File size: 7,449 Bytes
7231c7d 45711b1 7231c7d 45711b1 201b47f 45711b1 7231c7d 45711b1 7231c7d 45711b1 7231c7d 45711b1 7231c7d |
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 |
import os
import sys
import torch
from PIL import Image as PILImage
from PIL import ImageDraw, ImageFont
from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer, AutoProcessor
from loguru import logger
import gradio as gr
import spaces
# Note: The perceptron package needs to be installed or included in the Space
try:
from perceptron.tensorstream import VisionType
from perceptron.tensorstream.ops import tensor_stream_token_view, modality_mask
from perceptron.pointing.parser import extract_points
except ImportError:
logger.error("perceptron package not found. Please ensure it's installed in your Hugging Face Space.")
raise
# Load model at startup
hf_path = "PerceptronAI/Isaac-0.1"
logger.info(f"Loading processor and config from HF checkpoint: {hf_path}")
config = AutoConfig.from_pretrained(hf_path, trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(hf_path, trust_remote_code=True, use_fast=False)
processor = AutoProcessor.from_pretrained(hf_path, trust_remote_code=True)
processor.tokenizer = tokenizer # Ensure tokenizer is set
logger.info(f"Loading AutoModelForCausalLM from HF checkpoint: {hf_path}")
model = AutoModelForCausalLM.from_pretrained(hf_path, trust_remote_code=True)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
dtype = torch.bfloat16 if torch.cuda.is_available() else torch.float32
model = model.to(device=device, dtype=dtype)
model.eval()
logger.info(f"Model loaded on {device} with dtype {dtype}")
def document_to_messages(document, vision_token="<image>"):
messages = []
images = []
for item in document:
itype = item.get("type")
if itype == "text":
content = item.get("content")
if content:
messages.append({"role": item.get("role", "user"), "content": content})
elif itype == "image":
if "content" in item and item["content"] is not None:
img = PILImage.open(item["content"]).convert("RGB")
images.append(img)
messages.append({"role": item.get("role", "user"), "content": vision_token})
return messages, images
def decode_tensor_stream(tensor_stream, tokenizer):
token_view = tensor_stream_token_view(tensor_stream)
mod = modality_mask(tensor_stream)
text_tokens = token_view[(mod != VisionType.image.value)]
decoded = tokenizer.decode(text_tokens[0] if len(text_tokens.shape) > 1 else text_tokens)
return decoded
def visualize_predictions(generated_text, image, output_path="prediction.jpeg"):
boxes = extract_points(generated_text, expected="box")
if not boxes:
logger.info("No bounding boxes found in the generated text")
image.save(output_path)
return output_path
img_width, img_height = image.size
img_with_boxes = image.copy()
draw = ImageDraw.Draw(img_with_boxes)
try:
font = ImageFont.truetype("/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf", 16)
except:
font = ImageFont.load_default()
colors = ["red", "green", "blue", "yellow", "magenta", "cyan", "orange", "purple"]
for idx, box in enumerate(boxes):
color = colors[idx % len(colors)]
norm_x1, norm_y1 = box.top_left.x, box.top_left.y
norm_x2, norm_y2 = box.bottom_right.x, box.bottom_right.y
x1 = int((norm_x1 / 1000.0) * img_width)
y1 = int((norm_y1 / 1000.0) * img_height)
x2 = int((norm_x2 / 1000.0) * img_width)
y2 = int((norm_y2 / 1000.0) * img_height)
x1 = max(0, min(x1, img_width - 1))
y1 = max(0, min(y1, img_height - 1))
x2 = max(0, min(x2, img_width - 1))
y2 = max(0, min(y2, img_height - 1))
draw.rectangle([x1, y1, x2, y2], outline=color, width=3)
if box.mention:
text_y = max(y1 - 20, 5)
text_bbox = draw.textbbox((x1, text_y), box.mention, font=font)
draw.rectangle(text_bbox, fill=color)
draw.text((x1, text_y), box.mention, fill="white", font=font)
img_with_boxes.save(output_path, "JPEG")
return output_path
@spaces.GPU(duration=120)
def generate_response(image, prompt):
document = [
{"type": "text", "content": "<hint>BOX</hint>", "role": "user"},
{"type": "image", "content": image, "role": "user"},
{"type": "text", "content": prompt, "role": "user"},
]
messages, images = document_to_messages(document, vision_token=config.vision_token)
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=text, images=images, return_tensors="pt")
tensor_stream = inputs["tensor_stream"].to(device)
input_ids = inputs["input_ids"].to(device)
decoded_content = decode_tensor_stream(tensor_stream, processor.tokenizer)
with torch.no_grad():
generated_ids = model.generate(
tensor_stream=tensor_stream,
max_new_tokens=256,
do_sample=False,
pad_token_id=processor.tokenizer.eos_token_id,
eos_token_id=processor.tokenizer.eos_token_id,
)
generated_text = processor.tokenizer.decode(generated_ids[0], skip_special_tokens=False)
if images:
vis_path = visualize_predictions(generated_text, images[0])
return generated_text, vis_path
else:
return generated_text, None
# Example images and prompts
examples = [
["example.webp", "Determine whether it is safe to cross the street. Look for signage and moving traffic."],
]
with gr.Blocks(title="Perceptron Isaac Vision Model", theme=gr.themes.Soft()) as demo:
gr.Markdown("# 🔍 Perceptron Isaac Vision Model")
gr.Markdown("Built with [anycoder](https://huggingface.co/spaces/akhaliq/anycoder)")
gr.Markdown("""
This demo showcases the Perceptron Isaac-0.1 model for multimodal understanding with bounding box visualization.
Upload an image and provide a prompt to analyze the image and see detected objects with bounding boxes.
""")
with gr.Row():
with gr.Column(scale=1):
image_input = gr.Image(
type="filepath",
label="Upload Image",
sources=["upload", "webcam", "clipboard"],
height=400
)
prompt_input = gr.Textbox(
label="Prompt",
value="Determine whether it is safe to cross the street. Look for signage and moving traffic.",
lines=3,
placeholder="Enter your prompt here..."
)
generate_btn = gr.Button("🚀 Generate Response", variant="primary", size="lg")
with gr.Column(scale=1):
visualized_image = gr.Image(
label="Visualized Predictions (with Bounding Boxes)",
height=400
)
generated_text = gr.Textbox(
label="Generated Text",
lines=10,
max_lines=20
)
gr.Examples(
examples=examples,
inputs=[image_input, prompt_input],
outputs=[generated_text, visualized_image],
fn=generate_response,
cache_examples=False
)
generate_btn.click(
generate_response,
inputs=[image_input, prompt_input],
outputs=[generated_text, visualized_image]
)
if __name__ == "__main__":
demo.launch() |