Transforms need to handle list of channel-first data

Is your feature request related to a problem? Please describe. Currently, if using the default collate_fn, the interted data is a list of channel-first Tensor or numpy array, if we want to apply transforms on the data, most of the MONAI transforms can’t support it. For example, users want to use CopyItemsd, AsDiscreted, KeepLargestConnectedComponentd, etc. All these transforms can only support a batch-first Tensor or a channel-first numpy array.

Thanks.

[Nic-Ma]

Hi @vfdev-5 ,

You are right, no need to totally change ignite metrics, but may need your help to add some wrapper layer in ignite to support a list of channel-first tensors or support 1 channel-first tensor. We usually use ignite metrics in MONAI workflows directly, like this Accuracy: https://github.com/Project-MONAI/tutorials/blob/master/modules/engines/unet_evaluation_dict.py#L111 Actually, I already added support for list of channel-first tensors in MONAI metrics: https://github.com/Project-MONAI/MONAI/blob/dev/monai/handlers/iteration_metric.py#L91

Thanks.

[vfdev-5]

Hi @Nic-Ma ,

As I know, ignite metrics expect the data shape with batch dim, may I know some design reason for that?

In our practice it is the most common format for the models output and targets: (B, C, ...) and (B, ...). Also it helps to accelerate internal computations.

As I understand you would like to handle additionally two lists instead of tensors ?

metric.update((
    [ypred0, ypred1, ypred2, ...],
    [y0, y1, y2, ...],
))

where ypredX and yX have shapes [C, ...] and [...] or yX can be even a scalar. I thinks this could be a feature for ignite…