Xingang Pan

@OneMondy 1. SCNN is added between fc7 and fc8 in Fig.5. 2. Interpx8 means enlarge the featuremap by 8 times (along height and width dimensions) with bilinear interpolation. 3. Yes,...

@Neha-swan 1. "th: command not found" This is because you have not install torch successfully. Please install torch following this repo https://github.com/torch/distro 2. "Run.sh: ./evaluate: not found" This is because...

You may take a look at https://github.com/xxradon/PytorchToCaffe.

Could you email me the full network and solver prototxt scripts? [email protected]

@VadymBoikov You may try resetting to an earlier commit like 2cabaca76885d6167207a8e74edf2a7409e32379 and see if it works.

@Ning5195 For 2D input, instance normalization and instance whitening make no sense, thus should be removed. You need to modify the code accordingly if you want to apply to 2D...

You can split the parameters into 2 groups in the optimizer as below:   Then you can set the learning rate of SW to zero in the middle of...

@JiyueWang The initial version of SW is implemented with SVD decomposition, where the lack of early stop would lead to a decrease of about 0.5 scores. The current version of...

Sorry I haven't tried training with apex. It seems that there is a numerical issue with fp16, you may simply use fp32.

I have just uploaded the scripts for the GAN inversion part. You may refer to this example: https://github.com/XingangPan/ShadeGAN/blob/main/scripts/inversion.sh. You need to pull the new commit and download the model weights...