tf.data.Dataset: The `batch_size` argument must not be specified for the given input type

Question

I'm using Talos and Google colab TPU to run hyperparameter tuning of a Keras model. Note that I'm using Tensorflow 1.15.0 and Keras 2.2.4-tf.

import os
import tensorflow as tf
import talos as ta
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense
from tensorflow.keras.optimizers import Adam
from sklearn.model_selection import train_test_split

def iris_model(x_train, y_train, x_val, y_val, params):

    # Specify a distributed strategy to use TPU
    resolver = tf.contrib.cluster_resolver.TPUClusterResolver(tpu='grpc://' + os.environ['COLAB_TPU_ADDR'])
    tf.contrib.distribute.initialize_tpu_system(resolver)
    strategy = tf.contrib.distribute.TPUStrategy(resolver)

    # Use the strategy to create and compile a Keras model
    with strategy.scope():
      model = Sequential()
      model.add(Dense(32, input_shape=(4,), activation=tf.nn.relu, name="relu"))
      model.add(Dense(3, activation=tf.nn.softmax, name="softmax"))
      model.compile(optimizer=Adam(learning_rate=0.1), loss=params['losses'])

    # Convert data type to use TPU
    x_train = x_train.astype('float32')
    x_val = x_val.astype('float32')

    dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
    dataset = dataset.cache()
    dataset = dataset.shuffle(1000, reshuffle_each_iteration=True).repeat()
    dataset = dataset.batch(params['batch_size'], drop_remainder=True)

    # Fit the Keras model on the dataset
    out = model.fit(dataset, batch_size=params['batch_size'], epochs=params['epochs'], validation_data=[x_val, y_val], verbose=0, steps_per_epoch=2)

    return out, model

# Load dataset
X, y = ta.templates.datasets.iris()

# Train and test set
x_train, x_val, y_train, y_val = train_test_split(X, y, test_size=0.30, shuffle=False)

# Create a hyperparameter distributions 
p = {'losses': ['logcosh'], 'batch_size': [128, 256, 384, 512, 1024], 'epochs': [10, 20]}

# Use Talos to scan the best hyperparameters of the Keras model
scan_object = ta.Scan(x_train, y_train, params=p, model=iris_model, experiment_name='test', x_val=x_val, y_val=y_val, fraction_limit=0.1)

After converting the train set to a Dataset using tf.data.Dataset, I get the following error when fitting the model with out = model.fit:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-3-c812209b95d0> in <module>()
      8 
      9 # Use Talos to scan the best hyperparameters of the Keras model
---> 10 scan_object = ta.Scan(x_train, y_train, params=p, model=iris_model, experiment_name='test', x_val=x_val, y_val=y_val, fraction_limit=0.1)

8 frames
/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/keras/engine/training.py in _validate_or_infer_batch_size(self, batch_size, steps, x)
   1813             'The `batch_size` argument must not be specified for the given '
   1814             'input type. Received input: {}, batch_size: {}'.format(
-> 1815                 x, batch_size))
   1816       return
   1817 

ValueError: The `batch_size` argument must not be specified for the given input type. Received input: <DatasetV1Adapter shapes: ((512, 4), (512, 3)), types: (tf.float32, tf.float32)>, batch_size: 512

Then, if I follow those instructions and don't set the batch-size argument to model.fit. I get another error in:

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-3-c812209b95d0> in <module>()
      8 
      9 # Use Talos to scan the best hyperparameters of the Keras model
---> 10 scan_object = ta.Scan(x_train, y_train, params=p, model=iris_model, experiment_name='test', x_val=x_val, y_val=y_val, fraction_limit=0.1)

8 frames
/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/keras/engine/training.py in _distribution_standardize_user_data(self, x, y, sample_weight, class_weight, batch_size, validation_split, shuffle, epochs, allow_partial_batch)
   2307             strategy) and not drop_remainder:
   2308           dataset_size = first_x_value.shape[0]
-> 2309           if dataset_size % batch_size == 0:
   2310             drop_remainder = True
   2311 

TypeError: unsupported operand type(s) for %: 'int' and 'NoneType'

Answer 1

It looks to me that the problem with your code is that the training and validation data is not in the same format. You are batching the training data but not the validation examples.

You can ensure that they are in the same format by replacing the bottom half of your iris_model function with this:

def fix_data(x, y):
    x = x.astype('float32')
    ds = Dataset.from_tensor_slices((x, y))
    ds = ds.cache()
    ds = ds.shuffle(1000, reshuffle_each_iteration = True)
    ds = ds.repeat()
    ds = ds.batch(params['batch_size'], drop_remainder = True)
    return ds
train = fix_data(x_train, y_train)
val = fix_data(x_val, y_val)

# Fit the Keras model on the dataset
out = model.fit(x = train, epochs = params['epochs'],
                steps_per_epoch = 2,
                validation_data = val,
                validation_steps = 2)

At least this works for me and your code runs without error.

Answer 2

from github code :

ValueError will be raised if x is a generator or Sequence instance and batch_size is specified as we expect users to provide batched datasets.

Try using batch_size = None

Answer 3

Not sure if the following will fit your bill but something to try. All I have done is taken off the repeat() from dataset and batch_size=params['batch_size'] from model.fit

If the above is not what you are ready to sacrifice then please ignore the post.

import os
import tensorflow as tf
import talos as ta
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense

def iris_model(x_train, y_train, x_val, y_val, params):

    # Specify a distributed strategy to use TPU
    resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu='grpc://' + os.environ['COLAB_TPU_ADDR'])
    tf.config.experimental_connect_to_host(resolver.master())
    tf.tpu.experimental.initialize_tpu_system(resolver)
    strategy = tf.distribute.experimental.TPUStrategy(resolver)

    with strategy.scope():
        model = Sequential()
        model.add(Dense(32, input_dim=4, activation=params['activation']))
        model.add(Dense(3, activation='softmax'))
        model.compile(optimizer=params['optimizer'], loss=params['losses'])

    # Convert the train set to a Dataset to use TPU
    dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
    dataset = dataset.cache().shuffle(1000, reshuffle_each_iteration=True).batch(params['batch_size'], drop_remainder=True)

    out = model.fit(dataset, epochs=params['epochs'], validation_data=[x_val, y_val], verbose=0)

    return out, model

x, y = ta.templates.datasets.iris()

p = {'activation': ['relu', 'elu'],
       'optimizer': ['Nadam', 'Adam'],
       'losses': ['logcosh'],
       'batch_size': (20, 50, 5),
       'epochs': [10, 20]}

scan_object = ta.Scan(x, y, model=iris_model, params=p, fraction_limit=0.1, experiment_name='first_test')

Answer 4

That second error you get, in _distribution_standardize_user_data, when you don't pass the batch_size to fit.

The code you're running for that function is here:

https://github.com/tensorflow/tensorflow/blob/r1.15/tensorflow/python/keras/engine/training.py#L2192

You didn't post a trace-back, but I bet it's failing on line 2294, since that's the only place where batch_size is multiplied by something.

if shuffle:
          # We want a buffer size that is larger than the batch size provided by
          # the user and provides sufficient randomness. Note that larger
          # numbers introduce more memory usage based on the size of each
          # sample.
          ds = ds.shuffle(max(1024, batch_size * 8))

It looks like you can shut it off by setting shuffle=False.

fit(ds, shuffle=False,...)

Does that work?

Answer 5

Could you remove, these lines from your code and try:

    dataset = dataset.cache()
    dataset = dataset.shuffle(1000, reshuffle_each_iteration=True).repeat()
    dataset = dataset.batch(params['batch_size'], drop_remainder=True)
WITH THESE:
    dataset = dataset.repeat()
    dataset = dataset.batch(128, drop_remainder=True)
    dataset = dataset.prefetch(1)

Otherwise what you wrote in tf.data.Dataset.from_tensor_slices has something to do with the error.