MXNet.jl

MXNet.jl copied to clipboard

Custom output such as exponential function?

11

Hi there, I am trying to define the activation of the last layer as the exponential function. If x is the input vector to a node in the last layer,...

JockLawrie

Any plans to add custom loss functions?

3

The python version of MXNet has the capability to define custom loss functions, but this is missing from the Julia version. Are there any plans to add it?

Simplex-AP

There are a few methods missing in the KVStore API to setup a distributed setup. Ideally this should work with the distributed programming model in Julia.

vchuravy

Port Executor Manager to Julia

1

For full `Module` support we will need to port the Python Executor Managers to Julia. - https://github.com/dmlc/mxnet/blob/master/python/mxnet/executor_manager.py

vchuravy

Distributed MXNet with ClusterManager

3

I was thinking that since we now have the ability to run MXNet distributed it would be nice to integrate `ClusterManager.jl` and simplify the process for the Julia side.

vchuravy

[TODO] More optimizers

5

Currently we have SGD and Adam. Should be relatively simple to add more, like RMSProp, etc. Some references: - http://cs231n.github.io/neural-networks-3/#update - https://github.com/Lasagne/Lasagne/blob/master/lasagne/updates.py - https://github.com/fchollet/keras/blob/master/keras/optimizers.py

pluskid

Representing the network as a graph.

1

Is there any way besides parsing the json to work with the network as a graph? ``` julia using MXNet using LightGraphs using JSON net = @mx.chain mx.Variable(:data) => mx.Convolution(num_filter...

vchuravy

Julia garbage collector does not reclaim GPU memory

9

I'm trying to write Neural Style in MXNet.jl, and I keep running out of memory when I try to make new executors (and delete the old ones). My basic strategy...

bonsairobo

Reuse executor for repeated prediction with different checkpoints

5

Testing the different checkpoints of a training run requires loading checkpoints and predictions runs in a tight loop. ``` julia net = mx.load(archfile, mx.SymbolicNode) arch = mx.FeedForward(net, context=mx.gpu()) data =...

vchuravy

Automated adaption of batch_size to available GPU memory

1

I am running a different set of models upon different input sizes and so it is not very straightforward to choose an optimal batch-size for each model/input combination. I came...

vchuravy