| """Standalone Tiny Vial CNN v2 inference using Safetensors.""" |
| from __future__ import annotations |
|
|
| import argparse |
| import json |
| from pathlib import Path |
|
|
| import numpy as np |
| from PIL import Image, ImageDraw |
| from safetensors.torch import load_file |
| import torch |
| from torch import nn |
|
|
|
|
| class TinyVialDetectorV2(nn.Module): |
| def __init__(self): |
| super().__init__() |
| self.features = nn.Sequential( |
| nn.Conv2d(3, 12, 3, stride=2, padding=1), nn.ReLU(), |
| nn.Conv2d(12, 24, 3, stride=2, padding=1), nn.ReLU(), |
| nn.Conv2d(24, 32, 3, stride=2, padding=1), nn.ReLU(), |
| ) |
| self.shared = nn.Sequential(nn.Flatten(), nn.Linear(32 * 8 * 8, 64), nn.ReLU()) |
| self.objectness = nn.Linear(64, 1) |
| self.box = nn.Sequential(nn.Linear(64, 4), nn.Sigmoid()) |
|
|
| def forward(self, image): |
| hidden = self.shared(self.features(image)) |
| return self.objectness(hidden), self.box(hidden) |
|
|
|
|
| def pixel_box(parameters, width, height): |
| cx, cy, box_width, box_height = (float(value) for value in parameters) |
| x0 = max(0, min(width - 1, round((cx - box_width / 2) * width))) |
| y0 = max(0, min(height - 1, round((cy - box_height / 2) * height))) |
| x1 = max(x0, min(width - 1, round((cx + box_width / 2) * width) - 1)) |
| y1 = max(y0, min(height - 1, round((cy + box_height / 2) * height) - 1)) |
| return x0, y0, x1, y1 |
|
|
|
|
| def main(): |
| parser = argparse.ArgumentParser() |
| parser.add_argument("image", type=Path) |
| parser.add_argument("--weights", type=Path, default=Path("model.safetensors")) |
| parser.add_argument("--config", type=Path, default=Path("config.json")) |
| parser.add_argument("--output", type=Path, default=Path("prediction.png")) |
| args = parser.parse_args() |
| config = json.loads(args.config.read_text()) |
| model = TinyVialDetectorV2() |
| model.load_state_dict(load_file(args.weights)) |
| model.eval() |
| image = Image.open(args.image).convert("RGB") |
| resized = image.resize((config["image_size"], config["image_size"]), Image.Resampling.BILINEAR) |
| tensor = torch.from_numpy(np.asarray(resized, dtype=np.float32).copy()).permute(2, 0, 1).unsqueeze(0) / 255.0 |
| with torch.inference_mode(): |
| logits, parameters = model(tensor) |
| score = float(torch.sigmoid(logits[0, 0])) |
| detected = score >= float(config["objectness_threshold"]) |
| bbox = pixel_box(parameters[0], *image.size) if detected else None |
| if bbox: |
| ImageDraw.Draw(image).rectangle(bbox, outline=(118, 185, 0), width=max(1, image.width // 128)) |
| args.output.parent.mkdir(parents=True, exist_ok=True) |
| image.save(args.output) |
| print(json.dumps({"detected": detected, "objectness_score": round(score, 6), "threshold": config["objectness_threshold"], "bbox": list(bbox) if bbox else None, "output": str(args.output)}, indent=2)) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
