save the current model to the `FLAGS.channel_pruned_path`

Save a list of values to a plain text file. Args: * vals: list of values * file_path: file path

search activation operation in graph for quantization

search activation operation in graph for quantization

search matmul or Conv2D operations in graph for quantization

search matmul or Conv2D operations in graph for quantization

Setup the anchor bounding boxes' information.

Creates a split convolution. Split convolution splits the input and output into 'num_blocks' blocks of approximately the same size each, and on

Test the *.tflite model. Args: * file_path: file path to the *.tflite model * net_input_data: network's input node's data

Update running averages with a list of samples. Args: * x_new: np.array of list of samples (N x D)

constraint action during reinfocement learning search

Build actor & critic networks, replay buffer, and noise generators. Args: * s_dims: number of state vector's dimensions * a_dims: number

Build block-wise fine-tuning operations. Returns: * block_train_ops: list of training operations for each block * block_init_opt_ops: lis

Build the quantization points on [0, 1] and quantize 'x_normalized', the function applies for both type 'split' and 'channel'. Args:

Build a list of Conv2D operation's information. Args: * conv_krnls_prnd: list of convolutional kernels in the channel-pruned model Retur

Build the evaluation graph.

Build the evaluation graph.

Build the evaluation graph.

Build the evaluation graph for the 'optimal' protocol. Args: * model_helper: model helper with definitions of model & dataset

Build the evaluation graph.

Build the evaluation graph.

Build extra losses for regression. Returns: * reg_losses: list of regression losses (one per layer)

Build extra losses for regression & discrimination. Args: * labels: one-hot label vectors Returns: * reg_losses: list of regression

Build layer-wise fine-tuning operations. Returns: * layer_ops: list of training and initialization operations for each layer * lrn_rates_

Build layer-wise fine-tuning operations. Returns: * layer_train_ops: list of training operations for each layer * layer_init_opt_ops: lis

Build operations for layerwise regression. Returns: * init_op: initialization operation * train_ops: list of training operations, one per

build masks and corresponding operations for weight sparsification. Returns: * masks: list of masks for weight sparsification * prune_op:

Build pruning masks for all the maskable variables. Returns: * masks: list of pruning masks * pr_all: placeholder of all the pruning rati

Build a meta LASSO optimization problem.

Build a meta least-square optimization problem.

Build the minimal graph for 'uniform' & 'heurist' protocols. Args: * model_helper: model helper with definitions of model & dataset

Build operations for network fine-tuning. Args: * loss: loss function's value Returns: * init_op: initialization operation * tra

Build network training operations. Returns: * train_op: training operation of the whole network * init_opt_op: initialization operation o

Build network training operations. Returns: * train_op: training operation of the whole network * init_opt_op: initialization operation o

Build the quantization points on [0, 1] and quantize 'x_normalized', the function applies for no_bucket and channel-wise bucket. Args:

Build the channel pruning graph.

build a training model from pruned model

Build the RL helper and controller / agent. Args: * sess: TensorFlow session Returns: * rl_helper: RL helper * agent: RL control

Build the training graph.

Build the training graph.

Build the training graph.

Build the training graph for the 'optimal' protocol. Args: * model_helper: model helper with definitions of model & dataset

Build the training graph.

Build the training graph.

Calculate the mask-pruned gradients. Args: * grads_origin: list of original gradients Returns: * grads_pruned: list of mask-pruned g

Calculate the mask-pruned gradients. Args: * grads_origin: list of original gradients Returns: * grads_pruned: list of mask-pruned g

Calculate the pruned gradients Args: * grads_origin: the original gradient Return: * the pruned gradients * the corresponding mas

Calculate pruning ratios using the 'heurist' protocol. Returns: * var_names_n_prune_ratios: list of variable name & pruning ratio pairs

evaluate on the validation set

Calculate pruning ratios using the 'optimal' protocol. Returns: * var_names_n_prune_ratios: list of variable name & pruning ratio pairs

Calculate the dynamic pruning ratio. Args: * prune_ratio_fnl: final pruning ratio Returns: * prune_ratio_dyn: dynamic pruning ratio

Calculate the minimal & maximal pruning ratio for the chosen maskable variable. Args: * idx: index to the chosen maskable variable Retur

Calculate actions within one roll-out. Returns: * prune_ratios: list of pruning ratios * states_n_actions: list of state vector and actio

Calculate the reward of the current roll-out. Args: * prune_ratios: list of pruning ratios Returns: * reward: reward of the current

Calculate pruning ratios using the 'uniform' protocol. Returns: * var_names_n_prune_ratios: list of variable name & pruning ratio pairs

Compute the maximum bits used for layer idx

Compute the maximum bits used for layer idx

reshape weights according to bucket for 'channel' type. Note that for fc layers, buckets are created row-wisely. Args: w: A Tensor

Choose channels for all convolutional layers.

Choose channels for all convolutional layers.

Choose discrimination-aware channels.

create extracters which would be used to extract input of a convolution

Evaluate the model's loss & accuracy.

extract the input X (k_h, k_w, c) of a conv layer Args: conv: a convolution operation Returns: bathces of X (N, k_h, k_w, c)

extract new feature map via re-sampling some points

Extract output tensor name of convolution layers and sum layers in a residual block

Initialize a model for training with the distillation loss. Note: If the pre-trained model is not available on HDFS, then a new model will be tra

Inversed linear scale function Args: * w: A Tensor (weights or activation output) * alpha: A float value, scale factor * bete: A floa

Monitor the training progress. Args: * summary: summary protocol buffer * log_rslt: logging operations' results * idx_iter: index of

Monitor the training progress. Args: * summary: summary protocol buffer * log_rslt: logging operations' results * idx_iter: index of

Monitor the training progress. Args: * summary: summary protocol buffer * log_rslt: logging operations' results * idx_iter: index of

Monitor the training progress. Args: * summary: summary protocol buffer * log_rslt: logging operations' results * idx_iter: index of

Monitor the training progress. Args: * summary: summary protocol buffer * log_rslt: logging operations' results * idx_iter: index of

Monitor the training progress. Args: * summary: summary protocol buffer * log_rslt: logging operations' results * idx_iter: index of