Based on Venturillo JE Kernel , I tried out Fine Tuning last 5 Layers of Xception Net along with Dense Layers , and also added Dropouts , This was done just for experimentation , I have published the kernel which can be found here , Any Suggestions to improve this Xception Net architecture are most welcome.
", "messages": [ { "id": 1036971, "postDate": "2020-10-04T13:07:20.597Z", "content": "Based on Venturillo JE Kernel , I tried out Fine Tuning last 5 Layers of Xception Net along with Dense Layers , and also added Dropouts , This was done just for experimentation , I have published the kernel which can be found here , Any Suggestions to improve this Xception Net architecture are most welcome.
", "rawMarkdown": "Based on Venturillo JE [Kernel](https://www.kaggle.com/sayedathar11/xceptionnet-finetuning) , I tried out Fine Tuning last 5 Layers of Xception Net along with Dense Layers , and also added Dropouts , This was done just for experimentation , I have published the kernel which can be found [here ](https://www.kaggle.com/sayedathar11/xceptionnet-finetuning) , Any Suggestions to improve this Xception Net architecture are most welcome.", "votes": 3 }, { "id": 1037629, "postDate": "2020-10-05T07:19:02.697Z", "content": "The first problem you want to fix for improvement is the label balance. For example pe_present_on_image only have 5% of values as positive and the remaining 95% is negative. Training with such imbalance will lead to a biased model (Calling false all the time since it leads to high accuracy, etc.). More or less, all other labels are imbalanced. You can try and fix them as a test (like selective training on equal positive and negative classes for each target).
Okay Venturillo JE will try that !
", "rawMarkdown": "Okay Venturillo JE will try that !" } ] } ], "comments": [ { "id": 1037629, "author_name": "Venturillo JE", "author_url": "", "post_date": "2020-10-05T07:19:02.697000", "content": "The first problem you want to fix for improvement is the label balance. For example pe_present_on_image only have 5% of values as positive and the remaining 95% is negative. Training with such imbalance will lead to a biased model (Calling false all the time since it leads to high accuracy, etc.). More or less, all other labels are imbalanced. You can try and fix them as a test (like selective training on equal positive and negative classes for each target).
Okay Venturillo JE will try that !
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "1036971": "Based on Venturillo JE [Kernel](https://www.kaggle.com/sayedathar11/xceptionnet-finetuning) , I tried out Fine Tuning last 5 Layers of Xception Net along with Dense Layers , and also added Dropouts , This was done just for experimentation , I have published the kernel which can be found [here ](https://www.kaggle.com/sayedathar11/xceptionnet-finetuning) , Any Suggestions to improve this Xception Net architecture are most welcome.", "1037629": "The first problem you want to fix for improvement is the label balance. For example `pe_present_on_image` only have 5% of values as positive and the remaining 95% is negative. Training with such imbalance will lead to a biased model (Calling false all the time since it leads to high accuracy, etc.). More or less, all other labels are imbalanced. You can try and fix them as a test (like selective training on equal positive and negative classes for each target)." } }