Home

Awesome

Implement of MAS on MXNet

This is an implement of MAS on MXNet.

Origin MAS on pytorch

what does this project finish

environment

how to use

  1. clone the project
$ git clone https://github.com/mingzhang96/MAS-mxnet.git
$ cd MAS-mxnet
$ mkdir ckpt && mkdir data && mkdir reg_params
  1. We assume that you are in the $MAS-mxnet directory, and in $MAS-mxnet/data the mnist (.gz) data stays there.
python train_mnist.py

result

we use mlp instead of AlexNet as our base model.

notice: we use model trained on last task to test other tasks.

100 epoch, update_lr = 0.05, train_lr = 0.05

taskaccuracy
010.6274231678486998
230.9417238001958864
450.9797225186766275
670.972306143001007
890.9389813414019162

200 epoch, update_lr = 0.0001, train_lr = 0.0008

taskaccuracy
010.9952718676122931
230.8805093046033301
450.955709711846318
670.9823766364551864
890.9536056480080686

200 epoch, update_lr = 0.0001, train_lr = 0.005, fc2.output = 256

taskaccuracy
010.9933806146572104
230.9299706170421156
450.9802561366061899
670.9914400805639476
890.9646999495713565

tips