Upload joy_caption_jupyter (1).ipynb
Browse files- joy_caption_jupyter (1).ipynb +149 -0
joy_caption_jupyter (1).ipynb
ADDED
@@ -0,0 +1,149 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cells": [
|
3 |
+
{
|
4 |
+
"cell_type": "code",
|
5 |
+
"execution_count": null,
|
6 |
+
"metadata": {
|
7 |
+
"id": "VjYy0F2gZIPR"
|
8 |
+
},
|
9 |
+
"outputs": [],
|
10 |
+
"source": [
|
11 |
+
"!pip install gradio bitsandbytes transformers==4.43.3\n",
|
12 |
+
"\n",
|
13 |
+
"!wget https://huggingface.co/spaces/fancyfeast/joy-caption-pre-alpha/resolve/main/wpkklhc6/image_adapter.pt -O /content/image_adapter.pt\n",
|
14 |
+
"!wget https://huggingface.co/spaces/fancyfeast/joy-caption-pre-alpha/raw/main/wpkklhc6/config.yaml -O /content/config.yaml\n",
|
15 |
+
"\n",
|
16 |
+
"import gradio as gr\n",
|
17 |
+
"from huggingface_hub import InferenceClient\n",
|
18 |
+
"from torch import nn\n",
|
19 |
+
"from transformers import AutoModel, AutoProcessor, AutoTokenizer, PreTrainedTokenizer, PreTrainedTokenizerFast, AutoModelForCausalLM\n",
|
20 |
+
"from pathlib import Path\n",
|
21 |
+
"import torch\n",
|
22 |
+
"import torch.amp.autocast_mode\n",
|
23 |
+
"from PIL import Image\n",
|
24 |
+
"import os\n",
|
25 |
+
"\n",
|
26 |
+
"CLIP_PATH = \"google/siglip-so400m-patch14-384\"\n",
|
27 |
+
"VLM_PROMPT = \"A descriptive caption for this image:\\n\"\n",
|
28 |
+
"# MODEL_PATH = \"unsloth/Meta-Llama-3.1-8B\"\n",
|
29 |
+
"MODEL_PATH = \"unsloth/Meta-Llama-3.1-8B-bnb-4bit\"\n",
|
30 |
+
"CHECKPOINT_PATH = Path(\"wpkklhc6\")\n",
|
31 |
+
"\n",
|
32 |
+
"class ImageAdapter(nn.Module):\n",
|
33 |
+
"\tdef __init__(self, input_features: int, output_features: int):\n",
|
34 |
+
"\t\tsuper().__init__()\n",
|
35 |
+
"\t\tself.linear1 = nn.Linear(input_features, output_features)\n",
|
36 |
+
"\t\tself.activation = nn.GELU()\n",
|
37 |
+
"\t\tself.linear2 = nn.Linear(output_features, output_features)\n",
|
38 |
+
"\n",
|
39 |
+
"\tdef forward(self, vision_outputs: torch.Tensor):\n",
|
40 |
+
"\t\tx = self.linear1(vision_outputs)\n",
|
41 |
+
"\t\tx = self.activation(x)\n",
|
42 |
+
"\t\tx = self.linear2(x)\n",
|
43 |
+
"\t\treturn x\n",
|
44 |
+
"\n",
|
45 |
+
"# Load CLIP\n",
|
46 |
+
"print(\"Loading CLIP\")\n",
|
47 |
+
"clip_processor = AutoProcessor.from_pretrained(CLIP_PATH)\n",
|
48 |
+
"clip_model = AutoModel.from_pretrained(CLIP_PATH)\n",
|
49 |
+
"clip_model = clip_model.vision_model\n",
|
50 |
+
"clip_model.eval()\n",
|
51 |
+
"clip_model.requires_grad_(False)\n",
|
52 |
+
"clip_model.to(\"cuda\")\n",
|
53 |
+
"\n",
|
54 |
+
"# Tokenizer\n",
|
55 |
+
"print(\"Loading tokenizer\")\n",
|
56 |
+
"tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, load_in_4bit=True, use_fast=False)\n",
|
57 |
+
"assert isinstance(tokenizer, PreTrainedTokenizer) or isinstance(tokenizer, PreTrainedTokenizerFast), f\"Tokenizer is of type {type(tokenizer)}\"\n",
|
58 |
+
"\n",
|
59 |
+
"# LLM\n",
|
60 |
+
"print(\"Loading LLM\")\n",
|
61 |
+
"text_model = AutoModelForCausalLM.from_pretrained(MODEL_PATH, load_in_4bit=True, device_map=\"auto\", torch_dtype=torch.float16)\n",
|
62 |
+
"text_model.eval()\n",
|
63 |
+
"\n",
|
64 |
+
"# Image Adapter\n",
|
65 |
+
"print(\"Loading image adapter\")\n",
|
66 |
+
"image_adapter = ImageAdapter(clip_model.config.hidden_size, text_model.config.hidden_size)\n",
|
67 |
+
"image_adapter.load_state_dict(torch.load(\"/content/image_adapter.pt\", map_location=\"cpu\"))\n",
|
68 |
+
"image_adapter.eval()\n",
|
69 |
+
"image_adapter.to(\"cuda\")\n",
|
70 |
+
"\n",
|
71 |
+
"@torch.inference_mode()\n",
|
72 |
+
"def stream_chat(input_image: Image.Image):\n",
|
73 |
+
"\ttorch.cuda.empty_cache()\n",
|
74 |
+
"\n",
|
75 |
+
"\t# Preprocess image\n",
|
76 |
+
"\timage = clip_processor(images=input_image, return_tensors='pt').pixel_values\n",
|
77 |
+
"\timage = image.to('cuda')\n",
|
78 |
+
"\n",
|
79 |
+
"\t# Tokenize the prompt\n",
|
80 |
+
"\tprompt = tokenizer.encode(VLM_PROMPT, return_tensors='pt', padding=False, truncation=False, add_special_tokens=False)\n",
|
81 |
+
"\n",
|
82 |
+
"\t# Embed image\n",
|
83 |
+
"\twith torch.amp.autocast_mode.autocast('cuda', enabled=True):\n",
|
84 |
+
"\t\tvision_outputs = clip_model(pixel_values=image, output_hidden_states=True)\n",
|
85 |
+
"\t\timage_features = vision_outputs.hidden_states[-2]\n",
|
86 |
+
"\t\tembedded_images = image_adapter(image_features)\n",
|
87 |
+
"\t\tembedded_images = embedded_images.to('cuda')\n",
|
88 |
+
"\n",
|
89 |
+
"\t# Embed prompt\n",
|
90 |
+
"\tprompt_embeds = text_model.model.embed_tokens(prompt.to('cuda'))\n",
|
91 |
+
"\tassert prompt_embeds.shape == (1, prompt.shape[1], text_model.config.hidden_size), f\"Prompt shape is {prompt_embeds.shape}, expected {(1, prompt.shape[1], text_model.config.hidden_size)}\"\n",
|
92 |
+
"\tembedded_bos = text_model.model.embed_tokens(torch.tensor([[tokenizer.bos_token_id]], device=text_model.device, dtype=torch.int64))\n",
|
93 |
+
"\n",
|
94 |
+
"\t# Construct prompts\n",
|
95 |
+
"\tinputs_embeds = torch.cat([\n",
|
96 |
+
"\t\tembedded_bos.expand(embedded_images.shape[0], -1, -1),\n",
|
97 |
+
"\t\tembedded_images.to(dtype=embedded_bos.dtype),\n",
|
98 |
+
"\t\tprompt_embeds.expand(embedded_images.shape[0], -1, -1),\n",
|
99 |
+
"\t], dim=1)\n",
|
100 |
+
"\n",
|
101 |
+
"\tinput_ids = torch.cat([\n",
|
102 |
+
"\t\ttorch.tensor([[tokenizer.bos_token_id]], dtype=torch.long),\n",
|
103 |
+
"\t\ttorch.zeros((1, embedded_images.shape[1]), dtype=torch.long),\n",
|
104 |
+
"\t\tprompt,\n",
|
105 |
+
"\t], dim=1).to('cuda')\n",
|
106 |
+
"\tattention_mask = torch.ones_like(input_ids)\n",
|
107 |
+
"\n",
|
108 |
+
"\t#generate_ids = text_model.generate(input_ids, inputs_embeds=inputs_embeds, attention_mask=attention_mask, max_new_tokens=300, do_sample=False, suppress_tokens=None)\n",
|
109 |
+
"\tgenerate_ids = text_model.generate(input_ids, inputs_embeds=inputs_embeds, attention_mask=attention_mask, max_new_tokens=300, do_sample=True, top_k=10, temperature=0.5, suppress_tokens=None)\n",
|
110 |
+
"\n",
|
111 |
+
"\t# Trim off the prompt\n",
|
112 |
+
"\tgenerate_ids = generate_ids[:, input_ids.shape[1]:]\n",
|
113 |
+
"\tif generate_ids[0][-1] == tokenizer.eos_token_id:\n",
|
114 |
+
"\t\tgenerate_ids = generate_ids[:, :-1]\n",
|
115 |
+
"\n",
|
116 |
+
"\tcaption = tokenizer.batch_decode(generate_ids, skip_special_tokens=False, clean_up_tokenization_spaces=False)[0]\n",
|
117 |
+
"\n",
|
118 |
+
"\treturn caption.strip()\n",
|
119 |
+
"\n",
|
120 |
+
"\n",
|
121 |
+
"with gr.Blocks(css=\".gradio-container {max-width: 544px !important}\", analytics_enabled=False) as demo:\n",
|
122 |
+
"\twith gr.Row():\n",
|
123 |
+
"\t\twith gr.Column():\n",
|
124 |
+
"\t\t\tinput_image = gr.Image(type=\"pil\", label=\"Input Image\")\n",
|
125 |
+
"\t\t\trun_button = gr.Button(\"Caption\")\n",
|
126 |
+
"\t\t\toutput_caption = gr.Textbox(label=\"Caption\")\n",
|
127 |
+
"\trun_button.click(fn=stream_chat, inputs=[input_image], outputs=[output_caption])\n",
|
128 |
+
"\n",
|
129 |
+
"demo.queue().launch(share=True, inline=False, debug=True)"
|
130 |
+
]
|
131 |
+
}
|
132 |
+
],
|
133 |
+
"metadata": {
|
134 |
+
"accelerator": "GPU",
|
135 |
+
"colab": {
|
136 |
+
"gpuType": "T4",
|
137 |
+
"provenance": []
|
138 |
+
},
|
139 |
+
"kernelspec": {
|
140 |
+
"display_name": "Python 3",
|
141 |
+
"name": "python3"
|
142 |
+
},
|
143 |
+
"language_info": {
|
144 |
+
"name": "python"
|
145 |
+
}
|
146 |
+
},
|
147 |
+
"nbformat": 4,
|
148 |
+
"nbformat_minor": 0
|
149 |
+
}
|