Convolution kernel size issue.(Turns out to be an issue about different order of dimensions between Python and R)

[firearasi]

The image sizes are c(3,224,224) (with dataIterator created with im2rec)
And the current net structure is (I know the convolution layers are too few,but this is my debugging step)

net= 'data' %>% mx.symbol.Variable() %>%
  #224

  mx.symbol.Convolution(kernel=c(4,4),num.filter=40,stride=c(4,4)) %>%
  #56 
  mx.symbol.Activation(act.type='tanh') %>%

  mx.symbol.Flatten() %>%

  mx.symbol.FullyConnected(num.hidden=300) %>%
  mx.symbol.Activation(act.type='relu') %>%

  mx.symbol.FullyConnected(num.hidden=100) %>%
  mx.symbol.Activation(act.type='tanh') %>%

  mx.symbol.FullyConnected(num.hidden=2) %>%
  mx.symbol.SoftmaxOutput()

And when I trained the model, I immediately got the error message:
_" D:\chhong\mxnet\dmlc-core\include\dmlc/logging.h:241: [16:33:03] d:\chhong\mxnet\src\operator./convolution-inl.h:337: Check failed: ksize_x <= dshape[3] && ksize_y <= dshape[2] kernel size exceed input
Error: InferShape Error in convolution21: [16:33:03] d:\chhong\mxnet\src\operator./convolution-inl.h:337: Check failed: ksize_x <= dshape[3] && ksize_y <= dshape[2] kernel size exceed input "
_
So what part am I doing wrong?

[firearasi]

By trial and error, it seems that in R the data shape should be c(224,224,3),
need confirmation.

[thirdwing]

Yes, you are right.

Another example is here: https://github.com/dmlc/mxnet/blob/master/example/kaggle-ndsb2/Train.R#L96

https://github.com/dmlc/mxnet/blob/master/example/kaggle-ndsb2/Train.py#L114