Converting NHWC to NCHW

[antonio.bejar]

Converting a NHWC Tensorflow 2.8 model to NCHW ONNX model for later generating Tensor RT file

I’ve got a nice model with input [1,H,W,C], which works pretty well. Now, I want to convert it to ONNX for later using trtgen ang getting the final file for RT inferencing. (Yes, I know there’s another way to tranform but that’s the way I leant)

The problem is I never before realized TensorRT excepts channel first for properly working…
Is there any fast solution a can use to adapt my initial model to that format or, shall I train again the model with the desired format?

Windows 10

TensorRT Version: 7.2
GPU Type: RTX 3035 Ti
Nvidia Driver Version: 528.24
CUDA Version: 11.2
CUDNN Version:
Operating System + Version:
Python Version (if applicable):
TensorFlow Version (if applicable): 2.8
PyTorch Version (if applicable):
Baremetal or Container (if container which image + tag):

[antonio.bejar]

I found the answer.
The problem was I read it was necessary using changing the input layer and retraining the model from the begining, but it is not necessary.
When you have a trained model NHCW and want it to be INPUT NCHW, the only thing you have to do is to get the ‘old’ model and just add two layers before:

inputs = tf.keras.layers.Input((3,IMG_HEIGHT, IMG_WIDTH))#((IMG_HEIGHT, IMG_WIDTH, 3))

inputs = tf.keras.layers.Permute((2,3,1))(inputs)
Then, add the initial model, with the the trained weights and, that’s all.

Maybe there’s is some utility somewhere but, honestly, couldn’t find it.