Hello, I have recently started to test this package out. When browsing through the examples this struck me:
It would be convenient to have methods of the following type
tanh(x::mx.SymbolicNode) = mx.Activation(x, act_type=:tanh) relu(x::mx.SymbolicNode) = mx.Activation(x, act_type=:relu)
It would improve readability and make the code faster to write. I also believe it would make it easier for beginners
Groupfunction, but I have had no luck there. I also had the idea of creating a new network that ends at the hidden nodes that are interesting to me, and initialize this network with the weights of my bigger trained network, no luck here either though :/
I am following the R instructions for using the pre-trained Inception model. (http://mxnet.io/tutorials/r/classifyRealImageWithPretrainedModel.html)
However, whatever picture I send in, it seems to always send back "torch" or "school bus". Also, most scores are exactly zero. Anyone seen this before?
Convolution, there is no explanation of what
pad(although I found that term in Lasagne docs). Any suggestions on what to read? It seems there are many things that get added from separate papers and some list sources (like ADAM) and others don't.
My question has no relation with this topic. I want to know the reason why does people use ::testing::initGoogleTest instead of testing::initGoogleTest when people use gtest to do unit test. => from this, I am confused by the difference of ::testing::initGoogleTest and testing::initGoogleTest? Sorry to take up u guys time here in such stupid question.
Anyone can help me for this confusion?Thanks.
Traceback (most recent call last): File "neuralnet.py", line 1, in <module> import mxnet as mx ImportError: No module named mxnet