I am new to cv and this is my first cv competition. I am reading some public kernels about efficientnet, and I saw b3 and b4 were commonly used. I have a newbie question, why b3 or b4 are commonly used not b6 or b7? Is it chosen based on the validation score after trying all the versions or something else? Any help would be highly appreciated. Thanks!
", "messages": [ { "id": "1117958", "postDate": "12/18/2020 16:06:51", "content": "I am new to cv and this is my first cv competition. I am reading some public kernels about efficientnet, and I saw b3 and b4 were commonly used. I have a newbie question, why b3 or b4 are commonly used not b6 or b7? Is it chosen based on the validation score after trying all the versions or something else? Any help would be highly appreciated. Thanks!
", "rawMarkdown": "I am new to cv and this is my first cv competition. I am reading some public kernels about efficientnet, and I saw b3 and b4 were commonly used. I have a newbie question, why b3 or b4 are commonly used not b6 or b7? Is it chosen based on the validation score after trying all the versions or something else? Any help would be highly appreciated. Thanks!", "votes": null }, { "id": "1118072", "postDate": "12/18/2020 17:42:44", "content": "The number next to B represents the size of the model. So B4 is larger than B3. If some has to try a large model like B7 then they either need a GPU with larger memory or reduce the BS. it also takes lot of time to train the larger model. So you can end up seeing the larger models in some winning solutions and they mostly would have used it at the last few weeks.
", "rawMarkdown": "The number next to B represents the size of the model. So B4 is larger than B3. If some has to try a large model like B7 then they either need a GPU with larger memory or reduce the BS. it also takes lot of time to train the larger model. So you can end up seeing the larger models in some winning solutions and they mostly would have used it at the last few weeks.", "votes": null }, { "id": "1118258", "postDate": "12/18/2020 22:14:33", "content": "Right, to add some other reasons:
\nThanks for your replies!
", "rawMarkdown": "Thanks for your replies!", "votes": null }, { "id": "1119044", "postDate": "12/19/2020 17:08:10", "content": "\n\nI was in the same boat not very long ago. I did the MOOCs, I read the books. But I was entirely clueless on how to build a pipeline from where I load the data, do EDA, put it through some model, to create the submission file, and submit to get a score. I was entirely clueless.
\nHow I solved this problem, and what I suggest-
\n\n
\n- I took an active interest in closed competitions, looked at their data, read their descriptions, and so on.
\n- The most important step in ending my mystery was starting to look at others' Notebooks in those finished competitions. Since the competitions were finished, you could look at Notebooks which were really high on the Leaderboard. I could study winning solutions, and learn from them.
\n- Winners' notebooks will really seem overwhelming at first- nothing like you have seen in the MOOCs. But hold on to it. Copy winners' Notebooks and submit from that if you wish. Then learn by tweaking stuff in that notebook. I suggest that you copy stuff by hand, and try and understand each line. It will take a long time. But bear the pain. It will be worth it.
\n- Before doing all these, you have to choose a framework of your choice. In my case, it was PyTorch, and I studied through Notebooks written on PyTorch only. And sometimes the framework was fast.ai.
\n- Then dedicate some time to learn the framework of your choice.
\n- Actively monitor discussions and publicly available Notebooks related to this competition.
\nI think that will get you started.
\n
I wrote this in response to this post in Kaggle forums- First computer vision competition.
\nI hope it helps.
", "rawMarkdown": "> I was in the same boat not very long ago. I did the MOOCs, I read the books. But I was entirely clueless on how to build a pipeline from where I load the data, do EDA, put it through some model, to create the submission file, and submit to get a score. I was entirely clueless.\n\n> How I solved this problem, and what I suggest-\n* I took an active interest in closed competitions, looked at their data, read their descriptions, and so on.\n* The most important step in ending my mystery was starting *to look at others' Notebooks in those finished competitions*. Since the competitions were finished, you could look at Notebooks which were really high on the Leaderboard. I could study winning solutions, and learn from them.\n* Winners' notebooks will really seem overwhelming at first- nothing like you have seen in the MOOCs. But hold on to it. Copy winners' Notebooks and submit from that if you wish. Then learn by tweaking stuff in that notebook. I suggest that you copy stuff by hand, and try and understand each line. It will take a long time. But bear the pain. It will be worth it.\n* Before doing all these, you have to choose a framework of your choice. In my case, it was PyTorch, and I studied through Notebooks written on PyTorch only. And sometimes the framework was fast.ai.\n* Then dedicate some time to learn the framework of your choice.\n* Actively monitor discussions and publicly available Notebooks related to this competition.\n\n> I think that will get you started.\n\nI wrote this in response to this post in Kaggle forums- [First computer vision competition](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/204836#1116761).\n\nI hope it helps.", "votes": null }, { "id": "1119405", "postDate": "12/20/2020 02:51:59", "content": "Thanks a lot!😃
", "rawMarkdown": "Thanks a lot!😃", "votes": null } ], "comments": [ { "id": 1118072, "author_name": "vishnus", "author_url": "", "post_date": "12/18/2020 17:42:44", "content": "The number next to B represents the size of the model. So B4 is larger than B3. If some has to try a large model like B7 then they either need a GPU with larger memory or reduce the BS. it also takes lot of time to train the larger model. So you can end up seeing the larger models in some winning solutions and they mostly would have used it at the last few weeks.
", "votes": null, "replies": [ { "id": 1118258, "author_name": "aliabdin1", "author_url": "", "post_date": "12/18/2020 22:14:33", "content": "Right, to add some other reasons:
\nThanks for your replies!
", "votes": null, "replies": [] } ] }, { "id": 1119044, "author_name": "truthr", "author_url": "", "post_date": "12/19/2020 17:08:10", "content": "\n\nI was in the same boat not very long ago. I did the MOOCs, I read the books. But I was entirely clueless on how to build a pipeline from where I load the data, do EDA, put it through some model, to create the submission file, and submit to get a score. I was entirely clueless.
\nHow I solved this problem, and what I suggest-
\n\n
\n- I took an active interest in closed competitions, looked at their data, read their descriptions, and so on.
\n- The most important step in ending my mystery was starting to look at others' Notebooks in those finished competitions. Since the competitions were finished, you could look at Notebooks which were really high on the Leaderboard. I could study winning solutions, and learn from them.
\n- Winners' notebooks will really seem overwhelming at first- nothing like you have seen in the MOOCs. But hold on to it. Copy winners' Notebooks and submit from that if you wish. Then learn by tweaking stuff in that notebook. I suggest that you copy stuff by hand, and try and understand each line. It will take a long time. But bear the pain. It will be worth it.
\n- Before doing all these, you have to choose a framework of your choice. In my case, it was PyTorch, and I studied through Notebooks written on PyTorch only. And sometimes the framework was fast.ai.
\n- Then dedicate some time to learn the framework of your choice.
\n- Actively monitor discussions and publicly available Notebooks related to this competition.
\nI think that will get you started.
\n
I wrote this in response to this post in Kaggle forums- First computer vision competition.
\nI hope it helps.
", "votes": null, "replies": [ { "id": 1119405, "author_name": "josemori", "author_url": "", "post_date": "12/20/2020 02:51:59", "content": "Thanks a lot!😃
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1117958": "I am new to cv and this is my first cv competition. I am reading some public kernels about efficientnet, and I saw b3 and b4 were commonly used. I have a newbie question, why b3 or b4 are commonly used not b6 or b7? Is it chosen based on the validation score after trying all the versions or something else? Any help would be highly appreciated. Thanks!", "1118072": "The number next to B represents the size of the model. So B4 is larger than B3. If some has to try a large model like B7 then they either need a GPU with larger memory or reduce the BS. it also takes lot of time to train the larger model. So you can end up seeing the larger models in some winning solutions and they mostly would have used it at the last few weeks.", "1118258": "Right, to add some other reasons:\n- bigger models tend to overfit faster due to more parameters\n- inference also takes longer", "1118825": "Thanks for your replies!", "1119044": "> I was in the same boat not very long ago. I did the MOOCs, I read the books. But I was entirely clueless on how to build a pipeline from where I load the data, do EDA, put it through some model, to create the submission file, and submit to get a score. I was entirely clueless.\n\n> How I solved this problem, and what I suggest-\n* I took an active interest in closed competitions, looked at their data, read their descriptions, and so on.\n* The most important step in ending my mystery was starting *to look at others' Notebooks in those finished competitions*. Since the competitions were finished, you could look at Notebooks which were really high on the Leaderboard. I could study winning solutions, and learn from them.\n* Winners' notebooks will really seem overwhelming at first- nothing like you have seen in the MOOCs. But hold on to it. Copy winners' Notebooks and submit from that if you wish. Then learn by tweaking stuff in that notebook. I suggest that you copy stuff by hand, and try and understand each line. It will take a long time. But bear the pain. It will be worth it.\n* Before doing all these, you have to choose a framework of your choice. In my case, it was PyTorch, and I studied through Notebooks written on PyTorch only. And sometimes the framework was fast.ai.\n* Then dedicate some time to learn the framework of your choice.\n* Actively monitor discussions and publicly available Notebooks related to this competition.\n\n> I think that will get you started.\n\nI wrote this in response to this post in Kaggle forums- [First computer vision competition](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/204836#1116761).\n\nI hope it helps.", "1119405": "Thanks a lot!😃" }, "source": "meta" }