Data blob's shape incompatible with training's data shape during testing.

[omair18]

I trained my model using data shape = 10x1x32x32x32 in a hdf5 file (see 1st screenshot). But when I load my trained model in python using caffe.Net() and a deploy.prototxt file (screenshot 2) mentioning input_dim my data blob's shape returned is 10x1x32x32 which is a 4D vector whereas data loaded at the time of training was a 5D vector.
Why is there such difference? How can I correct this problem?

[dutran]

First question: in training, you have input
[data] = 10x1x32x32x32 I read as mini-batch =10 examples, channel=1, volume size = 32x32x32
[label] = 10x1x32x1x1 so that mean each label has size 1x32x1x1 what is that?

Could you point me to the python script you used? Is it #59, we haven't fully tested that PR yet. So potentially there are some dimension mismatch

[omair18]

There was a slight mistake in hdf5_data_layer.cpp. Lable's blob was mistakenly reshaped to data blobs length size. I have resolved this error and now label shape is 10x1x1x1x1.
I am simply loading Deploy.prototxt and trained model in caffe.Classifier() and trying to input data using net.forward(data= inputdata) function.

[omair18]

@dutran still having the same issue ( regarding data blob size ) even after correcting label's dimensions while training. Any suggestion please?

[dutran]

I think this issue caused by the python wrapper #163. So closing this and follow-up on the other issue.