"""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()