{ "id": 206368, "title": "Advice for Kaggle beginner with Deep Learning knowlege", "url": "/competitions/cassava-leaf-disease-classification/discussion/206368", "author_name": "", "post_date": "2020-12-24T08:52:43.986765600Z", "votes": 5, "comment_count": 1, "views": 0, "content": "

Hey everyone,

\n

I am a Computer Science student that has taken many Deep Learning courses but has no experience in Kaggle competitions. So I only have experience in achieving a reasonable benchmark performance, rather than squeezing out the maximum possible performance.

\n

I would appreciate it if someone could point me to resources that explains tricks typically used and how I should focus on to maximize my rank for this competition. For example:

\n
    \n
  1. I hadn't heard of using multiple Test Time Augmentations prior to this.
    2. \n
  2. How should I focus my limited compute resources on tweaking the architecture, hyper-parameters, data augmentation? The solution to last year's competition only mentions architecture and augmentation.
    3. \n
\n

So far, I have learned from the discussion board to:

\n
    \n
  1. Prioritize Cross-Validation over Public Leader Board
    2. \n
  2. Test Time Augmentation is important
    3. \n
  3. Use an ensemble of multiple networks
    4. \n
", "messages": [ { "id": "1124881", "postDate": "12/24/2020 08:52:43", "content": "

Hey everyone,

\n

I am a Computer Science student that has taken many Deep Learning courses but has no experience in Kaggle competitions. So I only have experience in achieving a reasonable benchmark performance, rather than squeezing out the maximum possible performance.

\n

I would appreciate it if someone could point me to resources that explains tricks typically used and how I should focus on to maximize my rank for this competition. For example:

\n
    \n
  1. I hadn't heard of using multiple Test Time Augmentations prior to this.
    2. \n
  2. How should I focus my limited compute resources on tweaking the architecture, hyper-parameters, data augmentation? The solution to last year's competition only mentions architecture and augmentation.
    3. \n
\n

So far, I have learned from the discussion board to:

\n
    \n
  1. Prioritize Cross-Validation over Public Leader Board
    2. \n
  2. Test Time Augmentation is important
    3. \n
  3. Use an ensemble of multiple networks
    4. \n
", "rawMarkdown": "Hey everyone,\n\nI am a Computer Science student that has taken many Deep Learning courses but has no experience in Kaggle competitions. So I only have experience in achieving a reasonable benchmark performance, rather than squeezing out the maximum possible performance.\n\nI would appreciate it if someone could point me to resources that explains tricks typically used and how I should focus on to maximize my rank for this competition. For example: \n\n1. I hadn't heard of using multiple Test Time Augmentations prior to this.\n2. How should I focus my limited compute resources on tweaking the architecture, hyper-parameters, data augmentation? The solution to [last year's competition](https://www.kaggle.com/c/cassava-disease/discussion/94114) only mentions architecture and augmentation.\n\nSo far, I have learned from the discussion board to:\n\n1. Prioritize Cross-Validation over Public Leader Board\n2. Test Time Augmentation is important\n3. Use an ensemble of multiple networks", "votes": null }, { "id": "1124926", "postDate": "12/24/2020 09:22:55", "content": "

Several resources needed to maximize rank.

\n
    \n
  1. Put a complete profile in your account. Where your live, school attending, etc. Be sure to include the social networking links that Kaggle provides (github, linkedin, etc). If you were a computer science student at the same university I attended I might extend you an invite as team member. Being on a team that is willing to help you learn Kaggle is the #1 way for a novice to maximize rank.
    2. \n
  2. Put an ad for yourself in the team wanted post. Per #1 above, being on a team. Be sure to repeat some of what you said in this post. (It is likely you will however end up on a team that sucks so don't take the first offer until you research your potential team member).
    3. \n
  3. Be a super smart or plan to spend lots of hours.
    4. \n
  4. Check out most of the previous (last year or so) vision competitions. Sort shared notebooks by best score. Pick your product of choice (keras or pytorch) and look at top scoring.
    5. \n
  5. This competition has some very noisy data - reach back into your Deep Learning course materials.
    6. \n
  6. The short answer - to get in the top 10% you will need to do it all.
    7. \n
", "rawMarkdown": "Several resources needed to maximize rank.\n\n1. Put a complete profile in your account. Where your live, school attending, etc. Be sure to include the social networking links that Kaggle provides (github, linkedin, etc). If you were a computer science student at the same university I attended I might extend you an invite as team member. Being on a team that is willing to help you learn Kaggle is the #1 way for a novice to maximize rank.\n2. Put an ad for yourself in the team wanted post. Per #1 above, being on a team. Be sure to repeat some of what you said in this post. (It is likely you will however end up on a team that sucks so don't take the first offer until you research your potential team member).\n3. Be a super smart or plan to spend lots of hours.\n4. Check out most of the previous (last year or so) vision competitions. Sort shared notebooks by best score. Pick your product of choice (keras or pytorch) and look at top scoring. \n5. This competition has some very noisy data - reach back into your Deep Learning course materials.\n6. The short answer - to get in the top 10% you will need to do it all.", "votes": null } ], "comments": [ { "id": 1124926, "author_name": "pcjimmmy", "author_url": "", "post_date": "12/24/2020 09:22:55", "content": "

Several resources needed to maximize rank.

\n
    \n
  1. Put a complete profile in your account. Where your live, school attending, etc. Be sure to include the social networking links that Kaggle provides (github, linkedin, etc). If you were a computer science student at the same university I attended I might extend you an invite as team member. Being on a team that is willing to help you learn Kaggle is the #1 way for a novice to maximize rank.
    2. \n
  2. Put an ad for yourself in the team wanted post. Per #1 above, being on a team. Be sure to repeat some of what you said in this post. (It is likely you will however end up on a team that sucks so don't take the first offer until you research your potential team member).
    3. \n
  3. Be a super smart or plan to spend lots of hours.
    4. \n
  4. Check out most of the previous (last year or so) vision competitions. Sort shared notebooks by best score. Pick your product of choice (keras or pytorch) and look at top scoring.
    5. \n
  5. This competition has some very noisy data - reach back into your Deep Learning course materials.
    6. \n
  6. The short answer - to get in the top 10% you will need to do it all.
    7. \n
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1124881": "Hey everyone,\n\nI am a Computer Science student that has taken many Deep Learning courses but has no experience in Kaggle competitions. So I only have experience in achieving a reasonable benchmark performance, rather than squeezing out the maximum possible performance.\n\nI would appreciate it if someone could point me to resources that explains tricks typically used and how I should focus on to maximize my rank for this competition. For example: \n\n1. I hadn't heard of using multiple Test Time Augmentations prior to this.\n2. How should I focus my limited compute resources on tweaking the architecture, hyper-parameters, data augmentation? The solution to [last year's competition](https://www.kaggle.com/c/cassava-disease/discussion/94114) only mentions architecture and augmentation.\n\nSo far, I have learned from the discussion board to:\n\n1. Prioritize Cross-Validation over Public Leader Board\n2. Test Time Augmentation is important\n3. Use an ensemble of multiple networks", "1124926": "Several resources needed to maximize rank.\n\n1. Put a complete profile in your account. Where your live, school attending, etc. Be sure to include the social networking links that Kaggle provides (github, linkedin, etc). If you were a computer science student at the same university I attended I might extend you an invite as team member. Being on a team that is willing to help you learn Kaggle is the #1 way for a novice to maximize rank.\n2. Put an ad for yourself in the team wanted post. Per #1 above, being on a team. Be sure to repeat some of what you said in this post. (It is likely you will however end up on a team that sucks so don't take the first offer until you research your potential team member).\n3. Be a super smart or plan to spend lots of hours.\n4. Check out most of the previous (last year or so) vision competitions. Sort shared notebooks by best score. Pick your product of choice (keras or pytorch) and look at top scoring. \n5. This competition has some very noisy data - reach back into your Deep Learning course materials.\n6. The short answer - to get in the top 10% you will need to do it all." }, "source": "meta" }