DynamicRNN Constructor

Applicability

Product	Supported
Atlas A3 training products/Atlas A3 inference products	√
Atlas A2 training products/Atlas A2 inference products	√
Atlas 200I/500 A2 inference products	☓
Atlas inference products	☓
Atlas training products	√

Description

Used for RNN training and inference with TensorFlow.

Prototype

class DynamicRNN(_DynamicBasic):
    def __init__(self,
                 hidden_size,
                 dtype,
                 cell_type="LSTM",
                 direction=DYNAMIC_RNN_UNIDIRECTION,
                 cell_depth=1,
                 use_peephole=False,
                 keep_prob=1.0,
                 cell_clip=-1.0,
                 num_proj=0,
                 time_major=True,
                 activation="tanh",
                 forget_bias=0.0,
                 is_training=True)

Parameters

Parameter	Input/Output	Description
hidden_size	Input	Dimension of the RNN hidden state h.
dtype	Input	Data type of initialized weight and bias. Note: The input data will be converted into the type supported by DynamicRNN.
cell_type	Input	(Optional) Only "LSTM" is supported.
direction	Input	(Optional) Only DYNAMIC_RNN_UNIDIRECTION is supported.
cell_depth	Input	(Optional) Only one layer is supported.
use_peephole	Input	(Optional) Peephole connections are not supported.
keep_prob	Input	(Optional) Dropout is not supported.
cell_clip	Input	(Optional) Cell state clipping is not supported.
num_proj	Input	(Optional) Projection calculation is not supported.
time_major	Input	(Optional) Only input x in [num_step, batch_size, embedding] is supported.
activation	Input	(Optional) Only "tanh" is supported.
forget_bias	Input	(Optional) Defaults to 0.0.
is_training	Input	(Optional) By default, the training mode is used.

Returns

output_y: RNN output tensor, with the shape of [num_step, batch_size, hidden_size].
output_h: RNN output tensor, with the shape of [num_step, batch_size, hidden_size].
output_c: RNN output tensor, with the shape of [num_step, batch_size, hidden_size].
i: RNN intermediate result, which is used for backward propagation.
j: RNN intermediate result, which is used for backward propagation.
f: RNN intermediate result, which is used for backward propagation.
o: RNN intermediate result, which is used for backward propagation.
tanhc: RNN intermediate result, which is used for backward propagation.

Restrictions

This API is the high-performance equivalent of tf.nn.dynamic_rnn. You are advised to use this API instead. Note the following:

The native TensorFlow API tf.nn.dynamic_rnn can add connections to cells (such as connecting the output layer to the input layer). DynamicRNN has no such native support and connection addition needs to be implemented separately.
Currently, there are restrictions on the API. See the following table for details.

Example

import tensorflow as tf
from npu_bridge.estimator.npu.npu_dynamic_rnn import DynamicRNN

inputs = tf.random.normal(shape=(25, 64, 128))
lstm = DynamicRNN(hidden_size=16, dtype=tf.float16, is_training=False)
y, output_h, output_c, i, j, f, o, tanhc = lstm(inputs, seq_length=None, init_h=None, init_c=None)

Parent topic: npu_bridge.estimator.npu.npu_dynamic_rnn