File size: 7,837 Bytes
b28fb05 697a4fe b28fb05 697a4fe b28fb05 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 |
import sys
from eval_dataset import SingleRegionCaptionDataset
from segment_anything import sam_model_registry, SamPredictor
import gradio as gr
import numpy as np
import cv2
import base64
import torch
from PIL import Image
import io
import argparse
from fastapi import FastAPI
from fastapi.staticfiles import StaticFiles
from transformers import AutoModel, AutoProcessor, GenerationConfig
from transformers import SamModel, SamProcessor
try:
from spaces import GPU
except ImportError:
print("Spaces not installed, using dummy GPU decorator")
def GPU(*args, **kwargs):
def decorator(fn):
return fn
return decorator
# Load SAM model
#device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
device = torch.device("cpu")
sam_model = SamModel.from_pretrained("facebook/sam-vit-huge").to(device)
sam_processor = SamProcessor.from_pretrained("facebook/sam-vit-huge")
print("sam ready")
model_path = "HaochenWang/GAR-1B"
# Initialize the captioning model and processor
model = AutoModel.from_pretrained(
model_path,
trust_remote_code=True,
torch_dtype=torch.bfloat16,
device_map="cpu",
use_flash_attn=False
).eval()
processor = AutoProcessor.from_pretrained(
model_path,
trust_remote_code=True,
)
@GPU(duration=75)
def image_to_sam_embedding(base64_image):
try:
# Decode base64 string to bytes
image_bytes = base64.b64decode(base64_image)
# Convert bytes to PIL Image
image = Image.open(io.BytesIO(image_bytes))
# Process image with SAM processor
inputs = sam_processor(image, return_tensors="pt").to(device)
# Get image embedding
with torch.no_grad():
image_embedding = sam_model.get_image_embeddings(inputs["pixel_values"])
# Convert to CPU and numpy
image_embedding = image_embedding.cpu().numpy()
# Encode the embedding as base64
embedding_bytes = image_embedding.tobytes()
embedding_base64 = base64.b64encode(embedding_bytes).decode('utf-8')
return embedding_base64
except Exception as e:
print(f"Error processing image: {str(e)}")
raise gr.Error(f"Failed to process image: {str(e)}")
@GPU(duration=75)
def describe(image_base64: str, mask_base64: str, query: str):
# Convert base64 to PIL Image
image_bytes = base64.b64decode(image_base64.split(',')[1] if ',' in image_base64 else image_base64)
img = Image.open(io.BytesIO(image_bytes))
mask_bytes = base64.b64decode(mask_base64.split(',')[1] if ',' in mask_base64 else mask_base64)
mask = Image.open(io.BytesIO(mask_bytes))
mask = np.array(mask.convert('L'))
prompt_number = model.config.prompt_numbers
prompt_tokens = [f"<Prompt{i_p}>" for i_p in range(prompt_number)] + ["<NO_Prompt>"]
# Assuming mask is given as a numpy array and the image is a PIL image
dataset = SingleRegionCaptionDataset(
image=img,
mask=mask,
processor=processor,
prompt_number=prompt_number,
visual_prompt_tokens=prompt_tokens,
data_dtype=torch.bfloat16,
)
data_sample = dataset[0]
# Generate the caption
with torch.no_grad():
generate_ids = model.generate(
**data_sample,
generation_config=GenerationConfig(
max_new_tokens=1024,
# do_sample= False,
eos_token_id=processor.tokenizer.eos_token_id,
pad_token_id=processor.tokenizer.pad_token_id,
),
return_dict=True,
)
output_caption = processor.tokenizer.decode(generate_ids.sequences[0], skip_special_tokens=True).strip()
# Stream the tokens
text = ""
for token in output_caption:
text += token
yield text
@GPU(duration=75)
def describe_without_streaming(image_base64: str, mask_base64: str, query: str):
# Convert base64 to PIL Image
image_bytes = base64.b64decode(image_base64.split(',')[1] if ',' in image_base64 else image_base64)
img = Image.open(io.BytesIO(image_bytes))
mask_bytes = base64.b64decode(mask_base64.split(',')[1] if ',' in mask_base64 else mask_base64)
mask = Image.open(io.BytesIO(mask_bytes))
mask = np.array(mask.convert('L'))
prompt_number = model.config.prompt_numbers
prompt_tokens = [f"<Prompt{i_p}>" for i_p in range(prompt_number)] + ["<NO_Prompt>"]
# Assuming mask is given as a numpy array and the image is a PIL image
dataset = SingleRegionCaptionDataset(
image=img,
mask=mask,
processor=processor,
prompt_number=prompt_number,
visual_prompt_tokens=prompt_tokens,
data_dtype=torch.bfloat16,
)
data_sample = dataset[0]
# Generate the caption
with torch.no_grad():
generate_ids = model.generate(
**data_sample,
generation_config=GenerationConfig(
max_new_tokens=1024,
# do_sample=False,
eos_token_id=processor.tokenizer.eos_token_id,
pad_token_id=processor.tokenizer.pad_token_id,
),
return_dict=True,
)
output_caption = processor.tokenizer.decode(generate_ids.sequences[0], skip_special_tokens=True).strip()
return output_caption
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Describe Anything gradio demo")
parser.add_argument("--server_addr", "--host", type=str, default=None, help="The server address to listen on.")
parser.add_argument("--server_port", "--port", type=int, default=None, help="The port to listen on.")
parser.add_argument("--model-path", type=str, default="HaochenWang/GAR-1B", help="Path to the model checkpoint")
parser.add_argument("--prompt-mode", type=str, default="full+focal_crop", help="Prompt mode")
parser.add_argument("--conv-mode", type=str, default="v1", help="Conversation mode")
parser.add_argument("--temperature", type=float, default=0.2, help="Sampling temperature")
parser.add_argument("--top_p", type=float, default=0.5, help="Top-p for sampling")
args = parser.parse_args()
# Create Gradio interface
with gr.Blocks() as demo:
gr.Interface(
fn=image_to_sam_embedding,
inputs=gr.Textbox(label="Image Base64"),
outputs=gr.Textbox(label="Embedding Base64"),
title="Image Embedding Generator",
api_name="image_to_sam_embedding"
)
gr.Interface(
fn=describe,
inputs=[
gr.Textbox(label="Image Base64"),
gr.Text(label="Mask Base64"),
gr.Text(label="Prompt")
],
outputs=[
gr.Text(label="Description")
],
title="Mask Description Generator",
api_name="describe"
)
gr.Interface(
fn=describe_without_streaming,
inputs=[
gr.Textbox(label="Image Base64"),
gr.Text(label="Mask Base64"),
gr.Text(label="Prompt")
],
outputs=[
gr.Text(label="Description")
],
title="Mask Description Generator (Non-Streaming)",
api_name="describe_without_streaming"
)
demo._block_thread = demo.block_thread
demo.block_thread = lambda: None
demo.launch(
share=True,
server_name=args.server_addr,
server_port=args.server_port,
ssr_mode=False,
)
for route in demo.app.routes:
if route.path == "/":
demo.app.routes.remove(route)
demo.app.mount("/", StaticFiles(directory="dist", html=True), name="demo")
demo._block_thread()
|