Compressor(Recommendation Compression)

Description

netspresso.compressor.v2.compressor.CompressorV2.recommendation_compression(compression_method, recommendation_method, recommendation_ratio, input_model_path, output_dir, input_shapes, framework=Framework.PYTORCH, options=RecommendationOptions(), dataset_path=None)

Compress a recommendation-based model using the given compression and recommendation methods.

Parameters:

Name	Type	Description	Default
compression_method	CompressionMethod	The selected compression method.	required
recommendation_method	RecommendationMethod	The selected recommendation method.	required
recommendation_ratio	float	The compression ratio recommended by the recommendation method.	required
input_model_path	str	The file path where the model is located.	required
output_dir	str	The local path to save the compressed model.	required
input_shapes	List[Dict[str, int]]	Input shapes of the model.	required
framework	Framework	The framework of the model.	PYTORCH
options(Options,	optional	The options for pruning method.	required
dataset_path	str	The path of the dataset used for nuclear norm compression method. Default is None.	None

Raises:

Type	Description
e	If an error occurs while performing recommendation compression.

Returns:

Name	Type	Description
CompressorMetadata	CompressorMetadata	Compress metadata.

Details of Parameters

Compression Method

Available Compression Method

Name	Description
PR_L2	L2 Norm Pruning
PR_GM	GM Pruning
PR_NN	Nuclear Norm Pruning
PR_SNP	Structured Neuron-level Pruning
FD_TK	Tucker Decomposition
FD_SVD	Singular Value Decomposition

Example

from netspresso.enums import CompressionMethod

COMPRESSION_METHOD = CompressionMethod.PR_L2

Warning

• Nuclear Norm is only supported in the Tensorflow-Keras framework.
• Structured Neuron-level is only supported in the PyTorch and ONNX frameworks.

Note

• Click on the link to learn more about the information. (Compression Method)

Recommendation Method

Available Recommendation Method

Name	Description
SLAMP	Structured Layer-adaptive Sparsity for the Magnitude-based Pruning
VBMF	Variational Bayesian Matrix Factorization

Example

from netspresso.enums import RecommendationMethod

RECOMMENDATION_METHOD = RecommendationMethod.SLAMP

Note

• If you selected PR_L2, PR_GM, PR_NN, PR_SNP for compression_method
  • The recommended_method available is SLAMP.
• If you selected FD_TK, FD_SVD for compression_method
  • The recommended_method available is VBMF.

Recommendation Ratio

• SLAMP (Pruning ratio)

  • Remove corresponding amounts of the filters. (e.g. 0.2 removes 20% of the filters in each layer)

  • Available ranges: 0 < ratio < 1

  • Click the link for more information. (SLAMP)

• VBMF (Calibration ratio)

  • This function control compression level of model if the result of recommendation doesn't meet the compression level user wants. Remained rank add or subtract (removed rank x calibration ratio) according to calibration ratio range.
    

  • Available ranges: -1 ≤ ratio ≤ 1

  • Click the link for more information. (VBMF)

Options

Example

from netspresso.enums import Policy, LayerNorm, GroupPolicy
from netspresso.clients.compressor.v2.schemas import Options

OPTIONS = Options(
    policy=Policy.AVERAGE,
    layer_norm=LayerNorm.TSS_NORM,
    group_policy=GroupPolicy.COUNT,
    reshape_channel_axis=-1
)

Note

• Click the link for more information. (Pruning Options)

Warning

• Nuclear Norm is only supported in the Tensorflow-Keras framework.
• Structured Neuron-level is only supported in the PyTorch and ONNX frameworks.

Example

from netspresso import NetsPresso
from netspresso.enums import CompressionMethod, RecommendationMethod


netspresso = NetsPresso(email="YOUR_EMAIL", password="YOUR_PASSWORD")

compressor = netspresso.compressor_v2()
compressed_model = compressor.recommendation_compression(
    compression_method=CompressionMethod.PR_L2,
    recommendation_method=RecommendationMethod.SLAMP,
    recommendation_ratio=0.5,
    input_model_path="./examples/sample_models/graphmodule.pt",
    output_dir="./outputs/compressed/graphmodule_recommend",
    input_shapes=[{"batch": 1, "channel": 3, "dimension": [224, 224]}],
)