SoftmaxCrossEntropy: Fix numerical stability.

http://cs231n.github.io/linear-classify/

Change labels to be another input layer.
The user needs to have two input layer to feed in the Graph.

Also, change the feed_in Dict to have objects in the layer.

Dropout ConvLayer Result

Right now, prediction for a model given an input is outputted by the loss function, which is unintuitive since loss function is technically not part of the net. We should consider tensorflow's implementation to output prediction else where.

Need to include test for numerical stability for the layer SoftMax.jl.

The current design of input layer easily leads to user's misuse. Input Layers must not have duplicate tags.

Case 1:
When a user creates two input layer without specifying tags, both layer would have the same tag "default".

It seems like max pooling layer could only accept size that's an integer multiplication of the pool size.

eg. if the pool size is (2,2), then the max polling layer could only accept image with size (2n,2m)

New cifar10 example.
Model using http://pytorch.org/tutorials/beginner/blitz/cifar10_tutorial.html.
Note:
Will first test on 0 padding and 1 stride

• Clean examples/mnist.jl.
• Benchmark: time and memory allocation
• Pytorch: comparison

Create new Julia type to wrap four common layer field: parents, children, has_init, id.
Also, due to change in interface, has_init is unnecessary now. This field should be removed in the future.

Weird Error message in CaffeConvLayer constructor.

ERROR: LoadError: MethodError: no method matching (::##call#11#13)(::Void, ::Int64, ::Int64, ::String, ::Array{Any,1}, ::Type{CaffeConvLayer}, ::InputLayer, ::Int64, ::Tuple{Int64,Int64})
Closest candidates are:
#call#11(!Matched::Dict{String,Any}, ::Any, ::Any, ::Any, ::Any, ::Any, ::Union{Layer, Void}, ::Int64, ::Tuple{Int64,Int64}) at /Users/tianyizhang/Documents/Grendel_NN.jl/NN.jl/test/../src/layers/CaffeConvLayer.jl:58
Stacktrace:
[1] CaffeConvLayer(::InputLayer, ::Int64, ::Tuple{Int64,Int64}) at /Users/tianyizhang/Documents/Grendel_NN.jl/NN.jl/test/../src/layers/CaffeConvLayer.jl:58
[2] include_from_node1(::String) at ./loading.jl:569
[3] include(::String) at ./sysimg.jl:14
[4] process_options(::Base.JLOptions) at ./client.jl:305
[5] _start() at ./client.jl:371
while loading /Users/tianyizhang/Documents/Grendel_NN.jl/NN.jl/test/GraphTest.jl, in expression starting on line 24

Last commit where GraphToposortTest is good is 933f29d.