inconsistent output from caffe and tf

[suhangpro]

caffe session
prob shape (1000,)
Top1:  n02123045 tabby, tabby cat
Top5:  ['n02123045 tabby, tabby cat', 'n02124075 Egyptian cat', 'n02123159 tiger cat', 'n02119789 kit fox, Vulpes macrotis', 'n02119022 red fox, Vulpes vulpes']

tensorflow session
Top1:  n02123159 tiger cat
Top5:  ['n02123159 tiger cat', 'n02124075 Egyptian cat', 'n02123045 tabby, tabby cat', 'n02113023 Pembroke, Pembroke Welsh corgi', 'n02094258 Norwich terrier']

Any idea why this happens? Awesome work by the way!

[ry]

Either numerical instability in the softmax layer because np.linalg.norm(caffe2tf_1d_blob('prob') - tf_activations['prob'] is a bit larger than the distance of other layers. And/or from the randomness in the dropout layers

[ry]

fixed in 97a9e1ff663ac7238f986d8b8247f4ab81f83432