from __future__ import annotations

import argparse
from pathlib import Path

import numpy as np
import torch
from torch import nn
from torch.utils.data import DataLoader
from torchvision import models, transforms

import ocular_preliminary_benchmark as base


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(
        description="Extract ImageNet EfficientNetB3 embeddings in benchmark sample order."
    )
    parser.add_argument("--data-root", type=Path, required=True)
    parser.add_argument("--output-dir", type=Path, default=Path("runs/phase1/features"))
    parser.add_argument("--input-size", type=int, default=384)
    parser.add_argument("--batch-size", type=int, default=32)
    parser.add_argument("--num-workers", type=int, default=0)
    parser.add_argument("--device", default="auto")
    parser.add_argument("--force", action="store_true")
    return parser.parse_args()


def build_model(device: torch.device) -> nn.Module:
    weights = models.EfficientNet_B3_Weights.DEFAULT
    model = models.efficientnet_b3(weights=weights)
    model.classifier = nn.Identity()
    model.eval()
    return model.to(device)


def main() -> None:
    args = parse_args()
    args.output_dir.mkdir(parents=True, exist_ok=True)
    feature_path = args.output_dir / f"efficientnet_b3_features_{args.input_size}.npy"
    label_path = args.output_dir / f"efficientnet_b3_labels_{args.input_size}.npy"
    if feature_path.exists() and label_path.exists() and not args.force:
        print(f"Feature cache already exists: {feature_path}")
        return

    device = base.get_device(args.device)
    samples = base.discover_samples(args.data_root.expanduser().resolve(), None)
    labels = np.asarray([sample.label for sample in samples], dtype=np.int64)
    transform = transforms.Compose(
        [
            transforms.Resize(int(args.input_size * 1.14)),
            transforms.CenterCrop(args.input_size),
            transforms.ToTensor(),
            transforms.Normalize(base.IMAGENET_MEAN, base.IMAGENET_STD),
        ]
    )
    dataset = base.FundusDataset(samples, transform)
    loader = DataLoader(
        dataset,
        batch_size=args.batch_size,
        shuffle=False,
        num_workers=args.num_workers,
        pin_memory=device.type == "cuda",
        persistent_workers=args.num_workers > 0,
    )
    model = build_model(device)
    features: list[np.ndarray] = []
    with torch.no_grad():
        for batch_index, (images, _) in enumerate(loader, start=1):
            images = images.to(device, non_blocking=device.type == "cuda")
            embedding = model(images)
            if device.type == "mps" and hasattr(torch, "mps"):
                torch.mps.synchronize()
            features.append(embedding.detach().cpu().numpy().astype(np.float32))
            if batch_index <= 3 or batch_index % 25 == 0:
                print(
                    f"extracted {min(batch_index * args.batch_size, len(samples))}/{len(samples)}",
                    flush=True,
                )

    matrix = np.concatenate(features, axis=0)
    np.save(feature_path, matrix)
    np.save(label_path, labels)
    print(f"saved {feature_path} shape={matrix.shape}")
    print(f"saved {label_path} shape={labels.shape}")


if __name__ == "__main__":
    main()