{ "id": 75047, "title": "Takeaways from this competition", "url": "/competitions/PLAsTiCC-2018/discussion/75047", "author_name": "", "post_date": "2018-12-18T05:49:57.216076Z", "votes": 5, "comment_count": 1, "views": 0, "content": "

Hi Kagglers,

\n\n

I wanted to start a topic where people share his/her takeaways from this competition. My takeaways are listed below:

\n\n

1) Importance of using Dask rather than Pandas.

\n\n

2) AutoEncoders didn't bring a gain while extracting features.

\n\n

Please share your takeaways to increase knowledge in the world.

", "messages": [ { "id": "440957", "postDate": "12/18/2018 05:49:57", "content": "

Hi Kagglers,

\n\n

I wanted to start a topic where people share his/her takeaways from this competition. My takeaways are listed below:

\n\n

1) Importance of using Dask rather than Pandas.

\n\n

2) AutoEncoders didn't bring a gain while extracting features.

\n\n

Please share your takeaways to increase knowledge in the world.

", "rawMarkdown": "Hi Kagglers,\n\nI wanted to start a topic where people share his/her takeaways from this competition. My takeaways are listed below:\n\n1) Importance of using [Dask](https://github.com/dask/dask) rather than [Pandas](https://pandas.pydata.org/). \n\n2) AutoEncoders didn't bring a gain while extracting features.\n\nPlease share your takeaways to increase knowledge in the world.", "votes": null }, { "id": "441768", "postDate": "12/19/2018 03:23:15", "content": "

Here are my main takeaways:\n- Augmenting and improving dataset can be very important as was shown with Kyle and other’s solutions. \n- RNN’s are not the only way to tackle time series (this was my first time series competition)\n- Parallel processing of large test set is essential to being able to iterate on ideas quickly. For the TrackML competition I had set up parallel processing of the test set, for this competition I did not and it cost me a lot of time.\n- Use your time wisely. I spent quite a bit of time building a PyTorch model that I didn’t end up using, this time would have been better spent on feature engineering but I learned a few new things about PyTorch so it wasn’t a complete waste of time.\n- Most of the time I spent working on this competition was in the evenings after work, I think I would have been better off getting up earlier and working on the competition in the morning when my mind was clearer. \n- Ask more questions in the discussions.

", "rawMarkdown": "Here are my main takeaways:\n- Augmenting and improving dataset can be very important as was shown with Kyle and other’s solutions. \n- RNN’s are not the only way to tackle time series (this was my first time series competition)\n- Parallel processing of large test set is essential to being able to iterate on ideas quickly. For the TrackML competition I had set up parallel processing of the test set, for this competition I did not and it cost me a lot of time.\n- Use your time wisely. I spent quite a bit of time building a PyTorch model that I didn’t end up using, this time would have been better spent on feature engineering but I learned a few new things about PyTorch so it wasn’t a complete waste of time.\n- Most of the time I spent working on this competition was in the evenings after work, I think I would have been better off getting up earlier and working on the competition in the morning when my mind was clearer. \n- Ask more questions in the discussions.", "votes": null } ], "comments": [ { "id": 441768, "author_name": "jackvial", "author_url": "", "post_date": "12/19/2018 03:23:15", "content": "

Here are my main takeaways:\n- Augmenting and improving dataset can be very important as was shown with Kyle and other’s solutions. \n- RNN’s are not the only way to tackle time series (this was my first time series competition)\n- Parallel processing of large test set is essential to being able to iterate on ideas quickly. For the TrackML competition I had set up parallel processing of the test set, for this competition I did not and it cost me a lot of time.\n- Use your time wisely. I spent quite a bit of time building a PyTorch model that I didn’t end up using, this time would have been better spent on feature engineering but I learned a few new things about PyTorch so it wasn’t a complete waste of time.\n- Most of the time I spent working on this competition was in the evenings after work, I think I would have been better off getting up earlier and working on the competition in the morning when my mind was clearer. \n- Ask more questions in the discussions.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "440957": "Hi Kagglers,\n\nI wanted to start a topic where people share his/her takeaways from this competition. My takeaways are listed below:\n\n1) Importance of using [Dask](https://github.com/dask/dask) rather than [Pandas](https://pandas.pydata.org/). \n\n2) AutoEncoders didn't bring a gain while extracting features.\n\nPlease share your takeaways to increase knowledge in the world.", "441768": "Here are my main takeaways:\n- Augmenting and improving dataset can be very important as was shown with Kyle and other’s solutions. \n- RNN’s are not the only way to tackle time series (this was my first time series competition)\n- Parallel processing of large test set is essential to being able to iterate on ideas quickly. For the TrackML competition I had set up parallel processing of the test set, for this competition I did not and it cost me a lot of time.\n- Use your time wisely. I spent quite a bit of time building a PyTorch model that I didn’t end up using, this time would have been better spent on feature engineering but I learned a few new things about PyTorch so it wasn’t a complete waste of time.\n- Most of the time I spent working on this competition was in the evenings after work, I think I would have been better off getting up earlier and working on the competition in the morning when my mind was clearer. \n- Ask more questions in the discussions." }, "source": "meta" }