Custom cost function is not being used for training

[xafilox]

Hi.

I have implemented my own cost function but I have realized it has only being used for printing, not for calculating the error for the gradient descend, and in fact the metric that is being used during the training is the default one (the accuracy).

The reason I am saying this is because I have tried returning always value 1 (return [(:EuclideanDist, 1)])in the mx.get method and I still get exactly the same results (when in this case it should not be able to even learn anything).

Thanks for your help.

using MXNet
using Distances

import MXNet.mx: get, reset!, update!

redirect_stderr(STDOUT)

srand(1234)

type EuclideanDist <: mx.AbstractEvalMetric
    loss_sum  :: Float64
    n_sample :: Int

    EuclideanDist() = new(0.0, 0)
end

function mx.update!(metric :: EuclideanDist, labels :: Vector{mx.NDArray}, preds :: Vector{mx.NDArray})
    preds  = copy(preds)
    labels  = copy(labels)
    
    @assert length(labels) == length(preds)
    
    loss = 0.0
    for (label, pred) in zip(labels, preds)
        @mx.nd_as_jl ro=(label, pred) begin
            for elem in 1:size(label)[2]
                _label = label[:, elem]
                _pred = pred[:, elem]
                _euc = euclidean([_label[1]/10000 * training_deg_to_m_lat, _label[2]/10000 * training_deg_to_m_long], [_pred[1]/10000 * training_deg_to_m_lat, _pred[2]/10000 * training_deg_to_m_long])
                loss += _euc
            end
        end
    end

    metric.loss_sum += loss
    metric.n_sample += size(labels[1])[2]
end

function mx.get(metric :: EuclideanDist)
    distance  = metric.loss_sum / metric.n_sample
    return [(:EuclideanDist, distance)]
end

function mx.reset!(metric :: EuclideanDist)
    metric.loss_sum  = 0.0
    metric.n_sample = 0
end

data = mx.Variable(:data)  # Do not change the name
lbl  = mx.Variable(:softmax_label) # Do not change the name
fc1  = mx.FullyConnected(data, name=:fc1, num_hidden=512)
act1 = mx.Activation(fc1, name=:relu1, act_type=:relu)
fc2  = mx.FullyConnected(act1, name=:fc2, num_hidden=512)
act2 = mx.Activation(fc2, name=:relu2, act_type=:relu)
fc3  = mx.FullyConnected(act2, name=:fc3, num_hidden=128)
act3 = mx.Activation(fc3, name=:relu3, act_type=:relu)
fc4  = mx.FullyConnected(act3, name=:fc4, num_hidden=32)
act4 = mx.Activation(fc4, name=:relu4, act_type=:relu)
fc5  = mx.FullyConnected(act4, name=:fc5, num_hidden=2)
mlp  = mx.LinearRegressionOutput(fc5, lbl, name=:linear)

# data provider
train_provider = mx.ArrayDataProvider(Array(training_data)', Array(training_labels)', batch_size = 100, shuffle = true)
eval_provider = mx.ArrayDataProvider(Array(validation_data)', Array(validation_labels)', batch_size = 100, shuffle = true)

# setup model
model = mx.FeedForward(mlp, context=mx.gpu(1))

# optimizer
optimizer = mx.ADAM()

# Initializer
#initializer = mx.XavierInitializer(distribution = mx.xv_uniform, regularization = mx.xv_avg, magnitude = 3)
initializer = mx.UniformInitializer(0.01)

# fit parameters
a = mx.fit(model, optimizer, train_provider, eval_data=eval_provider, initializer=initializer, n_epoch=200, eval_metric=EuclideanDist())

[iblislin]

I guess you need this: http://mxnet.io/how_to/new_op.html

[iblislin]

Note that mx.LinearRegressionOutput is used as a loss function. It will trigger the BP and update weigth, IIRC, so you need to write your own one and replace it.

[iblislin]

Just discovered that there is a MakeLoss helper for creating custom loss function, but seems it is still buggy now.
So... seems writing your own layer in Cpp is the only way to make it currently.

[xafilox]

Thanks a lot! Is it possible to create the new operators with Julia? I have tried mx.operator.CustomOp and mx.CustomOp and none of them seem to exist.

[Petterhg]

I have the same issue! Is there really no simpler way?
I have a multi-label regression issue where I need to specify my own loss function. Where can you find the implemented loss function for all the available output functions?

[iblislin]

Seems no wrapper for creating CustomOp .... 😕

[xafilox]

Ohh, what a pity :(
Thanks for your help @iblis17

[vchuravy]

CustomOp support is a longer project see #166 .
If anybody is interested in having this helping with #173 is a good place to start.