I'd like to share a brief overview of the work I've done to get to the 110th place. That was close to the medal zone, in which I actually spent a few weeks, but the competition really went wild on the last ~10 days and I couldn't spend much time on it, so at the end I didn't earn any medal, but still learnt a lot.
\nOverview:
\nI guess my solution lacked a voting system or a more careful look at the samples predicted to be in the medal zone (top 100).
\nThanks for everyone who competed and shared works and ideas!
", "messages": [ { "id": "1336394", "postDate": "06/04/2021 21:14:38", "content": "I'd like to share a brief overview of the work I've done to get to the 110th place. That was close to the medal zone, in which I actually spent a few weeks, but the competition really went wild on the last ~10 days and I couldn't spend much time on it, so at the end I didn't earn any medal, but still learnt a lot.
\nOverview:
\nI guess my solution lacked a voting system or a more careful look at the samples predicted to be in the medal zone (top 100).
\nThanks for everyone who competed and shared works and ideas!
", "rawMarkdown": "I'd like to share a brief overview of the work I've done to get to the 110th place. That was close to the medal zone, in which I actually spent a few weeks, but the competition really went wild on the last ~10 days and I couldn't spend much time on it, so at the end I didn't earn any medal, but still learnt a lot.\n\n\n**Overview:**\n- Used the data preprocessed in a few different image sizes ( (256,512,3), (512,512,3).... ) and max length (282, 282/2).\n- Uploaded all the preprocessed datasets to my personal account in Google Cloud Storage, which allowed me to use TPU in Colab Pro.\n- Trained some Densenet pytorch models using GPU (oof, that took some time) and some tensorflow models using TPU. The tensorflow models differed from each other mostly in hyperparameters choices and the dataset used.\n- Encoders tested: Densenet-169, EfficientNet-B0/B2/B4, EfficientNetV2-B2, and other [timm models](https://github.com/rwightman/pytorch-image-models/tree/master/timm) \n- Decoders tested: LSTM with attention, Transformer with attention\n- I've got the most effective results by manually tuning the learning rate scheduler\n- Post-processed ~10 submissions using @kirderf [notebook](https://www.kaggle.com/kirderf/basic-postp-with-different-submissions/)\n- RDKit normalization using @nofreewill [notebook](https://www.kaggle.com/nofreewill/normalize-your-predictions)\n\nI guess my solution lacked a voting system or a more careful look at the samples predicted to be in the medal zone (top 100).\n\nThanks for everyone who competed and shared works and ideas!", "votes": null }, { "id": "1336410", "postDate": "06/04/2021 21:47:36", "content": "Thanks for the write up @hinepo - it is interesting to learn about the method used by someone who finished close to my level in the rankings. I also found @kirderf's notebook and @nofreewill's method very useful.
", "rawMarkdown": "Thanks for the write up @hinepo - it is interesting to learn about the method used by someone who finished close to my level in the rankings. I also found @kirderf's notebook and @nofreewill's method very useful.", "votes": null }, { "id": "1337055", "postDate": "06/05/2021 11:36:56", "content": "Glad that it helped 😊 I also had less time training and tuning the solution in this competition, I put my effort in another one, had to prioritize the GPU and work limit.
", "rawMarkdown": "Glad that it helped 😊 I also had less time training and tuning the solution in this competition, I put my effort in another one, had to prioritize the GPU and work limit.", "votes": null } ], "comments": [ { "id": 1336410, "author_name": "jbomitchell", "author_url": "", "post_date": "06/04/2021 21:47:36", "content": "Thanks for the write up @hinepo - it is interesting to learn about the method used by someone who finished close to my level in the rankings. I also found @kirderf's notebook and @nofreewill's method very useful.
", "votes": null, "replies": [] }, { "id": 1337055, "author_name": "kirderf", "author_url": "", "post_date": "06/05/2021 11:36:56", "content": "Glad that it helped 😊 I also had less time training and tuning the solution in this competition, I put my effort in another one, had to prioritize the GPU and work limit.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1336394": "I'd like to share a brief overview of the work I've done to get to the 110th place. That was close to the medal zone, in which I actually spent a few weeks, but the competition really went wild on the last ~10 days and I couldn't spend much time on it, so at the end I didn't earn any medal, but still learnt a lot.\n\n\n**Overview:**\n- Used the data preprocessed in a few different image sizes ( (256,512,3), (512,512,3).... ) and max length (282, 282/2).\n- Uploaded all the preprocessed datasets to my personal account in Google Cloud Storage, which allowed me to use TPU in Colab Pro.\n- Trained some Densenet pytorch models using GPU (oof, that took some time) and some tensorflow models using TPU. The tensorflow models differed from each other mostly in hyperparameters choices and the dataset used.\n- Encoders tested: Densenet-169, EfficientNet-B0/B2/B4, EfficientNetV2-B2, and other [timm models](https://github.com/rwightman/pytorch-image-models/tree/master/timm) \n- Decoders tested: LSTM with attention, Transformer with attention\n- I've got the most effective results by manually tuning the learning rate scheduler\n- Post-processed ~10 submissions using @kirderf [notebook](https://www.kaggle.com/kirderf/basic-postp-with-different-submissions/)\n- RDKit normalization using @nofreewill [notebook](https://www.kaggle.com/nofreewill/normalize-your-predictions)\n\nI guess my solution lacked a voting system or a more careful look at the samples predicted to be in the medal zone (top 100).\n\nThanks for everyone who competed and shared works and ideas!", "1336410": "Thanks for the write up @hinepo - it is interesting to learn about the method used by someone who finished close to my level in the rankings. I also found @kirderf's notebook and @nofreewill's method very useful.", "1337055": "Glad that it helped 😊 I also had less time training and tuning the solution in this competition, I put my effort in another one, had to prioritize the GPU and work limit." }, "source": "meta" }