Setup:\n\nImage: 64x64 grayscale images\nValidation Split: 0.15\nModel: Resnet18\nEpochs: 30 epochs (Approximately 30 minutes to train on rtx2070 super)\nBatch Size: 256 \nOptimizer: Adam\nLearning Rate: 0.001\n\nResults (Validation Grapheme Recall):\n\n- Baseline (No augmentation): 0.88502\n- Flip: 0.88251 \n- Rotate: 0.89690\n- Shift: 0.90435\n- Scale: 0.89503\n- Cutout: 0.91844\n
Note: I only include dataset independent augmentation methods in this experiment. I haven't tried other methods like mixup or cutmix (coming soon).\nDataset independent augmentation methods only transform image without using other images in the dataset.
", "messages": [ { "id": "756736", "postDate": "02/26/2020 02:49:04", "content": "Setup:\n\nImage: 64x64 grayscale images\nValidation Split: 0.15\nModel: Resnet18\nEpochs: 30 epochs (Approximately 30 minutes to train on rtx2070 super)\nBatch Size: 256 \nOptimizer: Adam\nLearning Rate: 0.001\n\nResults (Validation Grapheme Recall):\n\n- Baseline (No augmentation): 0.88502\n- Flip: 0.88251 \n- Rotate: 0.89690\n- Shift: 0.90435\n- Scale: 0.89503\n- Cutout: 0.91844\n
Note: I only include dataset independent augmentation methods in this experiment. I haven't tried other methods like mixup or cutmix (coming soon).\nDataset independent augmentation methods only transform image without using other images in the dataset.
", "rawMarkdown": "Setup:\n```\nImage: 64x64 grayscale images\nValidation Split: 0.15\nModel: Resnet18\nEpochs: 30 epochs (Approximately 30 minutes to train on rtx2070 super)\nBatch Size: 256 \nOptimizer: Adam\nLearning Rate: 0.001\n```\nResults (Validation Grapheme Recall):\n```\n- Baseline (No augmentation): 0.88502\n- Flip: 0.88251 \n- Rotate: 0.89690\n- Shift: 0.90435\n- Scale: 0.89503\n- Cutout: 0.91844\n```\n\nNote: I only include dataset independent augmentation methods in this experiment. I haven't tried other methods like mixup or cutmix (coming soon).\nDataset independent augmentation methods only transform image without using other images in the dataset.", "votes": null }, { "id": "756769", "postDate": "02/26/2020 03:54:35", "content": "Cool.\nThis result is similar to my experiments.\nBtw, you only used Densenet121 with Cutout and ShiftScaleRotate?\nYou are placeholder, what is the difference between this experiment and your score?\ndo you know are there any magic?
", "rawMarkdown": "Cool.\nThis result is similar to my experiments.\nBtw, you only used Densenet121 with Cutout and ShiftScaleRotate?\nYou are placeholder, what is the difference between this experiment and your score?\ndo you know are there any magic?", "votes": null }, { "id": "756771", "postDate": "02/26/2020 03:59:39", "content": "This experiment helps me choose which augmentation methods to include in my augmentations mix. I used to believe people with score > 0.98 have some sort of magic. But there's no magic 😅. You just have to keep experimenting more and more.
", "rawMarkdown": "This experiment helps me choose which augmentation methods to include in my augmentations mix. I used to believe people with score > 0.98 have some sort of magic. But there's no magic 😅. You just have to keep experimenting more and more.", "votes": null }, { "id": "756773", "postDate": "02/26/2020 04:03:16", "content": "Thank you for replying =)\nyour reply motivate me!\nI keep experimenting!! =) =) =)
", "rawMarkdown": "Thank you for replying =)\nyour reply motivate me!\nI keep experimenting!! =) =) =)", "votes": null }, { "id": "756790", "postDate": "02/26/2020 04:42:34", "content": "you need some automatic way to search for best augmentation hyperprameters. you can check autoaugment and related papers (e.g. fast autoaugment or PBT autoaugment)
", "rawMarkdown": "you need some automatic way to search for best augmentation hyperprameters. you can check autoaugment and related papers (e.g. fast autoaugment or PBT autoaugment)", "votes": null }, { "id": "756792", "postDate": "02/26/2020 04:48:47", "content": "check this as well:\nAffinity and Diversity: Quantifying Mechanisms of Data Augmentation http://arxiv.org/abs/2002.08973
", "rawMarkdown": "check this as well:\nAffinity and Diversity: Quantifying Mechanisms of Data Augmentation http://arxiv.org/abs/2002.08973", "votes": null }, { "id": "757029", "postDate": "02/26/2020 11:13:22", "content": "@yuyuta Same for me. Been stuck @ bronze zone for nearly a month heading nowhere. No magic
", "rawMarkdown": "yuyuta Same for me. Been stuck @ bronze zone for nearly a month heading nowhere. No magic", "votes": null }, { "id": "757032", "postDate": "02/26/2020 11:14:45", "content": "@quandapro Just curious, you have not used mixup/cutmix for your current LB position? or is it just not included in the experiments. Yours is a very exceptional LB position in this case. Congrats :P
", "rawMarkdown": "quandapro Just curious, you have not used mixup/cutmix for your current LB position? or is it just not included in the experiments. Yours is a very exceptional LB position in this case. Congrats :P", "votes": null }, { "id": "757635", "postDate": "02/27/2020 01:13:20", "content": "Thanks, you too! \nI did not try mixup/cutmix. I would definately give it a shot!
", "rawMarkdown": "Thanks, you too! \nI did not try mixup/cutmix. I would definately give it a shot!", "votes": null } ], "comments": [ { "id": 756769, "author_name": "yuyuta", "author_url": "", "post_date": "02/26/2020 03:54:35", "content": "Cool.\nThis result is similar to my experiments.\nBtw, you only used Densenet121 with Cutout and ShiftScaleRotate?\nYou are placeholder, what is the difference between this experiment and your score?\ndo you know are there any magic?
", "votes": null, "replies": [ { "id": 756771, "author_name": "quandapro", "author_url": "", "post_date": "02/26/2020 03:59:39", "content": "This experiment helps me choose which augmentation methods to include in my augmentations mix. I used to believe people with score > 0.98 have some sort of magic. But there's no magic 😅. You just have to keep experimenting more and more.
", "votes": null, "replies": [] }, { "id": 756773, "author_name": "yuyuta", "author_url": "", "post_date": "02/26/2020 04:03:16", "content": "Thank you for replying =)\nyour reply motivate me!\nI keep experimenting!! =) =) =)
", "votes": null, "replies": [] }, { "id": 757029, "author_name": "roguekk007", "author_url": "", "post_date": "02/26/2020 11:13:22", "content": "@yuyuta Same for me. Been stuck @ bronze zone for nearly a month heading nowhere. No magic
", "votes": null, "replies": [] } ] }, { "id": 756790, "author_name": "hengck23", "author_url": "", "post_date": "02/26/2020 04:42:34", "content": "you need some automatic way to search for best augmentation hyperprameters. you can check autoaugment and related papers (e.g. fast autoaugment or PBT autoaugment)
", "votes": null, "replies": [] }, { "id": 756792, "author_name": "hengck23", "author_url": "", "post_date": "02/26/2020 04:48:47", "content": "check this as well:\nAffinity and Diversity: Quantifying Mechanisms of Data Augmentation http://arxiv.org/abs/2002.08973
", "votes": null, "replies": [] }, { "id": 757032, "author_name": "roguekk007", "author_url": "", "post_date": "02/26/2020 11:14:45", "content": "@quandapro Just curious, you have not used mixup/cutmix for your current LB position? or is it just not included in the experiments. Yours is a very exceptional LB position in this case. Congrats :P
", "votes": null, "replies": [ { "id": 757635, "author_name": "quandapro", "author_url": "", "post_date": "02/27/2020 01:13:20", "content": "Thanks, you too! \nI did not try mixup/cutmix. I would definately give it a shot!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "756736": "Setup:\n```\nImage: 64x64 grayscale images\nValidation Split: 0.15\nModel: Resnet18\nEpochs: 30 epochs (Approximately 30 minutes to train on rtx2070 super)\nBatch Size: 256 \nOptimizer: Adam\nLearning Rate: 0.001\n```\nResults (Validation Grapheme Recall):\n```\n- Baseline (No augmentation): 0.88502\n- Flip: 0.88251 \n- Rotate: 0.89690\n- Shift: 0.90435\n- Scale: 0.89503\n- Cutout: 0.91844\n```\n\nNote: I only include dataset independent augmentation methods in this experiment. I haven't tried other methods like mixup or cutmix (coming soon).\nDataset independent augmentation methods only transform image without using other images in the dataset.", "756769": "Cool.\nThis result is similar to my experiments.\nBtw, you only used Densenet121 with Cutout and ShiftScaleRotate?\nYou are placeholder, what is the difference between this experiment and your score?\ndo you know are there any magic?", "756771": "This experiment helps me choose which augmentation methods to include in my augmentations mix. I used to believe people with score > 0.98 have some sort of magic. But there's no magic 😅. You just have to keep experimenting more and more.", "756773": "Thank you for replying =)\nyour reply motivate me!\nI keep experimenting!! =) =) =)", "756790": "you need some automatic way to search for best augmentation hyperprameters. you can check autoaugment and related papers (e.g. fast autoaugment or PBT autoaugment)", "756792": "check this as well:\nAffinity and Diversity: Quantifying Mechanisms of Data Augmentation http://arxiv.org/abs/2002.08973", "757029": "yuyuta Same for me. Been stuck @ bronze zone for nearly a month heading nowhere. No magic", "757032": "quandapro Just curious, you have not used mixup/cutmix for your current LB position? or is it just not included in the experiments. Yours is a very exceptional LB position in this case. Congrats :P", "757635": "Thanks, you too! \nI did not try mixup/cutmix. I would definately give it a shot!" }, "source": "meta" }