Modifications to generation process

[aakashdp6548]

Making this issue to compile a list of minor tweaks we could/should make to the data generation process:

• Log the parameters used to generate the cached data to the directory where it is saved
• Instead of specifying the number of batches in the config, only require total number of images desired and the batch size, and the number of batches can be calculated at runtime
• Similarly, remove checks for certain amount of data - just load whatever is given
  

  bliss/bliss/simulator/simulated_dataset.py

  Lines 120 to 125 in 760e909

  	assert len(self.data) >= self.n_batches * self.batch_size, (
  	f"Insufficient cached data loaded; "
  	f"need at least {self.n_batches * self.batch_size} "
  	f"but only have {len(self.data)}. Re-run `generate.py` with "
  	f"different generation `n_batches` and/or `batch_size`."
  	)

• Don't require that all three splits are present when loading CachedSimulatedDataset - we might want to generate just a separate test split and test on that without train/validation data for example. We can instead raise an exception when the user tries to use data that wasn't loaded (e.g. in train/val/test_dataloader)
  

  bliss/bliss/simulator/simulated_dataset.py

  Lines 127 to 130 in 760e909

  	# fix validation set
  	assert os.path.exists(
  	f"{self.cached_data_path}/dataset_valid.pt"
  	), "No cached validation data found; run `generate.py` first"

[aakashdp6548]

Discussed with @zhixiangteoh - the reason for using num_batches instead of a number of images is 1) to keep things in multiples of batch size without ceil/flooring, and 2) consistency with SimulatedDataset, so the second and third points aren't necessary.