python main.py --net='ffa' --crop --crop_size=240 --blocks=19 --gps=3 --bs=2 --lr=0.0001 --trainset='rw2ah_train' --testset='rw2ah_test' --steps=60000 --eval_step=150 ( --skyl1 --ssim_loss --ssim_loss_type='region')| PSNR | SSIM |
|---|---|
 |
 |
Algorithm: Sky-weighted L1 + Region SSIM Loss
Input:
- ŷ: model output image [B, C, H, W]
- y: ground-truth clean image [B, C, H, W]
- weight_mask: sky region weight mask [B, 1, H, W], sky 1.5,others 1.0
- binary_mask: region mask, only buildings/trees, values {0,1} [B, 1, H, W]
- ssim_fn: function to compute region SSIM (returns per-image values)
- coefficient: SkyL1=1.0, RegionSSIM=0.5
Output:
- loss: scalar loss for backpropagation
Procedure:
- Per-pixel L1 (reduction='none') then mean
- l1_map = |ŷ - y| // [B, C, H, W]
- 3x3 smoothing of weight_mask (padding to keep size)
- weight_mask_smooth = AvgPool2d(weight_mask, kernel_size=3, stride=1, padding=1) // [B, 1, H, W]
- Sky-weighted L1 loss
- sky_l1_map = l1_map * weight_mask_smooth // [B, 1, H, W]
- sky_l1_loss = mean(sky_l1_map) // scalar
- Region SSIM loss
- masked_ŷ = ŷ * binary_mask
- masked_y = y * binary_mask
- ssim_val = ssim_fn(masked_ŷ, masked_y) // per-image values
- region_ssim_loss = 1.0 - mean(ssim_val) // scalar
- Total loss
- loss = sky_l1_loss + 0.5 * region_ssim_loss
- Backpropagation
- loss.backward()