I have published a pipeline on training using PyTorch. If anyone finds any improvement, please comment in the notebook! I have spent a lot of time to construct such a pipeline for image classification tasks, this is because I am a strong believer in code reusability. The purpose is to change a minimal amount of code every time you run a new training. Neat and clean brings one a long way as ML/DL is more than just knowing the theories, but also includes software engineering.
\nMuch thanks to @shonenkov as my Trainer Class is largely modified from his previous works! Also, thanks to the first Quadruple GM @abhishek as his youtube videos inspired me to (try my best to) write clean code.
PS: It does not look as neat in notebook format, but I made a GitHub link to store individual classes into respective files, which is a better practice. I will also further attach my detailed explanation of each line when I have the time.
\nFurther Edits: I used this Pipeline for almost all my training.
\nBest single model (5 folds) - effnetb5-ns - CV: 0.891, LB: 0.898 No TTA
I have published a pipeline on training using PyTorch. If anyone finds any improvement, please comment in the notebook! I have spent a lot of time to construct such a pipeline for image classification tasks, this is because I am a strong believer in code reusability. The purpose is to change a minimal amount of code every time you run a new training. Neat and clean brings one a long way as ML/DL is more than just knowing the theories, but also includes software engineering.
\nMuch thanks to @shonenkov as my Trainer Class is largely modified from his previous works! Also, thanks to the first Quadruple GM @abhishek as his youtube videos inspired me to (try my best to) write clean code.
PS: It does not look as neat in notebook format, but I made a GitHub link to store individual classes into respective files, which is a better practice. I will also further attach my detailed explanation of each line when I have the time.
\nFurther Edits: I used this Pipeline for almost all my training.
\nBest single model (5 folds) - effnetb5-ns - CV: 0.891, LB: 0.898 No TTA
Great notebook！Thanks for sharing.
\nBTW, how do you feel about the size of models.
\nIt there much difference from b3-b5 for you?
\nNot that much for me.
It wasn't that different for me.
\np.s. there may still be insufficient experiments.
\n@zlanan
@zlanan hey not much difference for me. I only used effnets so far but b5ns with a certain aug seems to score very well on LB. even without tta
", "rawMarkdown": "zlanan hey not much difference for me. I only used effnets so far but b5ns with a certain aug seems to score very well on LB. even without tta", "votes": null } ], "comments": [ { "id": 1128633, "author_name": "zlanan", "author_url": "", "post_date": "12/27/2020 15:57:10", "content": "Great notebook！Thanks for sharing.
\nBTW, how do you feel about the size of models.
\nIt there much difference from b3-b5 for you?
\nNot that much for me.
It wasn't that different for me.
\np.s. there may still be insufficient experiments.
\n@zlanan
@zlanan hey not much difference for me. I only used effnets so far but b5ns with a certain aug seems to score very well on LB. even without tta
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1128290": "I have published a pipeline on training using PyTorch. If anyone finds any improvement, please comment in the notebook! I have spent a lot of time to construct such a pipeline for image classification tasks, this is because I am a strong believer in code reusability. The purpose is to change a minimal amount of code every time you run a new training. Neat and clean brings one a long way as ML/DL is more than just knowing the theories, but also includes software engineering.\n\nMuch thanks to @shonenkov as my `Trainer Class` is largely modified from his previous works! Also, thanks to the first Quadruple GM @abhishek as his [youtube videos](https://www.youtube.com/user/abhisheksvnit) inspired me to (try my best to) write clean code.\n\nPS: It does not look as neat in notebook format, but I made a GitHub link to store individual classes into respective files, which is a better practice. I will also further attach my detailed explanation of each line when I have the time.\n\nFurther Edits: I used this Pipeline for almost all my training. \nBest single model (5 folds) - effnetb5-ns - CV: 0.891, LB: 0.898 No TTA\n\n[Reusable PyTorch Pipeline](https://www.kaggle.com/reighns/super-complete-and-reusable-pytorch-pipeline)", "1128633": "Great notebook！Thanks for sharing.\n\nBTW, how do you feel about the size of models. \nIt there much difference from b3-b5 for you?\nNot that much for me.", "1128642": "It wasn't that different for me.\n\np.s. there may still be insufficient experiments.\n@zlanan", "1128651": "zlanan hey not much difference for me. I only used effnets so far but b5ns with a certain aug seems to score very well on LB. even without tta" }, "source": "meta" }