How to convert between NHWC and NCHW in TensorFlow

In TensorFlow, NHWC and NCHW are two commonly used data formats representing different dimension orders: N denotes the batch size, H represents the image height, W represents the image width, and C represents the number of channels (e.g., RGB).

• NHWC: The data order is [batch, height, width, channels].
• NCHW: The data order is [batch, channels, height, width].

Conversion Methods

In TensorFlow, you can use the tf.transpose function to change the tensor's dimension order, enabling conversion between NHWC and NCHW formats.

1. From NHWC to NCHW

Assume you have a tensor input_tensor in NHWC format. To convert it to NCHW, use the following code:

python
nchw_tensor = tf.transpose(input_tensor, [0, 3, 1, 2])

Here, [0, 3, 1, 2] specifies the new dimension order, where 0 indicates the batch size remains unchanged, 3 moves the original channels dimension to the second position, and 1 and 2 correspond to the original height and width dimensions, respectively.

2. From NCHW to NHWC

Similarly, to convert from NCHW back to NHWC format, use:

python
nhwc_tensor = tf.transpose(input_tensor, [0, 2, 3, 1])

Here, [0, 2, 3, 1] defines the new dimension order, with 0 indicating the batch size remains unchanged, 2 and 3 corresponding to the original height and width dimensions, and 1 moving the original channels dimension to the last position.

Use Cases

Different hardware platforms may support these formats with varying efficiency. For instance, NVIDIA's CUDA often provides better performance with NCHW format due to specific optimizations for storage and computation. Therefore, when using GPUs, it is advisable to use NCHW format for optimal performance. Conversely, some CPUs or specific libraries may have better support for NHWC format.

Practical Example

Suppose you are working on an image classification task where the input data is a batch of images in NHWC format. To train on a CUDA-accelerated GPU, convert it to NCHW format:

python
import tensorflow as tf

# Assume input_images is a tensor of shape [batch_size, height, width, channels]
input_images = tf.random.normal([32, 224, 224, 3])

# Convert to NCHW format
images_nchw = tf.transpose(input_images, [0, 3, 1, 2])

# Then input images_nchw to the model for training

This conversion operation is common during data preprocessing, especially in deep learning training. By converting, you ensure compatibility with the hardware platform for optimal computational efficiency.