Search code examples
tensorflow

Convert between NHWC and NCHW in TensorFlow

What is the best way to convert a tensor from NHWC format to NCHW format, and vice versa?

Is there an op specifically that does this, or will I need to use some combination of the split/concat type operations?

Solution

  • All you need to do is a permutation of the dimensions from NHWC to NCHW (or the contrary).

    The meaning of each letter might help understand:

    • N: number of images in the batch
    • H: height of the image
    • W: width of the image
    • C: number of channels of the image (ex: 3 for RGB, 1 for grayscale...)

    From NHWC to NCHW

    The image shape is (N, H, W, C) and we want the output to have shape (N, C, H, W). Therefore we need to apply tf.transpose with a well chosen permutation perm.

    The returned tensor's dimension i will correspond to the input dimension perm[i]

    perm[0] = 0  # output dimension 0 will be 'N', which was dimension 0 in the input
perm[1] = 3  # output dimension 1 will be 'C', which was dimension 3 in the input
perm[2] = 1  # output dimension 2 will be 'H', which was dimension 1 in the input
perm[3] = 2  # output dimension 3 will be 'W', which was dimension 2 in the input

    In practice:

    images_nhwc = tf.placeholder(tf.float32, [None, 200, 300, 3])  # input batch
out = tf.transpose(images_nhwc, [0, 3, 1, 2])
print(out.get_shape())  # the shape of out is [None, 3, 200, 300]

    From NCHW to NHWC

    The image shape is (N, C, H, W) and we want the output to have shape (N, H, W, C). Therefore we need to apply tf.transpose with a well chosen permutation perm.

    The returned tensor's dimension i will correspond to the input dimension perm[i]

    perm[0] = 0  # output dimension 0 will be 'N', which was dimension 0 in the input
perm[1] = 2  # output dimension 1 will be 'H', which was dimension 2 in the input
perm[2] = 3  # output dimension 2 will be 'W', which was dimension 3 in the input
perm[3] = 1  # output dimension 3 will be 'C', which was dimension 1 in the input

    In practice:

    images_nchw = tf.placeholder(tf.float32, [None, 3, 200, 300])  # input batch
out = tf.transpose(images_nchw, [0, 2, 3, 1])
print(out.get_shape())  # the shape of out is [None, 200, 300, 3]