Converting a list of unequally shaped arrays to Tensorflow 2 Dataset: ValueError: Can't convert non-rectangular Python sequence to Tensor

Question

I have tokenized data in the form of a list of unequally shaped arrays:

array([array([1179,    6,  208,    2, 1625,   92,    9, 3870,    3, 2136,  435,
          5, 2453, 2180,   44,    1,  226,  166,    3, 4409,   49, 6728,
         ...
         10,   17, 1396,  106, 8002, 7968,  111,   33, 1130,   60,  181,
       7988, 7974, 7970])], dtype=object)

With their respective targets:

Out[74]: array([0, 0, 0, ..., 0, 0, 1], dtype=object)

I'm trying to transform them into a padded tf.data.Dataset(), but it won't let me convert unequal shapes to a tensor. I will get this error:

ValueError: Can't convert non-rectangular Python sequence to Tensor.

The full code is here. Assume that my starting point is after y = ...:

import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
import tensorflow as tf
import tensorflow_datasets as tfds
import numpy as np

(train_data, test_data) = tfds.load('imdb_reviews/subwords8k',
                                    split=(tfds.Split.TRAIN, tfds.Split.TEST),
                                    as_supervised=True)

x = np.array(list(train_data.as_numpy_iterator()))[:, 0]
y = np.array(list(train_data.as_numpy_iterator()))[:, 1]


train_tensor = tf.data.Dataset.from_tensor_slices((x.tolist(), y))\
    .padded_batch(batch_size=8, padded_shapes=([None], ()))

What are my options to turn this into a padded batch tensor?

Answer 1

If your data is stored in Numpy arrays or Python lists, then you can use tf.data.Dataset.from_generator method to create the dataset and then pad the batches:

train_batches = tf.data.Dataset.from_generator(
    lambda: iter(zip(x, y)), 
    output_types=(tf.int64, tf.int64)
).padded_batch(
    batch_size=32,
    padded_shapes=([None], ())
)

However, if you are using tensorflow_datasets.load function, then there is no need to use as_numpy_iterator to separate the data and the labels, and then put them back together in a dataset! That's redundant and inefficient. The objects returned by tensorflow_datasets.load are already an instance of tf.data.Dataset. So, you just need to use padded_batch on them:

train_batches = train_data.padded_batch(batch_size=32, padded_shapes=([None], []))
test_batches = test_data.padded_batch(batch_size=32, padded_shapes=([None], []))

Note that in TensorFlow 2.2 and above, you no longer need to provide the padded_shapes argument if you just want all the axes to be padded to the longest of the batch (i.e. default behavior).