Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,14 +1,129 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
import gradio as gr
|
| 2 |
import spaces
|
| 3 |
-
import
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 4 |
|
| 5 |
-
|
| 6 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 7 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 8 |
@spaces.GPU
|
| 9 |
-
def
|
| 10 |
-
|
| 11 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
|
| 13 |
-
|
| 14 |
-
demo.launch()
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
import re
|
| 3 |
+
import random
|
| 4 |
+
from dataclasses import dataclass
|
| 5 |
+
from functools import partial
|
| 6 |
+
|
| 7 |
+
import torch
|
| 8 |
import gradio as gr
|
| 9 |
import spaces
|
| 10 |
+
from datasets import load_dataset
|
| 11 |
+
from torch.utils.data import DataLoader
|
| 12 |
+
from transformers import AutoProcessor, Gemma3ForConditionalGeneration
|
| 13 |
+
from PIL import Image, ImageDraw
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
# --- Configuration ---
|
| 17 |
+
@dataclass
|
| 18 |
+
class Configuration:
|
| 19 |
+
dataset_id: str = "ariG23498/license-detection-paligemma"
|
| 20 |
+
model_id: str = "google/gemma-3-4b-pt"
|
| 21 |
+
checkpoint_id: str = "ariG23498/gemma-3-4b-pt-object-detection"
|
| 22 |
+
device: str = "cuda" if torch.cuda.is_available() else "cpu"
|
| 23 |
+
dtype: torch.dtype = torch.bfloat16
|
| 24 |
+
batch_size: int = 4
|
| 25 |
+
learning_rate: float = 2e-05
|
| 26 |
+
epochs: int = 1
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
# --- Utils ---
|
| 30 |
+
def parse_paligemma_label(label, width, height):
|
| 31 |
+
# Extract location codes
|
| 32 |
+
loc_pattern = r"<loc(\d{4})>"
|
| 33 |
+
locations = [int(loc) for loc in re.findall(loc_pattern, label)]
|
| 34 |
+
|
| 35 |
+
# Extract category (everything after the last location code)
|
| 36 |
+
category = label.split(">")[-1].strip()
|
| 37 |
+
|
| 38 |
+
# Order in PaliGemma format is: y1, x1, y2, x2
|
| 39 |
+
y1_norm, x1_norm, y2_norm, x2_norm = locations
|
| 40 |
+
|
| 41 |
+
# Convert normalized coordinates to image coordinates
|
| 42 |
+
x1 = (x1_norm / 1024) * width
|
| 43 |
+
y1 = (y1_norm / 1024) * height
|
| 44 |
+
x2 = (x2_norm / 1024) * width
|
| 45 |
+
y2 = (y2_norm / 1024) * height
|
| 46 |
+
|
| 47 |
+
return category, [x1, y1, x2, y2]
|
| 48 |
+
|
| 49 |
+
|
| 50 |
+
def visualize_bounding_boxes(image, label, width, height):
|
| 51 |
+
# Copy image for drawing
|
| 52 |
+
draw_image = image.copy()
|
| 53 |
+
draw = ImageDraw.Draw(draw_image)
|
| 54 |
+
|
| 55 |
+
category, bbox = parse_paligemma_label(label, width, height)
|
| 56 |
+
|
| 57 |
+
draw.rectangle(bbox, outline="red", width=2)
|
| 58 |
+
draw.text((bbox[0], max(0, bbox[1] - 10)), category, fill="red")
|
| 59 |
+
|
| 60 |
+
return draw_image
|
| 61 |
+
|
| 62 |
|
| 63 |
+
def test_collate_function(batch_of_samples, processor, dtype):
|
| 64 |
+
images = []
|
| 65 |
+
prompts = []
|
| 66 |
+
for sample in batch_of_samples:
|
| 67 |
+
images.append([sample["image"]])
|
| 68 |
+
prompts.append(f"{processor.tokenizer.boi_token} detect \n\n")
|
| 69 |
|
| 70 |
+
batch = processor(images=images, text=prompts, return_tensors="pt", padding=True)
|
| 71 |
+
batch["pixel_values"] = batch["pixel_values"].to(dtype)
|
| 72 |
+
return batch, images
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
# --- Initialize ---
|
| 76 |
+
cfg = Configuration()
|
| 77 |
+
|
| 78 |
+
processor = AutoProcessor.from_pretrained(cfg.checkpoint_id)
|
| 79 |
+
model = Gemma3ForConditionalGeneration.from_pretrained(
|
| 80 |
+
cfg.checkpoint_id,
|
| 81 |
+
torch_dtype=cfg.dtype,
|
| 82 |
+
device_map="cpu",
|
| 83 |
+
)
|
| 84 |
+
model.eval()
|
| 85 |
+
|
| 86 |
+
test_dataset = load_dataset(cfg.dataset_id, split="test")
|
| 87 |
+
|
| 88 |
+
|
| 89 |
+
def get_sample():
|
| 90 |
+
sample = random.choice(test_dataset)
|
| 91 |
+
images = [[sample["image"]]]
|
| 92 |
+
prompts = [f"{processor.tokenizer.boi_token} detect \n\n"]
|
| 93 |
+
|
| 94 |
+
batch = processor(images=images, text=prompts, return_tensors="pt", padding=True)
|
| 95 |
+
batch["pixel_values"] = batch["pixel_values"].to(cfg.dtype)
|
| 96 |
+
|
| 97 |
+
return batch, sample["image"]
|
| 98 |
+
|
| 99 |
+
|
| 100 |
+
# --- Prediction Logic ---
|
| 101 |
@spaces.GPU
|
| 102 |
+
def run_prediction():
|
| 103 |
+
model.to(cfg.device)
|
| 104 |
+
|
| 105 |
+
batch, raw_image = get_sample()
|
| 106 |
+
batch = {k: v.to(cfg.device) if isinstance(v, torch.Tensor) else v for k, v in batch.items()}
|
| 107 |
+
|
| 108 |
+
with torch.no_grad():
|
| 109 |
+
generation = model.generate(**batch, max_new_tokens=100)
|
| 110 |
+
decoded = processor.batch_decode(generation, skip_special_tokens=True)[0]
|
| 111 |
+
|
| 112 |
+
image = raw_image[0]
|
| 113 |
+
width, height = image.size
|
| 114 |
+
|
| 115 |
+
result_image = visualize_bounding_boxes(image, decoded, width, height)
|
| 116 |
+
return result_image
|
| 117 |
+
|
| 118 |
+
|
| 119 |
+
# --- Gradio Interface ---
|
| 120 |
+
demo = gr.Interface(
|
| 121 |
+
fn=run_prediction,
|
| 122 |
+
inputs=[],
|
| 123 |
+
outputs=gr.Image(type="pil", label="Detected Bounding Box"),
|
| 124 |
+
title="Gemma3 Object Detector",
|
| 125 |
+
description="Click 'Run' to visualize a prediction from a randomly sampled test image.",
|
| 126 |
+
)
|
| 127 |
|
| 128 |
+
if __name__ == "__main__":
|
| 129 |
+
demo.launch()
|