|
86 |
+
" frames.append((pil_image, timestamp))\n",
|
87 |
+
" vidcap.release()\n",
|
88 |
+
" return frames\n",
|
89 |
+
"\n",
|
90 |
+
"@spaces.GPU\n",
|
91 |
+
"def generate_image(text: str, image: Image.Image,\n",
|
92 |
+
" max_new_tokens: int = DEFAULT_MAX_NEW_TOKENS,\n",
|
93 |
+
" temperature: float = 0.6,\n",
|
94 |
+
" top_p: float = 0.9,\n",
|
95 |
+
" top_k: int = 50,\n",
|
96 |
+
" repetition_penalty: float = 1.2):\n",
|
97 |
+
"\n",
|
98 |
+
" if image is None:\n",
|
99 |
+
" yield \"Please upload an image.\"\n",
|
100 |
+
" return\n",
|
101 |
+
"\n",
|
102 |
+
" messages = [{\n",
|
103 |
+
" \"role\": \"user\",\n",
|
104 |
+
" \"content\": [\n",
|
105 |
+
" {\"type\": \"image\", \"image\": image},\n",
|
106 |
+
" {\"type\": \"text\", \"text\": text},\n",
|
107 |
+
" ]\n",
|
108 |
+
" }]\n",
|
109 |
+
" prompt_full = processor.apply_chat_template(\n",
|
110 |
+
" messages, tokenize=False, add_generation_prompt=True\n",
|
111 |
+
" )\n",
|
112 |
+
" inputs = processor(\n",
|
113 |
+
" text=[prompt_full],\n",
|
114 |
+
" images=[image],\n",
|
115 |
+
" return_tensors=\"pt\",\n",
|
116 |
+
" padding=True,\n",
|
117 |
+
" truncation=False # Disable truncation to keep image tokens intact\n",
|
118 |
+
" ).to(device)\n",
|
119 |
+
"\n",
|
120 |
+
" streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)\n",
|
121 |
+
" generation_kwargs = {\n",
|
122 |
+
" **inputs,\n",
|
123 |
+
" \"streamer\": streamer,\n",
|
124 |
+
" \"max_new_tokens\": max_new_tokens,\n",
|
125 |
+
" \"do_sample\": True,\n",
|
126 |
+
" \"temperature\": temperature,\n",
|
127 |
+
" \"top_p\": top_p,\n",
|
128 |
+
" \"top_k\": top_k,\n",
|
129 |
+
" \"repetition_penalty\": repetition_penalty,\n",
|
130 |
+
" }\n",
|
131 |
+
" thread = Thread(target=model_m.generate, kwargs=generation_kwargs)\n",
|
132 |
+
" thread.start()\n",
|
133 |
+
" buffer = \"\"\n",
|
134 |
+
" for new_text in streamer:\n",
|
135 |
+
" buffer += new_text.replace(\"<|im_end|>\", \"\")\n",
|
136 |
+
" time.sleep(0.01)\n",
|
137 |
+
" yield buffer\n",
|
138 |
+
"\n",
|
139 |
+
"@spaces.GPU\n",
|
140 |
+
"def generate_video(text: str, video_path: str,\n",
|
141 |
+
" max_new_tokens: int = DEFAULT_MAX_NEW_TOKENS,\n",
|
142 |
+
" temperature: float = 0.6,\n",
|
143 |
+
" top_p: float = 0.9,\n",
|
144 |
+
" top_k: int = 50,\n",
|
145 |
+
" repetition_penalty: float = 1.2):\n",
|
146 |
+
"\n",
|
147 |
+
" if video_path is None:\n",
|
148 |
+
" yield \"Please upload a video.\"\n",
|
149 |
+
" return\n",
|
150 |
+
"\n",
|
151 |
+
" frames = downsample_video(video_path)\n",
|
152 |
+
" messages = [\n",
|
153 |
+
" {\"role\": \"system\", \"content\": [{\"type\": \"text\", \"text\": \"You are a helpful assistant.\"}]},\n",
|
154 |
+
" {\"role\": \"user\", \"content\": [{\"type\": \"text\", \"text\": text}]}\n",
|
155 |
+
" ]\n",
|
156 |
+
" for image, timestamp in frames:\n",
|
157 |
+
" messages[1][\"content\"].extend([\n",
|
158 |
+
" {\"type\": \"text\", \"text\": f\"Frame {timestamp}:\"},\n",
|
159 |
+
" {\"type\": \"image\", \"image\": image}\n",
|
160 |
+
" ])\n",
|
161 |
+
"\n",
|
162 |
+
" # Use chat template with no truncation\n",
|
163 |
+
" inputs = processor.apply_chat_template(\n",
|
164 |
+
" messages,\n",
|
165 |
+
" tokenize=True,\n",
|
166 |
+
" add_generation_prompt=True,\n",
|
167 |
+
" return_dict=True,\n",
|
168 |
+
" return_tensors=\"pt\",\n",
|
169 |
+
" truncation=False\n",
|
170 |
+
" ).to(device)\n",
|
171 |
+
"\n",
|
172 |
+
" streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)\n",
|
173 |
+
" generation_kwargs = {\n",
|
174 |
+
" **inputs,\n",
|
175 |
+
" \"streamer\": streamer,\n",
|
176 |
+
" \"max_new_tokens\": max_new_tokens,\n",
|
177 |
+
" \"do_sample\": True,\n",
|
178 |
+
" \"temperature\": temperature,\n",
|
179 |
+
" \"top_p\": top_p,\n",
|
180 |
+
" \"top_k\": top_k,\n",
|
181 |
+
" \"repetition_penalty\": repetition_penalty,\n",
|
182 |
+
" }\n",
|
183 |
+
" thread = Thread(target=model_m.generate, kwargs=generation_kwargs)\n",
|
184 |
+
" thread.start()\n",
|
185 |
+
" buffer = \"\"\n",
|
186 |
+
" for new_text in streamer:\n",
|
187 |
+
" buffer += new_text.replace(\"<|im_end|>\", \"\")\n",
|
188 |
+
" time.sleep(0.01)\n",
|
189 |
+
" yield buffer\n",
|
190 |
+
"\n",
|
191 |
+
"# Gradio App Style and Layout\n",
|
192 |
+
"css = \"\"\"\n",
|
193 |
+
".submit-btn {\n",
|
194 |
+
" background-color: #2980b9 !important;\n",
|
195 |
+
" color: white !important;\n",
|
196 |
+
"}\n",
|
197 |
+
".submit-btn:hover {\n",
|
198 |
+
" background-color: #3498db !important;\n",
|
199 |
+
"}\n",
|
200 |
+
"\"\"\"\n",
|
201 |
+
"\n",
|
202 |
+
"with gr.Blocks(css=css, theme=\"bethecloud/storj_theme\") as demo:\n",
|
203 |
+
" gr.Markdown(\"# **prithivMLmods/Qwen2-VL-OCR-2B-Instruct**\")\n",
|
204 |
+
" with gr.Row():\n",
|
205 |
+
" with gr.Column():\n",
|
206 |
+
" with gr.Tabs():\n",
|
207 |
+
" with gr.TabItem(\"Image Inference\"):\n",
|
208 |
+
" image_query = gr.Textbox(label=\"Query Input\", placeholder=\"Enter your query here...\")\n",
|
209 |
+
" image_upload = gr.Image(type=\"pil\", label=\"Image\")\n",
|
210 |
+
" image_submit = gr.Button(\"Submit\", elem_classes=\"submit-btn\")\n",
|
211 |
+
"\n",
|
212 |
+
" with gr.TabItem(\"Video Inference\"):\n",
|
213 |
+
" video_query = gr.Textbox(label=\"Query Input\", placeholder=\"Enter your query here...\")\n",
|
214 |
+
" video_upload = gr.Video(label=\"Video\")\n",
|
215 |
+
" video_submit = gr.Button(\"Submit\", elem_classes=\"submit-btn\")\n",
|
216 |
+
"\n",
|
217 |
+
" with gr.Accordion(\"Advanced options\", open=False):\n",
|
218 |
+
" max_new_tokens = gr.Slider(label=\"Max new tokens\", minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1, value=DEFAULT_MAX_NEW_TOKENS)\n",
|
219 |
+
" temperature = gr.Slider(label=\"Temperature\", minimum=0.1, maximum=4.0, step=0.1, value=0.6)\n",
|
220 |
+
" top_p = gr.Slider(label=\"Top-p (nucleus sampling)\", minimum=0.05, maximum=1.0, step=0.05, value=0.9)\n",
|
221 |
+
" top_k = gr.Slider(label=\"Top-k\", minimum=1, maximum=1000, step=1, value=50)\n",
|
222 |
+
" repetition_penalty = gr.Slider(label=\"Repetition penalty\", minimum=1.0, maximum=2.0, step=0.05, value=1.2)\n",
|
223 |
+
" with gr.Column():\n",
|
224 |
+
" output = gr.Textbox(label=\"Output\", interactive=False)\n",
|
225 |
+
"\n",
|
226 |
+
" image_submit.click(\n",
|
227 |
+
" fn=generate_image,\n",
|
228 |
+
" inputs=[image_query, image_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],\n",
|
229 |
+
" outputs=output\n",
|
230 |
+
" )\n",
|
231 |
+
" video_submit.click(\n",
|
232 |
+
" fn=generate_video,\n",
|
233 |
+
" inputs=[video_query, video_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],\n",
|
234 |
+
" outputs=output\n",
|
235 |
+
" )\n",
|
236 |
+
"\n",
|
237 |
+
"if __name__ == \"__main__\":\n",
|
238 |
+
" demo.queue(max_size=30).launch(share=True, ssr_mode=False, show_error=True)\n"
|
239 |
+
],
|
240 |
+
"metadata": {
|
241 |
+
"id": "Y-NTbL1tdL9X"
|
242 |
+
},
|
243 |
+
"execution_count": null,
|
244 |
+
"outputs": []
|
245 |
+
}
|
246 |
+
]
|
247 |
+
}
|
Qwen2-VL/Qwen2_VL_2B_Instruct.ipynb
ADDED
@@ -0,0 +1,242 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cells": [
|
3 |
+
{
|
4 |
+
"cell_type": "code",
|
5 |
+
"execution_count": null,
|
6 |
+
"metadata": {
|
7 |
+
"id": "xL8y37Y6bORU"
|
8 |
+
},
|
9 |
+
"outputs": [],
|
10 |
+
"source": [
|
11 |
+
"%%capture\n",
|
12 |
+
"!pip install gradio spaces transformers accelerate numpy requests\n",
|
13 |
+
"!pip install torch torchvision qwen-vl-utils av hf_xet\n",
|
14 |
+
"!pip install pillow huggingface_hub opencv-python"
|
15 |
+
]
|
16 |
+
},
|
17 |
+
{
|
18 |
+
"cell_type": "code",
|
19 |
+
"execution_count": null,
|
20 |
+
"metadata": {
|
21 |
+
"id": "Y-NTbL1tdL9X"
|
22 |
+
},
|
23 |
+
"outputs": [],
|
24 |
+
"source": [
|
25 |
+
"import os\n",
|
26 |
+
"import random\n",
|
27 |
+
"import uuid\n",
|
28 |
+
"import json\n",
|
29 |
+
"import time\n",
|
30 |
+
"import asyncio\n",
|
31 |
+
"from threading import Thread\n",
|
32 |
+
"\n",
|
33 |
+
"import gradio as gr\n",
|
34 |
+
"import spaces\n",
|
35 |
+
"import torch\n",
|
36 |
+
"import numpy as np\n",
|
37 |
+
"from PIL import Image\n",
|
38 |
+
"import cv2\n",
|
39 |
+
"\n",
|
40 |
+
"from transformers import (\n",
|
41 |
+
" Qwen2VLForConditionalGeneration,\n",
|
42 |
+
" AutoProcessor,\n",
|
43 |
+
" TextIteratorStreamer,\n",
|
44 |
+
")\n",
|
45 |
+
"from transformers.image_utils import load_image\n",
|
46 |
+
"\n",
|
47 |
+
"# Constants for text generation\n",
|
48 |
+
"MAX_MAX_NEW_TOKENS = 2048\n",
|
49 |
+
"DEFAULT_MAX_NEW_TOKENS = 1024\n",
|
50 |
+
"# Increase or disable input truncation to avoid token mismatches\n",
|
51 |
+
"MAX_INPUT_TOKEN_LENGTH = int(os.getenv(\"MAX_INPUT_TOKEN_LENGTH\", \"8192\"))\n",
|
52 |
+
"\n",
|
53 |
+
"device = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n",
|
54 |
+
"\n",
|
55 |
+
"MODEL_ID = \"Qwen/Qwen2-VL-2B-Instruct\"\n",
|
56 |
+
"processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)\n",
|
57 |
+
"model_m = Qwen2VLForConditionalGeneration.from_pretrained(\n",
|
58 |
+
" MODEL_ID,\n",
|
59 |
+
" trust_remote_code=True,\n",
|
60 |
+
" torch_dtype=torch.float16\n",
|
61 |
+
").to(\"cuda\").eval()\n",
|
62 |
+
"\n",
|
63 |
+
"def downsample_video(video_path):\n",
|
64 |
+
" \"\"\"\n",
|
65 |
+
" Downsamples the video to evenly spaced frames.\n",
|
66 |
+
" Each frame is returned as a PIL image along with its timestamp.\n",
|
67 |
+
" \"\"\"\n",
|
68 |
+
" vidcap = cv2.VideoCapture(video_path)\n",
|
69 |
+
" total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))\n",
|
70 |
+
" fps = vidcap.get(cv2.CAP_PROP_FPS)\n",
|
71 |
+
" frames = []\n",
|
72 |
+
" # Sample 10 evenly spaced frames.\n",
|
73 |
+
" frame_indices = np.linspace(0, total_frames - 1, 10, dtype=int)\n",
|
74 |
+
" for i in frame_indices:\n",
|
75 |
+
" vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)\n",
|
76 |
+
" success, image = vidcap.read()\n",
|
77 |
+
" if success:\n",
|
78 |
+
" image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) # Convert BGR to RGB\n",
|
79 |
+
" pil_image = Image.fromarray(image)\n",
|
80 |
+
" timestamp = round(i / fps, 2)\n",
|
81 |
+
" frames.append((pil_image, timestamp))\n",
|
82 |
+
" vidcap.release()\n",
|
83 |
+
" return frames\n",
|
84 |
+
"\n",
|
85 |
+
"@spaces.GPU\n",
|
86 |
+
"def generate_image(text: str, image: Image.Image,\n",
|
87 |
+
" max_new_tokens: int = 1024,\n",
|
88 |
+
" temperature: float = 0.6,\n",
|
89 |
+
" top_p: float = 0.9,\n",
|
90 |
+
" top_k: int = 50,\n",
|
91 |
+
" repetition_penalty: float = 1.2):\n",
|
92 |
+
"\n",
|
93 |
+
" if image is None:\n",
|
94 |
+
" yield \"Please upload an image.\"\n",
|
95 |
+
" return\n",
|
96 |
+
"\n",
|
97 |
+
" messages = [{\n",
|
98 |
+
" \"role\": \"user\",\n",
|
99 |
+
" \"content\": [\n",
|
100 |
+
" {\"type\": \"image\", \"image\": image},\n",
|
101 |
+
" {\"type\": \"text\", \"text\": text},\n",
|
102 |
+
" ]\n",
|
103 |
+
" }]\n",
|
104 |
+
" prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)\n",
|
105 |
+
" inputs = processor(\n",
|
106 |
+
" text=[prompt_full],\n",
|
107 |
+
" images=[image],\n",
|
108 |
+
" return_tensors=\"pt\",\n",
|
109 |
+
" padding=True,\n",
|
110 |
+
" truncation=False,\n",
|
111 |
+
" max_length=MAX_INPUT_TOKEN_LENGTH\n",
|
112 |
+
" ).to(\"cuda\")\n",
|
113 |
+
" streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)\n",
|
114 |
+
" generation_kwargs = {**inputs, \"streamer\": streamer, \"max_new_tokens\": max_new_tokens}\n",
|
115 |
+
" thread = Thread(target=model_m.generate, kwargs=generation_kwargs)\n",
|
116 |
+
" thread.start()\n",
|
117 |
+
" buffer = \"\"\n",
|
118 |
+
" for new_text in streamer:\n",
|
119 |
+
" buffer += new_text\n",
|
120 |
+
" buffer = buffer.replace(\"<|im_end|>\", \"\")\n",
|
121 |
+
" time.sleep(0.01)\n",
|
122 |
+
" yield buffer\n",
|
123 |
+
"\n",
|
124 |
+
"@spaces.GPU\n",
|
125 |
+
"def generate_video(text: str, video_path: str,\n",
|
126 |
+
" max_new_tokens: int = 1024,\n",
|
127 |
+
" temperature: float = 0.6,\n",
|
128 |
+
" top_p: float = 0.9,\n",
|
129 |
+
" top_k: int = 50,\n",
|
130 |
+
" repetition_penalty: float = 1.2):\n",
|
131 |
+
"\n",
|
132 |
+
" if video_path is None:\n",
|
133 |
+
" yield \"Please upload a video.\"\n",
|
134 |
+
" return\n",
|
135 |
+
"\n",
|
136 |
+
" frames = downsample_video(video_path)\n",
|
137 |
+
" messages = [\n",
|
138 |
+
" {\"role\": \"system\", \"content\": [{\"type\": \"text\", \"text\": \"You are a helpful assistant.\"}]},\n",
|
139 |
+
" {\"role\": \"user\", \"content\": [{\"type\": \"text\", \"text\": text}]}\n",
|
140 |
+
" ]\n",
|
141 |
+
" # Append each frame with its timestamp.\n",
|
142 |
+
" for frame in frames:\n",
|
143 |
+
" image, timestamp = frame\n",
|
144 |
+
" messages[1][\"content\"].append({\"type\": \"text\", \"text\": f\"Frame {timestamp}:\"})\n",
|
145 |
+
" messages[1][\"content\"].append({\"type\": \"image\", \"image\": image})\n",
|
146 |
+
" inputs = processor.apply_chat_template(\n",
|
147 |
+
" messages,\n",
|
148 |
+
" tokenize=True,\n",
|
149 |
+
" add_generation_prompt=True,\n",
|
150 |
+
" return_dict=True,\n",
|
151 |
+
" return_tensors=\"pt\",\n",
|
152 |
+
" truncation=False,\n",
|
153 |
+
" max_length=MAX_INPUT_TOKEN_LENGTH\n",
|
154 |
+
" ).to(\"cuda\")\n",
|
155 |
+
" streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)\n",
|
156 |
+
" generation_kwargs = {\n",
|
157 |
+
" **inputs,\n",
|
158 |
+
" \"streamer\": streamer,\n",
|
159 |
+
" \"max_new_tokens\": max_new_tokens,\n",
|
160 |
+
" \"do_sample\": True,\n",
|
161 |
+
" \"temperature\": temperature,\n",
|
162 |
+
" \"top_p\": top_p,\n",
|
163 |
+
" \"top_k\": top_k,\n",
|
164 |
+
" \"repetition_penalty\": repetition_penalty,\n",
|
165 |
+
" }\n",
|
166 |
+
" thread = Thread(target=model_m.generate, kwargs=generation_kwargs)\n",
|
167 |
+
" thread.start()\n",
|
168 |
+
" buffer = \"\"\n",
|
169 |
+
" for new_text in streamer:\n",
|
170 |
+
" buffer += new_text\n",
|
171 |
+
" buffer = buffer.replace(\"<|im_end|>\", \"\")\n",
|
172 |
+
" time.sleep(0.01)\n",
|
173 |
+
" yield buffer\n",
|
174 |
+
"\n",
|
175 |
+
"css = \"\"\"\n",
|
176 |
+
".submit-btn {\n",
|
177 |
+
" background-color: #2980b9 !important;\n",
|
178 |
+
" color: white !important;\n",
|
179 |
+
"}\n",
|
180 |
+
".submit-btn:hover {\n",
|
181 |
+
" background-color: #3498db !important;\n",
|
182 |
+
"}\n",
|
183 |
+
"\"\"\"\n",
|
184 |
+
"\n",
|
185 |
+
"# Create the Gradio Interface\n",
|
186 |
+
"with gr.Blocks(css=css, theme=\"bethecloud/storj_theme\") as demo:\n",
|
187 |
+
" gr.Markdown(\"# **Qwen/Qwen2-VL-2B-Instruct**\")\n",
|
188 |
+
" with gr.Row():\n",
|
189 |
+
" with gr.Column():\n",
|
190 |
+
" with gr.Tabs():\n",
|
191 |
+
" with gr.TabItem(\"Image Inference\"):\n",
|
192 |
+
" image_query = gr.Textbox(label=\"Query Input\", placeholder=\"Enter your query here...\")\n",
|
193 |
+
" image_upload = gr.Image(type=\"pil\", label=\"Image\")\n",
|
194 |
+
" image_submit = gr.Button(\"Submit\", elem_classes=\"submit-btn\")\n",
|
195 |
+
"\n",
|
196 |
+
" with gr.TabItem(\"Video Inference\"):\n",
|
197 |
+
" video_query = gr.Textbox(label=\"Query Input\", placeholder=\"Enter your query here...\")\n",
|
198 |
+
" video_upload = gr.Video(label=\"Video\")\n",
|
199 |
+
" video_submit = gr.Button(\"Submit\", elem_classes=\"submit-btn\")\n",
|
200 |
+
"\n",
|
201 |
+
" with gr.Accordion(\"Advanced options\", open=False):\n",
|
202 |
+
" max_new_tokens = gr.Slider(label=\"Max new tokens\", minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1, value=DEFAULT_MAX_NEW_TOKENS)\n",
|
203 |
+
" temperature = gr.Slider(label=\"Temperature\", minimum=0.1, maximum=4.0, step=0.1, value=0.6)\n",
|
204 |
+
" top_p = gr.Slider(label=\"Top-p (nucleus sampling)\", minimum=0.05, maximum=1.0, step=0.05, value=0.9)\n",
|
205 |
+
" top_k = gr.Slider(label=\"Top-k\", minimum=1, maximum=1000, step=1, value=50)\n",
|
206 |
+
" repetition_penalty = gr.Slider(label=\"Repetition penalty\", minimum=1.0, maximum=2.0, step=0.05, value=1.2)\n",
|
207 |
+
" with gr.Column():\n",
|
208 |
+
" output = gr.Textbox(label=\"Output\", interactive=False)\n",
|
209 |
+
"\n",
|
210 |
+
" image_submit.click(\n",
|
211 |
+
" fn=generate_image,\n",
|
212 |
+
" inputs=[image_query, image_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],\n",
|
213 |
+
" outputs=output\n",
|
214 |
+
" )\n",
|
215 |
+
" video_submit.click(\n",
|
216 |
+
" fn=generate_video,\n",
|
217 |
+
" inputs=[video_query, video_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],\n",
|
218 |
+
" outputs=output\n",
|
219 |
+
" )\n",
|
220 |
+
"\n",
|
221 |
+
"if __name__ == \"__main__\":\n",
|
222 |
+
" demo.queue(max_size=30).launch(share=True, ssr_mode=False, show_error=True)"
|
223 |
+
]
|
224 |
+
}
|
225 |
+
],
|
226 |
+
"metadata": {
|
227 |
+
"accelerator": "GPU",
|
228 |
+
"colab": {
|
229 |
+
"gpuType": "T4",
|
230 |
+
"provenance": []
|
231 |
+
},
|
232 |
+
"kernelspec": {
|
233 |
+
"display_name": "Python 3",
|
234 |
+
"name": "python3"
|
235 |
+
},
|
236 |
+
"language_info": {
|
237 |
+
"name": "python"
|
238 |
+
}
|
239 |
+
},
|
240 |
+
"nbformat": 4,
|
241 |
+
"nbformat_minor": 0
|
242 |
+
}
|
Qwen2-VL/Qwen2_VL_7B_Instruct.ipynb
ADDED
@@ -0,0 +1,242 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cells": [
|
3 |
+
{
|
4 |
+
"cell_type": "code",
|
5 |
+
"execution_count": null,
|
6 |
+
"metadata": {
|
7 |
+
"id": "xL8y37Y6bORU"
|
8 |
+
},
|
9 |
+
"outputs": [],
|
10 |
+
"source": [
|
11 |
+
"%%capture\n",
|
12 |
+
"!pip install gradio spaces transformers accelerate numpy requests\n",
|
13 |
+
"!pip install torch torchvision qwen-vl-utils av hf_xet\n",
|
14 |
+
"!pip install pillow huggingface_hub opencv-python"
|
15 |
+
]
|
16 |
+
},
|
17 |
+
{
|
18 |
+
"cell_type": "code",
|
19 |
+
"execution_count": null,
|
20 |
+
"metadata": {
|
21 |
+
"id": "Y-NTbL1tdL9X"
|
22 |
+
},
|
23 |
+
"outputs": [],
|
24 |
+
"source": [
|
25 |
+
"import os\n",
|
26 |
+
"import random\n",
|
27 |
+
"import uuid\n",
|
28 |
+
"import json\n",
|
29 |
+
"import time\n",
|
30 |
+
"import asyncio\n",
|
31 |
+
"from threading import Thread\n",
|
32 |
+
"\n",
|
33 |
+
"import gradio as gr\n",
|
34 |
+
"import spaces\n",
|
35 |
+
"import torch\n",
|
36 |
+
"import numpy as np\n",
|
37 |
+
"from PIL import Image\n",
|
38 |
+
"import cv2\n",
|
39 |
+
"\n",
|
40 |
+
"from transformers import (\n",
|
41 |
+
" Qwen2VLForConditionalGeneration,\n",
|
42 |
+
" AutoProcessor,\n",
|
43 |
+
" TextIteratorStreamer,\n",
|
44 |
+
")\n",
|
45 |
+
"from transformers.image_utils import load_image\n",
|
46 |
+
"\n",
|
47 |
+
"# Constants for text generation\n",
|
48 |
+
"MAX_MAX_NEW_TOKENS = 2048\n",
|
49 |
+
"DEFAULT_MAX_NEW_TOKENS = 1024\n",
|
50 |
+
"# Increase or disable input truncation to avoid token mismatches\n",
|
51 |
+
"MAX_INPUT_TOKEN_LENGTH = int(os.getenv(\"MAX_INPUT_TOKEN_LENGTH\", \"8192\"))\n",
|
52 |
+
"\n",
|
53 |
+
"device = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n",
|
54 |
+
"\n",
|
55 |
+
"MODEL_ID = \"Qwen/Qwen2-VL-7B-Instruct\"\n",
|
56 |
+
"processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)\n",
|
57 |
+
"model_m = Qwen2VLForConditionalGeneration.from_pretrained(\n",
|
58 |
+
" MODEL_ID,\n",
|
59 |
+
" trust_remote_code=True,\n",
|
60 |
+
" torch_dtype=torch.float16\n",
|
61 |
+
").to(\"cuda\").eval()\n",
|
62 |
+
"\n",
|
63 |
+
"def downsample_video(video_path):\n",
|
64 |
+
" \"\"\"\n",
|
65 |
+
" Downsamples the video to evenly spaced frames.\n",
|
66 |
+
" Each frame is returned as a PIL image along with its timestamp.\n",
|
67 |
+
" \"\"\"\n",
|
68 |
+
" vidcap = cv2.VideoCapture(video_path)\n",
|
69 |
+
" total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))\n",
|
70 |
+
" fps = vidcap.get(cv2.CAP_PROP_FPS)\n",
|
71 |
+
" frames = []\n",
|
72 |
+
" # Sample 10 evenly spaced frames.\n",
|
73 |
+
" frame_indices = np.linspace(0, total_frames - 1, 10, dtype=int)\n",
|
74 |
+
" for i in frame_indices:\n",
|
75 |
+
" vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)\n",
|
76 |
+
" success, image = vidcap.read()\n",
|
77 |
+
" if success:\n",
|
78 |
+
" image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) # Convert BGR to RGB\n",
|
79 |
+
" pil_image = Image.fromarray(image)\n",
|
80 |
+
" timestamp = round(i / fps, 2)\n",
|
81 |
+
" frames.append((pil_image, timestamp))\n",
|
82 |
+
" vidcap.release()\n",
|
83 |
+
" return frames\n",
|
84 |
+
"\n",
|
85 |
+
"@spaces.GPU\n",
|
86 |
+
"def generate_image(text: str, image: Image.Image,\n",
|
87 |
+
" max_new_tokens: int = 1024,\n",
|
88 |
+
" temperature: float = 0.6,\n",
|
89 |
+
" top_p: float = 0.9,\n",
|
90 |
+
" top_k: int = 50,\n",
|
91 |
+
" repetition_penalty: float = 1.2):\n",
|
92 |
+
"\n",
|
93 |
+
" if image is None:\n",
|
94 |
+
" yield \"Please upload an image.\"\n",
|
95 |
+
" return\n",
|
96 |
+
"\n",
|
97 |
+
" messages = [{\n",
|
98 |
+
" \"role\": \"user\",\n",
|
99 |
+
" \"content\": [\n",
|
100 |
+
" {\"type\": \"image\", \"image\": image},\n",
|
101 |
+
" {\"type\": \"text\", \"text\": text},\n",
|
102 |
+
" ]\n",
|
103 |
+
" }]\n",
|
104 |
+
" prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)\n",
|
105 |
+
" inputs = processor(\n",
|
106 |
+
" text=[prompt_full],\n",
|
107 |
+
" images=[image],\n",
|
108 |
+
" return_tensors=\"pt\",\n",
|
109 |
+
" padding=True,\n",
|
110 |
+
" truncation=False,\n",
|
111 |
+
" max_length=MAX_INPUT_TOKEN_LENGTH\n",
|
112 |
+
" ).to(\"cuda\")\n",
|
113 |
+
" streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)\n",
|
114 |
+
" generation_kwargs = {**inputs, \"streamer\": streamer, \"max_new_tokens\": max_new_tokens}\n",
|
115 |
+
" thread = Thread(target=model_m.generate, kwargs=generation_kwargs)\n",
|
116 |
+
" thread.start()\n",
|
117 |
+
" buffer = \"\"\n",
|
118 |
+
" for new_text in streamer:\n",
|
119 |
+
" buffer += new_text\n",
|
120 |
+
" buffer = buffer.replace(\"<|im_end|>\", \"\")\n",
|
121 |
+
" time.sleep(0.01)\n",
|
122 |
+
" yield buffer\n",
|
123 |
+
"\n",
|
124 |
+
"@spaces.GPU\n",
|
125 |
+
"def generate_video(text: str, video_path: str,\n",
|
126 |
+
" max_new_tokens: int = 1024,\n",
|
127 |
+
" temperature: float = 0.6,\n",
|
128 |
+
" top_p: float = 0.9,\n",
|
129 |
+
" top_k: int = 50,\n",
|
130 |
+
" repetition_penalty: float = 1.2):\n",
|
131 |
+
"\n",
|
132 |
+
" if video_path is None:\n",
|
133 |
+
" yield \"Please upload a video.\"\n",
|
134 |
+
" return\n",
|
135 |
+
"\n",
|
136 |
+
" frames = downsample_video(video_path)\n",
|
137 |
+
" messages = [\n",
|
138 |
+
" {\"role\": \"system\", \"content\": [{\"type\": \"text\", \"text\": \"You are a helpful assistant.\"}]},\n",
|
139 |
+
" {\"role\": \"user\", \"content\": [{\"type\": \"text\", \"text\": text}]}\n",
|
140 |
+
" ]\n",
|
141 |
+
" # Append each frame with its timestamp.\n",
|
142 |
+
" for frame in frames:\n",
|
143 |
+
" image, timestamp = frame\n",
|
144 |
+
" messages[1][\"content\"].append({\"type\": \"text\", \"text\": f\"Frame {timestamp}:\"})\n",
|
145 |
+
" messages[1][\"content\"].append({\"type\": \"image\", \"image\": image})\n",
|
146 |
+
" inputs = processor.apply_chat_template(\n",
|
147 |
+
" messages,\n",
|
148 |
+
" tokenize=True,\n",
|
149 |
+
" add_generation_prompt=True,\n",
|
150 |
+
" return_dict=True,\n",
|
151 |
+
" return_tensors=\"pt\",\n",
|
152 |
+
" truncation=False,\n",
|
153 |
+
" max_length=MAX_INPUT_TOKEN_LENGTH\n",
|
154 |
+
" ).to(\"cuda\")\n",
|
155 |
+
" streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)\n",
|
156 |
+
" generation_kwargs = {\n",
|
157 |
+
" **inputs,\n",
|
158 |
+
" \"streamer\": streamer,\n",
|
159 |
+
" \"max_new_tokens\": max_new_tokens,\n",
|
160 |
+
" \"do_sample\": True,\n",
|
161 |
+
" \"temperature\": temperature,\n",
|
162 |
+
" \"top_p\": top_p,\n",
|
163 |
+
" \"top_k\": top_k,\n",
|
164 |
+
" \"repetition_penalty\": repetition_penalty,\n",
|
165 |
+
" }\n",
|
166 |
+
" thread = Thread(target=model_m.generate, kwargs=generation_kwargs)\n",
|
167 |
+
" thread.start()\n",
|
168 |
+
" buffer = \"\"\n",
|
169 |
+
" for new_text in streamer:\n",
|
170 |
+
" buffer += new_text\n",
|
171 |
+
" buffer = buffer.replace(\"<|im_end|>\", \"\")\n",
|
172 |
+
" time.sleep(0.01)\n",
|
173 |
+
" yield buffer\n",
|
174 |
+
"\n",
|
175 |
+
"css = \"\"\"\n",
|
176 |
+
".submit-btn {\n",
|
177 |
+
" background-color: #2980b9 !important;\n",
|
178 |
+
" color: white !important;\n",
|
179 |
+
"}\n",
|
180 |
+
".submit-btn:hover {\n",
|
181 |
+
" background-color: #3498db !important;\n",
|
182 |
+
"}\n",
|
183 |
+
"\"\"\"\n",
|
184 |
+
"\n",
|
185 |
+
"# Create the Gradio Interface\n",
|
186 |
+
"with gr.Blocks(css=css, theme=\"bethecloud/storj_theme\") as demo:\n",
|
187 |
+
" gr.Markdown(\"# **Qwen/Qwen2-VL-7B-Instruct**\")\n",
|
188 |
+
" with gr.Row():\n",
|
189 |
+
" with gr.Column():\n",
|
190 |
+
" with gr.Tabs():\n",
|
191 |
+
" with gr.TabItem(\"Image Inference\"):\n",
|
192 |
+
" image_query = gr.Textbox(label=\"Query Input\", placeholder=\"Enter your query here...\")\n",
|
193 |
+
" image_upload = gr.Image(type=\"pil\", label=\"Image\")\n",
|
194 |
+
" image_submit = gr.Button(\"Submit\", elem_classes=\"submit-btn\")\n",
|
195 |
+
"\n",
|
196 |
+
" with gr.TabItem(\"Video Inference\"):\n",
|
197 |
+
" video_query = gr.Textbox(label=\"Query Input\", placeholder=\"Enter your query here...\")\n",
|
198 |
+
" video_upload = gr.Video(label=\"Video\")\n",
|
199 |
+
" video_submit = gr.Button(\"Submit\", elem_classes=\"submit-btn\")\n",
|
200 |
+
"\n",
|
201 |
+
" with gr.Accordion(\"Advanced options\", open=False):\n",
|
202 |
+
" max_new_tokens = gr.Slider(label=\"Max new tokens\", minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1, value=DEFAULT_MAX_NEW_TOKENS)\n",
|
203 |
+
" temperature = gr.Slider(label=\"Temperature\", minimum=0.1, maximum=4.0, step=0.1, value=0.6)\n",
|
204 |
+
" top_p = gr.Slider(label=\"Top-p (nucleus sampling)\", minimum=0.05, maximum=1.0, step=0.05, value=0.9)\n",
|
205 |
+
" top_k = gr.Slider(label=\"Top-k\", minimum=1, maximum=1000, step=1, value=50)\n",
|
206 |
+
" repetition_penalty = gr.Slider(label=\"Repetition penalty\", minimum=1.0, maximum=2.0, step=0.05, value=1.2)\n",
|
207 |
+
" with gr.Column():\n",
|
208 |
+
" output = gr.Textbox(label=\"Output\", interactive=False)\n",
|
209 |
+
"\n",
|
210 |
+
" image_submit.click(\n",
|
211 |
+
" fn=generate_image,\n",
|
212 |
+
" inputs=[image_query, image_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],\n",
|
213 |
+
" outputs=output\n",
|
214 |
+
" )\n",
|
215 |
+
" video_submit.click(\n",
|
216 |
+
" fn=generate_video,\n",
|
217 |
+
" inputs=[video_query, video_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],\n",
|
218 |
+
" outputs=output\n",
|
219 |
+
" )\n",
|
220 |
+
"\n",
|
221 |
+
"if __name__ == \"__main__\":\n",
|
222 |
+
" demo.queue(max_size=30).launch(share=True, ssr_mode=False, show_error=True)"
|
223 |
+
]
|
224 |
+
}
|
225 |
+
],
|
226 |
+
"metadata": {
|
227 |
+
"accelerator": "GPU",
|
228 |
+
"colab": {
|
229 |
+
"gpuType": "T4",
|
230 |
+
"provenance": []
|
231 |
+
},
|
232 |
+
"kernelspec": {
|
233 |
+
"display_name": "Python 3",
|
234 |
+
"name": "python3"
|
235 |
+
},
|
236 |
+
"language_info": {
|
237 |
+
"name": "python"
|
238 |
+
}
|
239 |
+
},
|
240 |
+
"nbformat": 4,
|
241 |
+
"nbformat_minor": 0
|
242 |
+
}
|
Qwen2.5-VL/Qwen2_5VL_3B.ipynb
ADDED
@@ -0,0 +1,242 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cells": [
|
3 |
+
{
|
4 |
+
"cell_type": "code",
|
5 |
+
"execution_count": null,
|
6 |
+
"metadata": {
|
7 |
+
"id": "xL8y37Y6bORU"
|
8 |
+
},
|
9 |
+
"outputs": [],
|
10 |
+
"source": [
|
11 |
+
"%%capture\n",
|
12 |
+
"!pip install gradio spaces transformers accelerate numpy requests\n",
|
13 |
+
"!pip install torch torchvision qwen-vl-utils av hf_xet\n",
|
14 |
+
"!pip install pillow huggingface_hub opencv-python"
|
15 |
+
]
|
16 |
+
},
|
17 |
+
{
|
18 |
+
"cell_type": "code",
|
19 |
+
"execution_count": null,
|
20 |
+
"metadata": {
|
21 |
+
"id": "Y-NTbL1tdL9X"
|
22 |
+
},
|
23 |
+
"outputs": [],
|
24 |
+
"source": [
|
25 |
+
"import os\n",
|
26 |
+
"import random\n",
|
27 |
+
"import uuid\n",
|
28 |
+
"import json\n",
|
29 |
+
"import time\n",
|
30 |
+
"import asyncio\n",
|
31 |
+
"from threading import Thread\n",
|
32 |
+
"\n",
|
33 |
+
"import gradio as gr\n",
|
34 |
+
"import spaces\n",
|
35 |
+
"import torch\n",
|
36 |
+
"import numpy as np\n",
|
37 |
+
"from PIL import Image\n",
|
38 |
+
"import cv2\n",
|
39 |
+
"\n",
|
40 |
+
"from transformers import (\n",
|
41 |
+
" Qwen2_5_VLForConditionalGeneration,\n",
|
42 |
+
" AutoProcessor,\n",
|
43 |
+
" TextIteratorStreamer,\n",
|
44 |
+
")\n",
|
45 |
+
"from transformers.image_utils import load_image\n",
|
46 |
+
"\n",
|
47 |
+
"# Constants for text generation\n",
|
48 |
+
"MAX_MAX_NEW_TOKENS = 2048\n",
|
49 |
+
"DEFAULT_MAX_NEW_TOKENS = 1024\n",
|
50 |
+
"# Increase or disable input truncation to avoid token mismatches\n",
|
51 |
+
"MAX_INPUT_TOKEN_LENGTH = int(os.getenv(\"MAX_INPUT_TOKEN_LENGTH\", \"8192\"))\n",
|
52 |
+
"\n",
|
53 |
+
"device = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n",
|
54 |
+
"\n",
|
55 |
+
"MODEL_ID = \"Qwen/Qwen2.5-VL-3B-Instruct\"\n",
|
56 |
+
"processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)\n",
|
57 |
+
"model_m = Qwen2_5_VLForConditionalGeneration.from_pretrained(\n",
|
58 |
+
" MODEL_ID,\n",
|
59 |
+
" trust_remote_code=True,\n",
|
60 |
+
" torch_dtype=torch.float16\n",
|
61 |
+
").to(\"cuda\").eval()\n",
|
62 |
+
"\n",
|
63 |
+
"def downsample_video(video_path):\n",
|
64 |
+
" \"\"\"\n",
|
65 |
+
" Downsamples the video to evenly spaced frames.\n",
|
66 |
+
" Each frame is returned as a PIL image along with its timestamp.\n",
|
67 |
+
" \"\"\"\n",
|
68 |
+
" vidcap = cv2.VideoCapture(video_path)\n",
|
69 |
+
" total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))\n",
|
70 |
+
" fps = vidcap.get(cv2.CAP_PROP_FPS)\n",
|
71 |
+
" frames = []\n",
|
72 |
+
" # Sample 10 evenly spaced frames.\n",
|
73 |
+
" frame_indices = np.linspace(0, total_frames - 1, 10, dtype=int)\n",
|
74 |
+
" for i in frame_indices:\n",
|
75 |
+
" vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)\n",
|
76 |
+
" success, image = vidcap.read()\n",
|
77 |
+
" if success:\n",
|
78 |
+
" image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) # Convert BGR to RGB\n",
|
79 |
+
" pil_image = Image.fromarray(image)\n",
|
80 |
+
" timestamp = round(i / fps, 2)\n",
|
81 |
+
" frames.append((pil_image, timestamp))\n",
|
82 |
+
" vidcap.release()\n",
|
83 |
+
" return frames\n",
|
84 |
+
"\n",
|
85 |
+
"@spaces.GPU\n",
|
86 |
+
"def generate_image(text: str, image: Image.Image,\n",
|
87 |
+
" max_new_tokens: int = 1024,\n",
|
88 |
+
" temperature: float = 0.6,\n",
|
89 |
+
" top_p: float = 0.9,\n",
|
90 |
+
" top_k: int = 50,\n",
|
91 |
+
" repetition_penalty: float = 1.2):\n",
|
92 |
+
"\n",
|
93 |
+
" if image is None:\n",
|
94 |
+
" yield \"Please upload an image.\"\n",
|
95 |
+
" return\n",
|
96 |
+
"\n",
|
97 |
+
" messages = [{\n",
|
98 |
+
" \"role\": \"user\",\n",
|
99 |
+
" \"content\": [\n",
|
100 |
+
" {\"type\": \"image\", \"image\": image},\n",
|
101 |
+
" {\"type\": \"text\", \"text\": text},\n",
|
102 |
+
" ]\n",
|
103 |
+
" }]\n",
|
104 |
+
" prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)\n",
|
105 |
+
" inputs = processor(\n",
|
106 |
+
" text=[prompt_full],\n",
|
107 |
+
" images=[image],\n",
|
108 |
+
" return_tensors=\"pt\",\n",
|
109 |
+
" padding=True,\n",
|
110 |
+
" truncation=False,\n",
|
111 |
+
" max_length=MAX_INPUT_TOKEN_LENGTH\n",
|
112 |
+
" ).to(\"cuda\")\n",
|
113 |
+
" streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)\n",
|
114 |
+
" generation_kwargs = {**inputs, \"streamer\": streamer, \"max_new_tokens\": max_new_tokens}\n",
|
115 |
+
" thread = Thread(target=model_m.generate, kwargs=generation_kwargs)\n",
|
116 |
+
" thread.start()\n",
|
117 |
+
" buffer = \"\"\n",
|
118 |
+
" for new_text in streamer:\n",
|
119 |
+
" buffer += new_text\n",
|
120 |
+
" buffer = buffer.replace(\"<|im_end|>\", \"\")\n",
|
121 |
+
" time.sleep(0.01)\n",
|
122 |
+
" yield buffer\n",
|
123 |
+
"\n",
|
124 |
+
"@spaces.GPU\n",
|
125 |
+
"def generate_video(text: str, video_path: str,\n",
|
126 |
+
" max_new_tokens: int = 1024,\n",
|
127 |
+
" temperature: float = 0.6,\n",
|
128 |
+
" top_p: float = 0.9,\n",
|
129 |
+
" top_k: int = 50,\n",
|
130 |
+
" repetition_penalty: float = 1.2):\n",
|
131 |
+
"\n",
|
132 |
+
" if video_path is None:\n",
|
133 |
+
" yield \"Please upload a video.\"\n",
|
134 |
+
" return\n",
|
135 |
+
"\n",
|
136 |
+
" frames = downsample_video(video_path)\n",
|
137 |
+
" messages = [\n",
|
138 |
+
" {\"role\": \"system\", \"content\": [{\"type\": \"text\", \"text\": \"You are a helpful assistant.\"}]},\n",
|
139 |
+
" {\"role\": \"user\", \"content\": [{\"type\": \"text\", \"text\": text}]}\n",
|
140 |
+
" ]\n",
|
141 |
+
" # Append each frame with its timestamp.\n",
|
142 |
+
" for frame in frames:\n",
|
143 |
+
" image, timestamp = frame\n",
|
144 |
+
" messages[1][\"content\"].append({\"type\": \"text\", \"text\": f\"Frame {timestamp}:\"})\n",
|
145 |
+
" messages[1][\"content\"].append({\"type\": \"image\", \"image\": image})\n",
|
146 |
+
" inputs = processor.apply_chat_template(\n",
|
147 |
+
" messages,\n",
|
148 |
+
" tokenize=True,\n",
|
149 |
+
" add_generation_prompt=True,\n",
|
150 |
+
" return_dict=True,\n",
|
151 |
+
" return_tensors=\"pt\",\n",
|
152 |
+
" truncation=False,\n",
|
153 |
+
" max_length=MAX_INPUT_TOKEN_LENGTH\n",
|
154 |
+
" ).to(\"cuda\")\n",
|
155 |
+
" streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)\n",
|
156 |
+
" generation_kwargs = {\n",
|
157 |
+
" **inputs,\n",
|
158 |
+
" \"streamer\": streamer,\n",
|
159 |
+
" \"max_new_tokens\": max_new_tokens,\n",
|
160 |
+
" \"do_sample\": True,\n",
|
161 |
+
" \"temperature\": temperature,\n",
|
162 |
+
" \"top_p\": top_p,\n",
|
163 |
+
" \"top_k\": top_k,\n",
|
164 |
+
" \"repetition_penalty\": repetition_penalty,\n",
|
165 |
+
" }\n",
|
166 |
+
" thread = Thread(target=model_m.generate, kwargs=generation_kwargs)\n",
|
167 |
+
" thread.start()\n",
|
168 |
+
" buffer = \"\"\n",
|
169 |
+
" for new_text in streamer:\n",
|
170 |
+
" buffer += new_text\n",
|
171 |
+
" buffer = buffer.replace(\"<|im_end|>\", \"\")\n",
|
172 |
+
" time.sleep(0.01)\n",
|
173 |
+
" yield buffer\n",
|
174 |
+
"\n",
|
175 |
+
"css = \"\"\"\n",
|
176 |
+
".submit-btn {\n",
|
177 |
+
" background-color: #2980b9 !important;\n",
|
178 |
+
" color: white !important;\n",
|
179 |
+
"}\n",
|
180 |
+
".submit-btn:hover {\n",
|
181 |
+
" background-color: #3498db !important;\n",
|
182 |
+
"}\n",
|
183 |
+
"\"\"\"\n",
|
184 |
+
"\n",
|
185 |
+
"# Create the Gradio Interface\n",
|
186 |
+
"with gr.Blocks(css=css, theme=\"bethecloud/storj_theme\") as demo:\n",
|
187 |
+
" gr.Markdown(\"# **Qwen/Qwen2.5-VL-3B-Instruct**\")\n",
|
188 |
+
" with gr.Row():\n",
|
189 |
+
" with gr.Column():\n",
|
190 |
+
" with gr.Tabs():\n",
|
191 |
+
" with gr.TabItem(\"Image Inference\"):\n",
|
192 |
+
" image_query = gr.Textbox(label=\"Query Input\", placeholder=\"Enter your query here...\")\n",
|
193 |
+
" image_upload = gr.Image(type=\"pil\", label=\"Image\")\n",
|
194 |
+
" image_submit = gr.Button(\"Submit\", elem_classes=\"submit-btn\")\n",
|
195 |
+
"\n",
|
196 |
+
" with gr.TabItem(\"Video Inference\"):\n",
|
197 |
+
" video_query = gr.Textbox(label=\"Query Input\", placeholder=\"Enter your query here...\")\n",
|
198 |
+
" video_upload = gr.Video(label=\"Video\")\n",
|
199 |
+
" video_submit = gr.Button(\"Submit\", elem_classes=\"submit-btn\")\n",
|
200 |
+
"\n",
|
201 |
+
" with gr.Accordion(\"Advanced options\", open=False):\n",
|
202 |
+
" max_new_tokens = gr.Slider(label=\"Max new tokens\", minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1, value=DEFAULT_MAX_NEW_TOKENS)\n",
|
203 |
+
" temperature = gr.Slider(label=\"Temperature\", minimum=0.1, maximum=4.0, step=0.1, value=0.6)\n",
|
204 |
+
" top_p = gr.Slider(label=\"Top-p (nucleus sampling)\", minimum=0.05, maximum=1.0, step=0.05, value=0.9)\n",
|
205 |
+
" top_k = gr.Slider(label=\"Top-k\", minimum=1, maximum=1000, step=1, value=50)\n",
|
206 |
+
" repetition_penalty = gr.Slider(label=\"Repetition penalty\", minimum=1.0, maximum=2.0, step=0.05, value=1.2)\n",
|
207 |
+
" with gr.Column():\n",
|
208 |
+
" output = gr.Textbox(label=\"Output\", interactive=False)\n",
|
209 |
+
"\n",
|
210 |
+
" image_submit.click(\n",
|
211 |
+
" fn=generate_image,\n",
|
212 |
+
" inputs=[image_query, image_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],\n",
|
213 |
+
" outputs=output\n",
|
214 |
+
" )\n",
|
215 |
+
" video_submit.click(\n",
|
216 |
+
" fn=generate_video,\n",
|
217 |
+
" inputs=[video_query, video_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],\n",
|
218 |
+
" outputs=output\n",
|
219 |
+
" )\n",
|
220 |
+
"\n",
|
221 |
+
"if __name__ == \"__main__\":\n",
|
222 |
+
" demo.queue(max_size=30).launch(share=True, ssr_mode=False, show_error=True)"
|
223 |
+
]
|
224 |
+
}
|
225 |
+
],
|
226 |
+
"metadata": {
|
227 |
+
"accelerator": "GPU",
|
228 |
+
"colab": {
|
229 |
+
"gpuType": "T4",
|
230 |
+
"provenance": []
|
231 |
+
},
|
232 |
+
"kernelspec": {
|
233 |
+
"display_name": "Python 3",
|
234 |
+
"name": "python3"
|
235 |
+
},
|
236 |
+
"language_info": {
|
237 |
+
"name": "python"
|
238 |
+
}
|
239 |
+
},
|
240 |
+
"nbformat": 4,
|
241 |
+
"nbformat_minor": 0
|
242 |
+
}
|
Qwen2.5-VL/Qwen2_5VL_7B.ipynb
ADDED
@@ -0,0 +1,242 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cells": [
|
3 |
+
{
|
4 |
+
"cell_type": "code",
|
5 |
+
"execution_count": null,
|
6 |
+
"metadata": {
|
7 |
+
"id": "xL8y37Y6bORU"
|
8 |
+
},
|
9 |
+
"outputs": [],
|
10 |
+
"source": [
|
11 |
+
"%%capture\n",
|
12 |
+
"!pip install gradio spaces transformers accelerate numpy requests\n",
|
13 |
+
"!pip install torch torchvision qwen-vl-utils av hf_xet\n",
|
14 |
+
"!pip install pillow huggingface_hub opencv-python"
|
15 |
+
]
|
16 |
+
},
|
17 |
+
{
|
18 |
+
"cell_type": "code",
|
19 |
+
"execution_count": null,
|
20 |
+
"metadata": {
|
21 |
+
"id": "Y-NTbL1tdL9X"
|
22 |
+
},
|
23 |
+
"outputs": [],
|
24 |
+
"source": [
|
25 |
+
"import os\n",
|
26 |
+
"import random\n",
|
27 |
+
"import uuid\n",
|
28 |
+
"import json\n",
|
29 |
+
"import time\n",
|
30 |
+
"import asyncio\n",
|
31 |
+
"from threading import Thread\n",
|
32 |
+
"\n",
|
33 |
+
"import gradio as gr\n",
|
34 |
+
"import spaces\n",
|
35 |
+
"import torch\n",
|
36 |
+
"import numpy as np\n",
|
37 |
+
"from PIL import Image\n",
|
38 |
+
"import cv2\n",
|
39 |
+
"\n",
|
40 |
+
"from transformers import (\n",
|
41 |
+
" Qwen2_5_VLForConditionalGeneration,\n",
|
42 |
+
" AutoProcessor,\n",
|
43 |
+
" TextIteratorStreamer,\n",
|
44 |
+
")\n",
|
45 |
+
"from transformers.image_utils import load_image\n",
|
46 |
+
"\n",
|
47 |
+
"# Constants for text generation\n",
|
48 |
+
"MAX_MAX_NEW_TOKENS = 2048\n",
|
49 |
+
"DEFAULT_MAX_NEW_TOKENS = 1024\n",
|
50 |
+
"# Increase or disable input truncation to avoid token mismatches\n",
|
51 |
+
"MAX_INPUT_TOKEN_LENGTH = int(os.getenv(\"MAX_INPUT_TOKEN_LENGTH\", \"8192\"))\n",
|
52 |
+
"\n",
|
53 |
+
"device = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n",
|
54 |
+
"\n",
|
55 |
+
"MODEL_ID = \"Qwen/Qwen2.5-VL-7B-Instruct\"\n",
|
56 |
+
"processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)\n",
|
57 |
+
"model_m = Qwen2_5_VLForConditionalGeneration.from_pretrained(\n",
|
58 |
+
" MODEL_ID,\n",
|
59 |
+
" trust_remote_code=True,\n",
|
60 |
+
" torch_dtype=torch.float16\n",
|
61 |
+
").to(\"cuda\").eval()\n",
|
62 |
+
"\n",
|
63 |
+
"def downsample_video(video_path):\n",
|
64 |
+
" \"\"\"\n",
|
65 |
+
" Downsamples the video to evenly spaced frames.\n",
|
66 |
+
" Each frame is returned as a PIL image along with its timestamp.\n",
|
67 |
+
" \"\"\"\n",
|
68 |
+
" vidcap = cv2.VideoCapture(video_path)\n",
|
69 |
+
" total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))\n",
|
70 |
+
" fps = vidcap.get(cv2.CAP_PROP_FPS)\n",
|
71 |
+
" frames = []\n",
|
72 |
+
" # Sample 10 evenly spaced frames.\n",
|
73 |
+
" frame_indices = np.linspace(0, total_frames - 1, 10, dtype=int)\n",
|
74 |
+
" for i in frame_indices:\n",
|
75 |
+
" vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)\n",
|
76 |
+
" success, image = vidcap.read()\n",
|
77 |
+
" if success:\n",
|
78 |
+
" image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) # Convert BGR to RGB\n",
|
79 |
+
" pil_image = Image.fromarray(image)\n",
|
80 |
+
" timestamp = round(i / fps, 2)\n",
|
81 |
+
" frames.append((pil_image, timestamp))\n",
|
82 |
+
" vidcap.release()\n",
|
83 |
+
" return frames\n",
|
84 |
+
"\n",
|
85 |
+
"@spaces.GPU\n",
|
86 |
+
"def generate_image(text: str, image: Image.Image,\n",
|
87 |
+
" max_new_tokens: int = 1024,\n",
|
88 |
+
" temperature: float = 0.6,\n",
|
89 |
+
" top_p: float = 0.9,\n",
|
90 |
+
" top_k: int = 50,\n",
|
91 |
+
" repetition_penalty: float = 1.2):\n",
|
92 |
+
"\n",
|
93 |
+
" if image is None:\n",
|
94 |
+
" yield \"Please upload an image.\"\n",
|
95 |
+
" return\n",
|
96 |
+
"\n",
|
97 |
+
" messages = [{\n",
|
98 |
+
" \"role\": \"user\",\n",
|
99 |
+
" \"content\": [\n",
|
100 |
+
" {\"type\": \"image\", \"image\": image},\n",
|
101 |
+
" {\"type\": \"text\", \"text\": text},\n",
|
102 |
+
" ]\n",
|
103 |
+
" }]\n",
|
104 |
+
" prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)\n",
|
105 |
+
" inputs = processor(\n",
|
106 |
+
" text=[prompt_full],\n",
|
107 |
+
" images=[image],\n",
|
108 |
+
" return_tensors=\"pt\",\n",
|
109 |
+
" padding=True,\n",
|
110 |
+
" truncation=False,\n",
|
111 |
+
" max_length=MAX_INPUT_TOKEN_LENGTH\n",
|
112 |
+
" ).to(\"cuda\")\n",
|
113 |
+
" streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)\n",
|
114 |
+
" generation_kwargs = {**inputs, \"streamer\": streamer, \"max_new_tokens\": max_new_tokens}\n",
|
115 |
+
" thread = Thread(target=model_m.generate, kwargs=generation_kwargs)\n",
|
116 |
+
" thread.start()\n",
|
117 |
+
" buffer = \"\"\n",
|
118 |
+
" for new_text in streamer:\n",
|
119 |
+
" buffer += new_text\n",
|
120 |
+
" buffer = buffer.replace(\"<|im_end|>\", \"\")\n",
|
121 |
+
" time.sleep(0.01)\n",
|
122 |
+
" yield buffer\n",
|
123 |
+
"\n",
|
124 |
+
"@spaces.GPU\n",
|
125 |
+
"def generate_video(text: str, video_path: str,\n",
|
126 |
+
" max_new_tokens: int = 1024,\n",
|
127 |
+
" temperature: float = 0.6,\n",
|
128 |
+
" top_p: float = 0.9,\n",
|
129 |
+
" top_k: int = 50,\n",
|
130 |
+
" repetition_penalty: float = 1.2):\n",
|
131 |
+
"\n",
|
132 |
+
" if video_path is None:\n",
|
133 |
+
" yield \"Please upload a video.\"\n",
|
134 |
+
" return\n",
|
135 |
+
"\n",
|
136 |
+
" frames = downsample_video(video_path)\n",
|
137 |
+
" messages = [\n",
|
138 |
+
" {\"role\": \"system\", \"content\": [{\"type\": \"text\", \"text\": \"You are a helpful assistant.\"}]},\n",
|
139 |
+
" {\"role\": \"user\", \"content\": [{\"type\": \"text\", \"text\": text}]}\n",
|
140 |
+
" ]\n",
|
141 |
+
" # Append each frame with its timestamp.\n",
|
142 |
+
" for frame in frames:\n",
|
143 |
+
" image, timestamp = frame\n",
|
144 |
+
" messages[1][\"content\"].append({\"type\": \"text\", \"text\": f\"Frame {timestamp}:\"})\n",
|
145 |
+
" messages[1][\"content\"].append({\"type\": \"image\", \"image\": image})\n",
|
146 |
+
" inputs = processor.apply_chat_template(\n",
|
147 |
+
" messages,\n",
|
148 |
+
" tokenize=True,\n",
|
149 |
+
" add_generation_prompt=True,\n",
|
150 |
+
" return_dict=True,\n",
|
151 |
+
" return_tensors=\"pt\",\n",
|
152 |
+
" truncation=False,\n",
|
153 |
+
" max_length=MAX_INPUT_TOKEN_LENGTH\n",
|
154 |
+
" ).to(\"cuda\")\n",
|
155 |
+
" streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)\n",
|
156 |
+
" generation_kwargs = {\n",
|
157 |
+
" **inputs,\n",
|
158 |
+
" \"streamer\": streamer,\n",
|
159 |
+
" \"max_new_tokens\": max_new_tokens,\n",
|
160 |
+
" \"do_sample\": True,\n",
|
161 |
+
" \"temperature\": temperature,\n",
|
162 |
+
" \"top_p\": top_p,\n",
|
163 |
+
" \"top_k\": top_k,\n",
|
164 |
+
" \"repetition_penalty\": repetition_penalty,\n",
|
165 |
+
" }\n",
|
166 |
+
" thread = Thread(target=model_m.generate, kwargs=generation_kwargs)\n",
|
167 |
+
" thread.start()\n",
|
168 |
+
" buffer = \"\"\n",
|
169 |
+
" for new_text in streamer:\n",
|
170 |
+
" buffer += new_text\n",
|
171 |
+
" buffer = buffer.replace(\"<|im_end|>\", \"\")\n",
|
172 |
+
" time.sleep(0.01)\n",
|
173 |
+
" yield buffer\n",
|
174 |
+
"\n",
|
175 |
+
"css = \"\"\"\n",
|
176 |
+
".submit-btn {\n",
|
177 |
+
" background-color: #2980b9 !important;\n",
|
178 |
+
" color: white !important;\n",
|
179 |
+
"}\n",
|
180 |
+
".submit-btn:hover {\n",
|
181 |
+
" background-color: #3498db !important;\n",
|
182 |
+
"}\n",
|
183 |
+
"\"\"\"\n",
|
184 |
+
"\n",
|
185 |
+
"# Create the Gradio Interface\n",
|
186 |
+
"with gr.Blocks(css=css, theme=\"bethecloud/storj_theme\") as demo:\n",
|
187 |
+
" gr.Markdown(\"# **Qwen/Qwen2.5-VL-7B-Instruct**\")\n",
|
188 |
+
" with gr.Row():\n",
|
189 |
+
" with gr.Column():\n",
|
190 |
+
" with gr.Tabs():\n",
|
191 |
+
" with gr.TabItem(\"Image Inference\"):\n",
|
192 |
+
" image_query = gr.Textbox(label=\"Query Input\", placeholder=\"Enter your query here...\")\n",
|
193 |
+
" image_upload = gr.Image(type=\"pil\", label=\"Image\")\n",
|
194 |
+
" image_submit = gr.Button(\"Submit\", elem_classes=\"submit-btn\")\n",
|
195 |
+
"\n",
|
196 |
+
" with gr.TabItem(\"Video Inference\"):\n",
|
197 |
+
" video_query = gr.Textbox(label=\"Query Input\", placeholder=\"Enter your query here...\")\n",
|
198 |
+
" video_upload = gr.Video(label=\"Video\")\n",
|
199 |
+
" video_submit = gr.Button(\"Submit\", elem_classes=\"submit-btn\")\n",
|
200 |
+
"\n",
|
201 |
+
" with gr.Accordion(\"Advanced options\", open=False):\n",
|
202 |
+
" max_new_tokens = gr.Slider(label=\"Max new tokens\", minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1, value=DEFAULT_MAX_NEW_TOKENS)\n",
|
203 |
+
" temperature = gr.Slider(label=\"Temperature\", minimum=0.1, maximum=4.0, step=0.1, value=0.6)\n",
|
204 |
+
" top_p = gr.Slider(label=\"Top-p (nucleus sampling)\", minimum=0.05, maximum=1.0, step=0.05, value=0.9)\n",
|
205 |
+
" top_k = gr.Slider(label=\"Top-k\", minimum=1, maximum=1000, step=1, value=50)\n",
|
206 |
+
" repetition_penalty = gr.Slider(label=\"Repetition penalty\", minimum=1.0, maximum=2.0, step=0.05, value=1.2)\n",
|
207 |
+
" with gr.Column():\n",
|
208 |
+
" output = gr.Textbox(label=\"Output\", interactive=False)\n",
|
209 |
+
"\n",
|
210 |
+
" image_submit.click(\n",
|
211 |
+
" fn=generate_image,\n",
|
212 |
+
" inputs=[image_query, image_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],\n",
|
213 |
+
" outputs=output\n",
|
214 |
+
" )\n",
|
215 |
+
" video_submit.click(\n",
|
216 |
+
" fn=generate_video,\n",
|
217 |
+
" inputs=[video_query, video_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],\n",
|
218 |
+
" outputs=output\n",
|
219 |
+
" )\n",
|
220 |
+
"\n",
|
221 |
+
"if __name__ == \"__main__\":\n",
|
222 |
+
" demo.queue(max_size=30).launch(share=True, ssr_mode=False, show_error=True)"
|
223 |
+
]
|
224 |
+
}
|
225 |
+
],
|
226 |
+
"metadata": {
|
227 |
+
"accelerator": "GPU",
|
228 |
+
"colab": {
|
229 |
+
"gpuType": "T4",
|
230 |
+
"provenance": []
|
231 |
+
},
|
232 |
+
"kernelspec": {
|
233 |
+
"display_name": "Python 3",
|
234 |
+
"name": "python3"
|
235 |
+
},
|
236 |
+
"language_info": {
|
237 |
+
"name": "python"
|
238 |
+
}
|
239 |
+
},
|
240 |
+
"nbformat": 4,
|
241 |
+
"nbformat_minor": 0
|
242 |
+
}
|
RolmOCR-Qwen2.5-VL/reducto_RolmOCR_Qwen2_5VL_7B.ipynb
ADDED
@@ -0,0 +1,242 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cells": [
|
3 |
+
{
|
4 |
+
"cell_type": "code",
|
5 |
+
"execution_count": null,
|
6 |
+
"metadata": {
|
7 |
+
"id": "xL8y37Y6bORU"
|
8 |
+
},
|
9 |
+
"outputs": [],
|
10 |
+
"source": [
|
11 |
+
"%%capture\n",
|
12 |
+
"!pip install gradio spaces transformers accelerate numpy requests\n",
|
13 |
+
"!pip install torch torchvision qwen-vl-utils av hf_xet\n",
|
14 |
+
"!pip install pillow huggingface_hub opencv-python"
|
15 |
+
]
|
16 |
+
},
|
17 |
+
{
|
18 |
+
"cell_type": "code",
|
19 |
+
"execution_count": null,
|
20 |
+
"metadata": {
|
21 |
+
"id": "Y-NTbL1tdL9X"
|
22 |
+
},
|
23 |
+
"outputs": [],
|
24 |
+
"source": [
|
25 |
+
"import os\n",
|
26 |
+
"import random\n",
|
27 |
+
"import uuid\n",
|
28 |
+
"import json\n",
|
29 |
+
"import time\n",
|
30 |
+
"import asyncio\n",
|
31 |
+
"from threading import Thread\n",
|
32 |
+
"\n",
|
33 |
+
"import gradio as gr\n",
|
34 |
+
"import spaces\n",
|
35 |
+
"import torch\n",
|
36 |
+
"import numpy as np\n",
|
37 |
+
"from PIL import Image\n",
|
38 |
+
"import cv2\n",
|
39 |
+
"\n",
|
40 |
+
"from transformers import (\n",
|
41 |
+
" Qwen2_5_VLForConditionalGeneration,\n",
|
42 |
+
" AutoProcessor,\n",
|
43 |
+
" TextIteratorStreamer,\n",
|
44 |
+
")\n",
|
45 |
+
"from transformers.image_utils import load_image\n",
|
46 |
+
"\n",
|
47 |
+
"# Constants for text generation\n",
|
48 |
+
"MAX_MAX_NEW_TOKENS = 2048\n",
|
49 |
+
"DEFAULT_MAX_NEW_TOKENS = 1024\n",
|
50 |
+
"# Increase or disable input truncation to avoid token mismatches\n",
|
51 |
+
"MAX_INPUT_TOKEN_LENGTH = int(os.getenv(\"MAX_INPUT_TOKEN_LENGTH\", \"8192\"))\n",
|
52 |
+
"\n",
|
53 |
+
"device = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n",
|
54 |
+
"\n",
|
55 |
+
"MODEL_ID = \"reducto/RolmOCR\"\n",
|
56 |
+
"processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)\n",
|
57 |
+
"model_m = Qwen2_5_VLForConditionalGeneration.from_pretrained(\n",
|
58 |
+
" MODEL_ID,\n",
|
59 |
+
" trust_remote_code=True,\n",
|
60 |
+
" torch_dtype=torch.float16\n",
|
61 |
+
").to(\"cuda\").eval()\n",
|
62 |
+
"\n",
|
63 |
+
"def downsample_video(video_path):\n",
|
64 |
+
" \"\"\"\n",
|
65 |
+
" Downsamples the video to evenly spaced frames.\n",
|
66 |
+
" Each frame is returned as a PIL image along with its timestamp.\n",
|
67 |
+
" \"\"\"\n",
|
68 |
+
" vidcap = cv2.VideoCapture(video_path)\n",
|
69 |
+
" total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))\n",
|
70 |
+
" fps = vidcap.get(cv2.CAP_PROP_FPS)\n",
|
71 |
+
" frames = []\n",
|
72 |
+
" # Sample 10 evenly spaced frames.\n",
|
73 |
+
" frame_indices = np.linspace(0, total_frames - 1, 10, dtype=int)\n",
|
74 |
+
" for i in frame_indices:\n",
|
75 |
+
" vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)\n",
|
76 |
+
" success, image = vidcap.read()\n",
|
77 |
+
" if success:\n",
|
78 |
+
" image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) # Convert BGR to RGB\n",
|
79 |
+
" pil_image = Image.fromarray(image)\n",
|
80 |
+
" timestamp = round(i / fps, 2)\n",
|
81 |
+
" frames.append((pil_image, timestamp))\n",
|
82 |
+
" vidcap.release()\n",
|
83 |
+
" return frames\n",
|
84 |
+
"\n",
|
85 |
+
"@spaces.GPU\n",
|
86 |
+
"def generate_image(text: str, image: Image.Image,\n",
|
87 |
+
" max_new_tokens: int = 1024,\n",
|
88 |
+
" temperature: float = 0.6,\n",
|
89 |
+
" top_p: float = 0.9,\n",
|
90 |
+
" top_k: int = 50,\n",
|
91 |
+
" repetition_penalty: float = 1.2):\n",
|
92 |
+
"\n",
|
93 |
+
" if image is None:\n",
|
94 |
+
" yield \"Please upload an image.\"\n",
|
95 |
+
" return\n",
|
96 |
+
"\n",
|
97 |
+
" messages = [{\n",
|
98 |
+
" \"role\": \"user\",\n",
|
99 |
+
" \"content\": [\n",
|
100 |
+
" {\"type\": \"image\", \"image\": image},\n",
|
101 |
+
" {\"type\": \"text\", \"text\": text},\n",
|
102 |
+
" ]\n",
|
103 |
+
" }]\n",
|
104 |
+
" prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)\n",
|
105 |
+
" inputs = processor(\n",
|
106 |
+
" text=[prompt_full],\n",
|
107 |
+
" images=[image],\n",
|
108 |
+
" return_tensors=\"pt\",\n",
|
109 |
+
" padding=True,\n",
|
110 |
+
" truncation=False # Disable truncation to keep image tokens intact\n",
|
111 |
+
" max_length=MAX_INPUT_TOKEN_LENGTH\n",
|
112 |
+
" ).to(\"cuda\")\n",
|
113 |
+
" streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)\n",
|
114 |
+
" generation_kwargs = {**inputs, \"streamer\": streamer, \"max_new_tokens\": max_new_tokens}\n",
|
115 |
+
" thread = Thread(target=model_m.generate, kwargs=generation_kwargs)\n",
|
116 |
+
" thread.start()\n",
|
117 |
+
" buffer = \"\"\n",
|
118 |
+
" for new_text in streamer:\n",
|
119 |
+
" buffer += new_text\n",
|
120 |
+
" buffer = buffer.replace(\"<|im_end|>\", \"\")\n",
|
121 |
+
" time.sleep(0.01)\n",
|
122 |
+
" yield buffer\n",
|
123 |
+
"\n",
|
124 |
+
"@spaces.GPU\n",
|
125 |
+
"def generate_video(text: str, video_path: str,\n",
|
126 |
+
" max_new_tokens: int = 1024,\n",
|
127 |
+
" temperature: float = 0.6,\n",
|
128 |
+
" top_p: float = 0.9,\n",
|
129 |
+
" top_k: int = 50,\n",
|
130 |
+
" repetition_penalty: float = 1.2):\n",
|
131 |
+
"\n",
|
132 |
+
" if video_path is None:\n",
|
133 |
+
" yield \"Please upload a video.\"\n",
|
134 |
+
" return\n",
|
135 |
+
"\n",
|
136 |
+
" frames = downsample_video(video_path)\n",
|
137 |
+
" messages = [\n",
|
138 |
+
" {\"role\": \"system\", \"content\": [{\"type\": \"text\", \"text\": \"You are a helpful assistant.\"}]},\n",
|
139 |
+
" {\"role\": \"user\", \"content\": [{\"type\": \"text\", \"text\": text}]}\n",
|
140 |
+
" ]\n",
|
141 |
+
" # Append each frame with its timestamp.\n",
|
142 |
+
" for frame in frames:\n",
|
143 |
+
" image, timestamp = frame\n",
|
144 |
+
" messages[1][\"content\"].append({\"type\": \"text\", \"text\": f\"Frame {timestamp}:\"})\n",
|
145 |
+
" messages[1][\"content\"].append({\"type\": \"image\", \"image\": image})\n",
|
146 |
+
" inputs = processor.apply_chat_template(\n",
|
147 |
+
" messages,\n",
|
148 |
+
" tokenize=True,\n",
|
149 |
+
" add_generation_prompt=True,\n",
|
150 |
+
" return_dict=True,\n",
|
151 |
+
" return_tensors=\"pt\",\n",
|
152 |
+
" truncation=False,,\n",
|
153 |
+
" max_length=MAX_INPUT_TOKEN_LENGTH\n",
|
154 |
+
" ).to(\"cuda\")\n",
|
155 |
+
" streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)\n",
|
156 |
+
" generation_kwargs = {\n",
|
157 |
+
" **inputs,\n",
|
158 |
+
" \"streamer\": streamer,\n",
|
159 |
+
" \"max_new_tokens\": max_new_tokens,\n",
|
160 |
+
" \"do_sample\": True,\n",
|
161 |
+
" \"temperature\": temperature,\n",
|
162 |
+
" \"top_p\": top_p,\n",
|
163 |
+
" \"top_k\": top_k,\n",
|
164 |
+
" \"repetition_penalty\": repetition_penalty,\n",
|
165 |
+
" }\n",
|
166 |
+
" thread = Thread(target=model_m.generate, kwargs=generation_kwargs)\n",
|
167 |
+
" thread.start()\n",
|
168 |
+
" buffer = \"\"\n",
|
169 |
+
" for new_text in streamer:\n",
|
170 |
+
" buffer += new_text\n",
|
171 |
+
" buffer = buffer.replace(\"<|im_end|>\", \"\")\n",
|
172 |
+
" time.sleep(0.01)\n",
|
173 |
+
" yield buffer\n",
|
174 |
+
"\n",
|
175 |
+
"css = \"\"\"\n",
|
176 |
+
".submit-btn {\n",
|
177 |
+
" background-color: #2980b9 !important;\n",
|
178 |
+
" color: white !important;\n",
|
179 |
+
"}\n",
|
180 |
+
".submit-btn:hover {\n",
|
181 |
+
" background-color: #3498db !important;\n",
|
182 |
+
"}\n",
|
183 |
+
"\"\"\"\n",
|
184 |
+
"\n",
|
185 |
+
"# Create the Gradio Interface\n",
|
186 |
+
"with gr.Blocks(css=css, theme=\"bethecloud/storj_theme\") as demo:\n",
|
187 |
+
" gr.Markdown(\"# **reducto/RolmOCR**\")\n",
|
188 |
+
" with gr.Row():\n",
|
189 |
+
" with gr.Column():\n",
|
190 |
+
" with gr.Tabs():\n",
|
191 |
+
" with gr.TabItem(\"Image Inference\"):\n",
|
192 |
+
" image_query = gr.Textbox(label=\"Query Input\", placeholder=\"Enter your query here...\")\n",
|
193 |
+
" image_upload = gr.Image(type=\"pil\", label=\"Image\")\n",
|
194 |
+
" image_submit = gr.Button(\"Submit\", elem_classes=\"submit-btn\")\n",
|
195 |
+
"\n",
|
196 |
+
" with gr.TabItem(\"Video Inference\"):\n",
|
197 |
+
" video_query = gr.Textbox(label=\"Query Input\", placeholder=\"Enter your query here...\")\n",
|
198 |
+
" video_upload = gr.Video(label=\"Video\")\n",
|
199 |
+
" video_submit = gr.Button(\"Submit\", elem_classes=\"submit-btn\")\n",
|
200 |
+
"\n",
|
201 |
+
" with gr.Accordion(\"Advanced options\", open=False):\n",
|
202 |
+
" max_new_tokens = gr.Slider(label=\"Max new tokens\", minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1, value=DEFAULT_MAX_NEW_TOKENS)\n",
|
203 |
+
" temperature = gr.Slider(label=\"Temperature\", minimum=0.1, maximum=4.0, step=0.1, value=0.6)\n",
|
204 |
+
" top_p = gr.Slider(label=\"Top-p (nucleus sampling)\", minimum=0.05, maximum=1.0, step=0.05, value=0.9)\n",
|
205 |
+
" top_k = gr.Slider(label=\"Top-k\", minimum=1, maximum=1000, step=1, value=50)\n",
|
206 |
+
" repetition_penalty = gr.Slider(label=\"Repetition penalty\", minimum=1.0, maximum=2.0, step=0.05, value=1.2)\n",
|
207 |
+
" with gr.Column():\n",
|
208 |
+
" output = gr.Textbox(label=\"Output\", interactive=False)\n",
|
209 |
+
"\n",
|
210 |
+
" image_submit.click(\n",
|
211 |
+
" fn=generate_image,\n",
|
212 |
+
" inputs=[image_query, image_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],\n",
|
213 |
+
" outputs=output\n",
|
214 |
+
" )\n",
|
215 |
+
" video_submit.click(\n",
|
216 |
+
" fn=generate_video,\n",
|
217 |
+
" inputs=[video_query, video_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],\n",
|
218 |
+
" outputs=output\n",
|
219 |
+
" )\n",
|
220 |
+
"\n",
|
221 |
+
"if __name__ == \"__main__\":\n",
|
222 |
+
" demo.queue(max_size=30).launch(share=True, ssr_mode=False, show_error=True)"
|
223 |
+
]
|
224 |
+
}
|
225 |
+
],
|
226 |
+
"metadata": {
|
227 |
+
"accelerator": "GPU",
|
228 |
+
"colab": {
|
229 |
+
"gpuType": "T4",
|
230 |
+
"provenance": []
|
231 |
+
},
|
232 |
+
"kernelspec": {
|
233 |
+
"display_name": "Python 3",
|
234 |
+
"name": "python3"
|
235 |
+
},
|
236 |
+
"language_info": {
|
237 |
+
"name": "python"
|
238 |
+
}
|
239 |
+
},
|
240 |
+
"nbformat": 4,
|
241 |
+
"nbformat_minor": 0
|
242 |
+
}
|
olmOCR-Qwen2-VL/olmOCR_7B_0225.ipynb
ADDED
@@ -0,0 +1,242 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"nbformat": 4,
|
3 |
+
"nbformat_minor": 0,
|
4 |
+
"metadata": {
|
5 |
+
"colab": {
|
6 |
+
"provenance": [],
|
7 |
+
"gpuType": "T4"
|
8 |
+
},
|
9 |
+
"kernelspec": {
|
10 |
+
"name": "python3",
|
11 |
+
"display_name": "Python 3"
|
12 |
+
},
|
13 |
+
"language_info": {
|
14 |
+
"name": "python"
|
15 |
+
},
|
16 |
+
"accelerator": "GPU"
|
17 |
+
},
|
18 |
+
"cells": [
|
19 |
+
{
|
20 |
+
"cell_type": "code",
|
21 |
+
"execution_count": null,
|
22 |
+
"metadata": {
|
23 |
+
"id": "xL8y37Y6bORU"
|
24 |
+
},
|
25 |
+
"outputs": [],
|
26 |
+
"source": [
|
27 |
+
"%%capture\n",
|
28 |
+
"!pip install gradio spaces transformers accelerate numpy requests\n",
|
29 |
+
"!pip install torch torchvision qwen-vl-utils av hf_xet\n",
|
30 |
+
"!pip install pillow huggingface_hub opencv-python"
|
31 |
+
]
|
32 |
+
},
|
33 |
+
{
|
34 |
+
"cell_type": "code",
|
35 |
+
"source": [
|
36 |
+
"import os\n",
|
37 |
+
"import random\n",
|
38 |
+
"import uuid\n",
|
39 |
+
"import json\n",
|
40 |
+
"import time\n",
|
41 |
+
"import asyncio\n",
|
42 |
+
"from threading import Thread\n",
|
43 |
+
"\n",
|
44 |
+
"import gradio as gr\n",
|
45 |
+
"import spaces\n",
|
46 |
+
"import torch\n",
|
47 |
+
"import numpy as np\n",
|
48 |
+
"from PIL import Image\n",
|
49 |
+
"import cv2\n",
|
50 |
+
"\n",
|
51 |
+
"from transformers import (\n",
|
52 |
+
" Qwen2_5_VLForConditionalGeneration,\n",
|
53 |
+
" AutoProcessor,\n",
|
54 |
+
" TextIteratorStreamer,\n",
|
55 |
+
")\n",
|
56 |
+
"from transformers.image_utils import load_image\n",
|
57 |
+
"\n",
|
58 |
+
"# Constants for text generation\n",
|
59 |
+
"MAX_MAX_NEW_TOKENS = 2048\n",
|
60 |
+
"DEFAULT_MAX_NEW_TOKENS = 1024\n",
|
61 |
+
"# Increase or disable input truncation to avoid token mismatches\n",
|
62 |
+
"MAX_INPUT_TOKEN_LENGTH = int(os.getenv(\"MAX_INPUT_TOKEN_LENGTH\", \"8192\"))\n",
|
63 |
+
"\n",
|
64 |
+
"device = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n",
|
65 |
+
"\n",
|
66 |
+
"MODEL_ID = \"allenai/olmOCR-7B-0225-preview\"\n",
|
67 |
+
"processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)\n",
|
68 |
+
"model_m = Qwen2_5_VLForConditionalGeneration.from_pretrained(\n",
|
69 |
+
" MODEL_ID,\n",
|
70 |
+
" trust_remote_code=True,\n",
|
71 |
+
" torch_dtype=torch.float16\n",
|
72 |
+
").to(\"cuda\").eval()\n",
|
73 |
+
"\n",
|
74 |
+
"def downsample_video(video_path):\n",
|
75 |
+
" \"\"\"\n",
|
76 |
+
" Downsamples the video to evenly spaced frames.\n",
|
77 |
+
" Each frame is returned as a PIL image along with its timestamp.\n",
|
78 |
+
" \"\"\"\n",
|
79 |
+
" vidcap = cv2.VideoCapture(video_path)\n",
|
80 |
+
" total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))\n",
|
81 |
+
" fps = vidcap.get(cv2.CAP_PROP_FPS)\n",
|
82 |
+
" frames = []\n",
|
83 |
+
" # Sample 10 evenly spaced frames.\n",
|
84 |
+
" frame_indices = np.linspace(0, total_frames - 1, 10, dtype=int)\n",
|
85 |
+
" for i in frame_indices:\n",
|
86 |
+
" vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)\n",
|
87 |
+
" success, image = vidcap.read()\n",
|
88 |
+
" if success:\n",
|
89 |
+
" image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) # Convert BGR to RGB\n",
|
90 |
+
" pil_image = Image.fromarray(image)\n",
|
91 |
+
" timestamp = round(i / fps, 2)\n",
|
92 |
+
" frames.append((pil_image, timestamp))\n",
|
93 |
+
" vidcap.release()\n",
|
94 |
+
" return frames\n",
|
95 |
+
"\n",
|
96 |
+
"@spaces.GPU\n",
|
97 |
+
"def generate_image(text: str, image: Image.Image,\n",
|
98 |
+
" max_new_tokens: int = 1024,\n",
|
99 |
+
" temperature: float = 0.6,\n",
|
100 |
+
" top_p: float = 0.9,\n",
|
101 |
+
" top_k: int = 50,\n",
|
102 |
+
" repetition_penalty: float = 1.2):\n",
|
103 |
+
"\n",
|
104 |
+
" if image is None:\n",
|
105 |
+
" yield \"Please upload an image.\"\n",
|
106 |
+
" return\n",
|
107 |
+
"\n",
|
108 |
+
" messages = [{\n",
|
109 |
+
" \"role\": \"user\",\n",
|
110 |
+
" \"content\": [\n",
|
111 |
+
" {\"type\": \"image\", \"image\": image},\n",
|
112 |
+
" {\"type\": \"text\", \"text\": text},\n",
|
113 |
+
" ]\n",
|
114 |
+
" }]\n",
|
115 |
+
" prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)\n",
|
116 |
+
" inputs = processor(\n",
|
117 |
+
" text=[prompt_full],\n",
|
118 |
+
" images=[image],\n",
|
119 |
+
" return_tensors=\"pt\",\n",
|
120 |
+
" padding=True,\n",
|
121 |
+
" truncation=False,\n",
|
122 |
+
" max_length=MAX_INPUT_TOKEN_LENGTH\n",
|
123 |
+
" ).to(\"cuda\")\n",
|
124 |
+
" streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)\n",
|
125 |
+
" generation_kwargs = {**inputs, \"streamer\": streamer, \"max_new_tokens\": max_new_tokens}\n",
|
126 |
+
" thread = Thread(target=model_m.generate, kwargs=generation_kwargs)\n",
|
127 |
+
" thread.start()\n",
|
128 |
+
" buffer = \"\"\n",
|
129 |
+
" for new_text in streamer:\n",
|
130 |
+
" buffer += new_text\n",
|
131 |
+
" buffer = buffer.replace(\"<|im_end|>\", \"\")\n",
|
132 |
+
" time.sleep(0.01)\n",
|
133 |
+
" yield buffer\n",
|
134 |
+
"\n",
|
135 |
+
"@spaces.GPU\n",
|
136 |
+
"def generate_video(text: str, video_path: str,\n",
|
137 |
+
" max_new_tokens: int = 1024,\n",
|
138 |
+
" temperature: float = 0.6,\n",
|
139 |
+
" top_p: float = 0.9,\n",
|
140 |
+
" top_k: int = 50,\n",
|
141 |
+
" repetition_penalty: float = 1.2):\n",
|
142 |
+
"\n",
|
143 |
+
" if video_path is None:\n",
|
144 |
+
" yield \"Please upload a video.\"\n",
|
145 |
+
" return\n",
|
146 |
+
"\n",
|
147 |
+
" frames = downsample_video(video_path)\n",
|
148 |
+
" messages = [\n",
|
149 |
+
" {\"role\": \"system\", \"content\": [{\"type\": \"text\", \"text\": \"You are a helpful assistant.\"}]},\n",
|
150 |
+
" {\"role\": \"user\", \"content\": [{\"type\": \"text\", \"text\": text}]}\n",
|
151 |
+
" ]\n",
|
152 |
+
" # Append each frame with its timestamp.\n",
|
153 |
+
" for frame in frames:\n",
|
154 |
+
" image, timestamp = frame\n",
|
155 |
+
" messages[1][\"content\"].append({\"type\": \"text\", \"text\": f\"Frame {timestamp}:\"})\n",
|
156 |
+
" messages[1][\"content\"].append({\"type\": \"image\", \"image\": image})\n",
|
157 |
+
" inputs = processor.apply_chat_template(\n",
|
158 |
+
" messages,\n",
|
159 |
+
" tokenize=True,\n",
|
160 |
+
" add_generation_prompt=True,\n",
|
161 |
+
" return_dict=True,\n",
|
162 |
+
" return_tensors=\"pt\",\n",
|
163 |
+
" truncation=False,\n",
|
164 |
+
" max_length=MAX_INPUT_TOKEN_LENGTH\n",
|
165 |
+
" ).to(\"cuda\")\n",
|
166 |
+
" streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)\n",
|
167 |
+
" generation_kwargs = {\n",
|
168 |
+
" **inputs,\n",
|
169 |
+
" \"streamer\": streamer,\n",
|
170 |
+
" \"max_new_tokens\": max_new_tokens,\n",
|
171 |
+
" \"do_sample\": True,\n",
|
172 |
+
" \"temperature\": temperature,\n",
|
173 |
+
" \"top_p\": top_p,\n",
|
174 |
+
" \"top_k\": top_k,\n",
|
175 |
+
" \"repetition_penalty\": repetition_penalty,\n",
|
176 |
+
" }\n",
|
177 |
+
" thread = Thread(target=model_m.generate, kwargs=generation_kwargs)\n",
|
178 |
+
" thread.start()\n",
|
179 |
+
" buffer = \"\"\n",
|
180 |
+
" for new_text in streamer:\n",
|
181 |
+
" buffer += new_text\n",
|
182 |
+
" buffer = buffer.replace(\"<|im_end|>\", \"\")\n",
|
183 |
+
" time.sleep(0.01)\n",
|
184 |
+
" yield buffer\n",
|
185 |
+
"\n",
|
186 |
+
"css = \"\"\"\n",
|
187 |
+
".submit-btn {\n",
|
188 |
+
" background-color: #2980b9 !important;\n",
|
189 |
+
" color: white !important;\n",
|
190 |
+
"}\n",
|
191 |
+
".submit-btn:hover {\n",
|
192 |
+
" background-color: #3498db !important;\n",
|
193 |
+
"}\n",
|
194 |
+
"\"\"\"\n",
|
195 |
+
"\n",
|
196 |
+
"# Create the Gradio Interface\n",
|
197 |
+
"with gr.Blocks(css=css, theme=\"bethecloud/storj_theme\") as demo:\n",
|
198 |
+
" gr.Markdown(\"# **allenai/olmOCR-7B-0225-preview**\")\n",
|
199 |
+
" with gr.Row():\n",
|
200 |
+
" with gr.Column():\n",
|
201 |
+
" with gr.Tabs():\n",
|
202 |
+
" with gr.TabItem(\"Image Inference\"):\n",
|
203 |
+
" image_query = gr.Textbox(label=\"Query Input\", placeholder=\"Enter your query here...\")\n",
|
204 |
+
" image_upload = gr.Image(type=\"pil\", label=\"Image\")\n",
|
205 |
+
" image_submit = gr.Button(\"Submit\", elem_classes=\"submit-btn\")\n",
|
206 |
+
"\n",
|
207 |
+
" with gr.TabItem(\"Video Inference\"):\n",
|
208 |
+
" video_query = gr.Textbox(label=\"Query Input\", placeholder=\"Enter your query here...\")\n",
|
209 |
+
" video_upload = gr.Video(label=\"Video\")\n",
|
210 |
+
" video_submit = gr.Button(\"Submit\", elem_classes=\"submit-btn\")\n",
|
211 |
+
"\n",
|
212 |
+
" with gr.Accordion(\"Advanced options\", open=False):\n",
|
213 |
+
" max_new_tokens = gr.Slider(label=\"Max new tokens\", minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1, value=DEFAULT_MAX_NEW_TOKENS)\n",
|
214 |
+
" temperature = gr.Slider(label=\"Temperature\", minimum=0.1, maximum=4.0, step=0.1, value=0.6)\n",
|
215 |
+
" top_p = gr.Slider(label=\"Top-p (nucleus sampling)\", minimum=0.05, maximum=1.0, step=0.05, value=0.9)\n",
|
216 |
+
" top_k = gr.Slider(label=\"Top-k\", minimum=1, maximum=1000, step=1, value=50)\n",
|
217 |
+
" repetition_penalty = gr.Slider(label=\"Repetition penalty\", minimum=1.0, maximum=2.0, step=0.05, value=1.2)\n",
|
218 |
+
" with gr.Column():\n",
|
219 |
+
" output = gr.Textbox(label=\"Output\", interactive=False)\n",
|
220 |
+
"\n",
|
221 |
+
" image_submit.click(\n",
|
222 |
+
" fn=generate_image,\n",
|
223 |
+
" inputs=[image_query, image_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],\n",
|
224 |
+
" outputs=output\n",
|
225 |
+
" )\n",
|
226 |
+
" video_submit.click(\n",
|
227 |
+
" fn=generate_video,\n",
|
228 |
+
" inputs=[video_query, video_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],\n",
|
229 |
+
" outputs=output\n",
|
230 |
+
" )\n",
|
231 |
+
"\n",
|
232 |
+
"if __name__ == \"__main__\":\n",
|
233 |
+
" demo.queue(max_size=30).launch(share=True, ssr_mode=False, show_error=True)"
|
234 |
+
],
|
235 |
+
"metadata": {
|
236 |
+
"id": "Y-NTbL1tdL9X"
|
237 |
+
},
|
238 |
+
"execution_count": null,
|
239 |
+
"outputs": []
|
240 |
+
}
|
241 |
+
]
|
242 |
+
}
|
typhoon-ocr-7b-Qwen2.5VL/typhoon_ocr_7b.ipynb
ADDED
@@ -0,0 +1,242 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cells": [
|
3 |
+
{
|
4 |
+
"cell_type": "code",
|
5 |
+
"execution_count": null,
|
6 |
+
"metadata": {
|
7 |
+
"id": "xL8y37Y6bORU"
|
8 |
+
},
|
9 |
+
"outputs": [],
|
10 |
+
"source": [
|
11 |
+
"%%capture\n",
|
12 |
+
"!pip install gradio spaces transformers accelerate numpy requests\n",
|
13 |
+
"!pip install torch torchvision qwen-vl-utils av hf_xet\n",
|
14 |
+
"!pip install pillow huggingface_hub opencv-python"
|
15 |
+
]
|
16 |
+
},
|
17 |
+
{
|
18 |
+
"cell_type": "code",
|
19 |
+
"execution_count": null,
|
20 |
+
"metadata": {
|
21 |
+
"id": "Y-NTbL1tdL9X"
|
22 |
+
},
|
23 |
+
"outputs": [],
|
24 |
+
"source": [
|
25 |
+
"import os\n",
|
26 |
+
"import random\n",
|
27 |
+
"import uuid\n",
|
28 |
+
"import json\n",
|
29 |
+
"import time\n",
|
30 |
+
"import asyncio\n",
|
31 |
+
"from threading import Thread\n",
|
32 |
+
"\n",
|
33 |
+
"import gradio as gr\n",
|
34 |
+
"import spaces\n",
|
35 |
+
"import torch\n",
|
36 |
+
"import numpy as np\n",
|
37 |
+
"from PIL import Image\n",
|
38 |
+
"import cv2\n",
|
39 |
+
"\n",
|
40 |
+
"from transformers import (\n",
|
41 |
+
" Qwen2_5_VLForConditionalGeneration,\n",
|
42 |
+
" AutoProcessor,\n",
|
43 |
+
" TextIteratorStreamer,\n",
|
44 |
+
")\n",
|
45 |
+
"from transformers.image_utils import load_image\n",
|
46 |
+
"\n",
|
47 |
+
"# Constants for text generation\n",
|
48 |
+
"MAX_MAX_NEW_TOKENS = 2048\n",
|
49 |
+
"DEFAULT_MAX_NEW_TOKENS = 1024\n",
|
50 |
+
"# Increase or disable input truncation to avoid token mismatches\n",
|
51 |
+
"MAX_INPUT_TOKEN_LENGTH = int(os.getenv(\"MAX_INPUT_TOKEN_LENGTH\", \"8192\"))\n",
|
52 |
+
"\n",
|
53 |
+
"device = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n",
|
54 |
+
"\n",
|
55 |
+
"MODEL_ID = \"scb10x/typhoon-ocr-7b\"\n",
|
56 |
+
"processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)\n",
|
57 |
+
"model_m = Qwen2_5_VLForConditionalGeneration.from_pretrained(\n",
|
58 |
+
" MODEL_ID,\n",
|
59 |
+
" trust_remote_code=True,\n",
|
60 |
+
" torch_dtype=torch.float16\n",
|
61 |
+
").to(\"cuda\").eval()\n",
|
62 |
+
"\n",
|
63 |
+
"def downsample_video(video_path):\n",
|
64 |
+
" \"\"\"\n",
|
65 |
+
" Downsamples the video to evenly spaced frames.\n",
|
66 |
+
" Each frame is returned as a PIL image along with its timestamp.\n",
|
67 |
+
" \"\"\"\n",
|
68 |
+
" vidcap = cv2.VideoCapture(video_path)\n",
|
69 |
+
" total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))\n",
|
70 |
+
" fps = vidcap.get(cv2.CAP_PROP_FPS)\n",
|
71 |
+
" frames = []\n",
|
72 |
+
" # Sample 10 evenly spaced frames.\n",
|
73 |
+
" frame_indices = np.linspace(0, total_frames - 1, 10, dtype=int)\n",
|
74 |
+
" for i in frame_indices:\n",
|
75 |
+
" vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)\n",
|
76 |
+
" success, image = vidcap.read()\n",
|
77 |
+
" if success:\n",
|
78 |
+
" image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) # Convert BGR to RGB\n",
|
79 |
+
" pil_image = Image.fromarray(image)\n",
|
80 |
+
" timestamp = round(i / fps, 2)\n",
|
81 |
+
" frames.append((pil_image, timestamp))\n",
|
82 |
+
" vidcap.release()\n",
|
83 |
+
" return frames\n",
|
84 |
+
"\n",
|
85 |
+
"@spaces.GPU\n",
|
86 |
+
"def generate_image(text: str, image: Image.Image,\n",
|
87 |
+
" max_new_tokens: int = 1024,\n",
|
88 |
+
" temperature: float = 0.6,\n",
|
89 |
+
" top_p: float = 0.9,\n",
|
90 |
+
" top_k: int = 50,\n",
|
91 |
+
" repetition_penalty: float = 1.2):\n",
|
92 |
+
"\n",
|
93 |
+
" if image is None:\n",
|
94 |
+
" yield \"Please upload an image.\"\n",
|
95 |
+
" return\n",
|
96 |
+
"\n",
|
97 |
+
" messages = [{\n",
|
98 |
+
" \"role\": \"user\",\n",
|
99 |
+
" \"content\": [\n",
|
100 |
+
" {\"type\": \"image\", \"image\": image},\n",
|
101 |
+
" {\"type\": \"text\", \"text\": text},\n",
|
102 |
+
" ]\n",
|
103 |
+
" }]\n",
|
104 |
+
" prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)\n",
|
105 |
+
" inputs = processor(\n",
|
106 |
+
" text=[prompt_full],\n",
|
107 |
+
" images=[image],\n",
|
108 |
+
" return_tensors=\"pt\",\n",
|
109 |
+
" padding=True,\n",
|
110 |
+
" truncation=False,\n",
|
111 |
+
" max_length=MAX_INPUT_TOKEN_LENGTH\n",
|
112 |
+
" ).to(\"cuda\")\n",
|
113 |
+
" streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)\n",
|
114 |
+
" generation_kwargs = {**inputs, \"streamer\": streamer, \"max_new_tokens\": max_new_tokens}\n",
|
115 |
+
" thread = Thread(target=model_m.generate, kwargs=generation_kwargs)\n",
|
116 |
+
" thread.start()\n",
|
117 |
+
" buffer = \"\"\n",
|
118 |
+
" for new_text in streamer:\n",
|
119 |
+
" buffer += new_text\n",
|
120 |
+
" buffer = buffer.replace(\"<|im_end|>\", \"\")\n",
|
121 |
+
" time.sleep(0.01)\n",
|
122 |
+
" yield buffer\n",
|
123 |
+
"\n",
|
124 |
+
"@spaces.GPU\n",
|
125 |
+
"def generate_video(text: str, video_path: str,\n",
|
126 |
+
" max_new_tokens: int = 1024,\n",
|
127 |
+
" temperature: float = 0.6,\n",
|
128 |
+
" top_p: float = 0.9,\n",
|
129 |
+
" top_k: int = 50,\n",
|
130 |
+
" repetition_penalty: float = 1.2):\n",
|
131 |
+
"\n",
|
132 |
+
" if video_path is None:\n",
|
133 |
+
" yield \"Please upload a video.\"\n",
|
134 |
+
" return\n",
|
135 |
+
"\n",
|
136 |
+
" frames = downsample_video(video_path)\n",
|
137 |
+
" messages = [\n",
|
138 |
+
" {\"role\": \"system\", \"content\": [{\"type\": \"text\", \"text\": \"You are a helpful assistant.\"}]},\n",
|
139 |
+
" {\"role\": \"user\", \"content\": [{\"type\": \"text\", \"text\": text}]}\n",
|
140 |
+
" ]\n",
|
141 |
+
" # Append each frame with its timestamp.\n",
|
142 |
+
" for frame in frames:\n",
|
143 |
+
" image, timestamp = frame\n",
|
144 |
+
" messages[1][\"content\"].append({\"type\": \"text\", \"text\": f\"Frame {timestamp}:\"})\n",
|
145 |
+
" messages[1][\"content\"].append({\"type\": \"image\", \"image\": image})\n",
|
146 |
+
" inputs = processor.apply_chat_template(\n",
|
147 |
+
" messages,\n",
|
148 |
+
" tokenize=True,\n",
|
149 |
+
" add_generation_prompt=True,\n",
|
150 |
+
" return_dict=True,\n",
|
151 |
+
" return_tensors=\"pt\",\n",
|
152 |
+
" truncation=False,\n",
|
153 |
+
" max_length=MAX_INPUT_TOKEN_LENGTH\n",
|
154 |
+
" ).to(\"cuda\")\n",
|
155 |
+
" streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)\n",
|
156 |
+
" generation_kwargs = {\n",
|
157 |
+
" **inputs,\n",
|
158 |
+
" \"streamer\": streamer,\n",
|
159 |
+
" \"max_new_tokens\": max_new_tokens,\n",
|
160 |
+
" \"do_sample\": True,\n",
|
161 |
+
" \"temperature\": temperature,\n",
|
162 |
+
" \"top_p\": top_p,\n",
|
163 |
+
" \"top_k\": top_k,\n",
|
164 |
+
" \"repetition_penalty\": repetition_penalty,\n",
|
165 |
+
" }\n",
|
166 |
+
" thread = Thread(target=model_m.generate, kwargs=generation_kwargs)\n",
|
167 |
+
" thread.start()\n",
|
168 |
+
" buffer = \"\"\n",
|
169 |
+
" for new_text in streamer:\n",
|
170 |
+
" buffer += new_text\n",
|
171 |
+
" buffer = buffer.replace(\"<|im_end|>\", \"\")\n",
|
172 |
+
" time.sleep(0.01)\n",
|
173 |
+
" yield buffer\n",
|
174 |
+
"\n",
|
175 |
+
"css = \"\"\"\n",
|
176 |
+
".submit-btn {\n",
|
177 |
+
" background-color: #2980b9 !important;\n",
|
178 |
+
" color: white !important;\n",
|
179 |
+
"}\n",
|
180 |
+
".submit-btn:hover {\n",
|
181 |
+
" background-color: #3498db !important;\n",
|
182 |
+
"}\n",
|
183 |
+
"\"\"\"\n",
|
184 |
+
"\n",
|
185 |
+
"# Create the Gradio Interface\n",
|
186 |
+
"with gr.Blocks(css=css, theme=\"bethecloud/storj_theme\") as demo:\n",
|
187 |
+
" gr.Markdown(\"# **typhoon-ocr-7b**\")\n",
|
188 |
+
" with gr.Row():\n",
|
189 |
+
" with gr.Column():\n",
|
190 |
+
" with gr.Tabs():\n",
|
191 |
+
" with gr.TabItem(\"Image Inference\"):\n",
|
192 |
+
" image_query = gr.Textbox(label=\"Query Input\", placeholder=\"Enter your query here...\")\n",
|
193 |
+
" image_upload = gr.Image(type=\"pil\", label=\"Image\")\n",
|
194 |
+
" image_submit = gr.Button(\"Submit\", elem_classes=\"submit-btn\")\n",
|
195 |
+
"\n",
|
196 |
+
" with gr.TabItem(\"Video Inference\"):\n",
|
197 |
+
" video_query = gr.Textbox(label=\"Query Input\", placeholder=\"Enter your query here...\")\n",
|
198 |
+
" video_upload = gr.Video(label=\"Video\")\n",
|
199 |
+
" video_submit = gr.Button(\"Submit\", elem_classes=\"submit-btn\")\n",
|
200 |
+
"\n",
|
201 |
+
" with gr.Accordion(\"Advanced options\", open=False):\n",
|
202 |
+
" max_new_tokens = gr.Slider(label=\"Max new tokens\", minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1, value=DEFAULT_MAX_NEW_TOKENS)\n",
|
203 |
+
" temperature = gr.Slider(label=\"Temperature\", minimum=0.1, maximum=4.0, step=0.1, value=0.6)\n",
|
204 |
+
" top_p = gr.Slider(label=\"Top-p (nucleus sampling)\", minimum=0.05, maximum=1.0, step=0.05, value=0.9)\n",
|
205 |
+
" top_k = gr.Slider(label=\"Top-k\", minimum=1, maximum=1000, step=1, value=50)\n",
|
206 |
+
" repetition_penalty = gr.Slider(label=\"Repetition penalty\", minimum=1.0, maximum=2.0, step=0.05, value=1.2)\n",
|
207 |
+
" with gr.Column():\n",
|
208 |
+
" output = gr.Textbox(label=\"Output\", interactive=False)\n",
|
209 |
+
"\n",
|
210 |
+
" image_submit.click(\n",
|
211 |
+
" fn=generate_image,\n",
|
212 |
+
" inputs=[image_query, image_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],\n",
|
213 |
+
" outputs=output\n",
|
214 |
+
" )\n",
|
215 |
+
" video_submit.click(\n",
|
216 |
+
" fn=generate_video,\n",
|
217 |
+
" inputs=[video_query, video_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],\n",
|
218 |
+
" outputs=output\n",
|
219 |
+
" )\n",
|
220 |
+
"\n",
|
221 |
+
"if __name__ == \"__main__\":\n",
|
222 |
+
" demo.queue(max_size=30).launch(share=True, ssr_mode=False, show_error=True)"
|
223 |
+
]
|
224 |
+
}
|
225 |
+
],
|
226 |
+
"metadata": {
|
227 |
+
"accelerator": "GPU",
|
228 |
+
"colab": {
|
229 |
+
"gpuType": "T4",
|
230 |
+
"provenance": []
|
231 |
+
},
|
232 |
+
"kernelspec": {
|
233 |
+
"display_name": "Python 3",
|
234 |
+
"name": "python3"
|
235 |
+
},
|
236 |
+
"language_info": {
|
237 |
+
"name": "python"
|
238 |
+
}
|
239 |
+
},
|
240 |
+
"nbformat": 4,
|
241 |
+
"nbformat_minor": 0
|
242 |
+
}
|