elin24/pspl

The implementation of ICASSP 2020 paper "Pixel-level self-paced learning for super-resolution"

Hi there, I'm Wei Lin

• 🔭 I’m currently pursuing the Ph.D degree from Video, Image, and Sound Analysis Lab (VISAL) at the City University of Hong Kong.

🤝🏻  Connect with Me

Email    HomePage

pspl's Issues

Why the value of weight map is set to all 2?

I can't figure it out why all values will converge to 2 ( the hyper-parameter γ).

Question regarding the effect of the similarity map

I was curious as to what the effect of the similarity map was, so I added a few lines of code to the forward function of the Loss class, to write out sr[0] and hr[0] patches before and after multiplication by weight=gauss(ssim).detach(), for batch=1 of each epoch. My training command was:

python main.py --model EDSR --scale 4 --data_test Set5+Set14+B100+Urban100+DIV2K --n_GPUs 1 --epochs 300

For clarification, all arguments are:

Namespace(G0=64, RDNconfig='B', RDNkSize=3, act='relu', batch_size=16, betas=(0.9, 0.999), chop=False, cpu=False, data_range='1-800/801-900', data_test=['Set5', 'Set14', 'B100', 'Urban100', 'DIV2K'], data_train=['DIV2K'], debug=False, decay='200', dilation=False, dir_data='../x_imagedata', dir_demo='../test', disable_PSPL=False, epochs=300, epsilon=1e-08, ext='sep', extend='.', gamma=0.5, gan_k=1, gclip=0, load='', loss='1*L1', lr=0.0001, model='EDSR', momentum=0.9, n_GPUs=1, n_colors=3, n_feats=64, n_layers=8, n_resblocks=16, n_resgroups=10, n_threads=6, negative_slope=0.2, no_augment=False, optimizer='ADAM', patch_size=192, pre_train='', precision='single', print_every=250, reduction=16, res_scale=1, reset=False, resume=0, rgb_range=255, save='EDSR_04-08_22-15-40', save_gt=False, save_models=False, save_results=False, scale=[4], seed=1, self_ensemble=False, shift_mean=True, skip_threshold=100000000.0, splalpha=0.3, splbeta=0, split_batch=1, splval=2, template='.', test_every=1000, test_only=False, weight_decay=0)

My evaluation results were;

  [Set5 x4]     PSNR: 32.076 (Best: 32.134 @epoch 268)  ssim=0.896102
  [Set14 x4]    PSNR: 28.535 (Best: 28.568 @epoch 267)  ssim=0.785463
  [B100 x4]     PSNR: 27.539 (Best: 27.547 @epoch 257)  ssim=0.743243
  [Urban100 x4] PSNR: 25.956 (Best: 25.961 @epoch 293)  ssim=0.785183
  [DIV2K x4]    PSNR: 28.897 (Best: 28.903 @epoch 257)  ssim=0.837567

Here is what the images look like, as the epochs change, for just a few epochs. From left to right, these are sr[0], hr[0], sr[0]*weight, hr[0]*weight.

Is this about what you'd expect? They just seemed a little noisier to me than, e.g., Figure 2 in the paper. I can also try the training commands that you used in #1; the x2 case is running now, it looks like it'll take about 2.5 days...

Could you please specify the usage of your code? THX

Hi Elin

Thanks for your work, recently I want to reproduce the result of your method. Can you specify the work-flow of your paper, for convenient usage for others to use.

Thanks !

Documentation

@Elin24
Documentation on how to train and inference