inference/export_logits.py

import torch
import numpy as np
from tqdm import tqdm
from pathlib import Path
from typing import Dict, Tuple, List

def batch_sliding_window_logits(images: torch.Tensor, model: torch.nn.Module,
                                 device: torch.device,
                                 crop_size: Tuple[int, int], stride: Tuple[int, int]) -> torch.Tensor:
    """
    Version of sliding inference that returns average logits for export.
    """
    B, C, H, W = images.shape
    ph, pw = crop_size
    sh, sw = stride
    images = images.to(device)

    num_classes = model.config.num_labels
    full_logits = torch.zeros((B, num_classes, H, W), device=device)
    count_map = torch.zeros((H, W), device=device)

    with torch.no_grad():
        for top in range(0, H, sh):
            for left in range(0, W, sw):
                bottom = min(top + ph, H)
                right = min(left + pw, W)
                top0 = max(bottom - ph, 0)
                left0 = max(right - pw, 0)
                patch = images[:, :, top0:bottom, left0:right]
                logits = model(pixel_values=patch).logits
                full_logits[:, :, top0:bottom, left0:right] += logits
                count_map[top0:bottom, left0:right] += 1

    avg_logits = full_logits / count_map.unsqueeze(0).unsqueeze(0).clamp(min=1)
    return avg_logits


def export_logits_images(model, loader, device, crop_size, stride, output_dir: Path):
    """
    Applies batch sliding window and exports probs as .npy files.
    """
    model.eval()
    output_dir.mkdir(parents=True, exist_ok=True)

    for images, rel_paths in tqdm(loader, desc=f"Export logits to {output_dir}", leave=False):
        avg_logits = batch_sliding_window_logits(images, model, device, crop_size, stride)
        probs = torch.softmax(avg_logits, dim=1)
        probs = (probs * 255.0).clamp(0, 255).byte().cpu()

        B, C, H, W = probs.shape
        for b in range(B):
            arr = probs[b].permute(1, 2, 0).numpy()  # H×W×C
            out_path = output_dir / rel_paths[b]
            out_path.parent.mkdir(parents=True, exist_ok=True)
            np.save(out_path.with_suffix('.npy'), arr)