Tensorflow：如何将可变时间步长输入提供给 RNN

Question

. Answers to this question are eligible for a +50 reputation bounty. Fab wants to to this question.

I have a simple X_train and Y_train data:

x_train = [
  array([ 6,  1,  9, 10,  7,  7,  1,  9, 10,  3, 10,  1,  4]), 
  array([ 2,  8,  8,  1,  1,  4,  2,  5,  1,  2,  7,  2,  1,  1, 4,  5, 10, 4])
]
y_train = [23, 17]

Arrays are numpy arrays. I am now trying to use the tf.data.Dataset class to load these as tensors. Before I have done a similar thing successfully using the following code:

    dataset = data.Dataset.from_tensor_slices((x_train, y_train))

As this input is fed into a RNN, I have used the expand_dims method in the first RNN layer (the expand_dimension is passed as a function to overcome an apparent bug in tensorflow: see https://github.com/keras-team/keras/issues/5298#issuecomment-281914537 ):

def expand_dimension(x):
    from tensorflow import expand_dims
    return expand_dims(x, axis=-1)

model = models.Sequential(
    [
        layers.Lambda(expand_dimension,
                      input_shape=[None]),
        layers.LSTM(units=64, activation='tanh'),
        layers.Dense(units=1)
    ]
)

This worked although because I had arrays of equal length. In the example I posted instead the 1st array has 13 numbers and the 2nd one 18. In this case the method above doesn't work, and the recommended method seems to be using tf.data.Dataset.from_generator . Reading this How to use the Tensorflow Dataset Pipeline for Variable Length Inputs? , the accepted solution shows something like the following would work (where I am not caring here about y_train for simplicity):

dataset = tf.data.Dataset.from_generator(lambda: x_train, 
                                         tf.as_dtype(x_train[0].dtype),
                                         tf.TensorShape([None, ]))

However, the syntax in tensorflow has changed since this answer, and now it requires to use the output_signature argument (see https://www.tensorflow.org/api_docs/python/tf/data/Dataset#from_generator ).

I've tried different ways but I'm finding hard to understand from tensorflow documentation what the output_signature should exactly be in my case. Any help would be much appreciated.

Answer 1

Short answer is, you can define output_signature as follows.

import tensorflow as tf
import numpy as np
x_train = [
  np.array([ 6,  1,  9, 10,  7,  7,  1,  9, 10,  3, 10,  1,  4]), 
  np.array([ 2,  8,  8,  1,  1,  4,  2,  5,  1,  2,  7,  2,  1,  1, 4,  5, 10, 4])
]
y_train = [23, 17]

dataset = tf.data.Dataset.from_generator(
    lambda: x_train, 
    output_signature=tf.TensorSpec(
        [None, ], 
        dtype=tf.as_dtype(x_train[0].dtype)
    )
)

I'll also expand and improve on some things you're doing here to improve your pipeline.

Using both inputs and labels

dataset = tf.data.Dataset.from_generator(
    lambda: zip(x_train, y_train), 
    output_signature=(
        tf.TensorSpec([None, ], dtype=tf.as_dtype(x_train[0].dtype)),
        tf.TensorSpec([], dtype=tf.as_dtype(y_train.dtype))
    )
)

for x in dataset:
  print(x)

Which would output,

(<tf.Tensor: shape=(13,), dtype=int64, numpy=array([ 6,  1,  9, 10,  7,  7,  1,  9, 10,  3, 10,  1,  4])>, <tf.Tensor: shape=(), dtype=int64, numpy=23>)
(<tf.Tensor: shape=(18,), dtype=int64, numpy=
array([ 2,  8,  8,  1,  1,  4,  2,  5,  1,  2,  7,  2,  1,  1,  4,  5, 10,
        4])>, <tf.Tensor: shape=(), dtype=int64, numpy=17>)

Caveat : This can get slightly more complicated if you try to tf.data.Dataset.batch() items. Then you need to use RaggedTensorSpec instead of TensorSpec . Also, I haven't experimented too much with feeding in ragged tensors into a RNN. But I think those are out of scope for the question you've asked.