I think it's probably the most effortless and dumbest solution posted. I lazily started doing something 5 days before the deadline mostly because I got triggered by the word \"Arcface\" and wanted to try it.
\n\nData Loading: For each image in my training dataset I randomly chose between site 1 and site 2, created 6-channel image from .png's and normalized with np.mean() and np.std()\nModel: Single 512x512 resnet50 from torchvision, adapted to work with 6 channels as in this kernel \nTraining: 50 epochs, same starting parameters as in the kernel above. it took maybe 8-9 hours with single 1080Ti. \nAugmentations: I used D4 group augmentations from albumentations library (HorizotalFlip, VerticalFlip, Transpose, RandomRotate90)\nValidation: stratified split by sirna's with 0.05 left for validation (very smart)\nPrediction: d4_tta from pytorch toolbelt sor each site, then average sites and apply leak from [this kernel] (https://www.kaggle.com/zaharch/keras-model-boosted-with-plates-leak/) (~40 min test prediction time on single 1080Ti).
\n\nAnd that's it. I obtained my LB solution in the first day (it was supposed to be just a baseline) and for the rest 4 days tried to do something smart with arcface, but it simply didn't work (even after applying batchnorm in the beginning I could not get >10% training accuracy).
\n\nWithout leak and TTA, my model scores 0.52. Now I wonder if I could achieve higher results with the same straightforward approach.
", "messages": [ { "id": "635432", "postDate": "09/27/2019 14:19:35", "content": "I think it's probably the most effortless and dumbest solution posted. I lazily started doing something 5 days before the deadline mostly because I got triggered by the word \"Arcface\" and wanted to try it.
\n\nData Loading: For each image in my training dataset I randomly chose between site 1 and site 2, created 6-channel image from .png's and normalized with np.mean() and np.std()\nModel: Single 512x512 resnet50 from torchvision, adapted to work with 6 channels as in this kernel \nTraining: 50 epochs, same starting parameters as in the kernel above. it took maybe 8-9 hours with single 1080Ti. \nAugmentations: I used D4 group augmentations from albumentations library (HorizotalFlip, VerticalFlip, Transpose, RandomRotate90)\nValidation: stratified split by sirna's with 0.05 left for validation (very smart)\nPrediction: d4_tta from pytorch toolbelt sor each site, then average sites and apply leak from [this kernel] (https://www.kaggle.com/zaharch/keras-model-boosted-with-plates-leak/) (~40 min test prediction time on single 1080Ti).
\n\nAnd that's it. I obtained my LB solution in the first day (it was supposed to be just a baseline) and for the rest 4 days tried to do something smart with arcface, but it simply didn't work (even after applying batchnorm in the beginning I could not get >10% training accuracy).
\n\nWithout leak and TTA, my model scores 0.52. Now I wonder if I could achieve higher results with the same straightforward approach.
", "rawMarkdown": "I think it's probably the most effortless and dumbest solution posted. I lazily started doing something 5 days before the deadline mostly because I got triggered by the word \"Arcface\" and wanted to try it.\n\n**Data Loading**: For each image in my training dataset I randomly chose between site 1 and site 2, created 6-channel image from .png's and normalized with np.mean() and np.std()\n**Model**: Single 512x512 resnet50 from torchvision, adapted to work with 6 channels as in this [kernel](https://www.kaggle.com/yhn112/resnet18-baseline-pytorch-ignite) \n**Training**: 50 epochs, same starting parameters as in the kernel above. it took maybe 8-9 hours with single 1080Ti. \n**Augmentations**: I used D4 group augmentations from [albumentations](https://github.com/albu/albumentations) library (HorizotalFlip, VerticalFlip, Transpose, RandomRotate90)\n**Validation**: stratified split by sirna's with 0.05 left for validation (very smart)\n**Prediction**: d4_tta from [pytorch toolbelt](https://github.com/BloodAxe/pytorch-toolbelt) sor each site, then average sites and apply leak from [this kernel] (https://www.kaggle.com/zaharch/keras-model-boosted-with-plates-leak/) (~40 min test prediction time on single 1080Ti).\n\nAnd that's it. I obtained my LB solution in the first day (it was supposed to be just a baseline) and for the rest 4 days tried to do something smart with arcface, but it simply didn't work (even after applying batchnorm in the beginning I could not get >10% training accuracy).\n\nWithout leak and TTA, my model scores 0.52. Now I wonder if I could achieve higher results with the same straightforward approach.", "votes": null }, { "id": "635456", "postDate": "09/27/2019 15:12:58", "content": "Congratulations\nGreat Write-Up\nThanks for Sharing your Approach & Insights…!! @devvindan
", "rawMarkdown": "Congratulations\nGreat Write-Up\nThanks for Sharing your Approach & Insights…!! @devvindan", "votes": null }, { "id": "635527", "postDate": "09/27/2019 18:59:50", "content": "great!
", "rawMarkdown": "great!", "votes": null }, { "id": "635587", "postDate": "09/27/2019 21:57:35", "content": "Nice solution!\nDid you have problems with model.eval() during validation and inference? I wasn't able to use it because my model's accuracy just drops to random guessing
I experienced the problem when my model accuracy was ~0 during model.eval(), but it happened simply because I forgot to normalize input
", "rawMarkdown": "I experienced the problem when my model accuracy was ~0 during model.eval(), but it happened simply because I forgot to normalize input", "votes": null } ], "comments": [ { "id": 635456, "author_name": "veeralakrishna", "author_url": "", "post_date": "09/27/2019 15:12:58", "content": "Congratulations\nGreat Write-Up\nThanks for Sharing your Approach & Insights…!! @devvindan
", "votes": null, "replies": [] }, { "id": 635527, "author_name": "krishnakatyal", "author_url": "", "post_date": "09/27/2019 18:59:50", "content": "great!
", "votes": null, "replies": [] }, { "id": 635587, "author_name": "rafailfridman", "author_url": "", "post_date": "09/27/2019 21:57:35", "content": "Nice solution!\nDid you have problems with model.eval() during validation and inference? I wasn't able to use it because my model's accuracy just drops to random guessing
I experienced the problem when my model accuracy was ~0 during model.eval(), but it happened simply because I forgot to normalize input
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "635432": "I think it's probably the most effortless and dumbest solution posted. I lazily started doing something 5 days before the deadline mostly because I got triggered by the word \"Arcface\" and wanted to try it.\n\n**Data Loading**: For each image in my training dataset I randomly chose between site 1 and site 2, created 6-channel image from .png's and normalized with np.mean() and np.std()\n**Model**: Single 512x512 resnet50 from torchvision, adapted to work with 6 channels as in this [kernel](https://www.kaggle.com/yhn112/resnet18-baseline-pytorch-ignite) \n**Training**: 50 epochs, same starting parameters as in the kernel above. it took maybe 8-9 hours with single 1080Ti. \n**Augmentations**: I used D4 group augmentations from [albumentations](https://github.com/albu/albumentations) library (HorizotalFlip, VerticalFlip, Transpose, RandomRotate90)\n**Validation**: stratified split by sirna's with 0.05 left for validation (very smart)\n**Prediction**: d4_tta from [pytorch toolbelt](https://github.com/BloodAxe/pytorch-toolbelt) sor each site, then average sites and apply leak from [this kernel] (https://www.kaggle.com/zaharch/keras-model-boosted-with-plates-leak/) (~40 min test prediction time on single 1080Ti).\n\nAnd that's it. I obtained my LB solution in the first day (it was supposed to be just a baseline) and for the rest 4 days tried to do something smart with arcface, but it simply didn't work (even after applying batchnorm in the beginning I could not get >10% training accuracy).\n\nWithout leak and TTA, my model scores 0.52. Now I wonder if I could achieve higher results with the same straightforward approach.", "635456": "Congratulations\nGreat Write-Up\nThanks for Sharing your Approach & Insights…!! @devvindan", "635527": "great!", "635587": "Nice solution!\nDid you have problems with `model.eval()` during validation and inference? I wasn't able to use it because my model's accuracy just drops to random guessing", "635599": "I experienced the problem when my model accuracy was ~0 during model.eval(), but it happened simply because I forgot to normalize input" }, "source": "meta" }