AIMET Keras Compression API


AIMET supports the following model compression techniques for keras models
  • Spatial SVD

To learn more about these model compression techniques, please see Model Compression User Guide

For the Spatial SVD compression techniques, there are two modes in which you can invoke the AIMET API
  • Auto Mode: In Auto mode, AIMET will determine the optimal way to compress each layer of

    the model given an overall target compression ratio. Greedy Compression Ratio Selection Algorithm is used to pick appropriate compression ratios for each layer.

  • Manual Mode: In Manual mode, the user can pass in the desired compression-ratio per layer

    to AIMET. AIMET will apply the specified compression technique for each of the layers to achieve the desired compression-ratio per layer. It is recommended that the user start with Auto mode, and then tweak per-layer compression-ratios using Manual mode if desired.

Top-level API for Compression

class aimet_tensorflow.keras.compress.ModelCompressor[source]

aimet model compressor: Enables model compression using various schemes

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

Compress a given model using the specified parameters

  • model (Model) – Model, represented by a tf.keras.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.

  • 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 (SpatialSvdParameters) – Compression parameters specific to given compression scheme

  • trainer (Optional[Callable]) – Training function None: If per layer fine-tuning is not required while creating the final compressed model

  • visualization_url (Optional[str]) – url the user will need to input where visualizations will appear

Return type:

Tuple[Model, CompressionStats]


A tuple of the compressed model session, 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

  • 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.

Spatial SVD Configuration

class aimet_tensorflow.defs.SpatialSvdParameters(input_op_names, output_op_names, mode, params, multiplicity=1)[source]

Configuration parameters for spatial svd compression

  • input_op_names (List[str]) – list of input op names to the model

  • output_op_names (List[str]) – List of output op names of the model

  • 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(greedy_select_params, modules_to_ignore=None)[source]

Configuration parameters for auto-mode compression

  • greedy_select_params (GreedySelectionParameters) – Params for greedy comp-ratio selection algorithm

  • modules_to_ignore (Optional[List[Operation]]) – List of modules to ignore (None indicates nothing to ignore)

class ManualModeParams(list_of_module_comp_ratio_pairs)[source]

Configuration parameters for manual-mode spatial svd compression


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

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



memory = 2

Cost modeled for space requirements



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

Enumeration of compression schemes supported in aimet


Only Spatial SVD is supported for now.

channel_pruning = 3

Channel Pruning

spatial_svd = 2

Spatial SVD

weight_svd = 1

Weight SVD

class aimet_tensorflow.defs.ModuleCompRatioPair(module, comp_ratio)[source]

Pair of tf.Operation and a compression-ratio

  • module – Module of type tf.Operation

  • comp_ratio – Compression ratio. Compression ratio is the ratio of cost of compressed model to cost of the original model.

Code Examples

Required imports

from decimal import Decimal
from typing import Tuple

import numpy as np
import tensorflow as tf
from tensorflow.keras.applications.resnet import ResNet50, preprocess_input, decode_predictions

# imports for AIMET
import aimet_common.defs as aimet_common_defs
from aimet_tensorflow.keras.compress import ModelCompressor
import aimet_tensorflow.defs as aimet_tensorflow_defs

Evaluation function

def get_eval_func(dataset_dir, batch_size, num_iterations=50000):
    Sample Function which returns an evaluate function callback which can be
    called to evaluate a model on the provided dataset
    def func_wrapper(model, iterations):
        validation_ds = tf.keras.preprocessing.image_dataset_from_directory(
            image_size=(224, 224))
        # If no iterations specified, set to full validation set
        if not iterations:
            iterations = num_iterations
            iterations = iterations * batch_size
        top1 = 0
        total = 0
        inp_data = None
        for (img, label) in validation_ds:
            x = preprocess_input(img)
            inp_data = x if inp_data is None else inp_data
            preds = model.predict(x, batch_size=batch_size)
            label = np.where(label)[1]
            label = [validation_ds.class_names[int(i)] for i in label]
            cnt = sum([1 for a, b in zip(label, decode_predictions(preds, top=1)) if str(a) == b[0][0]])
            top1 += cnt
            total += len(label)
            if total >= iterations:

        return top1/total
    return func_wrapper

Compressing using Spatial SVD in auto mode

def aimet_spatial_svd(model, evaluator: aimet_common_defs.EvalFunction) -> Tuple[tf.keras.Model,
    Compresses the model using AIMET's Keras Spatial SVD auto mode compression scheme.

    :param model: The keras model to compress
    :param evaluator: Evaluator used during compression
    :return: A tuple of compressed sess graph and its statistics

    # Desired target compression ratio using Spatial SVD
    # This value denotes the desired compression % of the original model.
    # To compress the model to 20% of original model, use 0.2. This would
    # compress the model by 80%.
    # We are compressing the model by 50% here.
    target_comp_ratio = Decimal(0.5)

    # Number of compression ratio used by the API at each layer
    # API will evaluate 0.1, 0.2, ..., 0.9, 1.0 ratio (total 10 candidates)
    # at each layer
    num_comp_ratio_candidates = 10

    # Creating Greedy selection parameters:
    greedy_params = aimet_common_defs.GreedySelectionParameters(target_comp_ratio=target_comp_ratio,

    # Ignoring first convolutional layer of the model for compression
    modules_to_ignore = [model.layers[2]]

    # Creating Auto mode Parameters:
    auto_params = aimet_tensorflow_defs.SpatialSvdParameters.AutoModeParams(greedy_select_params=greedy_params,

    # Creating Spatial SVD parameters with Auto Mode:
    params = aimet_tensorflow_defs.SpatialSvdParameters(input_op_names=model.inputs,

    # Scheme is Spatial SVD:
    scheme = aimet_common_defs.CompressionScheme.spatial_svd

    # Cost metric is MAC, it can be MAC or Memory
    cost_metric = aimet_common_defs.CostMetric.mac

    # Calling model compression using Spatial SVD:
    # Here evaluator is passed which is used by the API to evaluate the
    # accuracy for various compression ratio of each layer. To speed up
    # the process, only 10 batches of data is being used inside evaluator
    # (by passing eval_iterations=10) instead of running evaluation on
    # complete dataset.
    results = ModelCompressor.compress_model(model=model,

    return results

Sample Driver Code for Spatial SVD using Resnet50

def compress():
    Example Driver Function Code in which we are compressing Resnet50 model.
    dataset_dir = '/path/to/dataset'
    model = ResNet50(weights='imagenet')
    eval_func = get_eval_func(dataset_dir, batch_size=16)
    compressed_model, stats = aimet_spatial_svd(model=model, evaluator=eval_func)