carnozhao / kaggle-uwmgit

code for kaggle: UW-Madison GI Tract Image Segmentation

Hi there, I'm Zhao Xun/赵洵 👋

Kaggle

CarnoZhao's Kaggle Profile

Rubik's Cube/魔方

养老魔方玩家粗饼主页

World Cube Association Zhao Xun's main page

kaggle-uwmgit's Issues

Loss

Is loss hardcoded to trversk in your SMP loss?

Using augments from Albumentation takes more longer training time.

Hello,
I tried to use your version of mmseg but I got some implementation and version errors. Then, I implement some features such as max normalization in ImageLoading and Albu from your source code. When I started training without Albu, it takes almost 2 mins for 50 iters, and with Albu, the training time for 50 iters gets doubled, 4+ mins. Is it OK or I do something wrong because your encoder_decoder.py code is a bit different from the original, and I did not use your style?

How to use swintransformer backbone in the mmsegmentation?

Only timm can be used by SMPUnet

Min max scaling

In your inference code at kaggle , I see we perform img/max(img)
but during the training we dont, or do we some where ?

Regards
Jaideep

Will there be any plan for the to do list on README.md?

Hi Carno,

Congratulations to the Championship! Thank you for sharing this elegant training pipeline during the competition. Will there be any plan to update the whole thing, like including the try and error part? (sorry...伸手党了）

这里为啥是 image.shape[1]

请问 这里不应该是 for i in range(msks.shape[0]): 吗？

而且 这样保存的 mask的shape是（h, w, 3）对吗？ 为啥不合到一个通道上呢，为啥mask的shape不合成（h, w）呢？

i find a problem in your code

Thanks for your sharing! i find a problem in your code /data/tract/baseline.ipynb : for i in range(msks.shape[1])->for i in range(msks.shape[0]). In addition, when I visualize, I find that the image and label do not match.