mixmo.loaders.dataset_wrapper.MixMoDataset

class mixmo.loaders.dataset_wrapper.MixMoDataset(dataset, num_classes, num_members, dict_config, properties)

Bases: mixmo.loaders.dataset_wrapper.MSDADataset

Dataset wrapper that returns dictionaries of multiple samples, and applies MSDA augmentations

__init__(dataset, num_classes, num_members, dict_config, properties)

Initialize self. See help(type(self)) for accurate signature.

Methods

__init__(dataset, num_classes, num_members, …)	Initialize self.
call_dataset(index[, seed])	Get target and image, apply AugMix if necessary and return dictionary
call_msda(index_0[, mixmo_mask, seed_da])	Get two samples and mix them.
get_mixmo_mix_method_at_ratio_epoch([batch_seed])	Select which mixing method should be used according to training scheduling.
set_ratio_epoch(ratioepoch)

Attributes

reverse_if_first_minor

_init_dict_output_mixmo(batch_seed)

Compute MixMo block variables (masks, lams) and prepare it as a dictionary output

_target_balancing(dict_output)

Final formatting of outputs with mixmo balancing

get_mixmo_mix_method_at_ratio_epoch(batch_seed=None)

Select which mixing method should be used according to training scheduling.

Procedure: Select self.dict_mixmo_mix_method[“method_name”] with proba self.dict_mixmo_mix_method[“prob”] that is linearly decreased towards 0 after 11/12 of training process. Otherwise, use self.dict_mixmo_mix_method[“replacement_method_name”] (in general mixup)