A fantastic analysis by Ross Wightman. Why new architectures aren't always better.
\n\nhttps://github.com/rwightman/pytorch-image-models/blob/master/notebooks/EffResNetComparison.ipynb
", "messages": [ { "id": "589559", "postDate": "08/01/2019 05:06:07", "content": "A fantastic analysis by Ross Wightman. Why new architectures aren't always better.
\n\nhttps://github.com/rwightman/pytorch-image-models/blob/master/notebooks/EffResNetComparison.ipynb
", "rawMarkdown": "A fantastic analysis by Ross Wightman. Why new architectures aren't always better. \n\nhttps://github.com/rwightman/pytorch-image-models/blob/master/notebooks/EffResNetComparison.ipynb", "votes": null }, { "id": "589876", "postDate": "08/01/2019 14:10:05", "content": "Very nice analysis work!! Thanks for sharing.
", "rawMarkdown": "Very nice analysis work!! Thanks for sharing.", "votes": null }, { "id": "589945", "postDate": "08/01/2019 15:36:56", "content": "Thanks for info👍
", "rawMarkdown": "Thanks for info👍", "votes": null }, { "id": "590222", "postDate": "08/01/2019 22:22:17", "content": "This is true but this analysis only applies to training from scratch (as mentioned here: https://forums.fast.ai/t/efficientnet/46978/90). Although I haven't used EfficientNets before, it seems based on some of the public kernels it is pretty decent for this competition.
", "rawMarkdown": "This is true but this analysis only applies to training from scratch (as mentioned here: https://forums.fast.ai/t/efficientnet/46978/90). Although I haven't used EfficientNets before, it seems based on some of the public kernels it is pretty decent for this competition.", "votes": null }, { "id": "590285", "postDate": "08/02/2019 02:32:33", "content": "Follow your message i decide to try the ig_resnext101_32*8d model
", "rawMarkdown": "Follow your message i decide to try the ig_resnext101_32*8d model", "votes": null }, { "id": "590593", "postDate": "08/02/2019 11:33:01", "content": "Very much true. Are you using classification or regression? I am only using classification.
", "rawMarkdown": "Very much true. Are you using classification or regression? I am only using classification.", "votes": null }, { "id": "590594", "postDate": "08/02/2019 11:33:15", "content": "Any success?
", "rawMarkdown": "Any success?", "votes": null }, { "id": "590595", "postDate": "08/02/2019 11:33:30", "content": "My pleasure. Are you using classification or regression?
", "rawMarkdown": "My pleasure. Are you using classification or regression?", "votes": null }, { "id": "590596", "postDate": "08/02/2019 11:33:46", "content": "It is not mine! :)
", "rawMarkdown": "It is not mine! :)", "votes": null }, { "id": "591973", "postDate": "08/04/2019 15:22:18", "content": "The effiency isn't just useful for train speed/inference, since there are much fewer parameters, your model is much less likely to overfit. Of course you can prevent overfitting with dropout, etc. but efficientnets are forgiving
", "rawMarkdown": "The effiency isn't just useful for train speed/inference, since there are much fewer parameters, your model is much less likely to overfit. Of course you can prevent overfitting with dropout, etc. but efficientnets are forgiving", "votes": null }, { "id": "593190", "postDate": "08/06/2019 09:26:09", "content": "Thanks for the notebook. I have a question about how does test time pooling work. Never heard of this technic before, and little info on the internet,
", "rawMarkdown": "Thanks for the notebook. I have a question about how does test time pooling work. Never heard of this technic before, and little info on the internet,", "votes": null }, { "id": "593200", "postDate": "08/06/2019 09:49:43", "content": "https://arxiv.org/pdf/1707.01629.pdf I found it at the last page of this papper. Maybe someone could try it instead of TTA.\nhttps://people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn.pdf And this one, converting CNN model to FCN
", "rawMarkdown": "https://arxiv.org/pdf/1707.01629.pdf I found it at the last page of this papper. Maybe someone could try it instead of TTA.\nhttps://people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn.pdf And this one, converting CNN model to FCN", "votes": null }, { "id": "593651", "postDate": "08/06/2019 23:46:29", "content": "Yes currently only classification
", "rawMarkdown": "Yes currently only classification", "votes": null }, { "id": "609945", "postDate": "08/28/2019 08:56:04", "content": "it works! thanks a lot
", "rawMarkdown": "it works! thanks a lot", "votes": null } ], "comments": [ { "id": 589876, "author_name": "suneelpatel", "author_url": "", "post_date": "08/01/2019 14:10:05", "content": "Very nice analysis work!! Thanks for sharing.
", "votes": null, "replies": [ { "id": 590596, "author_name": "solomonk", "author_url": "", "post_date": "08/02/2019 11:33:46", "content": "It is not mine! :)
", "votes": null, "replies": [] } ] }, { "id": 589945, "author_name": "vaishvik25", "author_url": "", "post_date": "08/01/2019 15:36:56", "content": "Thanks for info👍
", "votes": null, "replies": [ { "id": 590595, "author_name": "solomonk", "author_url": "", "post_date": "08/02/2019 11:33:30", "content": "My pleasure. Are you using classification or regression?
", "votes": null, "replies": [] } ] }, { "id": 590222, "author_name": "tanlikesmath", "author_url": "", "post_date": "08/01/2019 22:22:17", "content": "This is true but this analysis only applies to training from scratch (as mentioned here: https://forums.fast.ai/t/efficientnet/46978/90). Although I haven't used EfficientNets before, it seems based on some of the public kernels it is pretty decent for this competition.
", "votes": null, "replies": [ { "id": 590593, "author_name": "solomonk", "author_url": "", "post_date": "08/02/2019 11:33:01", "content": "Very much true. Are you using classification or regression? I am only using classification.
", "votes": null, "replies": [] }, { "id": 593651, "author_name": "tanlikesmath", "author_url": "", "post_date": "08/06/2019 23:46:29", "content": "Yes currently only classification
", "votes": null, "replies": [] } ] }, { "id": 590285, "author_name": "leixiang", "author_url": "", "post_date": "08/02/2019 02:32:33", "content": "Follow your message i decide to try the ig_resnext101_32*8d model
", "votes": null, "replies": [ { "id": 590594, "author_name": "solomonk", "author_url": "", "post_date": "08/02/2019 11:33:15", "content": "Any success?
", "votes": null, "replies": [] }, { "id": 609945, "author_name": "leixiang", "author_url": "", "post_date": "08/28/2019 08:56:04", "content": "it works! thanks a lot
", "votes": null, "replies": [] } ] }, { "id": 591973, "author_name": "sidhanthholalkere", "author_url": "", "post_date": "08/04/2019 15:22:18", "content": "The effiency isn't just useful for train speed/inference, since there are much fewer parameters, your model is much less likely to overfit. Of course you can prevent overfitting with dropout, etc. but efficientnets are forgiving
", "votes": null, "replies": [] }, { "id": 593190, "author_name": "ys19931006", "author_url": "", "post_date": "08/06/2019 09:26:09", "content": "Thanks for the notebook. I have a question about how does test time pooling work. Never heard of this technic before, and little info on the internet,
", "votes": null, "replies": [ { "id": 593200, "author_name": "ys19931006", "author_url": "", "post_date": "08/06/2019 09:49:43", "content": "https://arxiv.org/pdf/1707.01629.pdf I found it at the last page of this papper. Maybe someone could try it instead of TTA.\nhttps://people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn.pdf And this one, converting CNN model to FCN
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "589559": "A fantastic analysis by Ross Wightman. Why new architectures aren't always better. \n\nhttps://github.com/rwightman/pytorch-image-models/blob/master/notebooks/EffResNetComparison.ipynb", "589876": "Very nice analysis work!! Thanks for sharing.", "589945": "Thanks for info👍", "590222": "This is true but this analysis only applies to training from scratch (as mentioned here: https://forums.fast.ai/t/efficientnet/46978/90). Although I haven't used EfficientNets before, it seems based on some of the public kernels it is pretty decent for this competition.", "590285": "Follow your message i decide to try the ig_resnext101_32*8d model", "590593": "Very much true. Are you using classification or regression? I am only using classification.", "590594": "Any success?", "590595": "My pleasure. Are you using classification or regression?", "590596": "It is not mine! :)", "591973": "The effiency isn't just useful for train speed/inference, since there are much fewer parameters, your model is much less likely to overfit. Of course you can prevent overfitting with dropout, etc. but efficientnets are forgiving", "593190": "Thanks for the notebook. I have a question about how does test time pooling work. Never heard of this technic before, and little info on the internet,", "593200": "https://arxiv.org/pdf/1707.01629.pdf I found it at the last page of this papper. Maybe someone could try it instead of TTA.\nhttps://people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn.pdf And this one, converting CNN model to FCN", "593651": "Yes currently only classification", "609945": "it works! thanks a lot" }, "source": "meta" }