![\"https://i.ibb.co/sqK17hz/zebra-aug.gif\"](\"https://i.ibb.co/sqK17hz/zebra-aug.gif\")
I'm sharing my worked augmentation part here.
\nwave_new[:, 0:4096:2] = wave1[:, 0:4096:2]\nwave_new[:, 1:4096:2] = wave2[:, 1:4096:2]\nwave = wave[:, ::-1].copy()\nif np.random.uniform() > 0.5:\n wave1[0] = wave2[0]\nif np.random.uniform() > 0.5:\n wave1[1] = wave2[1]\nif np.random.uniform() > 0.5:\n wave1[2] = wave2[2]\nw0 = wave1[0].copy()\nwave1[0] = wave1[1]\nwave1[1] = w0\nI'm sharing my worked augmentation part here.
\nwave_new[:, 0:4096:2] = wave1[:, 0:4096:2]\nwave_new[:, 1:4096:2] = wave2[:, 1:4096:2]\nwave = wave[:, ::-1].copy()\nif np.random.uniform() > 0.5:\n wave1[0] = wave2[0]\nif np.random.uniform() > 0.5:\n wave1[1] = wave2[1]\nif np.random.uniform() > 0.5:\n wave1[2] = wave2[2]\nw0 = wave1[0].copy()\nwave1[0] = wave1[1]\nwave1[1] = w0\noh, i used the Malayan tapir instead:
\n if 1: \n index = torch.randperm(batch_size).cuda()\n wave1 = wave[index]\n\n for b in range(batch_size):\n if np.random.rand() >0.5:\n wave[b,:,:2048] = wave1[b,:,:2048]\ni should have try more stripes
", "rawMarkdown": "oh, i used the Malayan tapir instead:\n\n```\n if 1: \n index = torch.randperm(batch_size).cuda()\n wave1 = wave[index]\n\n for b in range(batch_size):\n if np.random.rand() >0.5:\n wave[b,:,:2048] = wave1[b,:,:2048]\n\n\n```\n\ni should have try more stripes", "votes": 5 }, { "id": 1529936, "postDate": "2021-09-30T19:18:41.150Z", "content": "here is the cqt of zebra augmented wave:
\n
comparison of zebra augmentation with others
\nred: no augmentation at all
\nblue: normal argumentation
\ngreen: normal + zebra argumentation
Note: the green curve will eventually oscillate if it is trained long enough and the model is large enough to overfit. If the current model doesn't overfits, we can try a larger model. This is Andrew Ng famous MLOps :
\ndata,model -->more data, larger model --> even more data, larger model …
![\"https://i.ibb.co/DLj0hyH/Selection-963.png\"](\"https://i.ibb.co/DLj0hyH/Selection-963.png\")
\n![\"https://i.ibb.co/XSyQ8kz/Selection-964.png\"](\"https://i.ibb.co/XSyQ8kz/Selection-964.png\")
explanation of oscillatory oberservation in overfitting can be found at
\nhttps://www.kaggle.com/c/g2net-gravitational-wave-detection/discussion/275476
\nhttps://www.kaggle.com/c/g2net-gravitational-wave-detection/discussion/269154
Thank you for these plots of experiments!
", "rawMarkdown": "Thank you for these plots of experiments!", "votes": 1 }, { "id": 1534575, "postDate": "2021-10-05T04:02:34.023Z", "content": "Thank you for the nice plots!
", "rawMarkdown": "Thank you for the nice plots!" } ] }, { "id": 1530196, "postDate": "2021-10-01T02:39:11.407Z", "content": "What is the intuition behind the Zebra Mixup? Do you mind sharing how the idea come up to you?
", "rawMarkdown": "What is the intuition behind the Zebra Mixup? Do you mind sharing how the idea come up to you?", "votes": 1, "replies": [ { "id": 1530398, "postDate": "2021-10-01T06:26:03.267Z", "content": "check this
\ngridmask data augmentation
\nhttps://arxiv.org/abs/2001.04086
![\"https://raw.githubusercontent.com/CrazyVertigo/awesome-data-augmentation/master/assets/GridMask.png\"](\"https://raw.githubusercontent.com/CrazyVertigo/awesome-data-augmentation/master/assets/GridMask.png\")
I am doing comparison experiments on zebra augmentation.
\nsee below.
(the effects are like dropout … you will see training loss increases, but overfitting is prevented. you will see the validation loss changes from U-shape to L-shape)
\ni use it as a noise generator
\n if noise is not None and np.random.rand() < 0.5: #random mix stripe\n wave[...,::2]=noise[...,::2]\n@richx86 We found this wave = wave1[:, 0:4096:2] got 0.80 on CV. So I just tried to mixup with it.
@hengck23 FYI, I used target = max(target1, target2)
check these as well:
\nImproved Mixed-Example Data Augmentation
\nhttps://arxiv.org/abs/1805.11272
![\"https://i.ibb.co/DwwLZ94/Selection-973.png\"](\"https://i.ibb.co/DwwLZ94/Selection-973.png\")
\n![\"https://i.ibb.co/dWV6YvH/Selection-974.png\"](\"https://i.ibb.co/dWV6YvH/Selection-974.png\")
I have read another paper that creates many images out of one by using, e.g.
\nnew image = image[:,::2]
\nnew image = image[:,::3]
\netc.
i forget the name of the paper.
\nthese may be useful for self-learning
", "rawMarkdown": "check these as well:\n\nImproved Mixed-Example Data Augmentation\nhttps://arxiv.org/abs/1805.11272\n\n\n\n\n\nI have read another paper that creates many images out of one by using, e.g.\nnew image = image[:,::2]\nnew image = image[:,::3]\netc.\n\ni forget the name of the paper.\n\nthese may be useful for self-learning", "votes": 5 }, { "id": 1530826, "postDate": "2021-10-01T13:04:02.700Z", "content": "@richx86 if your team did not use so much augmentation, there is a lot of room to improve with this. Please try out.
", "rawMarkdown": "@richx86 if your team did not use so much augmentation, there is a lot of room to improve with this. Please try out.", "votes": 1 }, { "id": 1534581, "postDate": "2021-10-05T04:16:22.413Z", "content": "\n\nWe found this wave = wave1[:, 0:4096:2] got 0.80 on CV. So I just tried to mixup with it.
\n
@onodera I assume this methodology can be generalized to other kinds of transformations, right?
\nI mostly work on 1D models and the augmentations are different, but we didn't think of a few augmentations here involving negative samples. Nicely done! Will definitely try later. Thanks!
\nDo you have a rough number for the total boost from augmentation? I saw diminishing effect from augmentation. For example, we have nice boost from vflip+gaussian noise, vflip+swap01 channels, but vflip+gaussian noise+swap01 channels doesn't further improve. Is it because there are too many augmentations and the augmented data is no longer resembling the original data? Do you see this diminishing/saturation behavior for 4(or more) composed augmentations?
", "rawMarkdown": "> We found this wave = wave1[:, 0:4096:2] got 0.80 on CV. So I just tried to mixup with it.\n\n@onodera I assume this methodology can be generalized to other kinds of transformations, right? \n\nI mostly work on 1D models and the augmentations are different, but we didn't think of a few augmentations here involving negative samples. Nicely done! Will definitely try later. Thanks!\n\nDo you have a rough number for the total boost from augmentation? I saw diminishing effect from augmentation. For example, we have nice boost from vflip+gaussian noise, vflip+swap01 channels, but vflip+gaussian noise+swap01 channels doesn't further improve. Is it because there are too many augmentations and the augmented data is no longer resembling the original data? Do you see this diminishing/saturation behavior for 4(or more) composed augmentations?", "votes": 1 }, { "id": 1534582, "postDate": "2021-10-05T04:17:14.660Z", "content": "@hengck23 Thanks! Your reply is very informative, as always.
", "rawMarkdown": "@hengck23 Thanks! Your reply is very informative, as always.", "votes": 1 }, { "id": 1536230, "postDate": "2021-10-06T14:38:53.033Z", "content": "@richx86 As you can see the plot of @hengck23, zebra is most powerful augmentation. I think if you use zebra, the uplift of other augmentation is very small.
\nRoughly,
@onodera I see. Thank you very much!
", "rawMarkdown": "@onodera I see. Thank you very much!", "votes": 1 } ] }, { "id": 1528828, "postDate": "2021-09-30T01:39:08.533Z", "content": "\n\nZebra Mixup
\n
LOL this is great. I wish I would have thought of doing that :-)
", "rawMarkdown": "> Zebra Mixup\n\nLOL this is great. I wish I would have thought of doing that :-)", "votes": 2 }, { "id": 1559910, "postDate": "2021-10-27T08:08:15.163Z", "content": "Hey All,
\nThank you all for taking part in our competition. The participation has been overwhelmingly positive. We are currently conducting a survey to gauge the demographic and outreach achieved. Kindly spare 2min and fill in this survey https://forms.gle/QP9L16niPexozyhu5.
\nThank you all,
\nRegards,
\nChris
thanks for sharing, may i ask the difference of score with and w/o these augmentations?
", "rawMarkdown": "thanks for sharing, may i ask the difference of score with and w/o these augmentations?", "replies": [ { "id": 1528873, "postDate": "2021-09-30T02:33:54.170Z", "content": "It's hard to describe, because it depends on other setup. For instance zebra is most powerful augmentation, so it decreases improvement of negative flip.
\nThat means we have 2**4 patterns.
oh, i used the Malayan tapir instead:
\n if 1: \n index = torch.randperm(batch_size).cuda()\n wave1 = wave[index]\n\n for b in range(batch_size):\n if np.random.rand() >0.5:\n wave[b,:,:2048] = wave1[b,:,:2048]\ni should have try more stripes
", "votes": 5, "replies": [] }, { "id": 1529936, "author_name": "hengck23", "author_url": "", "post_date": "2021-09-30T19:18:41.150000", "content": "here is the cqt of zebra augmented wave:
\n
comparison of zebra augmentation with others
\nred: no augmentation at all
\nblue: normal argumentation
\ngreen: normal + zebra argumentation
Note: the green curve will eventually oscillate if it is trained long enough and the model is large enough to overfit. If the current model doesn't overfits, we can try a larger model. This is Andrew Ng famous MLOps :
\ndata,model -->more data, larger model --> even more data, larger model …
\n
explanation of oscillatory oberservation in overfitting can be found at
\nhttps://www.kaggle.com/c/g2net-gravitational-wave-detection/discussion/275476
\nhttps://www.kaggle.com/c/g2net-gravitational-wave-detection/discussion/269154
Thank you for these plots of experiments!
", "votes": 1, "replies": [] }, { "id": 1534575, "author_name": "Richard Xing", "author_url": "", "post_date": "2021-10-05T04:02:34.023000", "content": "Thank you for the nice plots!
", "votes": 0, "replies": [] } ] }, { "id": 1530196, "author_name": "Richard Xing", "author_url": "", "post_date": "2021-10-01T02:39:11.407000", "content": "What is the intuition behind the Zebra Mixup? Do you mind sharing how the idea come up to you?
", "votes": 1, "replies": [ { "id": 1530398, "author_name": "hengck23", "author_url": "", "post_date": "2021-10-01T06:26:03.267000", "content": "check this
\ngridmask data augmentation
\nhttps://arxiv.org/abs/2001.04086
I am doing comparison experiments on zebra augmentation.
\nsee below.
(the effects are like dropout … you will see training loss increases, but overfitting is prevented. you will see the validation loss changes from U-shape to L-shape)
\ni use it as a noise generator
\n if noise is not None and np.random.rand() < 0.5: #random mix stripe\n wave[...,::2]=noise[...,::2]\n@richx86 We found this wave = wave1[:, 0:4096:2] got 0.80 on CV. So I just tried to mixup with it.
@hengck23 FYI, I used target = max(target1, target2)
check these as well:
\nImproved Mixed-Example Data Augmentation
\nhttps://arxiv.org/abs/1805.11272
\n
I have read another paper that creates many images out of one by using, e.g.
\nnew image = image[:,::2]
\nnew image = image[:,::3]
\netc.
i forget the name of the paper.
\nthese may be useful for self-learning
", "votes": 5, "replies": [] }, { "id": 1530826, "author_name": "ONODERA", "author_url": "", "post_date": "2021-10-01T13:04:02.700000", "content": "@richx86 if your team did not use so much augmentation, there is a lot of room to improve with this. Please try out.
", "votes": 1, "replies": [] }, { "id": 1534581, "author_name": "Richard Xing", "author_url": "", "post_date": "2021-10-05T04:16:22.413000", "content": "\n\nWe found this wave = wave1[:, 0:4096:2] got 0.80 on CV. So I just tried to mixup with it.
\n
@onodera I assume this methodology can be generalized to other kinds of transformations, right?
\nI mostly work on 1D models and the augmentations are different, but we didn't think of a few augmentations here involving negative samples. Nicely done! Will definitely try later. Thanks!
\nDo you have a rough number for the total boost from augmentation? I saw diminishing effect from augmentation. For example, we have nice boost from vflip+gaussian noise, vflip+swap01 channels, but vflip+gaussian noise+swap01 channels doesn't further improve. Is it because there are too many augmentations and the augmented data is no longer resembling the original data? Do you see this diminishing/saturation behavior for 4(or more) composed augmentations?
", "votes": 1, "replies": [] }, { "id": 1534582, "author_name": "Richard Xing", "author_url": "", "post_date": "2021-10-05T04:17:14.660000", "content": "@hengck23 Thanks! Your reply is very informative, as always.
", "votes": 1, "replies": [] }, { "id": 1536230, "author_name": "ONODERA", "author_url": "", "post_date": "2021-10-06T14:38:53.033000", "content": "@richx86 As you can see the plot of @hengck23, zebra is most powerful augmentation. I think if you use zebra, the uplift of other augmentation is very small.
\nRoughly,
@onodera I see. Thank you very much!
", "votes": 1, "replies": [] } ] }, { "id": 1528828, "author_name": "عثمان", "author_url": "", "post_date": "2021-09-30T01:39:08.533000", "content": "\n\nZebra Mixup
\n
LOL this is great. I wish I would have thought of doing that :-)
", "votes": 2, "replies": [] }, { "id": 1559910, "author_name": "ChristopherZerafa", "author_url": "", "post_date": "2021-10-27T08:08:15.163000", "content": "Hey All,
\nThank you all for taking part in our competition. The participation has been overwhelmingly positive. We are currently conducting a survey to gauge the demographic and outreach achieved. Kindly spare 2min and fill in this survey https://forms.gle/QP9L16niPexozyhu5.
\nThank you all,
\nRegards,
\nChris
thanks for sharing, may i ask the difference of score with and w/o these augmentations?
", "votes": 0, "replies": [ { "id": 1528873, "author_name": "ONODERA", "author_url": "", "post_date": "2021-09-30T02:33:54.170000", "content": "It's hard to describe, because it depends on other setup. For instance zebra is most powerful augmentation, so it decreases improvement of negative flip.
\nThat means we have 2**4 patterns.