Using Datasets from large numpy arrays in Tensorflow

Question

I’m trying to load a dataset, stored in two .npy files (for features and ground truth) on my drive, and use it to train a neural network.

print("loading features...")
data = np.load("[...]/features.npy")

print("loading labels...")
labels = np.load("[...]/groundtruth.npy") / 255

dataset = tf.data.Dataset.from_tensor_slices((data, labels))

throws a tensorflow.python.framework.errors_impl.InternalError: Failed copying input tensor from /job:localhost/replica:0/task:0/device:CPU:0 to /job:localhost/replica:0/task:0/device:GPU:0 in order to run _EagerConst: Dst tensor is not initialized. error when calling the from_tensor_slices() method.

The ground truth’s file is larger than 2.44GB and thus I encounter problems when creating a Dataset with it (see warnings here and here).

Possible solutions I found were either for TensorFlow 1.x (here and here, while I am running version 2.6) or to use numpy’s memmap (here), which I unfortunately don’t get to run, plus I wonder whether that slows down the computation?

I’d appreciate your help, thanks!

Asked By: babrs

||

Source

Answer 1

You need some kind of data generator, because your data is way too big to fit directly into tf.data.Dataset.from_tensor_slices. I don’t have your dataset, but here’s an example of how you could get data batches and train your model inside a custom training loop. The data is an NPZ NumPy archive from here:

import numpy as np

def load_data(file='dsprites_ndarray_co1sh3sc6or40x32y32_64x64.npz'):
    dataset_zip = np.load(file, encoding='latin1')

    images = dataset_zip['imgs']
    latents_classes = dataset_zip['latents_classes']

    return images, latents_classes

def get_batch(indices, train_images, train_categories):
    shapes_as_categories = np.array([train_categories[i][1] for i in indices])
    images = np.array([train_images[i] for i in indices])

    return [images.reshape((images.shape[0], 64, 64, 1)).astype('float32'), shapes_as_categories.reshape(
        shapes_as_categories.shape[0], 1).astype('float32')]

# Load your data once
train_images, train_categories = load_data()
indices = list(range(train_images.shape[0]))
random.shuffle(indices)

epochs = 2000
batch_size = 256
total_batch = train_images.shape[0] // batch_size

for epoch in range(epochs):
    for i in range(total_batch):
        batch_indices = indices[batch_size * i: batch_size * (i + 1)]
        batch = get_batch(batch_indices, train_images, train_categories)
        ...
        ...
        # Train your model with this batch.

Answered By: AloneTogether

Answer 2

The accepted answer (https://stackoverflow.com/a/69994287) loads the whole data into memory with the load_data function. That’s why our RAM is completely full.

You can try to make a npz file where each feature is its own npy file, then create a generator that loads this and use this generator like 1 to use it with tf.data.Dataset or build a data generator with keras like 2

or use the mmap method of numpy load while loading to stick to your one npy feature file like 3

Answered By: user20068036

Using Datasets from large numpy arrays in Tensorflow

Question:

Answers: