lin-honghui / tianchi_countyagriculturalbrain_top1

TIANCHI天池 2019县域农业大脑AI挑战赛 1/1520

tianchi_CountyAgriculturalBrain_top1

天池 2019县域农业大脑AI挑战赛 1/1520 冠军

0. 环境

conda env create -f env.yaml
conda activate pytorch

1. 数据预处理

sh prepare.sh

2. 代码目录

.
├── checkpoints     (保存训练过程权重)
├── configs         (配置文件)
├── exp             (推理结果保存)
├── jupyter
├── log
├── src
│   ├── data
│   ├── engine
│   ├── model
│   ├── solver
│   ├── tools
│   └── utils
├── prepare.sh
├── inference.sh
├── train.sh
└── README.md

3. 算法说明

详细方案请见 zhihu

线上demo 天池7号馆

tianchi_countyagriculturalbrain_top1's Issues

n

麻烦给一个数据集链接，[email protected],谢过了

请问 你训练测试loss miou指标达到了多少?

请问 你训练测试loss miou指标达到了多少?
我按照你的readme train loss 0.2左右，

谢谢

关于config/.py里数据制作制作流程

请问该项目里的数据制作流程是怎样的，比如deeplabv3plus_resnet101_StepLR_Adam.py里 ：
dataset=dict(type="PNG_Dataset",
csv_file=r'/home/LinHonghui/Datasets/tianchi_CountyAgricultural/Crop1024/train.csv',
image_dir=r'/home/LinHonghui/Datasets/tianchi_CountyAgricultural/Crop1024/image/',
mask_dir=r'/home/LinHonghui/Datasets/tianchi_CountyAgricultural/Crop1024/label/'),
上面对应的几个，然后训练时需要另外下载预训练模型么？ @lin-honghui

How to choose the value alpha in Label Smoothing loss function?

Appreciate your work.

 I know that the hard examples are optimized as label smoothing loss function, 

and the ground-truth label k distribution is replaced as q(k)=(1-alpha)+alpha/K. I noticed that the value of alpha in your code is chosen as the percentage of hard examples, but why set a upper limit 0.2?
In addition, I still have a basic problem, you used a 11*11 conv2d to compute the edge mask, is the conv2d called Laplacian operator, can you explain how it works(it would be better if you could provide related articles of the explanation).
Wait for your reply sincerely.

snapshot ensemble部分的代码本baseline有做吗

inference error ValueError: could not broadcast input array from shape (512,512) into shape (512,256)

请问inference代码，下面有个报错
ValueError: could not broadcast input array from shape (512,512) into shape (512,256)

if(buttomright_x-topleft_x)==1024 and (buttomright_y-topleft_y)==1024:
print("predict_png.shape",predict_png.shape) # predict_png.shape (43520, 22016)
print(topleft_x,topleft_y,buttomright_x,buttomright_y)
# 每次预测只保留图像中心(512,512)区域预测结果
# if predict_png[topleft_y+256:buttomright_y-256,topleft_x+256:buttomright_x-256].shape != (512,512):
# print(predict_png[topleft_y+256:buttomright_y-256,topleft_x+256:buttomright_x-256].shape) # ,21760:22272
predict_png[topleft_y+256:buttomright_y-256,topleft_x+256:buttomright_x-256] = predict[256:768,256:768]
应该如何修改才好，谢谢

前辈 可以问下你还有数据集吗 可以分享下吗

标签

博主您好，我想问一下伪标签的策略是怎么实现的？我看您提供的代码中并没有具体的实施

想请问你们是如何消除空洞的

想请问你们是如何消除空洞的

您好，我想问一下这个算法中的patch操作是在哪里？

不好意思，个人比较菜，能不能麻烦您告知一下哪部分是进行patch操作？我想学习一下。

数据集下载

能否提供下数据集下载链接，谢谢

如何训练自己的数据集

请问一下load_path是预训练模型吧，项目中没有给出，怎么训练自己的数据集，非常感谢，期待您们回复。

请问一下有人遇到分割结果有明显锯齿状吗？

感谢楼主分享代码，谢谢。我在训练自己数据过程中，发现小图片分割结果的边缘明显的锯齿状，有遇到过这种情况吗。训练与预测都是512×512。