The NN give NaN after 1/2 epochs with single precision. This happens when relu and softmax/sigmoid layer because the output goes out of bound. Probably some normalization is required.

See:
https://developer.nvidia.com/blog/mixed-precision-training-deep-neural-networks/

I am not sure if the transpose of the conv2d with stride different from 1 is doing what is suppose to do.

Compare with this: https://datascience.stackexchange.com/questions/6107/what-are-deconvolutional-layers

NilDa::layer* l0 = new NilDa::inputLayer({rI, cI, chI});

NilDa::layer* l1 = new NilDa::conv2DLayer(
nFilters,
{rF, cF},
{rS, cS},
padding,
"relu"
);

NilDa::layer* l2 = new NilDa::conv2DLayer(
nFilters,
{rF, cF},
{rS, cS},
padding,
"relu"
);

NilDa::layer* l3 = new NilDa::denseLayer(3, "softmax");

There is no check about they type of the last layer.

const bool padding = true;

NilDa::layer* l0 = new NilDa::inputLayer({rI, cI, chI});

NilDa::layer* l1 = new NilDa::conv2DLayer(
nFilters,
{rF, cF},
{rS, cS},
padding,
"relu"
);

NilDa::layer* l2 = new NilDa::conv2DLayer(
nFilters,
{rF, cF},
{rS, cS},
padding, // <-------- the problem is here
"relu"
);

NilDa::layer* l3 = new NilDa::denseLayer(3, "softmax");

After loading the model, the accuracy of the original model and the one loaded is different