YAHP interface for Timm.


These classes are used with yahp for YAML-based configuration.


YAHP interface for Timm.

class composer.models.timm.timm_hparams.TimmHparams(initializers=<factory>, num_classes=1000, model_name=None, pretrained=False, drop_rate=0.0, drop_path_rate=None, drop_block_rate=None, global_pool=None, bn_momentum=None, bn_eps=None)[source]#

Bases: composer.models.model_hparams.ModelHparams

YAHP interface for Timm.

  • model_name (str) โ€“ timm model name e.g: "resnet50". List of models can be found at PyTorch Image Models.

  • pretrained (bool, optional) โ€“ Imagenet pretrained. Default: False.

  • num_classes (int, optional) โ€“ The number of classes. Needed for classification tasks. Default: 1000.

  • drop_rate (float, optional) โ€“ Dropout rate. Default: 0.0.

  • drop_path_rate (float, optional) โ€“ Drop path rate (model default if None). Default: None.

  • drop_block_rate (float, optional) โ€“ Drop block rate (model default if None). Default: None.

  • global_pool (str, optional) โ€“ Global pool type, one of ("fast", "avg", "max", "avgmax", "avgmaxc"). Model default if None. Default: None.

  • bn_momentum (float, optional) โ€“ BatchNorm momentum override (model default if not None). Default: None.

  • bn_eps (float, optional) โ€“ BatchNorm epsilon override (model default if None). Default: None.