npu_dynamic_rnn
Applicability
Product |
Supported (√/x) |
|---|---|
Atlas 350 Accelerator Card |
√ |
√ |
|
√ |
|
X |
|
X |
|
√ |
Description
Creates a high-performance neural network specified by RNNCell.
Prototype
1 2 3 4 5 6 | def npu_dynamic_rnn(cell, inputs, initial_state=None, dtype=None, sequence_length=None, scope=None) |
Parameters
Parameter |
Input/Output |
Description |
|---|---|---|
cell |
Input |
RNNCell instance, which is the memory unit of long short-term memory (LSTM) and gated recurrent unit (GRU). |
inputs |
Input |
An input list whose length is T. Each input is a tuple whose shape is [max_time, batch_size, input_size], or a nested tuple of this shape. |
initial_state |
Input |
(Optional) Initial state of the recurrent neural network (RNN). If cell.state_size is an integer, it must be a tensor whose shape is [batch_size, cell.state_size]. If cell.state_size is a tuple, it must be a tuple of the [batch_size, s] tensor in cell.state_size. |
dtype |
Input |
(Optional) Data type of the initial state and expected output. This parameter is required if initial_state is empty, or there is a heterogeneous data type in the RNN state. |
sequence_length |
Input |
Length of each sequence for an input. This parameter is an int32 or int64 vector (tensor) whose size is [batch_size]. The value range is [0, T). |
scope |
Input |
VariableScope of the subgraph. Defaults to "rnn". |
Returns
Result tensor of the RNN.
Final state.
Restrictions
This API applies to the neural machine translation (NMT) network training script in the while_loop expansion scenario.
Example
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 | from npu_bridge.npu_init import * # Original code: inputs = npu_unstack(self.encoder_emb_inp, axis=0) encoder_outputs , encoder_state = static_rnn( cell, inputs, dtype= dtype, sequence_length = sequence_length ) encoder_outputs = npu_stack( encoder_outputs, axis=0 ) # Replace it with: encoder_outputs , encoder_state = npu_rnn.npu_dynamic_rnn( cell, inputs=self.encoder_emb_inp, dtype= dtype, sequence_length= sequence_length) |