feat: scaling factor and yt normalization [2301.10972]

[pnsuau]

Add the scaling factor and the noisy image normalization from https://arxiv.org/pdf/2301.10972.pdf.

New options :

• --alg_palette_scaling_factor : the scaling factor applied to ground truth image when added to noise at training time
• --alg_palette_normalize_yt: whether to normalize the noisy image before feeding it to the denoising neural network

Usage :

python3 train.py --alg_palette_scaling_factor 0.3 --alg_palette_normalize_yt

[beniz]

NameError: name 'y_t' is not defined

[beniz]

Breaks with --G_netG mobile_resnet_attn

  File "/data1/beniz/code/joliGAN/train.py", line 87, in train_gpu
    model = create_model(opt, rank)  # create a model given opt.model and other options
  File "/data1/beniz/code/joliGAN/models/__init__.py", line 84, in create_model
    instance = model(opt, rank)
  File "/data1/beniz/code/joliGAN/models/palette_model.py", line 114, in __init__
    self.netG_A = diffusion_networks.define_G(**vars(opt))
  File "/data1/beniz/code/joliGAN/models/diffusion_networks.py", line 111, in define_G
    denoise_fn = ResnetGenerator_attn_diff(
  File "/data1/beniz/code/joliGAN/models/modules/resnet_architecture/resnet_generator_diff.py", line 198, in __init__
    nn.Sequential(nn.Conv2d(input_nc, ngf, 7, 1, 0), normalization(ngf)),
TypeError: normalization() missing 1 required positional argument: 'norm'