I am a newbie to image classification competitions in Kaggle. Why is EfficientNet so popular in this competition? What is it about its architecture that is preferrable over other architectures like ResNet, ResNext, VGG etc?
", "messages": [ { "id": "951171", "postDate": "07/30/2020 00:43:37", "content": "I am a newbie to image classification competitions in Kaggle. Why is EfficientNet so popular in this competition? What is it about its architecture that is preferrable over other architectures like ResNet, ResNext, VGG etc?
", "rawMarkdown": "I am a newbie to image classification competitions in Kaggle. Why is EfficientNet so popular in this competition? What is it about its architecture that is preferrable over other architectures like ResNet, ResNext, VGG etc?", "votes": null }, { "id": "951205", "postDate": "07/30/2020 01:42:43", "content": "because it gives high number
", "rawMarkdown": "because it gives high number", "votes": null }, { "id": "951223", "postDate": "07/30/2020 02:08:47", "content": "It's scalable, so you can perform tests in small architecture and easily just switch from a smaller architecture with less parameters to a large one with more parameters. It's also quite efficient with it's resources so you would spend less time training when compared to non-efficient nets. In other words, it has high performance to parameter ratio.
", "rawMarkdown": "It's scalable, so you can perform tests in small architecture and easily just switch from a smaller architecture with less parameters to a large one with more parameters. It's also quite efficient with it's resources so you would spend less time training when compared to non-efficient nets. In other words, it has high performance to parameter ratio.", "votes": null }, { "id": "951285", "postDate": "07/30/2020 03:25:19", "content": "In my understanding, EfficientNet (released in Nov 2019) contains most of the innovations of the other networks like repeated convolutions (VGG19 2014), inception modules (Inception 2014), residual connections (ResNet 2015), Aggregation blocks (ResNext 2017), Squeeze and excite (SeNet and SeResNext 2017). And then EfficientNet's ratio of all these components has been optimized. So it's like the best of all networks.
", "rawMarkdown": "In my understanding, EfficientNet (released in Nov 2019) contains most of the innovations of the other networks like repeated convolutions (VGG19 2014), inception modules (Inception 2014), residual connections (ResNet 2015), Aggregation blocks (ResNext 2017), Squeeze and excite (SeNet and SeResNext 2017). And then EfficientNet's ratio of all these components has been optimized. So it's like the best of all networks.\n\n[1]: https://towardsdatascience.com/illustrated-10-cnn-architectures-95d78ace614d", "votes": null }, { "id": "951328", "postDate": "07/30/2020 04:25:04", "content": "I see. Thank you so much for replying @cdeotte . Will this always be the case, I wonder. Are there cases where EfficientNet may not perform as well as expected.
", "rawMarkdown": "I see. Thank you so much for replying @cdeotte . Will this always be the case, I wonder. Are there cases where EfficientNet may not perform as well as expected.", "votes": null }, { "id": "951330", "postDate": "07/30/2020 04:28:09", "content": "Thank you for your answer. I read up more on its benchmark performance. It is indeed impressive, higher accuracy all around with less parameters:
\n\nhttps://ai.googleblog.com/2019/05/efficientnet-improving-accuracy-and.html
", "rawMarkdown": "Thank you for your answer. I read up more on its benchmark performance. It is indeed impressive, higher accuracy all around with less parameters:\n\n https://ai.googleblog.com/2019/05/efficientnet-improving-accuracy-and.html", "votes": null }, { "id": "951332", "postDate": "07/30/2020 04:28:59", "content": "Short and sweet. Thanks! But more info would be useful :)
", "rawMarkdown": "Short and sweet. Thanks! But more info would be useful :)", "votes": null }, { "id": "951352", "postDate": "07/30/2020 04:52:40", "content": "So far, before EfficientNet, deciding the number of layers (depth), the number of neurons (width) of a layer in a neural net was an art. But EfficientNet came up with a logic to optimize these numbers, thus beating other hand-coded networks. I think this is going to be a goto model for CV problems, for some time to come, until a time when a new 'optimizing logic' gets published. I think gone are the days when you manually experiment with increasing width or depth.
", "rawMarkdown": "So far, before EfficientNet, deciding the number of layers (depth), the number of neurons (width) of a layer in a neural net was an art. But EfficientNet came up with a logic to optimize these numbers, thus beating other hand-coded networks. I think this is going to be a goto model for CV problems, for some time to come, until a time when a new 'optimizing logic' gets published. I think gone are the days when you manually experiment with increasing width or depth.", "votes": null }, { "id": "953844", "postDate": "08/01/2020 07:07:52", "content": "Yes. Why is it specifically used in this competition so much? I haven't seen this as frequent in any other competition.
", "rawMarkdown": "Yes. Why is it specifically used in this competition so much? I haven't seen this as frequent in any other competition.", "votes": null }, { "id": "954673", "postDate": "08/02/2020 01:49:09", "content": "What competitions have you seen another model more popular? The EfficientNet Arxiv paper was published in November 2019. Since then, I think EfficientNet has been winning all Kaggle image competitons.
", "rawMarkdown": "What competitions have you seen another model more popular? The EfficientNet Arxiv [paper][1] was published in November 2019. Since then, I think EfficientNet has been winning all Kaggle image competitons.\n\n[1]: https://arxiv.org/abs/1905.11946", "votes": null } ], "comments": [ { "id": 951205, "author_name": "moewie94", "author_url": "", "post_date": "07/30/2020 01:42:43", "content": "because it gives high number
", "votes": null, "replies": [ { "id": 951332, "author_name": "alanchn31", "author_url": "", "post_date": "07/30/2020 04:28:59", "content": "Short and sweet. Thanks! But more info would be useful :)
", "votes": null, "replies": [] } ] }, { "id": 951223, "author_name": "teeyee314", "author_url": "", "post_date": "07/30/2020 02:08:47", "content": "It's scalable, so you can perform tests in small architecture and easily just switch from a smaller architecture with less parameters to a large one with more parameters. It's also quite efficient with it's resources so you would spend less time training when compared to non-efficient nets. In other words, it has high performance to parameter ratio.
", "votes": null, "replies": [ { "id": 951330, "author_name": "alanchn31", "author_url": "", "post_date": "07/30/2020 04:28:09", "content": "Thank you for your answer. I read up more on its benchmark performance. It is indeed impressive, higher accuracy all around with less parameters:
\n\nhttps://ai.googleblog.com/2019/05/efficientnet-improving-accuracy-and.html
", "votes": null, "replies": [] } ] }, { "id": 951285, "author_name": "cdeotte", "author_url": "", "post_date": "07/30/2020 03:25:19", "content": "In my understanding, EfficientNet (released in Nov 2019) contains most of the innovations of the other networks like repeated convolutions (VGG19 2014), inception modules (Inception 2014), residual connections (ResNet 2015), Aggregation blocks (ResNext 2017), Squeeze and excite (SeNet and SeResNext 2017). And then EfficientNet's ratio of all these components has been optimized. So it's like the best of all networks.
", "votes": null, "replies": [ { "id": 951328, "author_name": "alanchn31", "author_url": "", "post_date": "07/30/2020 04:25:04", "content": "I see. Thank you so much for replying @cdeotte . Will this always be the case, I wonder. Are there cases where EfficientNet may not perform as well as expected.
", "votes": null, "replies": [] } ] }, { "id": 951352, "author_name": "krisho007", "author_url": "", "post_date": "07/30/2020 04:52:40", "content": "So far, before EfficientNet, deciding the number of layers (depth), the number of neurons (width) of a layer in a neural net was an art. But EfficientNet came up with a logic to optimize these numbers, thus beating other hand-coded networks. I think this is going to be a goto model for CV problems, for some time to come, until a time when a new 'optimizing logic' gets published. I think gone are the days when you manually experiment with increasing width or depth.
", "votes": null, "replies": [] }, { "id": 953844, "author_name": "jaseemck", "author_url": "", "post_date": "08/01/2020 07:07:52", "content": "Yes. Why is it specifically used in this competition so much? I haven't seen this as frequent in any other competition.
", "votes": null, "replies": [ { "id": 954673, "author_name": "cdeotte", "author_url": "", "post_date": "08/02/2020 01:49:09", "content": "What competitions have you seen another model more popular? The EfficientNet Arxiv paper was published in November 2019. Since then, I think EfficientNet has been winning all Kaggle image competitons.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "951171": "I am a newbie to image classification competitions in Kaggle. Why is EfficientNet so popular in this competition? What is it about its architecture that is preferrable over other architectures like ResNet, ResNext, VGG etc?", "951205": "because it gives high number", "951223": "It's scalable, so you can perform tests in small architecture and easily just switch from a smaller architecture with less parameters to a large one with more parameters. It's also quite efficient with it's resources so you would spend less time training when compared to non-efficient nets. In other words, it has high performance to parameter ratio.", "951285": "In my understanding, EfficientNet (released in Nov 2019) contains most of the innovations of the other networks like repeated convolutions (VGG19 2014), inception modules (Inception 2014), residual connections (ResNet 2015), Aggregation blocks (ResNext 2017), Squeeze and excite (SeNet and SeResNext 2017). And then EfficientNet's ratio of all these components has been optimized. So it's like the best of all networks.\n\n[1]: https://towardsdatascience.com/illustrated-10-cnn-architectures-95d78ace614d", "951328": "I see. Thank you so much for replying @cdeotte . Will this always be the case, I wonder. Are there cases where EfficientNet may not perform as well as expected.", "951330": "Thank you for your answer. I read up more on its benchmark performance. It is indeed impressive, higher accuracy all around with less parameters:\n\n https://ai.googleblog.com/2019/05/efficientnet-improving-accuracy-and.html", "951332": "Short and sweet. Thanks! But more info would be useful :)", "951352": "So far, before EfficientNet, deciding the number of layers (depth), the number of neurons (width) of a layer in a neural net was an art. But EfficientNet came up with a logic to optimize these numbers, thus beating other hand-coded networks. I think this is going to be a goto model for CV problems, for some time to come, until a time when a new 'optimizing logic' gets published. I think gone are the days when you manually experiment with increasing width or depth.", "953844": "Yes. Why is it specifically used in this competition so much? I haven't seen this as frequent in any other competition.", "954673": "What competitions have you seen another model more popular? The EfficientNet Arxiv [paper][1] was published in November 2019. Since then, I think EfficientNet has been winning all Kaggle image competitons.\n\n[1]: https://arxiv.org/abs/1905.11946" }, "source": "meta" }