I am a beginner in this competition,I have noticed that many people are using EfficientNet or ResNext but I really want to implement VGG16 in this task.However I could not get a satisfying result.I don’t know why.I guess it’s because the imbalanced label.But why other cnn architectures like EfficientNet can work?Is there anyone have ideas about it?
", "messages": [ { "id": "1092734", "postDate": "11/27/2020 05:58:11", "content": "I am a beginner in this competition,I have noticed that many people are using EfficientNet or ResNext but I really want to implement VGG16 in this task.However I could not get a satisfying result.I don’t know why.I guess it’s because the imbalanced label.But why other cnn architectures like EfficientNet can work?Is there anyone have ideas about it?
", "rawMarkdown": "I am a beginner in this competition,I have noticed that many people are using EfficientNet or ResNext but I really want to implement VGG16 in this task.However I could not get a satisfying result.I don’t know why.I guess it’s because the imbalanced label.But why other cnn architectures like EfficientNet can work?Is there anyone have ideas about it?", "votes": null }, { "id": "1092770", "postDate": "11/27/2020 06:39:24", "content": "There is no magic like \"X works and Y does not work\". Why people use EfficientNet, there are reasons:
\nBut like I say there is no magic that model X works and model Y does not. Try different architecture, different hyperparameters, etc. I remember in the past I joined a competition when EfficientNet did not work and DenseNet worked.
\nSide note: make sure that your code is the reason it does not work first. For example, when predicting I used shuffle and got only 4x. I changed it and got me 8x now.
", "rawMarkdown": "There is no magic like \"X works and Y does not work\". Why people use EfficientNet, there are reasons:\n- Unlike VGG scaling method when you add more layers to the network, EfficientNet scales both depth, resolution and width (aka compound scaling). Using this compound method, in the paper it showed very promising results compared to scale using only one method.\n- Scaling using the above method is easier to do that to change the CNN architecture.\n- EffcientNet use Mobile Conv Block, which was shown to be much more effective than VGG's normal Conv2D block.\n\nBut like I say there is no magic that model X works and model Y does not. Try different architecture, different hyperparameters, etc. I remember in the past I joined a competition when EfficientNet did not work and DenseNet worked.\n\nSide note: make sure that your code is the reason it does not work first. For example, when predicting I used shuffle and got only 4x. I changed it and got me 8x now.", "votes": null }, { "id": "1092844", "postDate": "11/27/2020 08:02:12", "content": "Thanks a lot！I will check my code carefully.
", "rawMarkdown": "Thanks a lot！I will check my code carefully.", "votes": null }, { "id": "1092852", "postDate": "11/27/2020 08:13:11", "content": "Also I think every model suffers the imbalance problem. There must be other tips & tricks to tackle this, and I think that depends on our experience.
", "rawMarkdown": "Also I think every model suffers the imbalance problem. There must be other tips & tricks to tackle this, and I think that depends on our experience.", "votes": null } ], "comments": [ { "id": 1092770, "author_name": "aeryss", "author_url": "", "post_date": "11/27/2020 06:39:24", "content": "There is no magic like \"X works and Y does not work\". Why people use EfficientNet, there are reasons:
\nBut like I say there is no magic that model X works and model Y does not. Try different architecture, different hyperparameters, etc. I remember in the past I joined a competition when EfficientNet did not work and DenseNet worked.
\nSide note: make sure that your code is the reason it does not work first. For example, when predicting I used shuffle and got only 4x. I changed it and got me 8x now.
", "votes": null, "replies": [ { "id": 1092844, "author_name": "jamestigers", "author_url": "", "post_date": "11/27/2020 08:02:12", "content": "Thanks a lot！I will check my code carefully.
", "votes": null, "replies": [] }, { "id": 1092852, "author_name": "aeryss", "author_url": "", "post_date": "11/27/2020 08:13:11", "content": "Also I think every model suffers the imbalance problem. There must be other tips & tricks to tackle this, and I think that depends on our experience.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1092734": "I am a beginner in this competition,I have noticed that many people are using EfficientNet or ResNext but I really want to implement VGG16 in this task.However I could not get a satisfying result.I don’t know why.I guess it’s because the imbalanced label.But why other cnn architectures like EfficientNet can work?Is there anyone have ideas about it?", "1092770": "There is no magic like \"X works and Y does not work\". Why people use EfficientNet, there are reasons:\n- Unlike VGG scaling method when you add more layers to the network, EfficientNet scales both depth, resolution and width (aka compound scaling). Using this compound method, in the paper it showed very promising results compared to scale using only one method.\n- Scaling using the above method is easier to do that to change the CNN architecture.\n- EffcientNet use Mobile Conv Block, which was shown to be much more effective than VGG's normal Conv2D block.\n\nBut like I say there is no magic that model X works and model Y does not. Try different architecture, different hyperparameters, etc. I remember in the past I joined a competition when EfficientNet did not work and DenseNet worked.\n\nSide note: make sure that your code is the reason it does not work first. For example, when predicting I used shuffle and got only 4x. I changed it and got me 8x now.", "1092844": "Thanks a lot！I will check my code carefully.", "1092852": "Also I think every model suffers the imbalance problem. There must be other tips & tricks to tackle this, and I think that depends on our experience." }, "source": "meta" }