composer.algorithms.stochastic_depth.sample_stochastic_layers

composer.algorithms.stochastic_depth.sample_stochastic_layers

Classes

Bottleneck	torchvision.models.resnet.Bottleneck
SampleStochasticBottleneck	Sample-wise stochastic ResNet Bottleneck block.

class composer.algorithms.stochastic_depth.sample_stochastic_layers.SampleStochasticBottleneck(drop_rate, **kwargs)

Bases: torchvision.models.resnet.Bottleneck

Sample-wise stochastic ResNet Bottleneck block.

This block has a probability of dropping samples before the identity connection, then adds back the untransformed samples using the identity connection.

Parameters

• sample_drop_rate (float) – The probability of dropping a sample within this block. Must be between 0.0 and 1.0.

• **kwargs – Used for the original Bottleneck initialization arguments.

static from_target_layer(module, module_index, module_count, drop_rate, drop_distribution)

Helper function to convert a ResNet Bottleneck block into a sample-wise stochastic block.