Testing 3D trained model using Python Script & Deploy file

[omair18]

I was successful in training my 3D dataset by giving 5D tensor ( 10x1x32x32x32) to the network. Now I'm using python script to load the trained model and deploy.prototxt and giving the input test dataset in the form of 5D tensor ( 10x1x32x32x32) but caffe outputs a ValueError saying "could not broadcast input array from shape (10,1,32,32,32) into shape (10,1,32,32).
Isn't it the right way to do? why my model is expecting a 10,1,32,32 shape?

[dutran]

need to modify HDF5data_output as well, please wait for another minor commit.

[dutran]

@omair18 please try again.

[omair18]

@dutran I can't see any change in the output, getting the very same error.
I have pulled the update and did make all, make pycaffe.
Secondly, I'm using caffe.Classifier( deploy, trainedfile) in my python script. Any suggestions?

[dutran]

Try this instead #59.

[omair18]

I'm slightly confused with the usage of above mentioned script for my specific case. Can you please guide me how to test a single 3D object? Object dimensions are 32x32x32 ( length x width x height) and i'm making a 5D tensor with dimensions batchsize x channels x length x width x height.

[dutran]

Is you problem is just simply classification? If so, you don't need to use python wrapper. Just use test_net to test (for final accuracy). If you prefer to use some intermediate output, use extract_image_features to extract them and further process them if needed.

[omair18]

My problem is somewhat related to scene classification. I want to crop a 3D piece from a scene and classify it with the help of my trained network. test_net won't help me in this case i guess.

[omair18]

@dutran I'm trying to look into this python wrapper ( pycaffe.py). If you get any success, do let me know please.

[omair18]

@dutran It seems like, changing Class CaffeBlobWrap in caffe.cpp solved the issue. I increased the dimensions in object_get_data() function from 4 to 5 and the ValueError is gone.
Also there was a check in pycaffe.py _Net_forward() function, I changed it from 4 to 5.

[dutran]

@omair18 can you send in a PR? Thanks

[dutran]

Glad to hear it works for you.

[omair18]

@dutran I'm rather new on github and unable to figure out how to send my modified project in a PR. Can you guide me a bit?
I've forked this project on my profile and committing changes into it.