decompose_network

Applicable Products

Description

Decomposes the input PyTorch model object based on the decomposition information file saved via the auto_decomposition API call and returns the decomposed model object and the layer names before and after decomposition.

Prototype

1

model, changes = decompose_network(model, decompose_info_path)

Parameters

Returns

• Returns the model object that is changed to the structure after tensor decomposition. The data type is torch.nn.Module.
• Dictionary consisting of the layer names before and after tensor decomposition, for example, {'conv1': ['conv1.0', 'conv1.1'], 'conv2': ['conv2.0', 'conv2.1'],...}.

Restrictions

• The input model must be an object of the torch.nn.Module type.
• This API function modifies only the structure of the convolution constructed by using torch.nn.Conv2d().
• The structure of the input model must be consistent with that in the generation of the decomposition information file by calling the auto_decomposition API.

Example

1
2
3
4
5
6

from amct_pytorch.tensor_decompose import decompose_network
net = Net()                                                      # Build a model object.
net, changes = decompose_network(                                # Load the decomposition information file to modify the model structure.
    model=net,
     decompose_info_path="decomposed_path/decompose_info.json"    # Specify the path of the decomposition information file saved through the auto_decomposition call.
)

1. If training is involved, this API must be called before the model parameters are passed to the optimizer; if torch.nn.parallel.DistributedDataParallel (DDP) is used, this API must be called before the model parameters are passed to the DDP.
2. This API modifies the input model object in place. That is, the model object input by the user is changed after decomposition (exception: The input model is a torch.nn.Conv2d object). In this case, this API does not modify it. The returned decomposed model is the newly constructed torch.nn.Module object.
3. This API only modifies the model structure and does not update the convolution weight after decomposition. The weight value is the default value built by torch.nn.Conv2d (). If fine-tuning is required, save the weight of the decomposed model after calling auto_decomposition, load the weight after calling this API, and then fine-tune the model.