create_prune_retrain_model

Applicability

Product

Supported

Atlas 350 Accelerator Card

  • Filter-level sparsity: √
  • 2:4 structured sparsity API: x

Atlas A3 training product/Atlas A3 inference product

  • Filter-level sparsity: √
  • 2:4 structured sparsity API: √

Atlas A2 training product/Atlas A2 inference product

  • Filter-level sparsity: √
  • 2:4 structured sparsity API: √

Atlas 200I/500 A2 inference product

  • Filter-level sparsity: √
  • 2:4 structured sparsity API: √

Atlas inference product

  • Filter-level sparsity: √
  • 2:4 structured sparsity API: x

Atlas training product

  • Filter-level sparsity: √
  • 2:4 structured sparsity API: x

Note: For the products marked with x, no error is reported when the API is called for the 2:4 structured sparsity feature, but no performance gains are obtained.

Description

Filter-level sparsity or 2:4 structured sparsity API. Only either of the two sparsity features can be enabled at a time. This API sparsifies the input graph based on the given sparsity configuration file, inserts or replaces related operators in the input graph, generates a sparsity record file record_file, and returns a resultant torch.nn.Module model that can be used for retraining.

Prototype

1
prune_retrain_model = create_prune_retrain_model (model, input_data, config_defination, record_file)

Parameters

Parameter

Input/Output

Description

model

Input

Model to be sparsified, with weights loaded.

Data type: torch.nn.Module

input_data

Input

Input data of the model. A torch.tensor is replaced with an equivalent tuple(torch.tensor).

A tuple.

config_defination

Input

Simplified configuration file.

The simplified configuration file prune.cfg is generated based on the retrain_config_pytorch.proto file. The *.proto file is stored in /amct_pytorch/proto/ under the AMCT installation directory. For details about the parameters in the *.proto file and the generated simplified quantization configuration file prune.cfg, see Simplified QAT Configuration File.

A string.

record_file

Input

Path (including the file name) of the file that records sparsity information. This file records the cascade correlations between filter-level sparsity nodes or 2:4 structured sparsity nodes.

A string.

Returns

Returns the resultant torch.nn.Module model that can be used for post-sparsity training.

Example

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
import amct_pytorch as amct
# Build a graph of the network for pruning.
model = build_model()
model.load_state_dict(torch.load(state_dict_path))
input_data = tuple([torch.randn(input_shape)])
 
# Call the API for sparsifying models.
record_file = os.path.join(TMP, 'scale_offset_record.txt')
cfg_file = './prune_config.cfg'
prune_retrain_model = amct.create_prune_retrain_model(
               model,
               input_data,
               cfg_file,
               record_file)