Hi, I am a newone to machine learning. Here is a question that I can't figure out and hope someone could share their experiences. I trained the 11 classes with 11 labels, so that the code return the inference of one class at a time. It looks normal when choose 'ETT - Normal', which give the accuracy around 0.91 after 60 epochs, however when I executed 'CVC - Borderline', my code gave no help on raising the accuracy beyond expectation number and val_acc didnt change at all.(The model returns only 0 in fact, so it might be c``aused by gradient vanishing??)
\nHere is the code :)
\ntargetCol==2 for 'ETT - Normal'
\ntargetCol==8 for 'CVC - Borderline'
\nhttps://github.com/shengwenyuan/Ranzcr
Hi, I am a newone to machine learning. Here is a question that I can't figure out and hope someone could share their experiences. I trained the 11 classes with 11 labels, so that the code return the inference of one class at a time. It looks normal when choose 'ETT - Normal', which give the accuracy around 0.91 after 60 epochs, however when I executed 'CVC - Borderline', my code gave no help on raising the accuracy beyond expectation number and val_acc didnt change at all.(The model returns only 0 in fact, so it might be c``aused by gradient vanishing??)
\nHere is the code :)
\ntargetCol==2 for 'ETT - Normal'
\ntargetCol==8 for 'CVC - Borderline'
\nhttps://github.com/shengwenyuan/Ranzcr
You need to provide more detail for the error.
", "rawMarkdown": "You need to provide more detail for the error.", "votes": null }, { "id": "1215790", "postDate": "02/24/2021 00:28:16", "content": "It's possible you need ReLU() and BatchNorm2d() layers after your initial convolution and before the resnet. Resnets, like most image models, expect their input to be within a certain range. This is range is typically 0-1, then normalised with mean = [0.485, 0.456, 0.406], std = [0.229, 0.224, 0.225].
If you have a raw convolution layer before the resnet, it could be receiving input values that are far too large/small. Personally, I do not recommend putting an additional convolution before the resnet input. If you want 3-channel input then you can either load the images as RGB or use feature engineering to fill the other 2 channels with something.
", "rawMarkdown": "It's possible you need `ReLU()` and `BatchNorm2d()` layers after your initial convolution and before the resnet. Resnets, like most image models, expect their input to be within a certain range. This is range is typically 0-1, then normalised with `mean = [0.485, 0.456, 0.406], std = [0.229, 0.224, 0.225]`.\n\nIf you have a raw convolution layer before the resnet, it could be receiving input values that are far too large/small. Personally, I do not recommend putting an additional convolution before the resnet input. If you want 3-channel input then you can either load the images as RGB or use feature engineering to fill the other 2 channels with something.", "votes": null } ], "comments": [ { "id": 1215664, "author_name": "amritpal333", "author_url": "", "post_date": "02/23/2021 21:00:01", "content": "You need to provide more detail for the error.
", "votes": null, "replies": [] }, { "id": 1215790, "author_name": "bigironsphere", "author_url": "", "post_date": "02/24/2021 00:28:16", "content": "It's possible you need ReLU() and BatchNorm2d() layers after your initial convolution and before the resnet. Resnets, like most image models, expect their input to be within a certain range. This is range is typically 0-1, then normalised with mean = [0.485, 0.456, 0.406], std = [0.229, 0.224, 0.225].
If you have a raw convolution layer before the resnet, it could be receiving input values that are far too large/small. Personally, I do not recommend putting an additional convolution before the resnet input. If you want 3-channel input then you can either load the images as RGB or use feature engineering to fill the other 2 channels with something.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1200467": "Hi, I am a newone to machine learning. Here is a question that I can't figure out and hope someone could share their experiences. I trained the 11 classes with 11 labels, so that the code return the inference of one class at a time. It looks normal when choose 'ETT - Normal', which give the accuracy around 0.91 after 60 epochs, however when I executed 'CVC - Borderline', my code gave no help on raising the accuracy beyond expectation number and val_acc didnt change at all.(The model returns only 0 in fact, so it might be c``aused by gradient vanishing??)\n\nHere is the code :)\ntargetCol==2 for 'ETT - Normal'\ntargetCol==8 for 'CVC - Borderline'\n[https://github.com/shengwenyuan/Ranzcr](url)", "1215664": "You need to provide more detail for the error.", "1215790": "It's possible you need `ReLU()` and `BatchNorm2d()` layers after your initial convolution and before the resnet. Resnets, like most image models, expect their input to be within a certain range. This is range is typically 0-1, then normalised with `mean = [0.485, 0.456, 0.406], std = [0.229, 0.224, 0.225]`.\n\nIf you have a raw convolution layer before the resnet, it could be receiving input values that are far too large/small. Personally, I do not recommend putting an additional convolution before the resnet input. If you want 3-channel input then you can either load the images as RGB or use feature engineering to fill the other 2 channels with something." }, "source": "meta" }