This code also requires the Synchronized-BatchNorm-PyTorch rep.

cd models/networks/
git clone https://github.com/vacancy/Synchronized-BatchNorm-PyTorch
cp -rf Synchronized-BatchNorm-PyTorch/sync_batchnorm .
cd ../../

To reproduce the results reported in the paper, you would need an NVIDIA DGX1 machine with 8 V100 GPUs.

---

Variable naming convention

Varible name should be a combination one or more parts below:

• normalized - normalized. If not indicated, then unnormalized
• pil or image - PIL.Image object. If not indicated, then torch.Tensor
• subnet - variable is for subnet, and should not be fed into the main network
• resized - resized from original image size (default:256x256) to the size of the correspondence matrix (default:64x64) with bilinear
• warped - warped output created according to the attention
• gt - should be used as ground truth when calculating loss

---

Specifics

Main network

• VGG perceptual loss

  • input: normalized RGB

• discriminator

  • input: LAB

• smoothness loss

  • input: LAB

• FID

  • input: RGB (not normalized)

Subnet:

• subnet_transformer:

  • input: image [wxh] PIL Image in RGB
  • output: (ref[256x256], ref_warp[64x64]),(target[256x256], target_gt[64x64]),(index_image[256x256], index_image_warp[64x64]) all of them are PIL.Image in RGB
    • we use indexes stored in index_image_warp[64x64] to index into ref_warp[64x64] and extract AB channels. For L channel, L from target_gt[64x64] is used.

• subnet

  • input: target_L & reference_LAB (256x256) normalized

• discriminator

  • input: warped_LAB, target_LAB (256x256) normalized

• FID

  • input: warped_RGB & target_RGB (not normalized)

• smoothness

  • input: warped_LAB (64x64)

• softmax

  • input: raw attention from subnet (64x64x64^2) & index tesnor (64x64)

• VGG

  • input: target_RGB & generated_RGB (64x64) (이미 네트워크의 인풋으로 normalized 를 넣어주는데 이때 genearted LAB 는 또 normalize 를 해줘야하나?)