{
  "id": 210875,
  "title": "[New Training] EfficientNetB2 on TPU",
  "url": "/competitions/rfcx-species-audio-detection/discussion/210875",
  "author_name": "Agastya Kumar",
  "post_date": "2021-01-12T19:02:21.375000",
  "votes": 4,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hi kagglers <br>\ni have recently created a notebook showing training of EffNetB2 [LB 0.807] based on my previous work on LYFT competition in different and clean way <strong>to get different perspective on how to train audio data on TPU.</strong><br>\nthe notebook also include :</p>\n<ul>\n<li>creation of mel spectrogram and patches of frames in  different way</li>\n<li>random_augmentation with no python overhead that includes specaug also  </li>\n<li>stepwise cosine decay with warm restarts using keras callback</li>\n<li>optimized training pipeline and one more way to improve training time signficantly on TPU</li>\n</ul>\n<p>here is the link of my recent work - [<a href=\"https://www.kaggle.com/ashusma/training-rfcx-tensorflow-tpu-effnet-b2\" target=\"_blank\">https://www.kaggle.com/ashusma/training-rfcx-tensorflow-tpu-effnet-b2</a>]<br>\nhere is the link of my previous work that uses custom training loop with TPU - [<a href=\"https://www.kaggle.com/ashusma/training-lyft-tensorflow-tpu-multi-mode\" target=\"_blank\">https://www.kaggle.com/ashusma/training-lyft-tensorflow-tpu-multi-mode</a>]</p>",
  "messages": [
    {
      "id": 1150682,
      "postDate": "2021-01-12T19:02:21.377Z",
      "content": "<p>Hi kagglers <br>\ni have recently created a notebook showing training of EffNetB2 [LB 0.807] based on my previous work on LYFT competition in different and clean way <strong>to get different perspective on how to train audio data on TPU.</strong><br>\nthe notebook also include :</p>\n<ul>\n<li>creation of mel spectrogram and patches of frames in  different way</li>\n<li>random_augmentation with no python overhead that includes specaug also  </li>\n<li>stepwise cosine decay with warm restarts using keras callback</li>\n<li>optimized training pipeline and one more way to improve training time signficantly on TPU</li>\n</ul>\n<p>here is the link of my recent work - [<a href=\"https://www.kaggle.com/ashusma/training-rfcx-tensorflow-tpu-effnet-b2\" target=\"_blank\">https://www.kaggle.com/ashusma/training-rfcx-tensorflow-tpu-effnet-b2</a>]<br>\nhere is the link of my previous work that uses custom training loop with TPU - [<a href=\"https://www.kaggle.com/ashusma/training-lyft-tensorflow-tpu-multi-mode\" target=\"_blank\">https://www.kaggle.com/ashusma/training-lyft-tensorflow-tpu-multi-mode</a>]</p>",
      "rawMarkdown": "Hi kagglers \ni have recently created a notebook showing training of EffNetB2 [LB 0.807] based on my previous work on LYFT competition in different and clean way **to get different perspective on how to train audio data on TPU.**\nthe notebook also include :\n- creation of mel spectrogram and patches of frames in  different way\n- random_augmentation with no python overhead that includes specaug also  \n- stepwise cosine decay with warm restarts using keras callback\n- optimized training pipeline and one more way to improve training time signficantly on TPU\n\nhere is the link of my recent work - [https://www.kaggle.com/ashusma/training-rfcx-tensorflow-tpu-effnet-b2]\nhere is the link of my previous work that uses custom training loop with TPU - [https://www.kaggle.com/ashusma/training-lyft-tensorflow-tpu-multi-mode]",
      "votes": 4
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1150682": "Hi kagglers \ni have recently created a notebook showing training of EffNetB2 [LB 0.807] based on my previous work on LYFT competition in different and clean way **to get different perspective on how to train audio data on TPU.**\nthe notebook also include :\n- creation of mel spectrogram and patches of frames in  different way\n- random_augmentation with no python overhead that includes specaug also  \n- stepwise cosine decay with warm restarts using keras callback\n- optimized training pipeline and one more way to improve training time signficantly on TPU\n\nhere is the link of my recent work - [https://www.kaggle.com/ashusma/training-rfcx-tensorflow-tpu-effnet-b2]\nhere is the link of my previous work that uses custom training loop with TPU - [https://www.kaggle.com/ashusma/training-lyft-tensorflow-tpu-multi-mode]"
  }
}