![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2383881%2F76cdb1d331eb7dee4f6c54e3dad4a328%2Floss.png?generation=1590761091889420&alt=media\")
Hi all, new to this. I've got my first time score. I have no idea is 0.48 a good or bad score for a baseline. What is the most recommended next step? I've only got a cross entropy loss like below:\n
Say it converges at epoch 6. Can I assume it's too easy to be overfitting?\nWhat else graph is recommended to print?\nThere are many tricks and models are suggested in Notebooks, but I don't know the priorities.\nI'm new to this stuff. So, any tip may be helpful.
\n\nIn baseline I've only taken efficientnet-b0 as my model, haven't done any tricks, augmentations, gleason_score..
", "messages": [ { "id": 866598, "postDate": "2020-05-29T14:29:01.833Z", "content": "Hi all, new to this. I've got my first time score. I have no idea is 0.48 a good or bad score for a baseline. What is the most recommended next step? I've only got a cross entropy loss like below:\n
Say it converges at epoch 6. Can I assume it's too easy to be overfitting?\nWhat else graph is recommended to print?\nThere are many tricks and models are suggested in Notebooks, but I don't know the priorities.\nI'm new to this stuff. So, any tip may be helpful.
\n\nIn baseline I've only taken efficientnet-b0 as my model, haven't done any tricks, augmentations, gleason_score..
", "rawMarkdown": "Hi all, new to this. I've got my first time score. I have no idea is 0.48 a good or bad score for a baseline. What is the most recommended next step? I've only got a cross entropy loss like below:\n\n\nSay it converges at epoch 6. Can I assume it's too easy to be overfitting?\nWhat else graph is recommended to print?\nThere are many tricks and models are suggested in Notebooks, but I don't know the priorities.\nI'm new to this stuff. So, any tip may be helpful.\n\nIn baseline I've only taken efficientnet-b0 as my model, haven't done any tricks, augmentations, gleason_score..", "votes": 2 }, { "id": 866841, "postDate": "2020-05-29T18:24:14.153Z", "content": "Why are you using a binary cross-entropy loss when you have more than 2 classes?
", "rawMarkdown": "Why are you using a binary cross-entropy loss when you have more than 2 classes?", "replies": [ { "id": 869526, "postDate": "2020-06-01T04:31:13.627Z", "content": "\"binary\" was mistyping.
", "rawMarkdown": "\"binary\" was mistyping." } ] }, { "id": 866696, "postDate": "2020-05-29T15:56:32.267Z", "content": "The most useful step is usually to check other people kernels and run through discussions. You can get to at least 0.80 with those.
", "rawMarkdown": "The most useful step is usually to check other people kernels and run through discussions. You can get to at least 0.80 with those.", "replies": [ { "id": 866715, "postDate": "2020-05-29T16:12:57.753Z", "content": "Thanks for suggesting, I've thought about this way, but having hard time to pick one of them. Instead of randomly pick a notebook, is there a good way to analyze for which one I should go first?
", "rawMarkdown": "Thanks for suggesting, I've thought about this way, but having hard time to pick one of them. Instead of randomly pick a notebook, is there a good way to analyze for which one I should go first?" }, { "id": 866867, "postDate": "2020-05-29T18:42:04.960Z", "content": "You can order them by votes or score. That's usually a good way to go. Also look for notebook with \"starter\" in title that use a framework you like. If you order by score though it can often be a fork of a notebook with a small tweak and no explanation. Usually the best notebooks have both good score and votes.
\n\nOnce you are more comfortable with the problem, look for niche notebooks that try different ideas.
", "rawMarkdown": "You can order them by votes or score. That's usually a good way to go. Also look for notebook with \"starter\" in title that use a framework you like. If you order by score though it can often be a fork of a notebook with a small tweak and no explanation. Usually the best notebooks have both good score and votes.\n\nOnce you are more comfortable with the problem, look for niche notebooks that try different ideas.", "votes": 1 }, { "id": 869486, "postDate": "2020-06-01T03:35:11.367Z", "content": "Thanks, Arnaud, that would be a nice introduction of using the community. I've got some clue. Voted for your comment.
", "rawMarkdown": "Thanks, Arnaud, that would be a nice introduction of using the community. I've got some clue. Voted for your comment." } ] } ], "comments": [ { "id": 866841, "author_name": "Pasquale", "author_url": "", "post_date": "2020-05-29T18:24:14.153000", "content": "Why are you using a binary cross-entropy loss when you have more than 2 classes?
", "votes": 0, "replies": [ { "id": 869526, "author_name": "Last Scene", "author_url": "", "post_date": "2020-06-01T04:31:13.627000", "content": "\"binary\" was mistyping.
", "votes": 0, "replies": [] } ] }, { "id": 866696, "author_name": "Arnaud Roussel", "author_url": "", "post_date": "2020-05-29T15:56:32.267000", "content": "The most useful step is usually to check other people kernels and run through discussions. You can get to at least 0.80 with those.
", "votes": 0, "replies": [ { "id": 866715, "author_name": "Last Scene", "author_url": "", "post_date": "2020-05-29T16:12:57.753000", "content": "Thanks for suggesting, I've thought about this way, but having hard time to pick one of them. Instead of randomly pick a notebook, is there a good way to analyze for which one I should go first?
", "votes": 0, "replies": [] }, { "id": 866867, "author_name": "Arnaud Roussel", "author_url": "", "post_date": "2020-05-29T18:42:04.960000", "content": "You can order them by votes or score. That's usually a good way to go. Also look for notebook with \"starter\" in title that use a framework you like. If you order by score though it can often be a fork of a notebook with a small tweak and no explanation. Usually the best notebooks have both good score and votes.
\n\nOnce you are more comfortable with the problem, look for niche notebooks that try different ideas.
", "votes": 1, "replies": [] }, { "id": 869486, "author_name": "Last Scene", "author_url": "", "post_date": "2020-06-01T03:35:11.367000", "content": "Thanks, Arnaud, that would be a nice introduction of using the community. I've got some clue. Voted for your comment.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "866598": "Hi all, new to this. I've got my first time score. I have no idea is 0.48 a good or bad score for a baseline. What is the most recommended next step? I've only got a cross entropy loss like below:\n\n\nSay it converges at epoch 6. Can I assume it's too easy to be overfitting?\nWhat else graph is recommended to print?\nThere are many tricks and models are suggested in Notebooks, but I don't know the priorities.\nI'm new to this stuff. So, any tip may be helpful.\n\nIn baseline I've only taken efficientnet-b0 as my model, haven't done any tricks, augmentations, gleason_score..", "866841": "Why are you using a binary cross-entropy loss when you have more than 2 classes?", "866696": "The most useful step is usually to check other people kernels and run through discussions. You can get to at least 0.80 with those." } }