My code can be found here:\nhttps://github.com/evagian/Kaggle-Recursion-Cellular-Image-Classification
\n\nYour help will be appreciated :)
", "messages": [ { "id": 619448, "postDate": "2019-09-06T07:59:25.880Z", "content": "What do you mean by \"not converging?\" Train loss not decreasing, test accuracy not increasing, or train - test gap is increasing?
", "rawMarkdown": "What do you mean by \"not converging?\" Train loss not decreasing, test accuracy not increasing, or train - test gap is increasing?", "votes": 1 }, { "id": 632595, "postDate": "2019-09-23T19:29:57.327Z", "content": "You're using resnet50 pretrained on imagenet, but setting layer.trainable=False for all layers.
\n\nImagenet is sadly deficient in cellular photomicrographs (I know I've never photographed one while on safari...). So any pretrained model will have difficulty recognising the structures it needs to - for example, because the 6 channels have been arbitrarily color coded in your unpack script. In particular, to provide a useful signal for a siRNA classifier, your model will need to learn about the basic cell structure and disruptions/deformities that appear in this particular dataset.
\n\nOne other problem: I think you'll need a lot more than 5 epochs to start seeing results.
", "rawMarkdown": "You're using resnet50 pretrained on imagenet, but setting layer.trainable=False for all layers.\n\nImagenet is sadly deficient in cellular photomicrographs (I know I've never photographed one while on safari...). So any pretrained model will have difficulty recognising the structures it needs to - for example, because the 6 channels have been arbitrarily color coded in your unpack script. In particular, to provide a useful signal for a siRNA classifier, your model will need to learn about the basic cell structure and disruptions/deformities that appear in this particular dataset.\n\nOne other problem: I think you'll need a lot more than 5 epochs to start seeing results.", "votes": 2, "replies": [ { "id": 632627, "postDate": "2019-09-23T20:15:41.003Z", "content": "Nice comment. I don't think she is following this thread though.
", "rawMarkdown": "Nice comment. I don't think she is following this thread though.", "votes": 1 } ] }, { "id": 622993, "postDate": "2019-09-10T10:50:44.377Z", "content": "Lots of arguable things inside your code (e.g. a glance on your test result image)
\n\nProbably, you should check public kernels.
", "rawMarkdown": "Lots of arguable things inside your code (e.g. a glance on your test result image)\n\nProbably, you should check public kernels. ", "votes": 2 }, { "id": 618506, "postDate": "2019-09-05T08:57:54.050Z", "content": "My code can be found here:\nhttps://github.com/evagian/Kaggle-Recursion-Cellular-Image-Classification
\n\nYour help will be appreciated :)
", "rawMarkdown": "My code can be found here:\nhttps://github.com/evagian/Kaggle-Recursion-Cellular-Image-Classification\n\nYour help will be appreciated :)", "votes": 2 }, { "id": 629559, "postDate": "2019-09-18T23:00:23.933Z", "rawMarkdown": "", "votes": -15, "isDeleted": true } ], "comments": [ { "id": 619448, "author_name": "🐢 Jun Koda", "author_url": "", "post_date": "2019-09-06T07:59:25.880000", "content": "What do you mean by \"not converging?\" Train loss not decreasing, test accuracy not increasing, or train - test gap is increasing?
", "votes": 1, "replies": [] }, { "id": 632595, "author_name": "Snorkle", "author_url": "", "post_date": "2019-09-23T19:29:57.327000", "content": "You're using resnet50 pretrained on imagenet, but setting layer.trainable=False for all layers.
\n\nImagenet is sadly deficient in cellular photomicrographs (I know I've never photographed one while on safari...). So any pretrained model will have difficulty recognising the structures it needs to - for example, because the 6 channels have been arbitrarily color coded in your unpack script. In particular, to provide a useful signal for a siRNA classifier, your model will need to learn about the basic cell structure and disruptions/deformities that appear in this particular dataset.
\n\nOne other problem: I think you'll need a lot more than 5 epochs to start seeing results.
", "votes": 2, "replies": [ { "id": 632627, "author_name": "nosound", "author_url": "", "post_date": "2019-09-23T20:15:41.003000", "content": "Nice comment. I don't think she is following this thread though.
", "votes": 1, "replies": [] } ] }, { "id": 622993, "author_name": "DmitryKustikov", "author_url": "", "post_date": "2019-09-10T10:50:44.377000", "content": "Lots of arguable things inside your code (e.g. a glance on your test result image)
\n\nProbably, you should check public kernels.
", "votes": 2, "replies": [] }, { "id": 629559, "author_name": "", "author_url": "", "post_date": "2019-09-18T23:00:23.933000", "content": "", "votes": -15, "replies": [] } ], "raw_markdown_by_id": { "619448": "What do you mean by \"not converging?\" Train loss not decreasing, test accuracy not increasing, or train - test gap is increasing?", "632595": "You're using resnet50 pretrained on imagenet, but setting layer.trainable=False for all layers.\n\nImagenet is sadly deficient in cellular photomicrographs (I know I've never photographed one while on safari...). So any pretrained model will have difficulty recognising the structures it needs to - for example, because the 6 channels have been arbitrarily color coded in your unpack script. In particular, to provide a useful signal for a siRNA classifier, your model will need to learn about the basic cell structure and disruptions/deformities that appear in this particular dataset.\n\nOne other problem: I think you'll need a lot more than 5 epochs to start seeing results.", "622993": "Lots of arguable things inside your code (e.g. a glance on your test result image)\n\nProbably, you should check public kernels. ", "618506": "My code can be found here:\nhttps://github.com/evagian/Kaggle-Recursion-Cellular-Image-Classification\n\nYour help will be appreciated :)", "629559": "" } }