No search space autotune error about scvi-tools HOT 8 CLOSED

Hi, sorry you're experiencing this issue. Could you paste the search space you are trying out?

from scvi-tools.

I also tried the default with just:

tuner.fit(adata, metric="validation_loss")

param_space = {
'n_hidden': {
'distribution': 'choice',
'choices': [64, 128, 256, 512],
},
'n_latent': {
'distribution': 'choice',
'choices': [5, 10, 20, 30],
},
'n_layers': {
'distribution': 'choice',
'choices': [1, 2, 3],
},
'dropout_rate': {
'distribution': 'uniform',
'low': 0.0,
'high': 0.5,
},
'gene_likelihood': {
'distribution': 'choice',
'choices': ['zinb', 'nb'],
}
}

from scvi-tools.

I am also a little confused because in the model.history there is no validation loss even after successful training. Only reconstruction_loss_train.

(only shows up when early_stopping = True)

from scvi-tools.

I checked earlier versions of scvi-tools:
1.0.4 did not work
0.20.3 does work

from scvi-tools.

It's possible that this is due to a new release of Ray. Could you pass in param_space formatted using the tune search API:

from ray import tune 

param_space = {
  'n_hidden': tune.choice([64, 128, 256, 512]),
  'n_latent': tune.choice([5, 10, 20, 30]),
  'n_layers': tune.choice([1, 2, 3]),
  ... # other params
}

from scvi-tools.

Regarding validation loss, our default is set to not run validation steps. This can be enabled with passing in check_val_every_n_epoch=1 to the train method.

from scvi-tools.

Ok, great. That fixed it. I was a little confused because the default search_space was returning the same error. Thanks!

from scvi-tools.

Yeah we format the default search space a little differently because of the way we parse it internally. That might have been a bit confusing, sorry.

from scvi-tools.