Implement of MAS on MXNet

This is an implement of MAS on MXNet.

Origin MAS on pytorch

what does this project finish

• standard setup and training on several task.
• finally calculate accuracy on each task.

environment

• mxnet-cu80 on version 1.1.0.post0
• python 2.7

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

2. 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

task	accuracy
01	0.6274231678486998
23	0.9417238001958864
45	0.9797225186766275
67	0.972306143001007
89	0.9389813414019162

200 epoch, update_lr = 0.0001, train_lr = 0.0008

task	accuracy
01	0.9952718676122931
23	0.8805093046033301
45	0.955709711846318
67	0.9823766364551864
89	0.9536056480080686

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

task	accuracy
01	0.9933806146572104
23	0.9299706170421156
45	0.9802561366061899
67	0.9914400805639476
89	0.9646999495713565

tips

• the more tasks are, the more epoch need to train.
• use small train_lr to finetune.
• the last fc performs well if it has large output.