{
"id": 220660,
"title": "My learnings so far - After the competition",
"url": "/competitions/cassava-leaf-disease-classification/discussion/220660",
"author_name": "Krishna Kishor Kammaje",
"post_date": "2021-02-19T05:51:37.336000",
"votes": 4,
"comment_count": 0,
"views": 0,
"content": "\n- Simple models (no multi-level ensembles, no fancy augs like cutmix) can also get a bronze/silver
\n- You might use EfficientNet B0, B3 for baselining, but for the final score ensure that you use the B7 version with 512 image size. B7 will most of the time improve the score compared to simpler B5, B3.
\n- Kaggle hardware alone is not enough. Get used to using Google Colab (still free)
\n- TTAs usually improve the score
\n- ViT can be used for ensemble (improves diversity) but not on its own
\n- fastai has come a long way and can be useful to create simple (less code) but powerful (augs like cutmix) training process.
\n- When you have two submissions with the same score, older submission is a better one. (if you decide to ignore training dynamics between the two)
\n
",
"messages": [
{
"id": 1209943,
"postDate": "2021-02-19T05:51:37.337Z",
"content": "\n- Simple models (no multi-level ensembles, no fancy augs like cutmix) can also get a bronze/silver
\n- You might use EfficientNet B0, B3 for baselining, but for the final score ensure that you use the B7 version with 512 image size. B7 will most of the time improve the score compared to simpler B5, B3.
\n- Kaggle hardware alone is not enough. Get used to using Google Colab (still free)
\n- TTAs usually improve the score
\n- ViT can be used for ensemble (improves diversity) but not on its own
\n- fastai has come a long way and can be useful to create simple (less code) but powerful (augs like cutmix) training process.
\n- When you have two submissions with the same score, older submission is a better one. (if you decide to ignore training dynamics between the two)
\n
",
"rawMarkdown": "- [Simple models](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/220632) (no multi-level ensembles, no fancy augs like cutmix) can also get a bronze/silver\n- You might use EfficientNet B0, B3 for baselining, but for the final score ensure that you use the B7 version with 512 image size. B7 will most of the time improve the score compared to simpler B5, B3. \n- Kaggle hardware alone is not enough. Get used to using Google Colab (still free)\n- TTAs usually improve the score\n- ViT can be used for ensemble (improves diversity) but not on its own\n- fastai has come a long way and can be useful to create simple (less code) but powerful (augs like cutmix) [training process](https://www.kaggle.com/andyjianzhou/cassava-training-fastai-efficnetnet-distill).\n- When you have two submissions with the same score, [older submission is a better one](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/220602). (if you decide to ignore training dynamics between the two)",
"votes": 4
}
],
"comments": [],
"raw_markdown_by_id": {
"1209943": "- [Simple models](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/220632) (no multi-level ensembles, no fancy augs like cutmix) can also get a bronze/silver\n- You might use EfficientNet B0, B3 for baselining, but for the final score ensure that you use the B7 version with 512 image size. B7 will most of the time improve the score compared to simpler B5, B3. \n- Kaggle hardware alone is not enough. Get used to using Google Colab (still free)\n- TTAs usually improve the score\n- ViT can be used for ensemble (improves diversity) but not on its own\n- fastai has come a long way and can be useful to create simple (less code) but powerful (augs like cutmix) [training process](https://www.kaggle.com/andyjianzhou/cassava-training-fastai-efficnetnet-distill).\n- When you have two submissions with the same score, [older submission is a better one](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/220602). (if you decide to ignore training dynamics between the two)"
}
}