Skip to content

mst6750623/weather-rectify

Repository files navigation

weather-rectify

Code for weather rectify competition

TODO:

  • 训练三个网络
  • ordinal网络trainer构建
  • validate模块构建
  • 跑测试集!
  • 重新构建一个最简单的UNET
  • 到底什么是正确的提交格式啊啊啊啊啊啊啊啊啊!
  • 加分支分类器预测时间(已做好dataloader返回)
  • 搞清楚那个把序回归转为具体降水值的公式!
  • validate.py里有几个TODO是我想不明白的问题,有空看看TAT

学长的建议

  • 先跑出一个简单baseline
  • OD加上降水概率
  • 加一个分支直接同时预测降水和气温
  • 加入时间监督信息,即加一个额外分类器层预测是几点的降水/气温!!及其重要!
  • 一个人调loss一个人调模型
  • 试试图网络(不用试了)
  • (试过图网络后)有余力的话可以试试可形变卷积

BEST记录

confidence lr=1e-4 100k降 encoder lr=1e-2 500k降 ODR 无预训练 lr=1e-3 50k降

气温:0.59, /mnt/pami23/zhengxin/projects/weather/unet_temperature/output/Pred_temperature_0_59 (ckpt: /mnt/pami23/zhengxin/projects/weather/unet_temperature/checkpoint/unet_lr04_600) 0.5966, /mnt/pami23/zhengxin/projects/temp/output/Pred_temperature_05966 (ckpt: /mnt/pami23/zhengxin/projects/temp/checkpoint/unet_best_05966.pth) 0.58, /mnt/pami23/zhengxin/projects/temp/output/Pred_temperature_058 (ckpt: '/mnt/pami23/zhengxin/projects/temp/checkpoint/unet_lr0405_05896.pth')

验证记录

  • ODA: confidence2.pth(没用), encoderwithodr2.pth, decoder.pth, odr2.pth | ts:0.07359,0.01593,0.00697,0 confidence2.pth(没用), encoderwithodr3.pth, decoder.pth, odr3.pth | ts:0.16698,0.08479,0.00267,0
  • unet: unet-5200.pth | ts: 0.52288, 0.30893, 0.13817, 0.04452 (1.19早第26名) unetwithtime200.pth | ts:0.55325 0.35189 0.15495 0.06053 unetwithtime700.pth | ts:0.55613 0.37411 0.17546 0.07844(1.20中午27名) unetwithtimeinit200.pth(随机初始化) | ts:0.54218 0.34381 0.14669 0.06306 unetwithtimeinit700.pth(随机初始化) | ts:0.55907 0.37914 0.18206 0.07829(1.20晚28名) unetwithtimeinitresample900.pth | ts:0.53944 0.37850 0.21543 0.11423 unetwithtimeinitresample900.pth(threshold=0.1) | ts:0.59304 0.41909 0.25539 0.16229

The correlation matrices between the features are as follows:

About

Code for weather rectify competition

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages