neural_compressor.compression.pruner.model_slim.weight_slim

Weight Squeezer.

Module Contents

Classes

PostCompressionUtils	Operations library related to weight compression.
LinearCompression	Class which automatically compresses two consecutive linear layers.
LinearCompressionIterator	Pruner of a sequence of consecutive linear patterns.

class neural_compressor.compression.pruner.model_slim.weight_slim.PostCompressionUtils

Operations library related to weight compression.

class neural_compressor.compression.pruner.model_slim.weight_slim.LinearCompression(root_linear, target_linears)

Class which automatically compresses two consecutive linear layers.

For two consecutive linear layer, when the second layer’s input channel is pruned, then the first layer’s output channel can also be pruned, while the second layer’s output hidden state value is identical. for example, two consecutive linears have following structure. x = layer_1(input) x = act_fn(x) x = layer_2(x)

Parameters:

• layer_1 (torch.nn.Linear) – the first Linear layer.

• layer_2 (torch.nn.Linear) – the second Linear layer.

layer_1

the first Linear layer.

Type:

torch.nn.Linear

layer_2

the second Linear layer.

Type:

torch.nn.Linear

device

the device of layers’ weights.

class neural_compressor.compression.pruner.model_slim.weight_slim.LinearCompressionIterator(linear_patterns)

Pruner of a sequence of consecutive linear patterns.

linear_patterns

a iterable object of consecutive linear patterns.

Type:

dict/list