Thank team @sggpls @vzaguskin @igorkrashenyi for hard work
\nAt the beggining of the competition tried different approaches (use all channels, only 0,2,4, with vstack and without it). So, the best option was to use only 0,2,4 channels with vstack.
\nAlso had a lot of experiments with img size and found that bigger resolution can improve results (512x512 and 640x640).
\nAugmentation:
\nAlso, during the training we tried mixup and fmix, both perform pretty same. Also, tried to increase the signal during mixup, worked but not always.
\nThanks to all participants, organizers, and hosts for their work and to recover competition after the leak.
", "messages": [ { "id": 1480555, "postDate": "2021-08-19T04:43:59.450Z", "content": "Thank team @sggpls @vzaguskin @igorkrashenyi for hard work
\nAt the beggining of the competition tried different approaches (use all channels, only 0,2,4, with vstack and without it). So, the best option was to use only 0,2,4 channels with vstack.
\nAlso had a lot of experiments with img size and found that bigger resolution can improve results (512x512 and 640x640).
\nAugmentation:
\nAlso, during the training we tried mixup and fmix, both perform pretty same. Also, tried to increase the signal during mixup, worked but not always.
\nThanks to all participants, organizers, and hosts for their work and to recover competition after the leak.
", "rawMarkdown": "**Thank team** @sggpls @vzaguskin @igorkrashenyi for hard work\n\n# Preprocessing\n\nAt the beggining of the competition tried different approaches (use all channels, only 0,2,4, with vstack and without it). So, the best option was to use only 0,2,4 channels with vstack.\n\nAlso had a lot of experiments with img size and found that bigger resolution can improve results (512x512 and 640x640).\n\nAugmentation:\n\n- HorizontalFlip, VerticalFlip\n- ShiflScaleRotate\n- RandomResizedCrop (scale 0.9)\n- IAAAdditiveGaussianNoise (scale 0.15)\n\nAlso, during the training we tried mixup and fmix, both perform pretty same. Also, tried to increase the signal during mixup, worked but not always.\n\n# Train\n\n##### Worked\n\n1. Big models (efficientnet_b7, efficientnet_v2m, resnet200d, nfnet_f3 and etc.)\n2. Mixup\n3. Big resolution (640x640)\n4. Pseudo labels\n5. TTA4 during inference\n6. OOF Stacking\n\n##### Didn't work\n1. Shuffle channels as augmentation\n2. Siamese Net for splitting channels\n3. LSTM Architectures to incorporate on/off\n4. VOLO Models (hard to train, need to tune params)\n\n# LB simplified roadmap\n\n- b0 single 0.74+LB\n- b0 oof 0.77+ LB\n- v2m oof 0.775+ LB\n- blend 0.785+ LB\n- pseudo 0.795 + LB\n\n**Thanks to all participants, organizers, and hosts** for their work and to recover competition after the leak.", "votes": 27 }, { "id": 1482222, "postDate": "2021-08-20T00:43:45.223Z", "content": "Congratulations! When using pseudo labels did you use the predicted values as soft targets, or did you convert predictions into 1s and 0s, hard targets?
", "rawMarkdown": "Congratulations! When using pseudo labels did you use the predicted values as soft targets, or did you convert predictions into 1s and 0s, hard targets?", "replies": [ { "id": 1482930, "postDate": "2021-08-20T11:00:26.543Z", "content": "Tried both, but hard targets worked better
", "rawMarkdown": "Tried both, but hard targets worked better", "votes": 1 } ] }, { "id": 1481072, "postDate": "2021-08-19T09:45:36.033Z", "content": "Thank you for sharing. I learned a lot.
", "rawMarkdown": "Thank you for sharing. I learned a lot." } ], "comments": [ { "id": 1482222, "author_name": "Chris Deotte", "author_url": "", "post_date": "2021-08-20T00:43:45.223000", "content": "Congratulations! When using pseudo labels did you use the predicted values as soft targets, or did you convert predictions into 1s and 0s, hard targets?
", "votes": 0, "replies": [ { "id": 1482930, "author_name": "YuryBolkonsky", "author_url": "", "post_date": "2021-08-20T11:00:26.543000", "content": "Tried both, but hard targets worked better
", "votes": 1, "replies": [] } ] }, { "id": 1481072, "author_name": "WOOSUNG YOON", "author_url": "", "post_date": "2021-08-19T09:45:36.033000", "content": "Thank you for sharing. I learned a lot.
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1480555": "**Thank team** @sggpls @vzaguskin @igorkrashenyi for hard work\n\n# Preprocessing\n\nAt the beggining of the competition tried different approaches (use all channels, only 0,2,4, with vstack and without it). So, the best option was to use only 0,2,4 channels with vstack.\n\nAlso had a lot of experiments with img size and found that bigger resolution can improve results (512x512 and 640x640).\n\nAugmentation:\n\n- HorizontalFlip, VerticalFlip\n- ShiflScaleRotate\n- RandomResizedCrop (scale 0.9)\n- IAAAdditiveGaussianNoise (scale 0.15)\n\nAlso, during the training we tried mixup and fmix, both perform pretty same. Also, tried to increase the signal during mixup, worked but not always.\n\n# Train\n\n##### Worked\n\n1. Big models (efficientnet_b7, efficientnet_v2m, resnet200d, nfnet_f3 and etc.)\n2. Mixup\n3. Big resolution (640x640)\n4. Pseudo labels\n5. TTA4 during inference\n6. OOF Stacking\n\n##### Didn't work\n1. Shuffle channels as augmentation\n2. Siamese Net for splitting channels\n3. LSTM Architectures to incorporate on/off\n4. VOLO Models (hard to train, need to tune params)\n\n# LB simplified roadmap\n\n- b0 single 0.74+LB\n- b0 oof 0.77+ LB\n- v2m oof 0.775+ LB\n- blend 0.785+ LB\n- pseudo 0.795 + LB\n\n**Thanks to all participants, organizers, and hosts** for their work and to recover competition after the leak.", "1482222": "Congratulations! When using pseudo labels did you use the predicted values as soft targets, or did you convert predictions into 1s and 0s, hard targets?", "1481072": "Thank you for sharing. I learned a lot." } }