Hi, everyone, I have tried ResNet18, ResNet34, ResNet50 with HPA external data.\nResNet50 performs best but it can only get 0.52 LB(without TTA/ensemble, leak). It has quite a big gap with the results in discussion threads, like results in https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/75640
\n\nHere is what I do:\n1. 512*512 + focal loss (I have also tried bce loss, but gives almost same result)\n2. oversample (from @brian)\n3. external data: download 4 channel jpg and process them by using methods in https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/75691#446302. i.e. straight copy of the corresponding channels, adding R+G for yellow\n4. read image by using opencv (BGR)\n5. use optimal threshold search on validation set\n6. data augmentation
\n\nHope someone can give me some hints about what I have been missing? How can I better train my model? Thanks.
", "messages": [ { "id": "450684", "postDate": "01/05/2019 14:24:07", "content": "Hi, everyone, I have tried ResNet18, ResNet34, ResNet50 with HPA external data.\nResNet50 performs best but it can only get 0.52 LB(without TTA/ensemble, leak). It has quite a big gap with the results in discussion threads, like results in https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/75640
\n\nHere is what I do:\n1. 512*512 + focal loss (I have also tried bce loss, but gives almost same result)\n2. oversample (from @brian)\n3. external data: download 4 channel jpg and process them by using methods in https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/75691#446302. i.e. straight copy of the corresponding channels, adding R+G for yellow\n4. read image by using opencv (BGR)\n5. use optimal threshold search on validation set\n6. data augmentation
\n\nHope someone can give me some hints about what I have been missing? How can I better train my model? Thanks.
", "rawMarkdown": "Hi, everyone, I have tried ResNet18, ResNet34, ResNet50 with HPA external data.\nResNet50 performs best but it can only get 0.52 LB(without TTA/ensemble, leak). It has quite a big gap with the results in discussion threads, like results in https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/75640\n\nHere is what I do:\n1. 512*512 + focal loss (I have also tried bce loss, but gives almost same result)\n2. oversample (from @brian)\n3. external data: download 4 channel jpg and process them by using methods in https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/75691#446302. i.e. straight copy of the corresponding channels, adding R+G for yellow\n4. read image by using opencv (BGR)\n5. use optimal threshold search on validation set\n6. data augmentation\n\nHope someone can give me some hints about what I have been missing? How can I better train my model? Thanks.", "votes": null }, { "id": "450846", "postDate": "01/05/2019 23:17:01", "content": "Did you try to threshold at 0.3? Because I was getting similar scores using thresholds from the validation and got a 0.02 boost using 0.3.
\n\nNot sure if it's a good idea to go with a threshold that performs better on LB instead of CV though
", "rawMarkdown": "Did you try to threshold at 0.3? Because I was getting similar scores using thresholds from the validation and got a 0.02 boost using 0.3.\n\nNot sure if it's a good idea to go with a threshold that performs better on LB instead of CV though", "votes": null }, { "id": "450877", "postDate": "01/06/2019 01:43:06", "content": "thanks for your reply. \nI have tried different flat threshold, but it decrease performance. I think that I must have missed something important.
", "rawMarkdown": "thanks for your reply. \nI have tried different flat threshold, but it decrease performance. I think that I must have missed something important.", "votes": null }, { "id": "450902", "postDate": "01/06/2019 03:21:41", "content": "Be careful about combining color channels, they are not really \"colors\" per se. \nSee these comments by Tilii, I found them very helpful
\n\nhttps://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/72895
", "rawMarkdown": "Be careful about combining color channels, they are not really \"colors\" per se. \nSee these comments by Tilii, I found them very helpful\n\nhttps://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/72895", "votes": null }, { "id": "450937", "postDate": "01/06/2019 05:21:41", "content": "why adding R+G for yellow since we have yellow channel?
", "rawMarkdown": "why adding R+G for yellow since we have yellow channel?", "votes": null } ], "comments": [ { "id": 450846, "author_name": "dr1t10", "author_url": "", "post_date": "01/05/2019 23:17:01", "content": "Did you try to threshold at 0.3? Because I was getting similar scores using thresholds from the validation and got a 0.02 boost using 0.3.
\n\nNot sure if it's a good idea to go with a threshold that performs better on LB instead of CV though
", "votes": null, "replies": [ { "id": 450877, "author_name": "yiwei2016", "author_url": "", "post_date": "01/06/2019 01:43:06", "content": "thanks for your reply. \nI have tried different flat threshold, but it decrease performance. I think that I must have missed something important.
", "votes": null, "replies": [] } ] }, { "id": 450902, "author_name": "bkosar1640", "author_url": "", "post_date": "01/06/2019 03:21:41", "content": "Be careful about combining color channels, they are not really \"colors\" per se. \nSee these comments by Tilii, I found them very helpful
\n\nhttps://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/72895
", "votes": null, "replies": [] }, { "id": 450937, "author_name": "zhangmiao", "author_url": "", "post_date": "01/06/2019 05:21:41", "content": "why adding R+G for yellow since we have yellow channel?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "450684": "Hi, everyone, I have tried ResNet18, ResNet34, ResNet50 with HPA external data.\nResNet50 performs best but it can only get 0.52 LB(without TTA/ensemble, leak). It has quite a big gap with the results in discussion threads, like results in https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/75640\n\nHere is what I do:\n1. 512*512 + focal loss (I have also tried bce loss, but gives almost same result)\n2. oversample (from @brian)\n3. external data: download 4 channel jpg and process them by using methods in https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/75691#446302. i.e. straight copy of the corresponding channels, adding R+G for yellow\n4. read image by using opencv (BGR)\n5. use optimal threshold search on validation set\n6. data augmentation\n\nHope someone can give me some hints about what I have been missing? How can I better train my model? Thanks.", "450846": "Did you try to threshold at 0.3? Because I was getting similar scores using thresholds from the validation and got a 0.02 boost using 0.3.\n\nNot sure if it's a good idea to go with a threshold that performs better on LB instead of CV though", "450877": "thanks for your reply. \nI have tried different flat threshold, but it decrease performance. I think that I must have missed something important.", "450902": "Be careful about combining color channels, they are not really \"colors\" per se. \nSee these comments by Tilii, I found them very helpful\n\nhttps://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/72895", "450937": "why adding R+G for yellow since we have yellow channel?" }, "source": "meta" }