Not so mini dev branch: various improvements

[vict0rsch]

Main additions:

• More comet logging:

  comet:
  display_size: 20
  rows_per_log: 5 # number of samples (rows) in a logged grid image. Number of total logged images: display_size // rows_per_log

• Better s image logging with utils.get_display_indices:

  ...
  if domain == "s":
      with temp_np_seed(123): # deterministic random
          display_indices = np.random.permutation(length)[:dsize]
  ...

• New SegmentationDecoder

• New option: data.max_samples=n to limit the size of the dataset (for instance for dev purposes, n=-1 for no limitation)

• Add tqdm logging

• Add self.G.eval() and self.D.eval() for the inference mode, especially for batch norm and dropout

• v0 for an almost automatic resume.py

• hyper parameter search

• Automatic comet experiment continuation:

  ...
  with comet_previous_path.open("r") as f:
    url = f.read().strip()
    comet_previous_id = comet_id_from_url(url)
  ...
  exp = ExistingExperiment(previous_experiment=comet_previous_id, **comet_kwargs)
  ...

[vict0rsch]

@alexrey88 I refactored the evaluation as you pointed out, do you see any other weird things?

[alexrey88]

@alexrey88 I refactored the evaluation as you pointed out, do you see any other weird things?

All seems good :)