Phosph0phyllite / tinymind_competition

tinymind_competition

My_Mxnet_toolkit

Use Mxnet do Classification

this code is for https://www.tinymind.cn/competitions/41

env

ubuntu16.04 mxnet 1.2.0, cuda9, cudnn7, python3

How to use

#1.release the data to the current directory

#2.python get_list.py --ratio 0.9     //produce train.lst and val.lst

#3.bash ./downmodel.sh //download the pretrained model， (mxnet model zoo)

#4.bash ./train.sh //start to train

when it converged, chose a good one with high top-5 acc

#python predict.py --epoch 10

then,get a result with 0.99+.

代码比较糙， :）

Update

Some improvements were made. Now, 0.991+ for single model,

Do as below:

python get_list.py --ratio 0.9

bash ./downmodel.sh

bash ./train.sh

......chose the best model with good validation top-5 acc, epoch 6 for example.

python predict.py --epoch 6

Update

now,it reached at least 0.991

show

run: python show.py

then, you can visualize the data after augmentation. And the data is random choose from test set.

Now, just play with it ：)

add view lrpolicy

it helps with visualize the lrscheduler, choose a ideal one

for example, in train.py #98 lr_scheduler = mx.lr_scheduler.PolyScheduler(8000,0.01, 3) run: python view_learnpolicy.py --base_lr 0.01 --max_update 8000 --power 3

if there is something wrong, contact me with e-mail: [e-mail]