Weight SVD¶

Context¶

Weight singular value decomposition (Weight SVD) is a technique that decomposes one large layer (in terms of multiply-accumulate (MAC) or memory) into two smaller layers.

Consider a convolution (Conv) layer with the kernel (m, n, h, w) where:

m is the input channels
n the output channels
h is the height of the kernel
w is the width of the kernel

Weight SVD decomposes the kernel into one of size (m, k, 1, 1) and another of size (k, n, h, w), where 𝑘 is called the rank. The smaller the value of k, larger the degree of compression.

The following figure illustrates how weight SVD decomposes the output channel dimension. Weight SVD is currently supported for convolution (Conv) and fully connected (FC) layers in AIMET.

Workflow¶

Code example¶

Setup¶

PyTorch

import os
from decimal import Decimal
import torch


# Compression-related imports
from aimet_common.defs import CostMetric, CompressionScheme, GreedySelectionParameters, RankSelectScheme
from aimet_torch.defs import WeightSvdParameters, SpatialSvdParameters, ChannelPruningParameters, \
    ModuleCompRatioPair
from aimet_torch.compress import ModelCompressor

def evaluate_model(model: torch.nn.Module, eval_iterations: int, use_cuda: bool = False) -> float:
    """
    This is intended to be the user-defined model evaluation function.
    AIMET requires the above signature. So if the user's eval function does not
    match this signature, please create a simple wrapper.

    Note: Honoring the number of iterations is not absolutely necessary.
    However if all evaluations run over an entire epoch of validation data,
    the runtime for AIMET compression will obviously be higher.

    :param model: Model to evaluate
    :param eval_iterations: Number of iterations to use for evaluation.
            None for entire epoch.
    :param use_cuda: If true, evaluate using gpu acceleration
    :return: single float number (accuracy) representing model's performance
    """
    return .5

Compression using Weight SVD¶

PyTorch

Compressing using Weight SVD in auto mode

def weight_svd_auto_mode():

    # Load trained MNIST model
    model = torch.load(os.path.join('../', 'data', 'mnist_trained_on_GPU.pth'))

    # Specify the necessary parameters
    greedy_params = GreedySelectionParameters(target_comp_ratio=Decimal(0.8),
                                              num_comp_ratio_candidates=10)
    rank_select = RankSelectScheme.greedy
    auto_params = WeightSvdParameters.AutoModeParams(rank_select_scheme=rank_select,
                                                     select_params=greedy_params,
                                                     modules_to_ignore=[model.conv1])

    params = WeightSvdParameters(mode=WeightSvdParameters.Mode.auto,
                                 params=auto_params)

    # Single call to compress the model
    results = ModelCompressor.compress_model(model,
                                             eval_callback=evaluate_model,
                                             eval_iterations=1000,
                                             input_shape=(1, 1, 28, 28),
                                             compress_scheme=CompressionScheme.weight_svd,
                                             cost_metric=CostMetric.mac,
                                             parameters=params)

    compressed_model, stats = results
    print(compressed_model)
    print(stats)     # Stats object can be pretty-printed easily

Compressing using Weight SVD in manual mode with multiplicity = 8 for rank rounding

def weight_svd_manual_mode():

    # Load a trained MNIST model
    model = torch.load(os.path.join('../', 'data', 'mnist_trained_on_GPU.pth'))

    # Specify the necessary parameters
    manual_params = WeightSvdParameters.ManualModeParams([ModuleCompRatioPair(model.conv1, 0.5),
                                                          ModuleCompRatioPair(model.conv2, 0.4)])
    params = WeightSvdParameters(mode=WeightSvdParameters.Mode.manual,
                                 params=manual_params, multiplicity=8)

    # Single call to compress the model
    results = ModelCompressor.compress_model(model,
                                             eval_callback=evaluate_model,
                                             eval_iterations=1000,
                                             input_shape=(1, 1, 28, 28),
                                             compress_scheme=CompressionScheme.weight_svd,
                                             cost_metric=CostMetric.mac,
                                             parameters=params)

    compressed_model, stats = results
    print(compressed_model)
    print(stats)    # Stats object can be pretty-printed easily

API¶

PyTorch

Top-level API for Compression

class aimet_torch.compress.ModelCompressor[source]: AIMET model compressor: Enables model compression using various schemes

static ModelCompressor.compress_model(model, eval_callback, eval_iterations, input_shape, compress_scheme, cost_metric, parameters, trainer=None, visualization_url=None)[source]

Compress a given model using the specified parameters

Parameters:

model (Module) – Model to compress
eval_callback (Callable[[Any, Optional[int], bool], float]) – Evaluation callback. Expected signature is evaluate(model, iterations, use_cuda). Expected to return an accuracy metric.
eval_iterations – Iterations to run evaluation for
trainer – Training Class: Contains a callable, train_model, which takes model, layer which is being fine tuned and an optional parameter train_flag as a parameter None: If per layer fine tuning is not required while creating the final compressed model
input_shape (Tuple) – Shape of the input tensor for model
compress_scheme (CompressionScheme) – Compression scheme. See the enum for allowed values
cost_metric (CostMetric) – Cost metric to use for the compression-ratio (either mac or memory)
parameters (Union[SpatialSvdParameters, WeightSvdParameters, ChannelPruningParameters]) – Compression parameters specific to given compression scheme
visualization_url – url the user will need to input where visualizations will appear

Return type:

Tuple[Module, CompressionStats]

Returns:

A tuple of the compressed model, and compression statistics

Greedy Selection Parameters

class aimet_common.defs.GreedySelectionParameters(target_comp_ratio, num_comp_ratio_candidates=10, use_monotonic_fit=False, saved_eval_scores_dict=None)[source]

Configuration parameters for the Greedy compression-ratio selection algorithm

Variables:

target_comp_ratio – Target compression ratio. Expressed as value between 0 and 1. Compression ratio is the ratio of cost of compressed model to cost of the original model.
num_comp_ratio_candidates – Number of comp-ratio candidates to analyze per-layer More candidates allows more granular distribution of compression at the cost of increased run-time during analysis. Default value=10. Value should be greater than 1.
use_monotonic_fit – If True, eval scores in the eval dictionary are fitted to a monotonically increasing function. This is useful if you see the eval dict scores for some layers are not monotonically increasing. By default, this option is set to False.
saved_eval_scores_dict – Path to the eval_scores dictionary pickle file that was saved in a previous run. This is useful to speed-up experiments when trying different target compression-ratios for example. aimet will save eval_scores dictionary pickle file automatically in a ./data directory relative to the current path. num_comp_ratio_candidates parameter will be ignored when this option is used.

Configuration Definitions

class aimet_common.defs.CostMetric(value)[source]

Enumeration of metrics to measure cost of a model/layer

mac = 1

Cost modeled for compute requirements

Type:: MAC

memory = 2

Cost modeled for space requirements

Type:: Memory

class aimet_common.defs.CompressionScheme(value)[source]

Enumeration of compression schemes supported in aimet

channel_pruning = 3: Channel Pruning

spatial_svd = 2: Spatial SVD

weight_svd = 1: Weight SVD

Weight SVD Configuration

class aimet_torch.defs.WeightSvdParameters(mode, params, multiplicity=1)[source]

Configuration parameters for weight svd compression

Parameters:

mode (Mode) – Either auto mode or manual mode
params (Union[ManualModeParams, AutoModeParams]) – Parameters for the mode selected
multiplicity – The multiplicity to which ranks/input channels will get rounded. Default: 1

class AutoModeParams(rank_select_scheme, select_params, modules_to_ignore=None)[source]

Configuration parameters for auto-mode compression

Parameters:

rank_select_scheme (RankSelectScheme) – supports two options greedy and tar
select_params (GreedySelectionParameters) – Params for greedy/TAR comp-ratio selection algorithm
modules_to_ignore (Optional[List[Module]]) – List of modules to ignore (None indicates nothing to ignore)

class ManualModeParams(list_of_module_comp_ratio_pairs)[source]

Configuration parameters for manual-mode weight svd compression

Parameters:: list_of_module_comp_ratio_pairs (List[ModuleCompRatioPair]) – List of (module, comp-ratio) pairs

class Mode(value)[source]

Mode enumeration

auto = 2: Auto mode

manual = 1: Manual mode