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?
Asked
Active
Viewed 5.5k times
3 Answers
88
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
iwill correspond to the input dimensionperm[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
iwill correspond to the input dimensionperm[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]
Olivier Moindrot
- 27,908
- 11
- 92
- 91
-
For the sake of completeness: an explication why these commands are needed would be supportive – user3085931 Jan 26 '18 at 08:40
-
2@user3085931 : you got it – Olivier Moindrot Jan 26 '18 at 09:08
-
Also, what is `perm` - or how is it defined? – nikk Jun 28 '18 at 22:06
-
`perm` is the permutation of the dimensions of the image to go from `(N, H, W, C)` to `(N, C, H, W)` for instance. – Olivier Moindrot Jun 29 '18 at 09:27
-
2where did you use `x` and `perm`? – Ruthvik Vaila Dec 03 '18 at 18:23
-
@OlivierMoindrot is x == images_nchw ? – Eric K. Jun 28 '21 at 15:02
0
To convert 'NCHW' to 'NHWC'
from keras import backend
backend.set_image_data_format('channels_last') #channels_first for NCHW
Rishi
- 17
- 1
-
4This does not do any conversion, it only changes how Keras interprets data. – Dr. Snoopy Nov 09 '21 at 04:57
-
Idd, deceptive answer of Rishi. This parameter pertains for example to `tensorflow.keras.preprocessing.image.load_img`. When you have a 128x128 RGB image, then `load_img` returns an array of shape (3,128,128) with `'channels_first'`. `load_img` returns an array of shape (128,128,3) with `'channels_last'`. – Bart Aug 29 '22 at 09:33
0
For the latest TF2 models, we have a functionality in tf2onnx package. tf2onnx.convert.from_keras(input_as_nchw = [List]) is the latest function update which can be used while converting the model from .pb format to .onnx also it successfully converts the NHWC to NCHW. https://github.com/onnx/tensorflow-onnx/blob/e896723e410a59a600d1a73657f9965a3cbf2c3b/tf2onnx/convert.py#L408
Sovik Gupta
- 147
- 2
- 13