Good evening.

I have faced the problem of the right network reconstruction. It seems that caffe converter makes mistakes with the input data dimensions.

I have trained the lenet-like network, using caffe and DIGITS interface. I have used my own dataset with the images like

They are 24 pixels wide and 40 pixels height.

However the caffe uses another dimension order(may be not that evident) - not the width x height but the other way around. This is described here:

The conventional blob dimensions for batches of image data are number N x channel K x height H x width W

in the chapter "Blob storage and communication".

The same I see in the deploy.prototxt file of my network:

input: "data"
input_shape {
dim: 1
dim: 1
dim: 40
dim: 24
}

However I see that in the caffe_converter the input picture is first resized (24,40) -> (40,24).
This is done in the:
void preprocess(...) in caffe_converter.cpp
called by void test(...) in caffe_converter.cpp
called by main(...) in caffe_converter.cpp

There I see that the size of the network stores in:
net->in_shape().width_

the "net" was assigned by the
create_net_from_caffe_prototxt(...) in layer_factory.h
that uses create_net_from_caffe_net(...) in in layer_factory.h

There I see:

int depth = static_cast(layer.input_shape(0).dim(1));
int width = static_cast(layer.input_shape(0).dim(2));
int height = static_cast(layer.input_shape(0).dim(3));

Could that be the place that caused the mistake?

I'm sorry if I'm wrong, I am just trying to make the model work.

P.S. it seems that the "power" layer is not supported by the caffe_converter. Is it planned to add it in the future?