Update auto3dseg algorithm templates

[dongyang0122]

Update auto3dseg algorithm templates for performance (accuracy) and efficiency improvement.

[mingxin-zheng]

I would like to hear @Nic-Ma 's thought on the ConfigParser. Since we already use num_epochs_per_validation in lots of places, such as the unit tests of MONAI, tutorials of Auto3DSeg, I feel hesitated to make such change.

[Nic-Ma]

Hi @dongyang0122 @mingxin-zheng ,

I tried to reproduce the issue with below config:

{
    "data": [1, 2],
    "data_per_batch": 2,
    "transform": {
        "_target_": "Compose",
        "transforms": [
            {"_target_": "LoadImaged", "keys": "image"},
            {"_target_": "RandTorchVisiond", "keys": "@##0#keys", "name": "ColorJitter", "brightness": 0.25},
        ],
    },
    "dataset": {"_target_": "Dataset", "data": "@data", "transform": "@transform"},
    "dataloader": {
        "_target_": "DataLoader",
        "dataset": "@##dataset",
        "batch_size": "@data_per_batch",
        "collate_fn": "$monai.data.list_data_collate",
    },
}

But seems the first 2 lines can be parsed OK:

"data": [1, 2],
"data_per_batch": 2

Could you please help provide a simple program to reproduce the issue you asked?

Thanks.

[mingxin-zheng]

Hi @dongyang0122 , if the fix in the dev branch has resolved the issue of num_epoch_*, can we revert the name of the variable back to num_epochs_per_validation?

[dongyang0122]

Hi @dongyang0122 , if the fix in the dev branch has resolved the issue of num_epoch_*, can we revert the name of the variable back to num_epochs_per_validation?

If we convert it back to previous name, then we have to inform users to install monai from the dev branch. I feel it adds another layer of complexity for general users. how about converting it back once the changes are reflected in the installation via pip install monai? and meanwhile we can merge the changes at the moment.