Key error: Prob pycaffe

[connor amorin]

Hello,

I am trying to use pycaffe on my test set. I am trying to roughly follow the classification notebook, but expanding it to a set rather than just the one image.

Enter cimport sys
import numpy as np
import os.path as osp
import caffe



CLASSES = 11
num2cls = ['AMP_Scratch', 'Coat_Defect','Coat_Flake', 'Etch_Pit', 'Mussel', 'Other', 'Pansy', 'Particle', 'Scratch', 'Splatter', 'TBD']

caffe.set_device(0)
caffe.set_mode_gpu()
net_def = '/path/to/prototext'
weights = '/path/to/weights'

net = caffe.Net(net_def, weights, caffe.TEST)

num_batches = 937 #number of files to be tested
output_file = '/g/g14/amorin1/DamageNet/Models/AlexNet/7.6.2017/pytest.out'

for t in range(num_batches):
     # next batch
    output = net.forward()
    label = net.blobs['label'].data
    output_prob = output['prob'][0]

    print 'label'
    print 'predicted class is:', output_prob.argmax()
    print 'output label:', num2cls[output_prob.argmax()]


ode here...

I get this error:

  output_prob = output['prob'].
KeyError: 'prob'

I also get this error before the net is loaded, but the test continues:

W0713 15:40:53.573498 125192 _caffe.cpp:135] DEPRECATION WARNING - deprecated use of Python interface
W0713 15:40:53.573940 125192 _caffe.cpp:136] Use this instead (with the named "weights" parameter):

Not sure if these are related?

Thanks!

[shruti sneha]

Hi,
after using net.forward( )
try to use net.blobs['prob'].data[0] for getting the probability value.

Plz let me knw if it helps and correct me if I am wrong :)

Best Regards,
Shruti

[C.amo]

So it turns out I was not understanding exactly what net.forward() does.  Output['prob'] does not exist because I  did not have a layer named 'prob'. I had assumed 'prob' was general caffe syntax, but it turns out the format is output['layer name']. It wasn't obvious from the pycaffe examples.

[shruti sneha]

So, does this mean, you resolved your problem? Great ..!!

[Przemek D]

It is actually output['blob name'] - you should not assume these two will always be equivalent.