Helper for training with distillation loss. Other learners can use calc_loss() (remember to call initialize() first to make sure that the pre-t

Actor network.

CIFAR-10 dataset.

ILSVRC-12 dataset.

Pascal VOC dataset.

Critic network.

Fashion-MNIST dataset.

Full-precision learner (no model compression applied).

Model helper for creating a ResNet model for the ILSVRC-12 dataset.

Class of non-uniform quantization

Running averages of mean value & standard deviation.

Class of uniform quantization

Adaptive parameter noise's specifications. To enable, set <ddpg_noise_type> to 'param' and <ddpg_noise_prtl> to 'adapt'.

Currently only weight bits are inferred via RL. Activations later.

Currently only weight bits are inferred via RL. Activations later.

Channel pruning learner with GPU-based optimization.

Learner with channel/filter pruning

Channel pruning learner - remastered.

The Channel Prunner

Discrimination-aware channel pruning learner.

Environment for the Move-to-target problem. The reward for moving from <x_curr> to <x_next> is defined as: reward := Dist(x_curr, target) - Dis

Helper class that provides TensorFlow image coding utilities.

The model wraper make it easier to do some operation on a tensorflow model

Model helper for creating a MobileNet model for the ILSVRC-12 dataset.

Model helper for creating a VGG model for the VOC dataset.

Model helper for creating a LeNet-like model for the CIFAR-10 dataset.

Model helper for creating a VGG model for the VOC dataset.

Model helper for creating a ResNet model for the CIFAR-10 dataset.

Model helper for creating a ConvNet model for the Fashion-MNIST dataset.

No-op context manager.

Nonuniform quantization for weights and activations

Pruning ratio optimizer for the weight sparsification learner.

Reinforcement learning helper for the weight sparsification learner.

The Helper Class for the DDPG algorithm. Making sure the states and actions satisfy the condition of constraints given the total number bits.

The Helper Class for the DDPG algorithm. Making sure the states and actions satisfy the condition of constraints given the total number bits.

Replay buffer for storing state-action-reward transitions. Each transition consists of five components: 1. state: current state 2. action: acti

Time-decaying action / parameter noise's specifications. To enable, set <ddpg_noise_type> to 'action' / 'param' and <ddpg_noise_prtl> to 'tdecy'.

Uniform quantization for weights and activations

Uniform quantization learner with TensorFlow's quantization APIs.

Weight sparsification learner.

Abstract class for datasets.

Abstract class for learners. A learner should take a ModelHelper object as input, which includes the data input pipeline and model definition,

Abstract class for model helpers. A model helper should define the following function interface: 1. Data input pipeline for training and evalua

DDPG (Deep Deterministic Policy Gradient) agent.

Base class for building the Resnet Model.

Abstract model for actor & critic networks.

Wrapper for multi-GPU training.