Run the optimizer to obtain pruning ratios for all maskable variables. Returns: * var_names_n_prune_ratios: list of variable name & pruning r

Append multiple transitions into the replay buffer. Args: * states: np.array of current state vectors * actions: np.array of action vecto

Check whether is the primary worker of all nodes (global) or the current node (local). Args: * scope: check scope ('global' OR 'local')

Calculate the distillation loss for the primary model's logits. Args: logits_pri: primary model's logits (shape: N x K) logits_dst: d

Get the number of workers at all nodes.

Automatically insert a barrier for multi-GPU training, or pass for single-GPU training.

Get the rank of current worker at the current node.

get the bit allocation strategy either with RL or not

Build iterator(s) for tf.data.Dataset() object. Args: * enbl_trn_val_split: whether to split into training & validation subsets Returns:

get the bit allocations either with RL or not

Get a distributed optimizer from the base optimizer.

Calculate the distilled model's logits for given images. Args: * sess: TensorFlow session to restore model weights * images: input images

Download remote model files and then uncompress. Note: All files in FLAGS.save_path will be removed and replaced by the pre-trained model.

Forward computation at training.

Performs a batch normalization using a standard set of parameters.

Forward computation at evaluation.

Check whether is the primary worker of all nodes (global) or the current node (local). Args: * scope: check scope ('global' OR 'local') Return

Build the data subset for training, usually with data augmentation.

Weather it is the primary worker

Strided 2-D convolution with explicit padding.

Dump the model's outputs to files and evaluate.

Build the data subset for evaluation, usually without data augmentation.

Get a TensorFlow operation to broadcast all the global variables.

Train the agent's actor & critic networks.

Wrapper for inserting int64 features into Example proto.

Create the learner as specified by FLAGS.learner. Args: * sm_writer: TensorFlow's summary writer * model_helper: model helper with definitions

Restore a model from the latest checkpoint files and then evaluate it.

Initialize the agent before all roll-outs. Actor & critic networks will be initialized and the replay buffer will be reset.

Setup the learning rate (and number of training iterations).

Calculate the overall pruning ratio for the given list of variables. Args: * vars_list: list of variables Returns: * prune_ratio: overall pr

Finalize the current roll-out (to update the baseline function). Args: * rewards: reward scalars (one per roll-out tick)

Initialize the agent before each roll-out. Parameter / action noise will be re-initialized for the upcoming roll-out.

if the op's input channels can be pruned

Append multiple transitions into the replay buffer. Args: * states: np.array of current state vectors * actions: np.array of action vecto

Calculate the accuracy of pruned model

get the definition of an operation which contains the following information: `ksizes`: kernel sizes `padding`: paddings `h`: height

Get the weights of an operation Args: op: An operation Return: The weights of a operation

Get the rank of current worker at all nodes.

In Default, we do not use BN to train resnet, since batch_size is too small. So is_training is False and trainable is False in the batch_nor

Setup the learning rate with piecewise constant strategy. Args: * global_step: training iteration counter * batch_size: number of samples in ea

Bind keys and values to dictionaries. Args: * keys: A list of op_names; * values: A Tensor with len(keys) elements; Returns: *

Bind keys and values to dictionaries. Args: * keys: A list of op_names * values: A Tensor with len(op_names) elements Returns:

Save the current model for training or evaluation. Args: * is_train: whether to save a model for training

Save the current model for training or evaluation. Args: * is_train: whether to save a model for training

Pads the input along the spatial dimensions independently of input size. Pads the input such that if it was used in a convolution with 'VALID' padd

Wrapper for inserting float features into Example proto.

Calculate the overall pruning ratio for the given list of variables. Args: * vars_list: list of variables Returns: * prune_ratio: overall pr

Calculate the overall pruning ratio for the given list of variables. Args: * vars_list: list of variables Returns: * prune_ratio: overall pr

Calculate the overall pruning ratio for the given list of variables. Args: * vars_list: list of variables Returns: * prune_ratio: overall pr

Calculate the reward. Args: * accuracy: classification accuracy after applying masks Returns: * reward: reward function's value

compute the flops of a certain convolution layer Args: operation: an convolution layer Return: The Flops

Get list of variables within certain name scope. Args: * scope: name scope Returns: * vars_dict: dictionary of list of all, trainable, and m

Get list of variables within certain name scope. Args: * scope: name scope Returns: * vars_dict: dictionary of list of all, trainable, and c

Reset the current position.

Move to the next position.

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

Inversed linear scale function Args: * w: A Tensor (weights or activations), origin shape if not use_buckets else [bucket_size, buck

Make an iterator from tf.data.Dataset. Args: * dataset: tf.data.Dataset object Returns: * iterator: iterator for the dataset

Make an iterator from tf.data.Dataset. Args: * dataset: tf.data.Dataset object Returns: * iterator: iterator for the dataset

Save the current model for training or evaluation. Args: * is_train: whether to save a model for training

Save the current model for training or evaluation. Args: * is_train: whether to save a model for training

Save the model to checkpoint files for training or evaluation. Args: * is_train: whether to save a model for training

linear scale function Args: * w: A Tensor (weights or activation output), the shape is [bucket_size, bucekt_num] if use_bucket else

Wrapper for inserting bytes features into Example proto.

:param bbox_pred: [-1, 4] in RPN. [-1, cls_num+1, 4] in Fast-rcnn :param bbox_targets: shape is same as bbox_pred :param sigma: :ret

Compress the model by channel pruning Args: action: preserving ratio

De-normalize with the given running averages of mean value & standard deviation. Args: * smpl_mat: normalized sample matrix * rms: running aver

A block of densely connected layers. Args: * inputs: input tensor to the block * units: number of neurons in each layer Returns: * inputs:

Restore a model from the latest checkpoint files and then evaluate it.

Restore a model from the latest checkpoint files and then evaluate it.

Extract feature-maps and do sampling for some convolutions with given images Args: names: convolution operation names init_fn: i

whether final layer reached

Forward pass function. Args: * inputs: inputs to the network's forward pass * data_format: data format ('channels_last' OR 'channels_first')

Get the input of an operation Args: op: An operation Return: The input of a operation

Get all the operations of a certain type within the model graph Args: otype: The type of returned operations Return: operation

Get list of operations within certain name scope and also matches the pattern. Args: * scope: name scope * pattern: name pattern to be matched

Get list of operations within certain name scope and also matches the pattern. Args: * scope: name scope * pattern: name pattern to be matched

Get list of operations within certain name scope and also matches the pattern. Args: * scope: name scope * pattern: name pattern to be matched

get output name by operation name

Get list of variables within certain name scope. Args: * scope: name scope Returns: * vars_dict: dictionary of list of all and trainable var

Get list of variables within certain name scope. Args: * scope: name scope Returns: * vars_dict: dictionary of list of all, trainable, and m

Get list of variables within certain name scope. Args: * scope: name scope Returns: * vars_dict: dictionary of list of all, trainable, and m