CIFAR10_mxnet
abstract
kaggle CIFAR10 compitition code, implement by mxnet gluon.
we got 0.9688 by merge some ideas from https://discuss.gluon.ai/t/topic/1545/423
directroy and file descriptor
file | descriptor --- | --- log | some train log file models | some trianed model params(weight) result | the forward result file on kaggle test set submission | the finnal kaggle submission result CIFAR10_train | main train and exp code plot | the visulization of train acc and valid acc and loss with epoch utils/netlib.py | ResNet18, ResNet164_v2, densenet, Focal Loss implement code by gluon, invoke by CIFAR10_train utils/utils.py | some tool function
models, reuslt, log can get from link: https://pan.baidu.com/s/1pLjzQWj key: f6p3
method description
the main idea is from mxnet topic,we merge the most ideas.
first, we train ResNet164_v2 in diffrent data argument policy sencondly, we use 'focal loss' replace 'softmax cross entropy loss' thirdly, we use densenet replace ResNet164_v2 lastly, we ensemble some net to get higher acc, we found this five models get best score:
policy | kaggle score --- | --- res164_v2 + DA1| 0.9529 res164_v2 + DA2| 0.9527 res164_v2 + focal loss + DA3| 0.9540 res164_v2 + focal loss + DA3(only use 90% train_data) | 0.9506 sherlock_densenet| 0.9539
DA1~DA3 is means diffrent data argument
DA | policy --- | --- DA1 | padding image to 40, and then random crop (32, 32)same as code in sherlock DA2 | resize image to a bigger sizeand then crop to (32, 32)and set data argument parma of HSI to 0.3,PCA noise to 0.01. DA3 | after DA2, clip the color of image to (0,1)make the generate image more friendly to human
we ensemble the five net, and got 0.9688 score.
| --- | --- log | lossacc models | () result | forward10output submission | CIFAR10_train | CIFAR10 plot | loss utils/netlib.py | ResNet18, ResNet164_v2, densenet, Focal Loss gluon, utils/utils.py |
modelsresultlog : https://pan.baidu.com/s/1pLjzQWj : f6p3
: (models)ensemble5
5
policy | kaggle --- | --- res164_v2 + DA1| 0.9529 res164_v2 + DA2| 0.9527 res164_v2 + focal loss + DA3| 0.9540 res164_v2 + focal loss + DA3 | train_data: 0.9506 sherlock_densenet| 0.9539
DA3:
DA | policy --- | --- DA1 | padding40,cropsherlock DA2 | resizecropHSI0.3,PCA0.01 DA3 | DA2clip0,1
0.9688