{ "id": 174494, "title": "Augmentation Methods", "url": "/competitions/birdsong-recognition/discussion/174494", "author_name": "", "post_date": "2020-08-13T19:06:37.393850700Z", "votes": 15, "comment_count": 6, "views": 0, "content": "

In additional to topic https://www.kaggle.com/c/birdsong-recognition/discussion/170913 with great visualizations I want to add some augmentation ideas with corresponding implementations.

\n

Audiomentations

\n

Great audio implementations in one of github repositories (link below): you can see about 20 augmentation techniques with short descriptions in front of each class. Some of those:

\n\n

code: https://github.com/iver56/audiomentations/blob/master/audiomentations/augmentations/transforms.py

\n

SpecAugment

\n

SpecAugment consists of three consecutive augmenatation methods:

\n\n

\"\"

\n

paper: https://arxiv.org/pdf/1904.08779.pdf.
\ncode: we've already had an implementation of this method inside one of a public kernels - https://www.kaggle.com/hidehisaarai1213/introduction-to-sound-event-detection#Train-SED-model-with-only-weak-supervision (another thanks for great kernel!).

\n

MixAugment

\n

The main contribution of this paper is an idea to use more than one sample inside an augmentation technique (but more than two ones don't increase a target metrics). For two samples (x1, y1) and (x2, y2) we can do it next way:

\n
x = lambda * x1  + (1 - lambda) * x2\ny = lambda * y1  + (1 - lambda) * y2\n
\n

It called mixups, but there are lots of other techniques of augmentations including two samples (see the paper for a details):

\n

\"\"

\n

paper: https://arxiv.org/pdf/1805.11272.pdf.
\ncode: https://github.com/ceciliaresearch/MixedExample/blob/master/mixed_example.py

\n

How change the final set of augmentations?

\n

In a last paper autors just applyed each augmentation separetly and saw on results:

\n

\"\"

\n

Top of these methods can be chosen as a final set of augmentation methods and can be applied to the train set together.

", "messages": [ { "id": "969548", "postDate": "08/13/2020 19:06:37", "content": "

In additional to topic https://www.kaggle.com/c/birdsong-recognition/discussion/170913 with great visualizations I want to add some augmentation ideas with corresponding implementations.

\n

Audiomentations

\n

Great audio implementations in one of github repositories (link below): you can see about 20 augmentation techniques with short descriptions in front of each class. Some of those:

\n\n

code: https://github.com/iver56/audiomentations/blob/master/audiomentations/augmentations/transforms.py

\n

SpecAugment

\n

SpecAugment consists of three consecutive augmenatation methods:

\n\n

\"\"

\n

paper: https://arxiv.org/pdf/1904.08779.pdf.
\ncode: we've already had an implementation of this method inside one of a public kernels - https://www.kaggle.com/hidehisaarai1213/introduction-to-sound-event-detection#Train-SED-model-with-only-weak-supervision (another thanks for great kernel!).

\n

MixAugment

\n

The main contribution of this paper is an idea to use more than one sample inside an augmentation technique (but more than two ones don't increase a target metrics). For two samples (x1, y1) and (x2, y2) we can do it next way:

\n
x = lambda * x1  + (1 - lambda) * x2\ny = lambda * y1  + (1 - lambda) * y2\n
\n

It called mixups, but there are lots of other techniques of augmentations including two samples (see the paper for a details):

\n

\"\"

\n

paper: https://arxiv.org/pdf/1805.11272.pdf.
\ncode: https://github.com/ceciliaresearch/MixedExample/blob/master/mixed_example.py

\n

How change the final set of augmentations?

\n

In a last paper autors just applyed each augmentation separetly and saw on results:

\n

\"\"

\n

Top of these methods can be chosen as a final set of augmentation methods and can be applied to the train set together.

", "rawMarkdown": "In additional to topic https://www.kaggle.com/c/birdsong-recognition/discussion/170913 with great visualizations I want to add some augmentation ideas with corresponding implementations.\n\n**Audiomentations**\n\nGreat audio implementations in one of github repositories (link below): you can see about 20 augmentation techniques with short descriptions in front of each class. Some of those:\n - *AddImpulseResponse, AddGaussianNoise, AddBackgroundNoise, AddShortNoises*.\n - *TimeMask, FrequencyMask*.\n - *Shift, PitchShift, TimeStretch*.\n\ncode: https://github.com/iver56/audiomentations/blob/master/audiomentations/augmentations/transforms.py\n\n**SpecAugment**\n\n SpecAugment consists of three consecutive augmenatation methods:\n\n - *Time warp*: exchange some tuple of points in a row audio signal.\n - *Frequency masking*: mask some horizontal line inside a mel-spectrogram.\n - *Time masking*: mask some vertical line inside a mel-spectrogram.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1659719%2F7a00e5b29b6ecc1c3e56da9b7cbd3012%2FSpecAugmentations.png?generation=1597342806077272&alt=media)\n\npaper: https://arxiv.org/pdf/1904.08779.pdf.\ncode: we've already had an implementation of this method inside one of a public kernels - https://www.kaggle.com/hidehisaarai1213/introduction-to-sound-event-detection#Train-SED-model-with-only-weak-supervision (another thanks for great kernel!).\n\n**MixAugment**\n\nThe main contribution of this paper is an idea to use more than one sample inside an augmentation technique (but more than two ones don't increase a target metrics). For two samples `(x1, y1)` and `(x2, y2)` we can do it next way:\n\n\n```\nx = lambda * x1 + (1 - lambda) * x2\ny = lambda * y1 + (1 - lambda) * y2\n```\n\nIt called *mixups*, but there are lots of other techniques of augmentations including two samples (see the paper for a details):\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1659719%2Fdc1510aa28daabe7651d42eed02a2cfe%2FMixAugment.png?generation=1597344047630223&alt=media)\n\npaper: https://arxiv.org/pdf/1805.11272.pdf. \ncode: https://github.com/ceciliaresearch/MixedExample/blob/master/mixed_example.py\n\n**How change the final set of augmentations?**\n\nIn a last paper autors just applyed each augmentation separetly and saw on results:\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1659719%2F57e2abe4418ca8046f0c411fdf00b572%2FAugmentResults.png?generation=1597345065735225&alt=media)\n\nTop of these methods can be chosen as a final set of augmentation methods and can be applied to the train set together.", "votes": null }, { "id": "970657", "postDate": "08/14/2020 16:37:56", "content": "

Great resources. Learned some new techniques of augmentations.

", "rawMarkdown": "Great resources. Learned some new techniques of augmentations.", "votes": null }, { "id": "970759", "postDate": "08/14/2020 18:41:57", "content": "

You're welcome!

", "rawMarkdown": "You're welcome!", "votes": null }, { "id": "971359", "postDate": "08/15/2020 12:56:04", "content": "

Thanks for sharing, great work Sir.

", "rawMarkdown": "Thanks for sharing, great work Sir.", "votes": null }, { "id": "971503", "postDate": "08/15/2020 15:42:35", "content": "

Thank you!

", "rawMarkdown": "Thank you!", "votes": null }, { "id": "1001826", "postDate": "09/07/2020 15:47:51", "content": "

thank you for sharing, useful topic 🙏

", "rawMarkdown": "thank you for sharing, useful topic 🙏", "votes": null }, { "id": "1001846", "postDate": "09/07/2020 16:04:51", "content": "

You're welcome!

", "rawMarkdown": "You're welcome!", "votes": null } ], "comments": [ { "id": 971359, "author_name": "palaksood97", "author_url": "", "post_date": "08/15/2020 12:56:04", "content": "

Thanks for sharing, great work Sir.

", "votes": null, "replies": [ { "id": 971503, "author_name": "koza4ukdmitrij", "author_url": "", "post_date": "08/15/2020 15:42:35", "content": "

Thank you!

", "votes": null, "replies": [] } ] }, { "id": 1001826, "author_name": "", "author_url": "", "post_date": "09/07/2020 15:47:51", "content": "

thank you for sharing, useful topic 🙏

", "votes": null, "replies": [ { "id": 1001846, "author_name": "koza4ukdmitrij", "author_url": "", "post_date": "09/07/2020 16:04:51", "content": "

You're welcome!

", "votes": null, "replies": [] } ] }, { "id": 970657, "author_name": "azaemon", "author_url": "", "post_date": "08/14/2020 16:37:56", "content": "

Great resources. Learned some new techniques of augmentations.

", "votes": null, "replies": [ { "id": 970759, "author_name": "koza4ukdmitrij", "author_url": "", "post_date": "08/14/2020 18:41:57", "content": "

You're welcome!

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "969548": "In additional to topic https://www.kaggle.com/c/birdsong-recognition/discussion/170913 with great visualizations I want to add some augmentation ideas with corresponding implementations.\n\n**Audiomentations**\n\nGreat audio implementations in one of github repositories (link below): you can see about 20 augmentation techniques with short descriptions in front of each class. Some of those:\n - *AddImpulseResponse, AddGaussianNoise, AddBackgroundNoise, AddShortNoises*.\n - *TimeMask, FrequencyMask*.\n - *Shift, PitchShift, TimeStretch*.\n\ncode: https://github.com/iver56/audiomentations/blob/master/audiomentations/augmentations/transforms.py\n\n**SpecAugment**\n\n SpecAugment consists of three consecutive augmenatation methods:\n\n - *Time warp*: exchange some tuple of points in a row audio signal.\n - *Frequency masking*: mask some horizontal line inside a mel-spectrogram.\n - *Time masking*: mask some vertical line inside a mel-spectrogram.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1659719%2F7a00e5b29b6ecc1c3e56da9b7cbd3012%2FSpecAugmentations.png?generation=1597342806077272&alt=media)\n\npaper: https://arxiv.org/pdf/1904.08779.pdf.\ncode: we've already had an implementation of this method inside one of a public kernels - https://www.kaggle.com/hidehisaarai1213/introduction-to-sound-event-detection#Train-SED-model-with-only-weak-supervision (another thanks for great kernel!).\n\n**MixAugment**\n\nThe main contribution of this paper is an idea to use more than one sample inside an augmentation technique (but more than two ones don't increase a target metrics). For two samples `(x1, y1)` and `(x2, y2)` we can do it next way:\n\n\n```\nx = lambda * x1 + (1 - lambda) * x2\ny = lambda * y1 + (1 - lambda) * y2\n```\n\nIt called *mixups*, but there are lots of other techniques of augmentations including two samples (see the paper for a details):\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1659719%2Fdc1510aa28daabe7651d42eed02a2cfe%2FMixAugment.png?generation=1597344047630223&alt=media)\n\npaper: https://arxiv.org/pdf/1805.11272.pdf. \ncode: https://github.com/ceciliaresearch/MixedExample/blob/master/mixed_example.py\n\n**How change the final set of augmentations?**\n\nIn a last paper autors just applyed each augmentation separetly and saw on results:\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1659719%2F57e2abe4418ca8046f0c411fdf00b572%2FAugmentResults.png?generation=1597345065735225&alt=media)\n\nTop of these methods can be chosen as a final set of augmentation methods and can be applied to the train set together.", "970657": "Great resources. Learned some new techniques of augmentations.", "970759": "You're welcome!", "971359": "Thanks for sharing, great work Sir.", "971503": "Thank you!", "1001826": "thank you for sharing, useful topic 🙏", "1001846": "You're welcome!" }, "source": "meta" }