Spaces:
Sleeping
Sleeping
Commit
·
ae365ab
0
Parent(s):
fix: Update Gradio to 4.44.1 and remove example images
Browse files- .gitattributes +35 -0
- .gitignore +60 -0
- README.md +52 -0
- app.py +148 -0
- models/__init__.py +3 -0
- models/llava.py +312 -0
- requirements.txt +26 -0
- utils/__init__.py +37 -0
- utils/data_utils.py +175 -0
- utils/eval_utils.py +180 -0
- utils/visualization.py +178 -0
.gitattributes
ADDED
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
*.7z filter=lfs diff=lfs merge=lfs -text
|
2 |
+
*.arrow filter=lfs diff=lfs merge=lfs -text
|
3 |
+
*.bin filter=lfs diff=lfs merge=lfs -text
|
4 |
+
*.bz2 filter=lfs diff=lfs merge=lfs -text
|
5 |
+
*.ckpt filter=lfs diff=lfs merge=lfs -text
|
6 |
+
*.ftz filter=lfs diff=lfs merge=lfs -text
|
7 |
+
*.gz filter=lfs diff=lfs merge=lfs -text
|
8 |
+
*.h5 filter=lfs diff=lfs merge=lfs -text
|
9 |
+
*.joblib filter=lfs diff=lfs merge=lfs -text
|
10 |
+
*.lfs.* filter=lfs diff=lfs merge=lfs -text
|
11 |
+
*.mlmodel filter=lfs diff=lfs merge=lfs -text
|
12 |
+
*.model filter=lfs diff=lfs merge=lfs -text
|
13 |
+
*.msgpack filter=lfs diff=lfs merge=lfs -text
|
14 |
+
*.npy filter=lfs diff=lfs merge=lfs -text
|
15 |
+
*.npz filter=lfs diff=lfs merge=lfs -text
|
16 |
+
*.onnx filter=lfs diff=lfs merge=lfs -text
|
17 |
+
*.ot filter=lfs diff=lfs merge=lfs -text
|
18 |
+
*.parquet filter=lfs diff=lfs merge=lfs -text
|
19 |
+
*.pb filter=lfs diff=lfs merge=lfs -text
|
20 |
+
*.pickle filter=lfs diff=lfs merge=lfs -text
|
21 |
+
*.pkl filter=lfs diff=lfs merge=lfs -text
|
22 |
+
*.pt filter=lfs diff=lfs merge=lfs -text
|
23 |
+
*.pth filter=lfs diff=lfs merge=lfs -text
|
24 |
+
*.rar filter=lfs diff=lfs merge=lfs -text
|
25 |
+
*.safetensors filter=lfs diff=lfs merge=lfs -text
|
26 |
+
saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
27 |
+
*.tar.* filter=lfs diff=lfs merge=lfs -text
|
28 |
+
*.tar filter=lfs diff=lfs merge=lfs -text
|
29 |
+
*.tflite filter=lfs diff=lfs merge=lfs -text
|
30 |
+
*.tgz filter=lfs diff=lfs merge=lfs -text
|
31 |
+
*.wasm filter=lfs diff=lfs merge=lfs -text
|
32 |
+
*.xz filter=lfs diff=lfs merge=lfs -text
|
33 |
+
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
+
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
+
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
.gitignore
ADDED
@@ -0,0 +1,60 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Python
|
2 |
+
__pycache__/
|
3 |
+
*.py[cod]
|
4 |
+
*$py.class
|
5 |
+
*.so
|
6 |
+
.Python
|
7 |
+
env/
|
8 |
+
build/
|
9 |
+
develop-eggs/
|
10 |
+
dist/
|
11 |
+
downloads/
|
12 |
+
eggs/
|
13 |
+
.eggs/
|
14 |
+
lib/
|
15 |
+
lib64/
|
16 |
+
parts/
|
17 |
+
sdist/
|
18 |
+
var/
|
19 |
+
wheels/
|
20 |
+
*.egg-info/
|
21 |
+
.installed.cfg
|
22 |
+
*.egg
|
23 |
+
|
24 |
+
# Virtual Environment
|
25 |
+
venv/
|
26 |
+
ENV/
|
27 |
+
|
28 |
+
# IDE
|
29 |
+
.idea/
|
30 |
+
.vscode/
|
31 |
+
*.swp
|
32 |
+
*.swo
|
33 |
+
|
34 |
+
# OS
|
35 |
+
.DS_Store
|
36 |
+
Thumbs.db
|
37 |
+
|
38 |
+
# Model files
|
39 |
+
*.bin
|
40 |
+
*.pt
|
41 |
+
*.pth
|
42 |
+
*.ckpt
|
43 |
+
*.safetensors
|
44 |
+
|
45 |
+
# Logs
|
46 |
+
*.log
|
47 |
+
logs/
|
48 |
+
|
49 |
+
# Frontend
|
50 |
+
frontend/
|
51 |
+
node_modules/
|
52 |
+
npm-debug.log*
|
53 |
+
yarn-debug.log*
|
54 |
+
yarn-error.log*
|
55 |
+
|
56 |
+
# Temporary files
|
57 |
+
*.tmp
|
58 |
+
*.temp
|
59 |
+
temp/
|
60 |
+
tmp/
|
README.md
ADDED
@@ -0,0 +1,52 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
title: LLaVA Chat
|
3 |
+
emoji: 🖼️
|
4 |
+
colorFrom: blue
|
5 |
+
colorTo: indigo
|
6 |
+
sdk: gradio
|
7 |
+
sdk_version: 3.50.2
|
8 |
+
app_file: app.py
|
9 |
+
pinned: false
|
10 |
+
license: mit
|
11 |
+
---
|
12 |
+
|
13 |
+
# LLaVA Chat
|
14 |
+
|
15 |
+
A powerful multimodal AI assistant that can understand and discuss images. Upload any image and chat with LLaVA about it!
|
16 |
+
|
17 |
+
## Features
|
18 |
+
|
19 |
+
- 🖼️ Upload and analyze any image
|
20 |
+
- 💬 Natural conversation about image content
|
21 |
+
- ⚙️ Adjustable generation parameters
|
22 |
+
- 🎯 High-quality image understanding
|
23 |
+
- 🚀 Fast and responsive interface
|
24 |
+
|
25 |
+
## How to Use
|
26 |
+
|
27 |
+
1. Upload an image using the image uploader
|
28 |
+
2. Type your question or prompt about the image
|
29 |
+
3. (Optional) Adjust the generation parameters:
|
30 |
+
- Max New Tokens: Control response length
|
31 |
+
- Temperature: Adjust response creativity
|
32 |
+
- Top P: Fine-tune response diversity
|
33 |
+
4. Click "Generate Response" to get LLaVA's analysis
|
34 |
+
|
35 |
+
## Example Prompts
|
36 |
+
|
37 |
+
- "What can you see in this image?"
|
38 |
+
- "Describe this scene in detail"
|
39 |
+
- "What emotions does this image convey?"
|
40 |
+
- "What's happening in this picture?"
|
41 |
+
- "Can you identify any objects or people in this image?"
|
42 |
+
|
43 |
+
## Model Details
|
44 |
+
|
45 |
+
This Space uses the LLaVA (Large Language and Vision Assistant) model, which combines:
|
46 |
+
- CLIP ViT-L/14 vision encoder
|
47 |
+
- Vicuna-7B language model
|
48 |
+
- Advanced multimodal understanding capabilities
|
49 |
+
|
50 |
+
## License
|
51 |
+
|
52 |
+
This project is licensed under the MIT License.
|
app.py
ADDED
@@ -0,0 +1,148 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from fastapi import FastAPI, UploadFile, File, HTTPException
|
2 |
+
from fastapi.middleware.cors import CORSMiddleware
|
3 |
+
from fastapi.responses import JSONResponse
|
4 |
+
import os
|
5 |
+
import tempfile
|
6 |
+
from typing import Optional
|
7 |
+
from pydantic import BaseModel
|
8 |
+
import torch
|
9 |
+
import gradio as gr
|
10 |
+
from models.llava import LLaVA
|
11 |
+
|
12 |
+
# Initialize model globally
|
13 |
+
model = None
|
14 |
+
|
15 |
+
def initialize_model():
|
16 |
+
global model
|
17 |
+
try:
|
18 |
+
model = LLaVA(
|
19 |
+
vision_model_path="openai/clip-vit-large-patch14-336",
|
20 |
+
language_model_path="lmsys/vicuna-7b-v1.5",
|
21 |
+
device="cuda" if torch.cuda.is_available() else "cpu",
|
22 |
+
load_in_8bit=True
|
23 |
+
)
|
24 |
+
print(f"Model initialized on {model.device}")
|
25 |
+
return True
|
26 |
+
except Exception as e:
|
27 |
+
print(f"Error initializing model: {e}")
|
28 |
+
return False
|
29 |
+
|
30 |
+
def process_image(image, prompt, max_new_tokens=256, temperature=0.7, top_p=0.9):
|
31 |
+
if not model:
|
32 |
+
return "Error: Model not initialized"
|
33 |
+
|
34 |
+
try:
|
35 |
+
# Save the uploaded image temporarily
|
36 |
+
with tempfile.NamedTemporaryFile(delete=False, suffix='.png') as temp_file:
|
37 |
+
image.save(temp_file.name)
|
38 |
+
temp_path = temp_file.name
|
39 |
+
|
40 |
+
# Generate response
|
41 |
+
response = model.generate_from_image(
|
42 |
+
image_path=temp_path,
|
43 |
+
prompt=prompt,
|
44 |
+
max_new_tokens=max_new_tokens,
|
45 |
+
temperature=temperature,
|
46 |
+
top_p=top_p
|
47 |
+
)
|
48 |
+
|
49 |
+
# Clean up temporary file
|
50 |
+
os.unlink(temp_path)
|
51 |
+
return response
|
52 |
+
|
53 |
+
except Exception as e:
|
54 |
+
return f"Error processing image: {str(e)}"
|
55 |
+
|
56 |
+
# Create Gradio interface
|
57 |
+
def create_interface():
|
58 |
+
with gr.Blocks(title="LLaVA Chat", theme=gr.themes.Soft()) as demo:
|
59 |
+
gr.Markdown("""
|
60 |
+
# LLaVA Chat
|
61 |
+
Upload an image and chat with LLaVA about it. This model can understand and describe images, answer questions about them, and engage in visual conversations.
|
62 |
+
""")
|
63 |
+
|
64 |
+
with gr.Row():
|
65 |
+
with gr.Column(scale=1):
|
66 |
+
image_input = gr.Image(type="pil", label="Upload Image")
|
67 |
+
prompt_input = gr.Textbox(
|
68 |
+
label="Ask about the image",
|
69 |
+
placeholder="What can you see in this image?",
|
70 |
+
lines=3
|
71 |
+
)
|
72 |
+
|
73 |
+
with gr.Accordion("Advanced Settings", open=False):
|
74 |
+
max_tokens = gr.Slider(
|
75 |
+
minimum=32,
|
76 |
+
maximum=512,
|
77 |
+
value=256,
|
78 |
+
step=32,
|
79 |
+
label="Max New Tokens"
|
80 |
+
)
|
81 |
+
temperature = gr.Slider(
|
82 |
+
minimum=0.1,
|
83 |
+
maximum=1.0,
|
84 |
+
value=0.7,
|
85 |
+
step=0.1,
|
86 |
+
label="Temperature"
|
87 |
+
)
|
88 |
+
top_p = gr.Slider(
|
89 |
+
minimum=0.1,
|
90 |
+
maximum=1.0,
|
91 |
+
value=0.9,
|
92 |
+
step=0.1,
|
93 |
+
label="Top P"
|
94 |
+
)
|
95 |
+
|
96 |
+
submit_btn = gr.Button("Generate Response", variant="primary")
|
97 |
+
|
98 |
+
with gr.Column(scale=1):
|
99 |
+
output = gr.Textbox(
|
100 |
+
label="Model Response",
|
101 |
+
lines=10,
|
102 |
+
show_copy_button=True
|
103 |
+
)
|
104 |
+
|
105 |
+
# Set up the submit action
|
106 |
+
submit_btn.click(
|
107 |
+
fn=process_image,
|
108 |
+
inputs=[image_input, prompt_input, max_tokens, temperature, top_p],
|
109 |
+
outputs=output
|
110 |
+
)
|
111 |
+
|
112 |
+
# Add examples
|
113 |
+
gr.Examples(
|
114 |
+
examples=[
|
115 |
+
["examples/cat.jpg", "What can you see in this image?"],
|
116 |
+
["examples/landscape.jpg", "Describe this scene in detail."],
|
117 |
+
["examples/food.jpg", "What kind of food is this and how would you describe it?"]
|
118 |
+
],
|
119 |
+
inputs=[image_input, prompt_input]
|
120 |
+
)
|
121 |
+
|
122 |
+
return demo
|
123 |
+
|
124 |
+
# Create FastAPI app
|
125 |
+
app = FastAPI(title="LLaVA Web Interface")
|
126 |
+
|
127 |
+
# Configure CORS
|
128 |
+
app.add_middleware(
|
129 |
+
CORSMiddleware,
|
130 |
+
allow_origins=["*"],
|
131 |
+
allow_credentials=True,
|
132 |
+
allow_methods=["*"],
|
133 |
+
allow_headers=["*"],
|
134 |
+
)
|
135 |
+
|
136 |
+
# Create Gradio app
|
137 |
+
demo = create_interface()
|
138 |
+
|
139 |
+
# Mount Gradio app
|
140 |
+
app = gr.mount_gradio_app(app, demo, path="/")
|
141 |
+
|
142 |
+
if __name__ == "__main__":
|
143 |
+
# Initialize model
|
144 |
+
if initialize_model():
|
145 |
+
import uvicorn
|
146 |
+
uvicorn.run(app, host="0.0.0.0", port=7860) # Hugging Face Spaces uses port 7860
|
147 |
+
else:
|
148 |
+
print("Failed to initialize model. Exiting...")
|
models/__init__.py
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
from .llava import LLaVA, MLP, StoppingCriteriaSub
|
2 |
+
|
3 |
+
__all__ = ['LLaVA', 'MLP', 'StoppingCriteriaSub']
|
models/llava.py
ADDED
@@ -0,0 +1,312 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
LLaVA: Large Language and Vision Assistant
|
3 |
+
Implementation based on the paper "Visual Instruction Tuning" (NeurIPS 2023)
|
4 |
+
"""
|
5 |
+
|
6 |
+
import torch
|
7 |
+
import torch.nn as nn
|
8 |
+
from transformers import (
|
9 |
+
CLIPVisionModel,
|
10 |
+
CLIPImageProcessor,
|
11 |
+
AutoTokenizer,
|
12 |
+
AutoModelForCausalLM,
|
13 |
+
StoppingCriteria,
|
14 |
+
StoppingCriteriaList
|
15 |
+
)
|
16 |
+
from PIL import Image
|
17 |
+
import os
|
18 |
+
from typing import List, Dict, Optional, Tuple, Union
|
19 |
+
|
20 |
+
|
21 |
+
class StoppingCriteriaSub(StoppingCriteria):
|
22 |
+
"""Custom stopping criteria for text generation."""
|
23 |
+
|
24 |
+
def __init__(self, stops=None, encounters=1):
|
25 |
+
super().__init__()
|
26 |
+
self.stops = stops or []
|
27 |
+
self.encounters = encounters
|
28 |
+
self.counter = 0
|
29 |
+
|
30 |
+
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
|
31 |
+
"""Check if stopping criteria are met."""
|
32 |
+
for stop_id in self.stops:
|
33 |
+
if stop_id in input_ids[0][-1:]:
|
34 |
+
self.counter += 1
|
35 |
+
if self.counter >= self.encounters:
|
36 |
+
return True
|
37 |
+
return False
|
38 |
+
|
39 |
+
|
40 |
+
class MLP(nn.Module):
|
41 |
+
"""MLP projection layer to connect vision and language models."""
|
42 |
+
|
43 |
+
def __init__(self, input_dim: int, hidden_dim: int, output_dim: int, dropout_rate: float = 0.1):
|
44 |
+
super().__init__()
|
45 |
+
self.fc1 = nn.Linear(input_dim, hidden_dim)
|
46 |
+
self.fc2 = nn.Linear(hidden_dim, output_dim)
|
47 |
+
self.act = nn.GELU()
|
48 |
+
self.dropout = nn.Dropout(dropout_rate)
|
49 |
+
|
50 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
51 |
+
"""Forward pass through the MLP."""
|
52 |
+
x = self.fc1(x)
|
53 |
+
x = self.act(x)
|
54 |
+
x = self.dropout(x)
|
55 |
+
x = self.fc2(x)
|
56 |
+
return x
|
57 |
+
|
58 |
+
|
59 |
+
class LLaVA(nn.Module):
|
60 |
+
"""
|
61 |
+
LLaVA: Large Language and Vision Assistant
|
62 |
+
A multimodal model that connects a vision encoder with a language model.
|
63 |
+
"""
|
64 |
+
|
65 |
+
def __init__(
|
66 |
+
self,
|
67 |
+
vision_model_path: str = "openai/clip-vit-large-patch14-336",
|
68 |
+
language_model_path: str = "lmsys/vicuna-7b-v1.5",
|
69 |
+
projection_hidden_dim: int = 4096,
|
70 |
+
device: str = None,
|
71 |
+
load_in_8bit: bool = False,
|
72 |
+
load_in_4bit: bool = False,
|
73 |
+
):
|
74 |
+
"""
|
75 |
+
Initialize the LLaVA model.
|
76 |
+
|
77 |
+
Args:
|
78 |
+
vision_model_path: Path or name of the vision model
|
79 |
+
language_model_path: Path or name of the language model
|
80 |
+
projection_hidden_dim: Hidden dimension of the projection layer
|
81 |
+
device: Device to load the model on ('cuda', 'cpu', etc.)
|
82 |
+
load_in_8bit: Whether to load the language model in 8-bit precision
|
83 |
+
load_in_4bit: Whether to load the language model in 4-bit precision
|
84 |
+
"""
|
85 |
+
super().__init__()
|
86 |
+
|
87 |
+
self.device = device or ('cuda' if torch.cuda.is_available() else 'cpu')
|
88 |
+
|
89 |
+
# Load vision model
|
90 |
+
self.vision_model = CLIPVisionModel.from_pretrained(vision_model_path)
|
91 |
+
self.image_processor = CLIPImageProcessor.from_pretrained(vision_model_path)
|
92 |
+
|
93 |
+
# Load language model
|
94 |
+
kwargs = {}
|
95 |
+
if load_in_8bit:
|
96 |
+
kwargs['load_in_8bit'] = True
|
97 |
+
elif load_in_4bit:
|
98 |
+
kwargs['load_in_4bit'] = True
|
99 |
+
kwargs['bnb_4bit_compute_dtype'] = torch.float16
|
100 |
+
|
101 |
+
self.tokenizer = AutoTokenizer.from_pretrained(language_model_path)
|
102 |
+
self.language_model = AutoModelForCausalLM.from_pretrained(
|
103 |
+
language_model_path,
|
104 |
+
torch_dtype=torch.float16,
|
105 |
+
**kwargs
|
106 |
+
)
|
107 |
+
|
108 |
+
# Set padding token if not set
|
109 |
+
if self.tokenizer.pad_token is None:
|
110 |
+
self.tokenizer.pad_token = self.tokenizer.eos_token
|
111 |
+
|
112 |
+
# Get dimensions
|
113 |
+
vision_hidden_size = self.vision_model.config.hidden_size
|
114 |
+
language_hidden_size = self.language_model.config.hidden_size
|
115 |
+
|
116 |
+
# Create projection layer
|
117 |
+
self.projection = MLP(
|
118 |
+
input_dim=vision_hidden_size,
|
119 |
+
hidden_dim=projection_hidden_dim,
|
120 |
+
output_dim=language_hidden_size
|
121 |
+
)
|
122 |
+
|
123 |
+
# Move models to device
|
124 |
+
self.vision_model.to(self.device)
|
125 |
+
self.language_model.to(self.device)
|
126 |
+
self.projection.to(self.device)
|
127 |
+
|
128 |
+
# Set to evaluation mode
|
129 |
+
self.vision_model.eval()
|
130 |
+
self.language_model.eval()
|
131 |
+
self.projection.eval()
|
132 |
+
|
133 |
+
# Template for conversation
|
134 |
+
self.conv_template = [
|
135 |
+
{"role": "system", "content": "You are a helpful assistant that can understand images and answer questions about them."},
|
136 |
+
]
|
137 |
+
|
138 |
+
def encode_image(self, image_path: str) -> torch.Tensor:
|
139 |
+
"""
|
140 |
+
Encode an image using the vision model.
|
141 |
+
|
142 |
+
Args:
|
143 |
+
image_path: Path to the image file
|
144 |
+
|
145 |
+
Returns:
|
146 |
+
Tensor containing the image features
|
147 |
+
"""
|
148 |
+
image = Image.open(image_path).convert('RGB')
|
149 |
+
inputs = self.image_processor(images=image, return_tensors="pt").to(self.device)
|
150 |
+
|
151 |
+
with torch.no_grad():
|
152 |
+
outputs = self.vision_model(**inputs)
|
153 |
+
image_features = outputs.pooler_output # [1, hidden_size]
|
154 |
+
|
155 |
+
return image_features
|
156 |
+
|
157 |
+
def project_image_features(self, image_features: torch.Tensor) -> torch.Tensor:
|
158 |
+
"""
|
159 |
+
Project image features to the language model's embedding space.
|
160 |
+
|
161 |
+
Args:
|
162 |
+
image_features: Image features from the vision model
|
163 |
+
|
164 |
+
Returns:
|
165 |
+
Projected image features
|
166 |
+
"""
|
167 |
+
with torch.no_grad():
|
168 |
+
projected_features = self.projection(image_features)
|
169 |
+
|
170 |
+
return projected_features
|
171 |
+
|
172 |
+
def format_prompt(self, prompt: str, conversation: List[Dict[str, str]] = None) -> str:
|
173 |
+
"""
|
174 |
+
Format the prompt for the language model.
|
175 |
+
|
176 |
+
Args:
|
177 |
+
prompt: The text prompt
|
178 |
+
conversation: Optional conversation history
|
179 |
+
|
180 |
+
Returns:
|
181 |
+
Formatted prompt string
|
182 |
+
"""
|
183 |
+
if conversation is None:
|
184 |
+
conversation = self.conv_template.copy()
|
185 |
+
|
186 |
+
conversation.append({"role": "user", "content": prompt})
|
187 |
+
|
188 |
+
formatted_prompt = ""
|
189 |
+
for message in conversation:
|
190 |
+
if message["role"] == "system":
|
191 |
+
formatted_prompt += f"<s>[INST] <<SYS>>\n{message['content']}\n<</SYS>>\n\n"
|
192 |
+
elif message["role"] == "user":
|
193 |
+
if formatted_prompt:
|
194 |
+
formatted_prompt += f"{message['content']} [/INST]"
|
195 |
+
else:
|
196 |
+
formatted_prompt += f"<s>[INST] {message['content']} [/INST]"
|
197 |
+
elif message["role"] == "assistant":
|
198 |
+
formatted_prompt += f" {message['content']} </s><s>[INST] "
|
199 |
+
|
200 |
+
return formatted_prompt
|
201 |
+
|
202 |
+
def generate_from_image(
|
203 |
+
self,
|
204 |
+
image_path: str,
|
205 |
+
prompt: str,
|
206 |
+
max_new_tokens: int = 512,
|
207 |
+
temperature: float = 0.7,
|
208 |
+
top_p: float = 0.9,
|
209 |
+
conversation: List[Dict[str, str]] = None
|
210 |
+
) -> str:
|
211 |
+
"""
|
212 |
+
Generate text based on an image and a prompt.
|
213 |
+
|
214 |
+
Args:
|
215 |
+
image_path: Path to the image file
|
216 |
+
prompt: Text prompt
|
217 |
+
max_new_tokens: Maximum number of tokens to generate
|
218 |
+
temperature: Sampling temperature
|
219 |
+
top_p: Top-p sampling parameter
|
220 |
+
conversation: Optional conversation history
|
221 |
+
|
222 |
+
Returns:
|
223 |
+
Generated text response
|
224 |
+
"""
|
225 |
+
# Encode image
|
226 |
+
image_features = self.encode_image(image_path)
|
227 |
+
projected_features = self.project_image_features(image_features)
|
228 |
+
|
229 |
+
# Format prompt
|
230 |
+
formatted_prompt = self.format_prompt(prompt, conversation)
|
231 |
+
|
232 |
+
# Tokenize prompt
|
233 |
+
inputs = self.tokenizer(formatted_prompt, return_tensors="pt").to(self.device)
|
234 |
+
input_ids = inputs.input_ids
|
235 |
+
|
236 |
+
# Prepare for generation
|
237 |
+
stopping_criteria = StoppingCriteriaList([
|
238 |
+
StoppingCriteriaSub(stops=[self.tokenizer.eos_token_id], encounters=1)
|
239 |
+
])
|
240 |
+
|
241 |
+
# Generate response
|
242 |
+
with torch.no_grad():
|
243 |
+
# Prepare the inputs for the language model
|
244 |
+
# Here we would normally inject the image features into the language model
|
245 |
+
# This is a simplified version - in the actual LLaVA, this is done by modifying
|
246 |
+
# the language model's forward pass to accept image features
|
247 |
+
|
248 |
+
# For demonstration purposes, we'll just use the language model directly
|
249 |
+
outputs = self.language_model.generate(
|
250 |
+
input_ids=input_ids,
|
251 |
+
max_new_tokens=max_new_tokens,
|
252 |
+
temperature=temperature,
|
253 |
+
top_p=top_p,
|
254 |
+
stopping_criteria=stopping_criteria,
|
255 |
+
do_sample=True
|
256 |
+
)
|
257 |
+
|
258 |
+
# Decode the generated text
|
259 |
+
generated_text = self.tokenizer.decode(outputs[0][input_ids.shape[1]:], skip_special_tokens=True)
|
260 |
+
|
261 |
+
return generated_text.strip()
|
262 |
+
|
263 |
+
def save_model(self, output_dir: str):
|
264 |
+
"""
|
265 |
+
Save the model to the specified directory.
|
266 |
+
|
267 |
+
Args:
|
268 |
+
output_dir: Directory to save the model
|
269 |
+
"""
|
270 |
+
os.makedirs(output_dir, exist_ok=True)
|
271 |
+
|
272 |
+
# Save vision model
|
273 |
+
self.vision_model.save_pretrained(os.path.join(output_dir, "vision_model"))
|
274 |
+
self.image_processor.save_pretrained(os.path.join(output_dir, "vision_model"))
|
275 |
+
|
276 |
+
# Save language model
|
277 |
+
self.language_model.save_pretrained(os.path.join(output_dir, "language_model"))
|
278 |
+
self.tokenizer.save_pretrained(os.path.join(output_dir, "language_model"))
|
279 |
+
|
280 |
+
# Save projection layer
|
281 |
+
torch.save(self.projection.state_dict(), os.path.join(output_dir, "projection.pt"))
|
282 |
+
|
283 |
+
@classmethod
|
284 |
+
def from_pretrained(cls, model_path: str, device: str = None):
|
285 |
+
"""
|
286 |
+
Load a pretrained LLaVA model.
|
287 |
+
|
288 |
+
Args:
|
289 |
+
model_path: Path to the saved model
|
290 |
+
device: Device to load the model on
|
291 |
+
|
292 |
+
Returns:
|
293 |
+
Loaded LLaVA model
|
294 |
+
"""
|
295 |
+
# Load vision model
|
296 |
+
vision_model_path = os.path.join(model_path, "vision_model")
|
297 |
+
|
298 |
+
# Load language model
|
299 |
+
language_model_path = os.path.join(model_path, "language_model")
|
300 |
+
|
301 |
+
# Create model instance
|
302 |
+
model = cls(
|
303 |
+
vision_model_path=vision_model_path,
|
304 |
+
language_model_path=language_model_path,
|
305 |
+
device=device
|
306 |
+
)
|
307 |
+
|
308 |
+
# Load projection layer
|
309 |
+
projection_path = os.path.join(model_path, "projection.pt")
|
310 |
+
model.projection.load_state_dict(torch.load(projection_path, map_location=model.device))
|
311 |
+
|
312 |
+
return model
|
requirements.txt
ADDED
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
torch>=2.0.0
|
2 |
+
torchvision>=0.15.0
|
3 |
+
transformers>=4.30.0
|
4 |
+
accelerate>=0.20.0
|
5 |
+
pillow>=9.0.0
|
6 |
+
numpy>=1.24.0
|
7 |
+
tqdm>=4.65.0
|
8 |
+
matplotlib>=3.7.0
|
9 |
+
opencv-python>=4.7.0
|
10 |
+
einops>=0.6.0
|
11 |
+
timm>=0.9.0
|
12 |
+
sentencepiece>=0.1.99
|
13 |
+
gradio>=3.35.0
|
14 |
+
peft>=0.4.0
|
15 |
+
bitsandbytes>=0.40.0
|
16 |
+
safetensors>=0.3.1
|
17 |
+
fastapi==0.104.1
|
18 |
+
uvicorn==0.24.0
|
19 |
+
python-multipart==0.0.6
|
20 |
+
pydantic==2.5.2
|
21 |
+
python-jose==3.3.0
|
22 |
+
passlib==1.7.4
|
23 |
+
bcrypt==4.0.1
|
24 |
+
aiofiles==23.2.1
|
25 |
+
python-dotenv==1.0.0
|
26 |
+
httpx==0.25.2
|
utils/__init__.py
ADDED
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from .data_utils import (
|
2 |
+
load_image,
|
3 |
+
process_image,
|
4 |
+
pad_image,
|
5 |
+
load_conversation_data,
|
6 |
+
format_conversation,
|
7 |
+
create_image_text_pair
|
8 |
+
)
|
9 |
+
|
10 |
+
from .eval_utils import (
|
11 |
+
evaluate_vqa,
|
12 |
+
visualize_results,
|
13 |
+
compute_metrics
|
14 |
+
)
|
15 |
+
|
16 |
+
from .visualization import (
|
17 |
+
display_image_with_caption,
|
18 |
+
visualize_attention,
|
19 |
+
create_comparison_grid,
|
20 |
+
add_caption_to_image
|
21 |
+
)
|
22 |
+
|
23 |
+
__all__ = [
|
24 |
+
'load_image',
|
25 |
+
'process_image',
|
26 |
+
'pad_image',
|
27 |
+
'load_conversation_data',
|
28 |
+
'format_conversation',
|
29 |
+
'create_image_text_pair',
|
30 |
+
'evaluate_vqa',
|
31 |
+
'visualize_results',
|
32 |
+
'compute_metrics',
|
33 |
+
'display_image_with_caption',
|
34 |
+
'visualize_attention',
|
35 |
+
'create_comparison_grid',
|
36 |
+
'add_caption_to_image'
|
37 |
+
]
|
utils/data_utils.py
ADDED
@@ -0,0 +1,175 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
Utility functions for data processing in LLaVA.
|
3 |
+
"""
|
4 |
+
|
5 |
+
import os
|
6 |
+
import json
|
7 |
+
import torch
|
8 |
+
from PIL import Image
|
9 |
+
from typing import List, Dict, Tuple, Optional, Union
|
10 |
+
import numpy as np
|
11 |
+
from transformers import CLIPImageProcessor
|
12 |
+
|
13 |
+
|
14 |
+
def load_image(image_path: str) -> Image.Image:
|
15 |
+
"""
|
16 |
+
Load an image from a file path.
|
17 |
+
|
18 |
+
Args:
|
19 |
+
image_path: Path to the image file
|
20 |
+
|
21 |
+
Returns:
|
22 |
+
PIL Image object
|
23 |
+
"""
|
24 |
+
if not os.path.exists(image_path):
|
25 |
+
raise FileNotFoundError(f"Image file not found: {image_path}")
|
26 |
+
|
27 |
+
try:
|
28 |
+
image = Image.open(image_path).convert('RGB')
|
29 |
+
return image
|
30 |
+
except Exception as e:
|
31 |
+
raise ValueError(f"Error loading image: {e}")
|
32 |
+
|
33 |
+
|
34 |
+
def process_image(
|
35 |
+
image: Union[str, Image.Image],
|
36 |
+
image_processor: CLIPImageProcessor,
|
37 |
+
device: str = "cuda"
|
38 |
+
) -> torch.Tensor:
|
39 |
+
"""
|
40 |
+
Process an image for input to the vision model.
|
41 |
+
|
42 |
+
Args:
|
43 |
+
image: PIL Image object or path to image file
|
44 |
+
image_processor: CLIP image processor
|
45 |
+
device: Device to load the processed image on
|
46 |
+
|
47 |
+
Returns:
|
48 |
+
Processed image tensor
|
49 |
+
"""
|
50 |
+
if isinstance(image, str):
|
51 |
+
image = load_image(image)
|
52 |
+
|
53 |
+
inputs = image_processor(images=image, return_tensors="pt").to(device)
|
54 |
+
return inputs
|
55 |
+
|
56 |
+
|
57 |
+
def pad_image(image: Image.Image, target_size: Tuple[int, int] = (336, 336)) -> Image.Image:
|
58 |
+
"""
|
59 |
+
Pad an image to the target size while maintaining aspect ratio.
|
60 |
+
|
61 |
+
Args:
|
62 |
+
image: PIL Image object
|
63 |
+
target_size: Target size (width, height)
|
64 |
+
|
65 |
+
Returns:
|
66 |
+
Padded image
|
67 |
+
"""
|
68 |
+
width, height = image.size
|
69 |
+
target_width, target_height = target_size
|
70 |
+
|
71 |
+
# Calculate padding
|
72 |
+
ratio = min(target_width / width, target_height / height)
|
73 |
+
new_width = int(width * ratio)
|
74 |
+
new_height = int(height * ratio)
|
75 |
+
|
76 |
+
# Resize image
|
77 |
+
resized_image = image.resize((new_width, new_height), Image.LANCZOS)
|
78 |
+
|
79 |
+
# Create new image with padding
|
80 |
+
new_image = Image.new("RGB", target_size, (0, 0, 0))
|
81 |
+
paste_x = (target_width - new_width) // 2
|
82 |
+
paste_y = (target_height - new_height) // 2
|
83 |
+
new_image.paste(resized_image, (paste_x, paste_y))
|
84 |
+
|
85 |
+
return new_image
|
86 |
+
|
87 |
+
|
88 |
+
def load_conversation_data(json_path: str) -> List[Dict]:
|
89 |
+
"""
|
90 |
+
Load conversation data from a JSON file.
|
91 |
+
|
92 |
+
Args:
|
93 |
+
json_path: Path to the JSON file
|
94 |
+
|
95 |
+
Returns:
|
96 |
+
List of conversation dictionaries
|
97 |
+
"""
|
98 |
+
if not os.path.exists(json_path):
|
99 |
+
raise FileNotFoundError(f"JSON file not found: {json_path}")
|
100 |
+
|
101 |
+
try:
|
102 |
+
with open(json_path, 'r', encoding='utf-8') as f:
|
103 |
+
data = json.load(f)
|
104 |
+
return data
|
105 |
+
except Exception as e:
|
106 |
+
raise ValueError(f"Error loading JSON data: {e}")
|
107 |
+
|
108 |
+
|
109 |
+
def format_conversation(
|
110 |
+
conversation: List[Dict[str, str]],
|
111 |
+
system_prompt: Optional[str] = None
|
112 |
+
) -> List[Dict[str, str]]:
|
113 |
+
"""
|
114 |
+
Format a conversation for the LLaVA model.
|
115 |
+
|
116 |
+
Args:
|
117 |
+
conversation: List of conversation messages
|
118 |
+
system_prompt: Optional system prompt to prepend
|
119 |
+
|
120 |
+
Returns:
|
121 |
+
Formatted conversation
|
122 |
+
"""
|
123 |
+
formatted_conv = []
|
124 |
+
|
125 |
+
# Add system prompt if provided
|
126 |
+
if system_prompt:
|
127 |
+
formatted_conv.append({"role": "system", "content": system_prompt})
|
128 |
+
|
129 |
+
# Add conversation messages
|
130 |
+
for message in conversation:
|
131 |
+
if "role" in message and "content" in message:
|
132 |
+
formatted_conv.append({
|
133 |
+
"role": message["role"],
|
134 |
+
"content": message["content"]
|
135 |
+
})
|
136 |
+
|
137 |
+
return formatted_conv
|
138 |
+
|
139 |
+
|
140 |
+
def create_image_text_pair(
|
141 |
+
image_path: str,
|
142 |
+
text: str,
|
143 |
+
image_processor: CLIPImageProcessor,
|
144 |
+
tokenizer,
|
145 |
+
max_length: int = 512,
|
146 |
+
device: str = "cuda"
|
147 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
148 |
+
"""
|
149 |
+
Create an image-text pair for training or inference.
|
150 |
+
|
151 |
+
Args:
|
152 |
+
image_path: Path to the image file
|
153 |
+
text: Text prompt
|
154 |
+
image_processor: CLIP image processor
|
155 |
+
tokenizer: Language model tokenizer
|
156 |
+
max_length: Maximum text length
|
157 |
+
device: Device to load tensors on
|
158 |
+
|
159 |
+
Returns:
|
160 |
+
Tuple of (image_tensor, text_tensor)
|
161 |
+
"""
|
162 |
+
# Process image
|
163 |
+
image = load_image(image_path)
|
164 |
+
image_inputs = image_processor(images=image, return_tensors="pt").to(device)
|
165 |
+
|
166 |
+
# Process text
|
167 |
+
text_inputs = tokenizer(
|
168 |
+
text,
|
169 |
+
return_tensors="pt",
|
170 |
+
padding="max_length",
|
171 |
+
max_length=max_length,
|
172 |
+
truncation=True
|
173 |
+
).to(device)
|
174 |
+
|
175 |
+
return image_inputs, text_inputs
|
utils/eval_utils.py
ADDED
@@ -0,0 +1,180 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
Utility functions for evaluating LLaVA models.
|
3 |
+
"""
|
4 |
+
|
5 |
+
import os
|
6 |
+
import json
|
7 |
+
import torch
|
8 |
+
import numpy as np
|
9 |
+
from typing import List, Dict, Tuple, Optional, Union
|
10 |
+
from PIL import Image
|
11 |
+
import matplotlib.pyplot as plt
|
12 |
+
from tqdm import tqdm
|
13 |
+
|
14 |
+
|
15 |
+
def evaluate_vqa(
|
16 |
+
model,
|
17 |
+
questions_file: str,
|
18 |
+
image_folder: str,
|
19 |
+
output_file: Optional[str] = None,
|
20 |
+
max_new_tokens: int = 100
|
21 |
+
) -> Dict:
|
22 |
+
"""
|
23 |
+
Evaluate the model on visual question answering.
|
24 |
+
|
25 |
+
Args:
|
26 |
+
model: LLaVA model
|
27 |
+
questions_file: Path to the questions JSON file
|
28 |
+
image_folder: Path to the folder containing images
|
29 |
+
output_file: Optional path to save results
|
30 |
+
max_new_tokens: Maximum number of tokens to generate
|
31 |
+
|
32 |
+
Returns:
|
33 |
+
Dictionary with evaluation results
|
34 |
+
"""
|
35 |
+
# Load questions
|
36 |
+
with open(questions_file, 'r', encoding='utf-8') as f:
|
37 |
+
questions = json.load(f)
|
38 |
+
|
39 |
+
results = []
|
40 |
+
|
41 |
+
# Process each question
|
42 |
+
for q in tqdm(questions, desc="Evaluating VQA"):
|
43 |
+
image_path = os.path.join(image_folder, q['image'])
|
44 |
+
question_text = q['question']
|
45 |
+
|
46 |
+
# Generate answer
|
47 |
+
try:
|
48 |
+
answer = model.generate_from_image(
|
49 |
+
image_path=image_path,
|
50 |
+
prompt=question_text,
|
51 |
+
max_new_tokens=max_new_tokens
|
52 |
+
)
|
53 |
+
|
54 |
+
result = {
|
55 |
+
'question_id': q.get('question_id', None),
|
56 |
+
'image': q['image'],
|
57 |
+
'question': question_text,
|
58 |
+
'answer': answer,
|
59 |
+
'gt_answer': q.get('answer', None)
|
60 |
+
}
|
61 |
+
|
62 |
+
results.append(result)
|
63 |
+
except Exception as e:
|
64 |
+
print(f"Error processing question {q.get('question_id', '')}: {e}")
|
65 |
+
|
66 |
+
# Save results if output file is provided
|
67 |
+
if output_file:
|
68 |
+
with open(output_file, 'w', encoding='utf-8') as f:
|
69 |
+
json.dump(results, f, indent=2)
|
70 |
+
|
71 |
+
# Calculate accuracy if ground truth answers are available
|
72 |
+
accuracy = None
|
73 |
+
if all('gt_answer' in r and r['gt_answer'] is not None for r in results):
|
74 |
+
correct = 0
|
75 |
+
for r in results:
|
76 |
+
# Simple exact match accuracy
|
77 |
+
if r['answer'].lower() == r['gt_answer'].lower():
|
78 |
+
correct += 1
|
79 |
+
|
80 |
+
accuracy = correct / len(results) if results else 0
|
81 |
+
|
82 |
+
return {
|
83 |
+
'results': results,
|
84 |
+
'accuracy': accuracy,
|
85 |
+
'num_questions': len(results)
|
86 |
+
}
|
87 |
+
|
88 |
+
|
89 |
+
def visualize_results(
|
90 |
+
results: List[Dict],
|
91 |
+
num_examples: int = 5,
|
92 |
+
figsize: Tuple[int, int] = (15, 10),
|
93 |
+
image_folder: str = None
|
94 |
+
) -> None:
|
95 |
+
"""
|
96 |
+
Visualize VQA results.
|
97 |
+
|
98 |
+
Args:
|
99 |
+
results: List of result dictionaries
|
100 |
+
num_examples: Number of examples to visualize
|
101 |
+
figsize: Figure size
|
102 |
+
image_folder: Path to the folder containing images
|
103 |
+
"""
|
104 |
+
# Select a subset of results
|
105 |
+
if len(results) > num_examples:
|
106 |
+
indices = np.random.choice(len(results), num_examples, replace=False)
|
107 |
+
selected_results = [results[i] for i in indices]
|
108 |
+
else:
|
109 |
+
selected_results = results
|
110 |
+
|
111 |
+
# Create figure
|
112 |
+
fig, axes = plt.subplots(len(selected_results), 1, figsize=figsize)
|
113 |
+
if len(selected_results) == 1:
|
114 |
+
axes = [axes]
|
115 |
+
|
116 |
+
# Plot each example
|
117 |
+
for i, result in enumerate(selected_results):
|
118 |
+
# Load image
|
119 |
+
if image_folder:
|
120 |
+
image_path = os.path.join(image_folder, result['image'])
|
121 |
+
img = Image.open(image_path).convert('RGB')
|
122 |
+
axes[i].imshow(img)
|
123 |
+
|
124 |
+
# Set title and text
|
125 |
+
title = f"Q: {result['question']}"
|
126 |
+
text = f"A: {result['answer']}"
|
127 |
+
if 'gt_answer' in result and result['gt_answer']:
|
128 |
+
text += f"\nGT: {result['gt_answer']}"
|
129 |
+
|
130 |
+
axes[i].set_title(title)
|
131 |
+
axes[i].text(0, -0.5, text, transform=axes[i].transAxes, fontsize=12)
|
132 |
+
axes[i].axis('off')
|
133 |
+
|
134 |
+
plt.tight_layout()
|
135 |
+
plt.show()
|
136 |
+
|
137 |
+
|
138 |
+
def compute_metrics(results: List[Dict]) -> Dict:
|
139 |
+
"""
|
140 |
+
Compute evaluation metrics.
|
141 |
+
|
142 |
+
Args:
|
143 |
+
results: List of result dictionaries
|
144 |
+
|
145 |
+
Returns:
|
146 |
+
Dictionary with metrics
|
147 |
+
"""
|
148 |
+
metrics = {}
|
149 |
+
|
150 |
+
# Check if ground truth answers are available
|
151 |
+
has_gt = all('gt_answer' in r and r['gt_answer'] is not None for r in results)
|
152 |
+
|
153 |
+
if has_gt:
|
154 |
+
# Exact match accuracy
|
155 |
+
correct = 0
|
156 |
+
for r in results:
|
157 |
+
if r['answer'].lower() == r['gt_answer'].lower():
|
158 |
+
correct += 1
|
159 |
+
|
160 |
+
metrics['exact_match_accuracy'] = correct / len(results) if results else 0
|
161 |
+
|
162 |
+
# Token overlap (simple BLEU-like metric)
|
163 |
+
total_overlap = 0
|
164 |
+
for r in results:
|
165 |
+
pred_tokens = set(r['answer'].lower().split())
|
166 |
+
gt_tokens = set(r['gt_answer'].lower().split())
|
167 |
+
|
168 |
+
if gt_tokens: # Avoid division by zero
|
169 |
+
overlap = len(pred_tokens.intersection(gt_tokens)) / len(gt_tokens)
|
170 |
+
total_overlap += overlap
|
171 |
+
|
172 |
+
metrics['token_overlap'] = total_overlap / len(results) if results else 0
|
173 |
+
|
174 |
+
# Response length statistics
|
175 |
+
lengths = [len(r['answer'].split()) for r in results]
|
176 |
+
metrics['avg_response_length'] = sum(lengths) / len(lengths) if lengths else 0
|
177 |
+
metrics['min_response_length'] = min(lengths) if lengths else 0
|
178 |
+
metrics['max_response_length'] = max(lengths) if lengths else 0
|
179 |
+
|
180 |
+
return metrics
|
utils/visualization.py
ADDED
@@ -0,0 +1,178 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
Visualization utilities for LLaVA.
|
3 |
+
"""
|
4 |
+
|
5 |
+
import matplotlib.pyplot as plt
|
6 |
+
import numpy as np
|
7 |
+
from PIL import Image, ImageDraw, ImageFont
|
8 |
+
import torch
|
9 |
+
from typing import List, Dict, Tuple, Optional, Union
|
10 |
+
import cv2
|
11 |
+
|
12 |
+
|
13 |
+
def display_image_with_caption(
|
14 |
+
image_path: str,
|
15 |
+
caption: str,
|
16 |
+
figsize: Tuple[int, int] = (10, 10)
|
17 |
+
) -> None:
|
18 |
+
"""
|
19 |
+
Display an image with a caption.
|
20 |
+
|
21 |
+
Args:
|
22 |
+
image_path: Path to the image file
|
23 |
+
caption: Caption text
|
24 |
+
figsize: Figure size
|
25 |
+
"""
|
26 |
+
image = Image.open(image_path).convert('RGB')
|
27 |
+
|
28 |
+
plt.figure(figsize=figsize)
|
29 |
+
plt.imshow(image)
|
30 |
+
plt.axis('off')
|
31 |
+
plt.title(caption)
|
32 |
+
plt.tight_layout()
|
33 |
+
plt.show()
|
34 |
+
|
35 |
+
|
36 |
+
def visualize_attention(
|
37 |
+
image_path: str,
|
38 |
+
attention_weights: torch.Tensor,
|
39 |
+
figsize: Tuple[int, int] = (15, 5)
|
40 |
+
) -> None:
|
41 |
+
"""
|
42 |
+
Visualize attention weights on an image.
|
43 |
+
|
44 |
+
Args:
|
45 |
+
image_path: Path to the image file
|
46 |
+
attention_weights: Attention weights tensor
|
47 |
+
figsize: Figure size
|
48 |
+
"""
|
49 |
+
# Load image
|
50 |
+
image = Image.open(image_path).convert('RGB')
|
51 |
+
image_np = np.array(image)
|
52 |
+
|
53 |
+
# Normalize attention weights
|
54 |
+
if attention_weights.dim() > 2:
|
55 |
+
# Average across heads and layers if necessary
|
56 |
+
attention_weights = attention_weights.mean(dim=(0, 1))
|
57 |
+
|
58 |
+
attention_weights = attention_weights.detach().cpu().numpy()
|
59 |
+
attention_weights = (attention_weights - attention_weights.min()) / (attention_weights.max() - attention_weights.min())
|
60 |
+
|
61 |
+
# Resize attention map to image size
|
62 |
+
attention_map = cv2.resize(attention_weights, (image.width, image.height))
|
63 |
+
|
64 |
+
# Create heatmap
|
65 |
+
heatmap = cv2.applyColorMap(np.uint8(255 * attention_map), cv2.COLORMAP_JET)
|
66 |
+
heatmap = cv2.cvtColor(heatmap, cv2.COLOR_BGR2RGB)
|
67 |
+
|
68 |
+
# Overlay heatmap on image
|
69 |
+
alpha = 0.5
|
70 |
+
overlay = heatmap * alpha + image_np * (1 - alpha)
|
71 |
+
overlay = overlay.astype(np.uint8)
|
72 |
+
|
73 |
+
# Display original image and attention overlay
|
74 |
+
fig, axes = plt.subplots(1, 3, figsize=figsize)
|
75 |
+
|
76 |
+
axes[0].imshow(image_np)
|
77 |
+
axes[0].set_title('Original Image')
|
78 |
+
axes[0].axis('off')
|
79 |
+
|
80 |
+
axes[1].imshow(heatmap)
|
81 |
+
axes[1].set_title('Attention Map')
|
82 |
+
axes[1].axis('off')
|
83 |
+
|
84 |
+
axes[2].imshow(overlay)
|
85 |
+
axes[2].set_title('Overlay')
|
86 |
+
axes[2].axis('off')
|
87 |
+
|
88 |
+
plt.tight_layout()
|
89 |
+
plt.show()
|
90 |
+
|
91 |
+
|
92 |
+
def create_comparison_grid(
|
93 |
+
image_path: str,
|
94 |
+
responses: List[Dict[str, str]],
|
95 |
+
output_path: Optional[str] = None,
|
96 |
+
figsize: Tuple[int, int] = (12, 10)
|
97 |
+
) -> None:
|
98 |
+
"""
|
99 |
+
Create a comparison grid of different model responses.
|
100 |
+
|
101 |
+
Args:
|
102 |
+
image_path: Path to the image file
|
103 |
+
responses: List of dictionaries with 'model' and 'response' keys
|
104 |
+
output_path: Optional path to save the figure
|
105 |
+
figsize: Figure size
|
106 |
+
"""
|
107 |
+
# Load image
|
108 |
+
image = Image.open(image_path).convert('RGB')
|
109 |
+
|
110 |
+
# Create figure
|
111 |
+
fig = plt.figure(figsize=figsize)
|
112 |
+
|
113 |
+
# Add image
|
114 |
+
ax1 = plt.subplot2grid((len(responses) + 1, 3), (0, 0), colspan=3)
|
115 |
+
ax1.imshow(image)
|
116 |
+
ax1.set_title('Input Image')
|
117 |
+
ax1.axis('off')
|
118 |
+
|
119 |
+
# Add responses
|
120 |
+
for i, resp in enumerate(responses):
|
121 |
+
ax = plt.subplot2grid((len(responses) + 1, 3), (i + 1, 0), colspan=3)
|
122 |
+
ax.text(0.5, 0.5, f"{resp['model']}: {resp['response']}",
|
123 |
+
wrap=True, horizontalalignment='center',
|
124 |
+
verticalalignment='center', fontsize=10)
|
125 |
+
ax.axis('off')
|
126 |
+
|
127 |
+
plt.tight_layout()
|
128 |
+
|
129 |
+
# Save figure if output path is provided
|
130 |
+
if output_path:
|
131 |
+
plt.savefig(output_path, bbox_inches='tight')
|
132 |
+
|
133 |
+
plt.show()
|
134 |
+
|
135 |
+
|
136 |
+
def add_caption_to_image(
|
137 |
+
image_path: str,
|
138 |
+
caption: str,
|
139 |
+
output_path: str,
|
140 |
+
font_size: int = 20,
|
141 |
+
font_color: Tuple[int, int, int] = (255, 255, 255),
|
142 |
+
bg_color: Tuple[int, int, int] = (0, 0, 0)
|
143 |
+
) -> None:
|
144 |
+
"""
|
145 |
+
Add a caption to an image and save it.
|
146 |
+
|
147 |
+
Args:
|
148 |
+
image_path: Path to the input image
|
149 |
+
caption: Caption text
|
150 |
+
output_path: Path to save the output image
|
151 |
+
font_size: Font size
|
152 |
+
font_color: Font color (RGB)
|
153 |
+
bg_color: Background color (RGB)
|
154 |
+
"""
|
155 |
+
# Load image
|
156 |
+
image = Image.open(image_path).convert('RGB')
|
157 |
+
|
158 |
+
# Create a new image with space for the caption
|
159 |
+
caption_height = font_size + 20 # Add some padding
|
160 |
+
new_image = Image.new('RGB', (image.width, image.height + caption_height), bg_color)
|
161 |
+
new_image.paste(image, (0, 0))
|
162 |
+
|
163 |
+
# Add caption
|
164 |
+
draw = ImageDraw.Draw(new_image)
|
165 |
+
try:
|
166 |
+
font = ImageFont.truetype("arial.ttf", font_size)
|
167 |
+
except IOError:
|
168 |
+
font = ImageFont.load_default()
|
169 |
+
|
170 |
+
# Calculate text position
|
171 |
+
text_width = draw.textlength(caption, font=font)
|
172 |
+
text_position = ((image.width - text_width) // 2, image.height + 10)
|
173 |
+
|
174 |
+
# Draw text
|
175 |
+
draw.text(text_position, caption, font=font, fill=font_color)
|
176 |
+
|
177 |
+
# Save image
|
178 |
+
new_image.save(output_path)
|