In my kernel batch implementation of more data augmentations, you can find my batch implementation for CutMix, MixUp (see Chris Deotte's kernel) and GridMask -- see GridMask data augmentation with tensorflow (by Xie29)
", "messages": [ { "id": "765676", "postDate": "03/07/2020 00:24:35", "content": "In my kernel batch implementation of more data augmentations, you can find my batch implementation for CutMix, MixUp (see Chris Deotte's kernel) and GridMask -- see GridMask data augmentation with tensorflow (by Xie29)
", "rawMarkdown": "In my kernel [batch implementation of more data augmentations](https://www.kaggle.com/yihdarshieh/batch-implementation-of-more-data-augmentations?scriptVersionId=29767726), you can find my batch implementation for CutMix, MixUp (see Chris Deotte's kernel) and GridMask -- see [GridMask data augmentation with tensorflow (by Xie29)](https://www.kaggle.com/c/flower-classification-with-tpus/discussion/132986)", "votes": null }, { "id": "767877", "postDate": "03/10/2020 07:38:38", "content": "Great work. It looks like you sped up all three augmentations by 5x each. And now we can add them to tf.keras.layers() if we want. Thanks!
Hi, @cdeotte , I think it still slows down if we put the computation in layers, as you benchmarked before, because it is just batch form as i did for your rotation code. Also, the timing here is on cpu, without tf dataset iteration involved.
", "rawMarkdown": "Hi, @cdeotte , I think it still slows down if we put the computation in layers, as you benchmarked before, because it is just batch form as i did for your rotation code. Also, the timing here is on cpu, without tf dataset iteration involved.", "votes": null }, { "id": "768384", "postDate": "03/10/2020 18:16:34", "content": "Fantastic. Thank you for the contribution @yihdarshieh
", "rawMarkdown": "Fantastic. Thank you for the contribution @yihdarshieh", "votes": null }, { "id": "777783", "postDate": "03/17/2020 23:47:27", "content": "I checked the kernel and your implementation is great. I do have a question though. What would be the correct way to use your implementation of cutmix when used for training ?
", "rawMarkdown": "I checked the kernel and your implementation is great. I do have a question though. What would be the correct way to use your implementation of cutmix when used for training ?", "votes": null } ], "comments": [ { "id": 767877, "author_name": "cdeotte", "author_url": "", "post_date": "03/10/2020 07:38:38", "content": "Great work. It looks like you sped up all three augmentations by 5x each. And now we can add them to tf.keras.layers() if we want. Thanks!
Hi, @cdeotte , I think it still slows down if we put the computation in layers, as you benchmarked before, because it is just batch form as i did for your rotation code. Also, the timing here is on cpu, without tf dataset iteration involved.
", "votes": null, "replies": [] } ] }, { "id": 768384, "author_name": "mgorner", "author_url": "", "post_date": "03/10/2020 18:16:34", "content": "Fantastic. Thank you for the contribution @yihdarshieh
", "votes": null, "replies": [] }, { "id": 777783, "author_name": "ibrahimsherify", "author_url": "", "post_date": "03/17/2020 23:47:27", "content": "I checked the kernel and your implementation is great. I do have a question though. What would be the correct way to use your implementation of cutmix when used for training ?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "765676": "In my kernel [batch implementation of more data augmentations](https://www.kaggle.com/yihdarshieh/batch-implementation-of-more-data-augmentations?scriptVersionId=29767726), you can find my batch implementation for CutMix, MixUp (see Chris Deotte's kernel) and GridMask -- see [GridMask data augmentation with tensorflow (by Xie29)](https://www.kaggle.com/c/flower-classification-with-tpus/discussion/132986)", "767877": "Great work. It looks like you sped up all three augmentations by 5x each. And now we can add them to `tf.keras.layers()` if we want. Thanks!", "767913": "Hi, @cdeotte , I think it still slows down if we put the computation in layers, as you benchmarked before, because it is just batch form as i did for your rotation code. Also, the timing here is on cpu, without tf dataset iteration involved.", "768384": "Fantastic. Thank you for the contribution @yihdarshieh", "777783": "I checked the kernel and your implementation is great. I do have a question though. What would be the correct way to use your implementation of cutmix when used for training ?" }, "source": "meta" }