Quickstart tutorial and DEFAULT_ONE_EVENT_TRAINING_CONFIG : removing a param from the config file which was set default/training.yml
I wanted to change the default network to SPIDNA (because it is more flexible, eg it can take batches of different sizes, so we do not need to control the nb of SNPs in the input (it will be padded within each batch to the max nb in this batch and wont raise an error when moving to the next batch that has different size. CustomCNN on the other end needs all batches to be of the same size.
So I went on and changed the defaults/training.yml to
network:
name: SPIDNA
params:
n_blocks: 7
n_features: 50
(see b8f708cd)
However the quickstart tuto has its own default values, with parameters set up for CustomCNN to work, which is good. So I wanted to keep the default to CustomCNN for the quickstart tuto, and to SPIDNA in a second tuto with a given toy dataset (requested by the reviewers).
I went on to change the defaults in examples/one_event.py but the issue is that it is overriding default parameters, but cannot remove parameters that were defined in the default config and should be removed from the new config (SPIDNA params n_blocks and n_features that raise an error for CustomCNN. I tried null, None, empty list but no luck.
DEFAULT_ONE_EVENT_TRAINING_CONFIG = Config({
# This keeps the default params n_features and n_blocks, which makes the training fail later
'network': {
'name': 'CustomCNN',
'params': {
},
},
# The following would correctly replace the default values by 18 and 19, training still fails later
# 'network': {
# 'name': 'CustomCNN',
# 'params': {
# 'n_features': 18,
# 'n_blocks': 19,
# },
# },
'dataset_transforms': [
{'crop': {'max_snp': 500}}
],
'n_epochs': 5,
'batch_size': 20,
'evaluation_interval': 10,
'loader_num_workers': 4
})
If it is currently impossible, I'll go back to previous settings