'tf.data()' throwing 你的输入数据用完了； 中断训练

Question

I'm seeing weird issues when trying to use tf.data() to generate data in batches with keras api. It keeps throwing errors saying it's running out of training_data.

TensorFlow 2.1

import numpy as np
import nibabel
import tensorflow as tf
from tensorflow.keras.layers import Conv3D, MaxPooling3D
from tensorflow.keras.layers import Dense
from tensorflow.keras.layers import Dropout
from tensorflow.keras.layers import Flatten
from tensorflow.keras import Model
import os
import random


"""Configure GPUs to prevent OOM errors"""
gpus = tf.config.experimental.list_physical_devices('GPU')
for gpu in gpus:
    tf.config.experimental.set_memory_growth(gpu, True)

"""Retrieve file names"""
ad_files = os.listdir("/home/asdf/OASIS/3D/ad/")
cn_files = os.listdir("/home/asdf/OASIS/3D/cn/")

sub_id_ad = []
sub_id_cn = []

"""OASIS AD: 178 Subjects, 278 3T MRIs"""
"""OASIS CN: 588 Subjects, 1640 3T MRIs"""
"""Down-sampling CN to 278 MRIs"""
random.Random(129).shuffle(ad_files)
random.Random(129).shuffle(cn_files)

"""Split files for training"""
ad_train = ad_files[0:276]
cn_train = cn_files[0:276]

"""Shuffle Train data and Train labels"""
train = ad_train + cn_train
labels = np.concatenate((np.ones(len(ad_train)), np.zeros(len(cn_train))), axis=None)
random.Random(129).shuffle(train)
random.Random(129).shuffle(labels)
print(len(train))
print(len(labels))

"""Change working directory to OASIS/3D/all/"""
os.chdir("/home/asdf/OASIS/3D/all/")

"""Create tf data pipeline"""


def load_image(file, label):
    nifti = np.asarray(nibabel.load(file.numpy().decode('utf-8')).get_fdata())

    xs, ys, zs = np.where(nifti != 0)
    nifti = nifti[min(xs):max(xs) + 1, min(ys):max(ys) + 1, min(zs):max(zs) + 1]
    nifti = nifti[0:100, 0:100, 0:100]
    nifti = np.reshape(nifti, (100, 100, 100, 1))
    nifti = tf.convert_to_tensor(nifti, np.float64)
    return nifti, label


@tf.autograph.experimental.do_not_convert
def load_image_wrapper(file, labels):
    return tf.py_function(load_image, [file, labels], [tf.float64, tf.float64])


dataset = tf.data.Dataset.from_tensor_slices((train, labels))
dataset = dataset.shuffle(6, 129)
dataset = dataset.repeat(50)
dataset = dataset.map(load_image_wrapper, num_parallel_calls=6)
dataset = dataset.batch(6)
dataset = dataset.prefetch(buffer_size=1)
iterator = iter(dataset)
batch_images, batch_labels = iterator.get_next()

########################################################################################
with tf.device("/cpu:0"):
    with tf.device("/gpu:0"):
        model = tf.keras.Sequential()

        model.add(Conv3D(64,
                         input_shape=(100, 100, 100, 1),
                         data_format='channels_last',
                         kernel_size=(7, 7, 7),
                         strides=(2, 2, 2),
                         padding='valid',
                         activation='relu'))

    with tf.device("/gpu:1"):
        model.add(Conv3D(64,
                         kernel_size=(3, 3, 3),
                         padding='valid',
                         activation='relu'))

    with tf.device("/gpu:2"):
        model.add(Conv3D(128,
                         kernel_size=(3, 3, 3),
                         padding='valid',
                         activation='relu'))

        model.add(MaxPooling3D(pool_size=(2, 2, 2),
                               padding='valid'))

        model.add(Flatten())

        model.add(Dense(256, activation='relu'))
        model.add(Dense(1, activation='sigmoid'))


model.compile(loss=tf.keras.losses.binary_crossentropy,
              optimizer=tf.keras.optimizers.Adagrad(0.01),
              metrics=['accuracy'])


########################################################################################
model.fit(batch_images, batch_labels, steps_per_epoch=92, epochs=50)

After creating the dataset, I'm shuffling and adding the repeat parameter to the num_of_epochs , ie 50 in this case. This works, but it crashes after the 3rd epoch, and I can't seem to figure out what I'm doing wrong in this particular instance. Am I supossed to declare the repeat and shuffle statements at the top of the pipeline?

Here is the error:

Epoch 3/50
92/6 [============================================================================================================================================================================================================================================================================================================================================================================================================================================================================] - 3s 36ms/sample - loss: 0.1902 - accuracy: 0.8043
Epoch 4/50
5/6 [========================>.....] - ETA: 0s - loss: 0.2216 - accuracy: 0.80002020-03-06 15:18:17.804126: W tensorflow/core/common_runtime/base_collective_executor.cc:217] BaseCollectiveExecutor::StartAbort Out of range: End of sequence
         [[{{node IteratorGetNext}}]]
         [[BiasAddGrad_3/_54]]
2020-03-06 15:18:17.804137: W tensorflow/core/common_runtime/base_collective_executor.cc:217] BaseCollectiveExecutor::StartAbort Out of range: End of sequence
         [[{{node IteratorGetNext}}]]
         [[sequential/conv3d_3/Conv3D/ReadVariableOp/_21]]
2020-03-06 15:18:17.804140: W tensorflow/core/common_runtime/base_collective_executor.cc:217] BaseCollectiveExecutor::StartAbort Out of range: End of sequence
         [[{{node IteratorGetNext}}]]
         [[Conv3DBackpropFilterV2_3/_68]]
2020-03-06 15:18:17.804263: W tensorflow/core/common_runtime/base_collective_executor.cc:217] BaseCollectiveExecutor::StartAbort Out of range: End of sequence
         [[{{node IteratorGetNext}}]]
         [[sequential/dense/MatMul/ReadVariableOp/_30]]
2020-03-06 15:18:17.804364: W tensorflow/core/common_runtime/base_collective_executor.cc:217] BaseCollectiveExecutor::StartAbort Out of range: End of sequence
         [[{{node IteratorGetNext}}]]
         [[BiasAddGrad_5/_62]]
2020-03-06 15:18:17.804561: W tensorflow/core/common_runtime/base_collective_executor.cc:217] BaseCollectiveExecutor::StartAbort Out of range: End of sequence
         [[{{node IteratorGetNext}}]]
WARNING:tensorflow:Your input ran out of data; interrupting training. Make sure that your dataset or generator can generate at least `steps_per_epoch * epochs` batches (in this case, 4600 batches). You may need to use the repeat() f24/6 [========================================================================================================================] - 1s 36ms/sample - loss: 0.1673 - accuracy: 0.8750
Traceback (most recent call last):
  File "python_scripts/gpu_farm/tf_data_generator/3D_tf_data_generator.py", line 181, in <module>
    evaluation_ad = model.evaluate(ad_test, ad_test_labels, verbose=0)
  File "/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/keras/engine/training.py", line 930, in evaluate
    use_multiprocessing=use_multiprocessing)
  File "/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/keras/engine/training_v2.py", line 490, in evaluate
    use_multiprocessing=use_multiprocessing, **kwargs)
  File "/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/keras/engine/training_v2.py", line 426, in _model_iteration
    use_multiprocessing=use_multiprocessing)
  File "/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/keras/engine/training_v2.py", line 646, in _process_inputs
    x, y, sample_weight=sample_weights)
  File "/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/keras/engine/training.py", line 2383, in _standardize_user_data
    batch_size=batch_size)
  File "/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/keras/engine/training.py", line 2489, in _standardize_tensors
    y, self._feed_loss_fns, feed_output_shapes)
  File "/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/keras/engine/training_utils.py", line 810, in check_loss_and_target_compatibility
    ' while using as loss `' + loss_name + '`. '
ValueError: A target array with shape (5, 2) was passed for an output of shape (None, 1) while using as loss `binary_crossentropy`. This loss expects targets to have the same shape as the output.

Update: So model.fit() should be supplied with model.fit(x=data, y=labels) , when using tf.data() because of a weird problem. This removes the list out of index error. And now I'm back to my original error. However it looks like this could be a tensorflow problem: https://github.com/tensorflow/tensorflow/issues/32

So when I increase the batch size from 6 to higher numbers, and decrease the steps_per_epoch , it goes through more epochs without throwing the StartAbort: Out of range errors

Update2: As per @jkjung13 suggestion, model.fit() takes one parameter when using a dataset, model.fit(x=batch) . This is the correct implementation.

But, you are supposed to supply the dataset instead of an iterable object if you're only using the x parameter in model.fit() .

So, it should be: model.fit(dataset, epochs=50, steps_per_epoch=46, validation_data=(v, v_labels))

And with that I get a new error: GitHub Issue

Now to overcome this, I'm converting the dataset to a numpy_iterator(): model.fit(dataset.as_numpy_iterator(), epochs=50, steps_per_epoch=46, validation_data=(v, v_labels))

This solves the problem, however, the performance is appaling, similar to the old keras model.fit_generator without multiprocessing. So this defeats the whole purpose of 'tf.data'.

Answer 1

TF 2.1

This is now working with the following parameters:

def load_image(file, label):
    nifti = np.asarray(nibabel.load(file.numpy().decode('utf-8')).get_fdata()).astype(np.float32)

    xs, ys, zs = np.where(nifti != 0)
    nifti = nifti[min(xs):max(xs) + 1, min(ys):max(ys) + 1, min(zs):max(zs) + 1]
    nifti = nifti[0:100, 0:100, 0:100]
    nifti = np.reshape(nifti, (100, 100, 100, 1))
    return nifti, label


@tf.autograph.experimental.do_not_convert
def load_image_wrapper(file, label):
    return tf.py_function(load_image, [file, label], [tf.float64, tf.float64])


dataset = tf.data.Dataset.from_tensor_slices((train, labels))
dataset = dataset.map(load_image_wrapper, num_parallel_calls=32)
dataset = dataset.prefetch(buffer_size=1)
dataset = dataset.apply(tf.data.experimental.prefetch_to_device('/device:GPU:0', 1))

# So, my dataset size is 522, i.e. 522 MRI images.
# I need to load the entire dataset as a batch.
# This should exceed 60GiBs of RAM, but it doesn't go over 12GiB of RAM.
# I'm not sure how tf.data batch() stores the data, maybe a custom file?
# And also add a repeat parameter to iterate with each epoch.
dataset = dataset.batch(522, drop_remainder=True).repeat()

# Now initialise an iterator
iterator = iter(dataset)

# Create two objects, x & y, from batch
batch_image, batch_label = iterator.get_next()

##################################################################################
with tf.device("/cpu:0"):
    with tf.device("/gpu:0"):
        model = tf.keras.Sequential()

        model.add(Conv3D(64,
                         input_shape=(100, 100, 100, 1),
                         data_format='channels_last',
                         kernel_size=(7, 7, 7),
                         strides=(2, 2, 2),
                         padding='valid',
                         activation='relu'))

    with tf.device("/gpu:1"):
        model.add(Conv3D(64,
                         kernel_size=(3, 3, 3),
                         padding='valid',
                         activation='relu'))

    with tf.device("/gpu:2"):
        model.add(Conv3D(128,
                         kernel_size=(3, 3, 3),
                         padding='valid',
                         activation='relu'))

        model.add(MaxPooling3D(pool_size=(2, 2, 2),
                               padding='valid'))

        model.add(Flatten())

        model.add(Dense(256, activation='relu'))
        model.add(Dropout(0.7))
        model.add(Dense(1, activation='sigmoid'))

model.compile(loss=tf.keras.losses.binary_crossentropy,
              optimizer=tf.keras.optimizers.Adagrad(0.01),
              metrics=['accuracy'])
##################################################################################

# Now supply x=batch_image, y= batch_label to Keras' model.fit()
# And finally, supply your batchs_size here!
model.fit(batch_image, batch_label, epochs=100, batch_size=12)

##################################################################################

With this, it takes around 8 Minutes for the training to start. But once training starts, I'm seeing incredible speeds!

Epoch 30/100
522/522 [==============================] - 14s 26ms/sample - loss: 0.3526 - accuracy: 0.8640
Epoch 31/100
522/522 [==============================] - 15s 28ms/sample - loss: 0.3334 - accuracy: 0.8448
Epoch 32/100
522/522 [==============================] - 16s 31ms/sample - loss: 0.3308 - accuracy: 0.8697
Epoch 33/100
522/522 [==============================] - 14s 26ms/sample - loss: 0.2936 - accuracy: 0.8755
Epoch 34/100
522/522 [==============================] - 14s 26ms/sample - loss: 0.2935 - accuracy: 0.8851
Epoch 35/100
522/522 [==============================] - 14s 28ms/sample - loss: 0.3157 - accuracy: 0.8889
Epoch 36/100
522/522 [==============================] - 16s 31ms/sample - loss: 0.2910 - accuracy: 0.8851
Epoch 37/100
522/522 [==============================] - 14s 26ms/sample - loss: 0.2810 - accuracy: 0.8697
Epoch 38/100
522/522 [==============================] - 14s 26ms/sample - loss: 0.2536 - accuracy: 0.8966
Epoch 39/100
522/522 [==============================] - 16s 31ms/sample - loss: 0.2506 - accuracy: 0.9004
Epoch 40/100
522/522 [==============================] - 15s 28ms/sample - loss: 0.2353 - accuracy: 0.8927
Epoch 41/100
522/522 [==============================] - 14s 26ms/sample - loss: 0.2336 - accuracy: 0.9042
Epoch 42/100
522/522 [==============================] - 14s 26ms/sample - loss: 0.2243 - accuracy: 0.9234
Epoch 43/100
522/522 [==============================] - 15s 29ms/sample - loss: 0.2181 - accuracy: 0.9176

15 seconds per epoch compared to the old 12 minutes per epoch!

I will do further testing to see if this is actually working, and what impact it has on my test data. If there are any errors, I will come back and update this post.

Why does this work? I have no idea. I couldn't find anything in the Keras documentation.