zhangboshen / Alibaba-Cloud-German-AI-Challenge-2018-Rank-17th-Solution

Alibaba Cloud German AI Challenge 2018, 17th place solution. https://tianchi.aliyun.com/competition/entrance/231683/introduction
10 stars 2 forks source link
cloud-german-challenge tianchi

Alibaba-Cloud-German-AI-Challenge-2018-Rank-17th-Solution

Alibaba Cloud German AI Challenge 2018, 17th place solution. https://tianchi.aliyun.com/competition/entrance/231683/introduction


Development tools:

GPU: NVIDIA GTX-1080
Code: Python3.6
Framework: Pytorch1.0
Optimizer: Adam
Loss: CrossEntropyLoss and sphereFaceAngleLoss
Schedule: StepLR with step_size=8, gamma=0.2


Solution:

Usage:

train.py is training code, we use both training set and validation set for training, four .npy are mean and std computed through the whole training set.

model.py is the models that we used, resnet.py is a modified version of original resnet.py (torchvision), it consists of only 4 times downsampling instead of 32. sphere_model.py is used when we train our model with sphereFaceAngleLoss, netvlad.py is used in one of our models -- ResNet18 with VLAD.

test.py is testing code, it loads trained models and get the final prediction and gives us a submission.