CODEWITHSUNDEEP
tensorflowtensorflow2.0
Benny K
Benny K

Reputation: 2087

Add None dimension in tensorflow 2.0

I have a tensor xx with shape:

>>> xx.shape
TensorShape([32, 32, 256])

How can I add a leading None dimension to get:

>>> xx.shape
TensorShape([None, 32, 32, 256])

I have seen many answers here but all are related to TF 1.x

What is the straight forward way for TF 2.0?

Upvotes: 10

Views: 11349

Answers (3)

Nathan Van Woudenberg
Nathan Van Woudenberg

Reputation: 11

I do not think it is possible to simply add a "None" dimension.

However, assuming you are trying to prepend a variable-size batch dimension to your tensor, you can just tf.repeat your tensor using the tf.shape() of another tensor.

y = tf.keras.layers.Input(shape=(32, 32, 3))   # Shape: [None, 32, 32, 3]
...
batch_size = tf.shape(y)[0]                    # Will be None initially, but is mutable
xx = tf.ones(shape=(32, 32, 356))              # Shape: [32, 32, 356]
xx = tf.expand_dims(xx, 0)                     # Shape: [1, 32, 32, 356]
xx = tf.repeat(xx, repeats=batch_size, axis=0) # Shape: [None, 32, 32, 356]

This will likely be more useful than just hard-coding the first dimension to None since what you probably actually want to be doing is copying it along that first dimension based on batch size.

Upvotes: 1

Gerard Carbó
Gerard Carbó

Reputation: 1905

In TF2 you can use tf.expand_dims:

xx = tf.expand_dims(xx, 0)
xx.shape
> TensorShape([1, 32, 32, 256])

Upvotes: 2

user8879803
user8879803

Reputation:

You can either use "None" or numpy's "newaxis" to create the new dimension.

General Tip: You can also use None in place of np.newaxis; These are in fact the same objects.

Below is the code that explains both the options.

try:
  %tensorflow_version 2.x
except Exception:
  pass
import tensorflow as tf

print(tf.__version__)

# TensorFlow and tf.keras
from tensorflow import keras

# Helper libraries
import numpy as np

#### Import the Fashion MNIST dataset
fashion_mnist = keras.datasets.fashion_mnist
(train_images, train_labels), (test_images, test_labels) = fashion_mnist.load_data()

#Original Dimension
print(train_images.shape)

train_images1 = train_images[None,:,:,:]
#Add Dimension using None
print(train_images1.shape)

train_images2 = train_images[np.newaxis is None,:,:,:]
#Add dimension using np.newaxis
print(train_images2.shape)

#np.newaxis and none are same
np.newaxis is None

The Output of the above code is

2.1.0
(60000, 28, 28)
(1, 60000, 28, 28)
(1, 60000, 28, 28)
True

Upvotes: 2

Related Questions