Hi Everyone, \nI'm really excited to be interviewing another person from the team at H2O.ai: GrandMaster @rohanrao, this is an AMA Questions invite (A huge Thanks to GM Rohan for agreeing to it) Thread for the interview.
\n\nPlease feel free to post any/all questions, if you like here. The date of the interview is yet to TBD so I'll keep collecting questions till then, I'll also post here once the date is finalised so that all of the questions can be collected before it. (It'll be in the first week-ish of January)
\n\nThis interview will be released on the Chai Time Data Science Podcast, available both as video, audio and blogposts of all of the interviews will be coming really soon here which I'll also announce on my Newsletter-so if you're excited to read the interviews, please feel free to sign up or just keep an eye out on this thread.
\n\nThemes that I'd be interested in talking about (For you to suggest more ideas similar to these, but of course any questions are welcome)
\n\nThank You in Advance for the Questions!\nAnd Thanks to GM Rohan! 🍵
", "messages": [ { "id": "705791", "postDate": "12/29/2019 13:28:15", "content": "Hi Everyone, \nI'm really excited to be interviewing another person from the team at H2O.ai: GrandMaster @rohanrao, this is an AMA Questions invite (A huge Thanks to GM Rohan for agreeing to it) Thread for the interview.
\n\nPlease feel free to post any/all questions, if you like here. The date of the interview is yet to TBD so I'll keep collecting questions till then, I'll also post here once the date is finalised so that all of the questions can be collected before it. (It'll be in the first week-ish of January)
\n\nThis interview will be released on the Chai Time Data Science Podcast, available both as video, audio and blogposts of all of the interviews will be coming really soon here which I'll also announce on my Newsletter-so if you're excited to read the interviews, please feel free to sign up or just keep an eye out on this thread.
\n\nThemes that I'd be interested in talking about (For you to suggest more ideas similar to these, but of course any questions are welcome)
\n\nThank You in Advance for the Questions!\nAnd Thanks to GM Rohan! 🍵
", "rawMarkdown": "Hi Everyone, \nI'm really excited to be interviewing another person from the team at H2O.ai: GrandMaster @rohanrao, this is an AMA Questions invite (A huge Thanks to GM Rohan for agreeing to it) Thread for the interview. \n\nPlease feel free to post any/all questions, if you like here. The date of the interview is yet to TBD so I'll keep collecting questions till then, I'll also post here once the date is finalised so that all of the questions can be collected before it. (It'll be in the first week-ish of January)\n\nThis interview will be released on the [Chai Time Data Science Podcast](https://www.youtube.com/channel/UCRjtBP-o5FbgRzX2BHQEFtQ), available both as [video](https://www.youtube.com/playlist?list=PLLvvXm0q8zUbiNdoIazGzlENMXvZ9bd3x), [audio](https://anchor.fm/chaitimedatascience) and blogposts of **all** of the interviews will be coming **really** soon [here](https://sanyambhutani.com) which I'll also announce on my [Newsletter](https://tinyletter.com/sanyambhutani)-so if you're excited to read the interviews, please feel free to sign up or just keep an eye out on this thread.\n\nThemes that I'd be interested in talking about (For you to suggest more ideas similar to these, but of course any questions are welcome)\n\n- Rohan is one of the few GMs that have been active over the span of a few years on Kaggle. His Kaggle Journey and approaches over the few years. \n- Ofcourse the 2nd place solution overview [Read it here if you haven't](https://www.kaggle.com/c/ashrae-energy-prediction/discussion/123481)\n- Rohan has done many awesome talks, 2 of which I've witnessed about Data Science off Kaggle and the differences of things you learn on the platform VS off the platform \n- His current work at H2O.\n\n\nThank You in Advance for the Questions!\nAnd Thanks to GM Rohan! 🍵", "votes": null }, { "id": "705806", "postDate": "12/29/2019 13:53:28", "content": "Can you ask rohan
\n\n1.How does he balance between study and practice of data science ?
", "rawMarkdown": "Can you ask rohan\n\n1.How does he balance between study and practice of data science ?", "votes": null }, { "id": "706309", "postDate": "12/30/2019 08:27:07", "content": "Data Science is a vast and fast growing subject. How can one manage to upskill in this fast pace?
", "rawMarkdown": "Data Science is a vast and fast growing subject. How can one manage to upskill in this fast pace?", "votes": null }, { "id": "706512", "postDate": "12/30/2019 13:45:33", "content": "I have a similar question with @anandpuntambekar. \nKaggle's competitions are very good chance to learn, but the competitions also cost a lot of time. So I wonder how he arrange the time between job and Kaggle.
", "rawMarkdown": "I have a similar question with @anandpuntambekar. \nKaggle's competitions are very good chance to learn, but the competitions also cost a lot of time. So I wonder how he arrange the time between job and Kaggle.", "votes": null }, { "id": "706513", "postDate": "12/30/2019 13:48:15", "content": "A question about the ASHRAE solution: Did you and your team try to build time-series models like CNN/LSTM's for ASHRAE? If so, why were they not included in the solution?
", "rawMarkdown": "A question about the ASHRAE solution: Did you and your team try to build time-series models like CNN/LSTM's for ASHRAE? If so, why were they not included in the solution?", "votes": null }, { "id": "706624", "postDate": "12/30/2019 16:18:36", "content": "Q1. How did they train on that massive data? (Loading the entire data on RAM was not possible. Down-sampling also didn't help and had the issue of losing out information of the data not included in the sample.)\nQ2. How to keep track if the training is working fine? What was the average training duration for each model and what were the maximum and minimum training duration? (Sometimes, my kernel ran for 8 hours and still didn't conclude. I had to force shut the training since I couldn't print anything and check if process is still working fine.)\nQ3. How did they deal with missing data?\nQ4. Did they perform any hyperparameter tuning? If yes, any technique that helped them perform the tuning conveniently? (Looking at the architecture, large number of models were trained. Doing hyperparameter tuning for all of them would be an arduous task.)\nQ5. How many total hours were put into the entire process, starting from data cleaning to model prediction accuracy?
", "rawMarkdown": "Q1. How did they train on that massive data? (Loading the entire data on RAM was not possible. Down-sampling also didn't help and had the issue of losing out information of the data not included in the sample.)\nQ2. How to keep track if the training is working fine? What was the average training duration for each model and what were the maximum and minimum training duration? (Sometimes, my kernel ran for 8 hours and still didn't conclude. I had to force shut the training since I couldn't print anything and check if process is still working fine.)\nQ3. How did they deal with missing data?\nQ4. Did they perform any hyperparameter tuning? If yes, any technique that helped them perform the tuning conveniently? (Looking at the architecture, large number of models were trained. Doing hyperparameter tuning for all of them would be an arduous task.)\nQ5. How many total hours were put into the entire process, starting from data cleaning to model prediction accuracy?", "votes": null }, { "id": "706831", "postDate": "12/30/2019 22:17:12", "content": "Brilliant! Looking forward to this podcast.\nSome suggestions for questions:
\n\nWhat did he struggle with during this competition? what unique challenges did he face in this competition that he did not experience in previous competitions. Did he find any similarities between this competition and previous Kaggle competitions?
He mentionned that his team spent lot of time cleaning the data manually for each building.\nDid they try automating the data cleaning first? Any ideas on how to automate the data cleaning as this could be beneficial if the dataset grows a lot in terms of the number of buildings. \nhow did they go about experimenting before cleaning up all data? perhaps, just clealing small part of the data and see if that helps LB score?
As asked previously, did he use Kaggle kernels for this competition or other alternatives?\nIf Kaggle kernels were used, how to make sure the kernel does not run more than the 9 hours limit?\nWas he forced to make any compromise in model training/validation due to the large dataset?
Estimated average time spent per week or per day on this competition. Classic question on how to deal with daily work and Kaggle.
How did he join his team? How did they communicate and coordinate their work together?
He was very active during this competition and has shared lot of valuable kernels. I've heard some people avoid sharing too much information during competitions to avoid losing medals.. What does he think about this? Will he continue sharing knowledge during competitions after winning in the money for the first time?
Small story on his impressive history of Sudoku/Puzzle championship :) Does he juggle between practicing Sudoku and Kaggle as well? xD
Brilliant! Thanks for sharing @init27 ! I learn a lot with your Interviews
", "rawMarkdown": "Brilliant! Thanks for sharing @init27 ! I learn a lot with your Interviews", "votes": null }, { "id": "708069", "postDate": "01/01/2020 23:26:10", "content": "Congrats @@rohanrao !
", "rawMarkdown": "Congrats @@rohanrao !", "votes": null }, { "id": "708179", "postDate": "01/02/2020 03:40:08", "content": "Can you ask rohan?
\n\nHow much time do I have to spend daily to be a grandmaster?\nSometimes I spend a lot of time but I don't get any results, maybe while studying I get distracted\nand I feel that I am studying but I don't xd, so what might be a reasonable daily amount of time to get gold medals?\nThanks :D
", "rawMarkdown": "Can you ask rohan?\n\nHow much time do I have to spend daily to be a grandmaster?\nSometimes I spend a lot of time but I don't get any results, maybe while studying I get distracted\nand I feel that I am studying but I don't xd, so what might be a reasonable daily amount of time to get gold medals?\nThanks :D", "votes": null }, { "id": "709442", "postDate": "01/03/2020 14:17:18", "content": "Thank you for sharing, very nice solution indeed :)
", "rawMarkdown": "Thank you for sharing, very nice solution indeed :)", "votes": null } ], "comments": [ { "id": 705806, "author_name": "anandpuntambekar", "author_url": "", "post_date": "12/29/2019 13:53:28", "content": "Can you ask rohan
\n\n1.How does he balance between study and practice of data science ?
", "votes": null, "replies": [] }, { "id": 706309, "author_name": "poornaramakrishnan", "author_url": "", "post_date": "12/30/2019 08:27:07", "content": "Data Science is a vast and fast growing subject. How can one manage to upskill in this fast pace?
", "votes": null, "replies": [] }, { "id": 706512, "author_name": "lsyiverson1991", "author_url": "", "post_date": "12/30/2019 13:45:33", "content": "I have a similar question with @anandpuntambekar. \nKaggle's competitions are very good chance to learn, but the competitions also cost a lot of time. So I wonder how he arrange the time between job and Kaggle.
", "votes": null, "replies": [] }, { "id": 706513, "author_name": "carlolepelaars", "author_url": "", "post_date": "12/30/2019 13:48:15", "content": "A question about the ASHRAE solution: Did you and your team try to build time-series models like CNN/LSTM's for ASHRAE? If so, why were they not included in the solution?
", "votes": null, "replies": [] }, { "id": 706624, "author_name": "shibashrit", "author_url": "", "post_date": "12/30/2019 16:18:36", "content": "Q1. How did they train on that massive data? (Loading the entire data on RAM was not possible. Down-sampling also didn't help and had the issue of losing out information of the data not included in the sample.)\nQ2. How to keep track if the training is working fine? What was the average training duration for each model and what were the maximum and minimum training duration? (Sometimes, my kernel ran for 8 hours and still didn't conclude. I had to force shut the training since I couldn't print anything and check if process is still working fine.)\nQ3. How did they deal with missing data?\nQ4. Did they perform any hyperparameter tuning? If yes, any technique that helped them perform the tuning conveniently? (Looking at the architecture, large number of models were trained. Doing hyperparameter tuning for all of them would be an arduous task.)\nQ5. How many total hours were put into the entire process, starting from data cleaning to model prediction accuracy?
", "votes": null, "replies": [] }, { "id": 706831, "author_name": "nanowa", "author_url": "", "post_date": "12/30/2019 22:17:12", "content": "Brilliant! Looking forward to this podcast.\nSome suggestions for questions:
\n\nWhat did he struggle with during this competition? what unique challenges did he face in this competition that he did not experience in previous competitions. Did he find any similarities between this competition and previous Kaggle competitions?
He mentionned that his team spent lot of time cleaning the data manually for each building.\nDid they try automating the data cleaning first? Any ideas on how to automate the data cleaning as this could be beneficial if the dataset grows a lot in terms of the number of buildings. \nhow did they go about experimenting before cleaning up all data? perhaps, just clealing small part of the data and see if that helps LB score?
As asked previously, did he use Kaggle kernels for this competition or other alternatives?\nIf Kaggle kernels were used, how to make sure the kernel does not run more than the 9 hours limit?\nWas he forced to make any compromise in model training/validation due to the large dataset?
Estimated average time spent per week or per day on this competition. Classic question on how to deal with daily work and Kaggle.
How did he join his team? How did they communicate and coordinate their work together?
He was very active during this competition and has shared lot of valuable kernels. I've heard some people avoid sharing too much information during competitions to avoid losing medals.. What does he think about this? Will he continue sharing knowledge during competitions after winning in the money for the first time?
Small story on his impressive history of Sudoku/Puzzle championship :) Does he juggle between practicing Sudoku and Kaggle as well? xD
Brilliant! Thanks for sharing @init27 ! I learn a lot with your Interviews
", "votes": null, "replies": [ { "id": 708069, "author_name": "caesarlupum", "author_url": "", "post_date": "01/01/2020 23:26:10", "content": "Congrats @@rohanrao !
", "votes": null, "replies": [] } ] }, { "id": 708179, "author_name": "johnt666", "author_url": "", "post_date": "01/02/2020 03:40:08", "content": "Can you ask rohan?
\n\nHow much time do I have to spend daily to be a grandmaster?\nSometimes I spend a lot of time but I don't get any results, maybe while studying I get distracted\nand I feel that I am studying but I don't xd, so what might be a reasonable daily amount of time to get gold medals?\nThanks :D
", "votes": null, "replies": [] }, { "id": 709442, "author_name": "vladlee", "author_url": "", "post_date": "01/03/2020 14:17:18", "content": "Thank you for sharing, very nice solution indeed :)
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "705791": "Hi Everyone, \nI'm really excited to be interviewing another person from the team at H2O.ai: GrandMaster @rohanrao, this is an AMA Questions invite (A huge Thanks to GM Rohan for agreeing to it) Thread for the interview. \n\nPlease feel free to post any/all questions, if you like here. The date of the interview is yet to TBD so I'll keep collecting questions till then, I'll also post here once the date is finalised so that all of the questions can be collected before it. (It'll be in the first week-ish of January)\n\nThis interview will be released on the [Chai Time Data Science Podcast](https://www.youtube.com/channel/UCRjtBP-o5FbgRzX2BHQEFtQ), available both as [video](https://www.youtube.com/playlist?list=PLLvvXm0q8zUbiNdoIazGzlENMXvZ9bd3x), [audio](https://anchor.fm/chaitimedatascience) and blogposts of **all** of the interviews will be coming **really** soon [here](https://sanyambhutani.com) which I'll also announce on my [Newsletter](https://tinyletter.com/sanyambhutani)-so if you're excited to read the interviews, please feel free to sign up or just keep an eye out on this thread.\n\nThemes that I'd be interested in talking about (For you to suggest more ideas similar to these, but of course any questions are welcome)\n\n- Rohan is one of the few GMs that have been active over the span of a few years on Kaggle. His Kaggle Journey and approaches over the few years. \n- Ofcourse the 2nd place solution overview [Read it here if you haven't](https://www.kaggle.com/c/ashrae-energy-prediction/discussion/123481)\n- Rohan has done many awesome talks, 2 of which I've witnessed about Data Science off Kaggle and the differences of things you learn on the platform VS off the platform \n- His current work at H2O.\n\n\nThank You in Advance for the Questions!\nAnd Thanks to GM Rohan! 🍵", "705806": "Can you ask rohan\n\n1.How does he balance between study and practice of data science ?", "706309": "Data Science is a vast and fast growing subject. How can one manage to upskill in this fast pace?", "706512": "I have a similar question with @anandpuntambekar. \nKaggle's competitions are very good chance to learn, but the competitions also cost a lot of time. So I wonder how he arrange the time between job and Kaggle.", "706513": "A question about the ASHRAE solution: Did you and your team try to build time-series models like CNN/LSTM's for ASHRAE? If so, why were they not included in the solution?", "706624": "Q1. How did they train on that massive data? (Loading the entire data on RAM was not possible. Down-sampling also didn't help and had the issue of losing out information of the data not included in the sample.)\nQ2. How to keep track if the training is working fine? What was the average training duration for each model and what were the maximum and minimum training duration? (Sometimes, my kernel ran for 8 hours and still didn't conclude. I had to force shut the training since I couldn't print anything and check if process is still working fine.)\nQ3. How did they deal with missing data?\nQ4. Did they perform any hyperparameter tuning? If yes, any technique that helped them perform the tuning conveniently? (Looking at the architecture, large number of models were trained. Doing hyperparameter tuning for all of them would be an arduous task.)\nQ5. How many total hours were put into the entire process, starting from data cleaning to model prediction accuracy?", "706831": "Brilliant! Looking forward to this podcast.\nSome suggestions for questions:\n\n1. What did he struggle with during this competition? what unique challenges did he face in this competition that he did not experience in previous competitions. Did he find any similarities between this competition and previous Kaggle competitions?\n\n2. He mentionned that his team spent lot of time cleaning the data manually for each building.\nDid they try automating the data cleaning first? Any ideas on how to automate the data cleaning as this could be beneficial if the dataset grows a lot in terms of the number of buildings. \nhow did they go about experimenting before cleaning up all data? perhaps, just clealing small part of the data and see if that helps LB score? \n\n3. As asked previously, did he use Kaggle kernels for this competition or other alternatives?\nIf Kaggle kernels were used, how to make sure the kernel does not run more than the 9 hours limit?\nWas he forced to make any compromise in model training/validation due to the large dataset?\n\n4. Estimated average time spent per week or per day on this competition. Classic question on how to deal with daily work and Kaggle.\n\n5. How did he join his team? How did they communicate and coordinate their work together?\n\n6. He was very active during this competition and has shared lot of valuable kernels. I've heard some people avoid sharing too much information during competitions to avoid losing medals.. What does he think about this? Will he continue sharing knowledge during competitions after winning in the money for the first time?\n\n7. Small story on his impressive history of Sudoku/Puzzle championship :) Does he juggle between practicing Sudoku and Kaggle as well? xD", "708068": "Brilliant! Thanks for sharing @init27 ! I learn a lot with your Interviews", "708069": "Congrats @@rohanrao !", "708179": "Can you ask rohan?\n\nHow much time do I have to spend daily to be a grandmaster?\nSometimes I spend a lot of time but I don't get any results, maybe while studying I get distracted\nand I feel that I am studying but I don't xd, so what might be a reasonable daily amount of time to get gold medals?\nThanks :D", "709442": "Thank you for sharing, very nice solution indeed :)" }, "source": "meta" }