{ "id": 106915, "title": "Does the computation resources are crucial in this competition?", "url": "/competitions/recursion-cellular-image-classification/discussion/106915", "author_name": "", "post_date": "2019-08-31T19:01:40.778736400Z", "votes": 6, "comment_count": 9, "views": 0, "content": "

Could you point might someone was able to pass 0.5 with constraints in computation resources. Your experience are more than welcome. Just wanna clarify how much effort I gonna put toward this.

", "messages": [ { "id": "614612", "postDate": "08/31/2019 19:01:40", "content": "

Could you point might someone was able to pass 0.5 with constraints in computation resources. Your experience are more than welcome. Just wanna clarify how much effort I gonna put toward this.

", "rawMarkdown": "Could you point might someone was able to pass 0.5 with constraints in computation resources. Your experience are more than welcome. Just wanna clarify how much effort I gonna put toward this.", "votes": null }, { "id": "614624", "postDate": "08/31/2019 19:43:13", "content": "

I only used a MacBook Pro and google cloud for my submissions !

", "rawMarkdown": "I only used a MacBook Pro and google cloud for my submissions !", "votes": null }, { "id": "614652", "postDate": "08/31/2019 20:57:17", "content": "

For example, ResNet34 training with un-frozen layer4 and last_linear layers with 6-channel tensors takes about 3-5 minutes per epoch on 1080Ti, depending on your batch size, augmentation methods, etc. So I would say that this competition is not too resources consuming, taking into account plenty of ​cloud providers and Kaggle kernels.

", "rawMarkdown": "For example, ResNet34 training with un-frozen `layer4` and `last_linear` layers with 6-channel tensors takes about 3-5 minutes per epoch on 1080Ti, depending on your batch size, augmentation methods, etc. So I would say that this competition is not *too* resources consuming, taking into account plenty of ​cloud providers and Kaggle kernels.", "votes": null }, { "id": "614848", "postDate": "09/01/2019 06:41:38", "content": "

I can prove you can pass 0.5 easily using kaggle kernels

", "rawMarkdown": "I can prove you can pass 0.5 easily using kaggle kernels", "votes": null }, { "id": "614953", "postDate": "09/01/2019 09:37:23", "content": "

Interesting! With or without using the 277 group leak?

", "rawMarkdown": "Interesting! With or without using the 277 group leak?", "votes": null }, { "id": "615073", "postDate": "09/01/2019 12:36:34", "content": "

only used kaggle kernels so far. takes some dedication though ;)

", "rawMarkdown": "only used kaggle kernels so far. takes some dedication though ;)", "votes": null }, { "id": "615487", "postDate": "09/02/2019 02:35:45", "content": "

Can you elaborate on the easy part of your comment @yaroshevskiy ? What's the general idea or approach you have taken to arrive at such high results with Kaggle Kernels only?

", "rawMarkdown": "Can you elaborate on the *easy* part of your comment @yaroshevskiy ? What's the general idea or approach you have taken to arrive at such high results with Kaggle Kernels only?", "votes": null }, { "id": "615700", "postDate": "09/02/2019 09:18:55", "content": "

@seesee With and without.\n@michelml \"easy\" means almost any cnn model can achieve that score if you understand experiment design properly (rxrx.ai)

", "rawMarkdown": "seesee With and without.\n@michelml \"easy\" means almost any cnn model can achieve that score if you understand experiment design properly (rxrx.ai)", "votes": null }, { "id": "616136", "postDate": "09/02/2019 18:17:10", "content": "

Thanks @yaroshevskiy , and what does \"easy\" mean in terms of training time with Kaggle Kernels only?

", "rawMarkdown": "Thanks @yaroshevskiy , and what does \"easy\" mean in terms of training time with Kaggle Kernels only?", "votes": null }, { "id": "616143", "postDate": "09/02/2019 18:30:22", "content": "

@michelml probably 2 or 3 kaggle kernel session to achieve good score

", "rawMarkdown": "michelml probably 2 or 3 kaggle kernel session to achieve good score", "votes": null } ], "comments": [ { "id": 614624, "author_name": "olaflegrand", "author_url": "", "post_date": "08/31/2019 19:43:13", "content": "

I only used a MacBook Pro and google cloud for my submissions !

", "votes": null, "replies": [] }, { "id": 614652, "author_name": "purplejester", "author_url": "", "post_date": "08/31/2019 20:57:17", "content": "

For example, ResNet34 training with un-frozen layer4 and last_linear layers with 6-channel tensors takes about 3-5 minutes per epoch on 1080Ti, depending on your batch size, augmentation methods, etc. So I would say that this competition is not too resources consuming, taking into account plenty of ​cloud providers and Kaggle kernels.

", "votes": null, "replies": [] }, { "id": 614848, "author_name": "yaroshevskiy", "author_url": "", "post_date": "09/01/2019 06:41:38", "content": "

I can prove you can pass 0.5 easily using kaggle kernels

", "votes": null, "replies": [ { "id": 614953, "author_name": "seesee", "author_url": "", "post_date": "09/01/2019 09:37:23", "content": "

Interesting! With or without using the 277 group leak?

", "votes": null, "replies": [] }, { "id": 615487, "author_name": "michelml", "author_url": "", "post_date": "09/02/2019 02:35:45", "content": "

Can you elaborate on the easy part of your comment @yaroshevskiy ? What's the general idea or approach you have taken to arrive at such high results with Kaggle Kernels only?

", "votes": null, "replies": [] }, { "id": 615700, "author_name": "yaroshevskiy", "author_url": "", "post_date": "09/02/2019 09:18:55", "content": "

@seesee With and without.\n@michelml \"easy\" means almost any cnn model can achieve that score if you understand experiment design properly (rxrx.ai)

", "votes": null, "replies": [] }, { "id": 616136, "author_name": "michelml", "author_url": "", "post_date": "09/02/2019 18:17:10", "content": "

Thanks @yaroshevskiy , and what does \"easy\" mean in terms of training time with Kaggle Kernels only?

", "votes": null, "replies": [] }, { "id": 616143, "author_name": "yaroshevskiy", "author_url": "", "post_date": "09/02/2019 18:30:22", "content": "

@michelml probably 2 or 3 kaggle kernel session to achieve good score

", "votes": null, "replies": [] } ] }, { "id": 615073, "author_name": "context", "author_url": "", "post_date": "09/01/2019 12:36:34", "content": "

only used kaggle kernels so far. takes some dedication though ;)

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "614612": "Could you point might someone was able to pass 0.5 with constraints in computation resources. Your experience are more than welcome. Just wanna clarify how much effort I gonna put toward this.", "614624": "I only used a MacBook Pro and google cloud for my submissions !", "614652": "For example, ResNet34 training with un-frozen `layer4` and `last_linear` layers with 6-channel tensors takes about 3-5 minutes per epoch on 1080Ti, depending on your batch size, augmentation methods, etc. So I would say that this competition is not *too* resources consuming, taking into account plenty of ​cloud providers and Kaggle kernels.", "614848": "I can prove you can pass 0.5 easily using kaggle kernels", "614953": "Interesting! With or without using the 277 group leak?", "615073": "only used kaggle kernels so far. takes some dedication though ;)", "615487": "Can you elaborate on the *easy* part of your comment @yaroshevskiy ? What's the general idea or approach you have taken to arrive at such high results with Kaggle Kernels only?", "615700": "seesee With and without.\n@michelml \"easy\" means almost any cnn model can achieve that score if you understand experiment design properly (rxrx.ai)", "616136": "Thanks @yaroshevskiy , and what does \"easy\" mean in terms of training time with Kaggle Kernels only?", "616143": "michelml probably 2 or 3 kaggle kernel session to achieve good score" }, "source": "meta" }