recurrent Namespace Reference

Module caffe2.python.recurrent. More...

Functions
def	recurrent_net (net, cell_net, inputs, initial_cell_inputs, links, timestep=None, scope=None, outputs_with_grads=(0,), recompute_blobs_on_backward=None)
def	MILSTM (model, input_blob, seq_lengths, initial_states, dim_in, dim_out, scope, outputs_with_grads=(0,), memory_optimization=False, forget_bias=0.0)

Detailed Description

Module caffe2.python.recurrent.

Function Documentation

MILSTM()

def recurrent.MILSTM	(		model,
			input_blob,
			seq_lengths,
			initial_states,
			dim_in,
			dim_out,
			scope,
			outputs_with_grads = (0,),
			memory_optimization = False,
			forget_bias = 0.0
	)

Adds MI flavor of standard LSTM recurrent network operator to a model.
See https://arxiv.org/pdf/1606.06630.pdf

model: CNNModelHelper object new operators would be added to

input_blob: the input sequence in a format T x N x D
where T is sequence size, N - batch size and D - input dimention

seq_lengths: blob containing sequence lengths which would be passed to
LSTMUnit operator

initial_states: a tupple of (hidden_input_blob, cell_input_blob)
which are going to be inputs to the cell net on the first iteration

dim_in: input dimention

dim_out: output dimention

outputs_with_grads : position indices of output blobs which will receive
external error gradient during backpropagation

memory_optimization: if enabled, the LSTM step is recomputed on backward step
               so that we don't need to store forward activations for each
               timestep. Saves memory with cost of computation.

Definition at line 252 of file recurrent.py.

recurrent_net()

def recurrent.recurrent_net	(		net,
			cell_net,
			inputs,
			initial_cell_inputs,
			links,
			timestep = None,
			scope = None,
			outputs_with_grads = (0,),
			recompute_blobs_on_backward = None
	)

net: the main net operator should be added to

cell_net: cell_net which is executed in a recurrent fasion

inputs: sequences to be fed into the recurrent net. Currently only one input
is supported. It has to be in a format T x N x (D1...Dk) where T is lengths
of the sequence. N is a batch size and (D1...Dk) are the rest of dimentions

initial_cell_inputs: inputs of the cell_net for the 0 timestamp.
Format for each input is:
    (cell_net_input_name, external_blob_with_data)

links: a dictionary from cell_net input names in moment t+1 and
output names of moment t. Currently we assume that each output becomes
an input for the next timestep.

timestep: name of the timestep blob to be used. If not provided "timestep"
is used.

scope: Internal blobs are going to be scoped in a format
<scope_name>/<blob_name>
If not provided we generate a scope name automatically

outputs_with_grads : position indices of output blobs which will receive
error gradient (from outside recurrent network) during backpropagation

recompute_blobs_on_backward: specify a list of blobs that will be
             recomputed for backward pass, and thus need not to be
             stored for each forward timestep.

Definition at line 18 of file recurrent.py.