*you may consider to use mixed precision to train your model faster and gain some extra epoch to improve it *🤔 \nhow to implement it ?\nhttps://www.tensorflow.org/guide/keras/mixed_precision\nhttps://arxiv.org/abs/1710.03740
", "messages": [ { "id": "801910", "postDate": "04/08/2020 23:46:14", "content": "*you may consider to use mixed precision to train your model faster and gain some extra epoch to improve it *🤔 \nhow to implement it ?\nhttps://www.tensorflow.org/guide/keras/mixed_precision\nhttps://arxiv.org/abs/1710.03740
", "rawMarkdown": "**you may consider to use mixed precision to train your model faster and gain some extra epoch to improve it **🤔 \n[how to implement it ?](https://stackoverflow.com/a/57656007/8736575)\n[https://www.tensorflow.org/guide/keras/mixed_precision](https://www.tensorflow.org/guide/keras/mixed_precision)\n[https://arxiv.org/abs/1710.03740](https://arxiv.org/abs/1710.03740)", "votes": null }, { "id": "803726", "postDate": "04/10/2020 19:38:34", "content": "Not useful with P100 GPUs which doe not have mixed precision hardware. It works really well with V100 GPUs but unfortunately those are not the ones available on Kaggle. I have a couple of benchmarks with and without mixed precision.\nCNNs for image recognition: https://www.kaggle.com/c/flower-classification-with-tpus/discussion/137983#784972\nBERT for NLP: https://www.kaggle.com/c/jigsaw-multilingual-toxic-comment-classification/discussion/140022#793524
", "rawMarkdown": "Not useful with P100 GPUs which doe not have mixed precision hardware. It works really well with V100 GPUs but unfortunately those are not the ones available on Kaggle. I have a couple of benchmarks with and without mixed precision.\nCNNs for image recognition: https://www.kaggle.com/c/flower-classification-with-tpus/discussion/137983#784972\nBERT for NLP: https://www.kaggle.com/c/jigsaw-multilingual-toxic-comment-classification/discussion/140022#793524", "votes": null }, { "id": "804197", "postDate": "04/11/2020 10:43:05", "content": "thank you it is really helpful
", "rawMarkdown": "thank you it is really helpful", "votes": null } ], "comments": [ { "id": 803726, "author_name": "mgorner", "author_url": "", "post_date": "04/10/2020 19:38:34", "content": "Not useful with P100 GPUs which doe not have mixed precision hardware. It works really well with V100 GPUs but unfortunately those are not the ones available on Kaggle. I have a couple of benchmarks with and without mixed precision.\nCNNs for image recognition: https://www.kaggle.com/c/flower-classification-with-tpus/discussion/137983#784972\nBERT for NLP: https://www.kaggle.com/c/jigsaw-multilingual-toxic-comment-classification/discussion/140022#793524
", "votes": null, "replies": [ { "id": 804197, "author_name": "hatemamine", "author_url": "", "post_date": "04/11/2020 10:43:05", "content": "thank you it is really helpful
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "801910": "**you may consider to use mixed precision to train your model faster and gain some extra epoch to improve it **🤔 \n[how to implement it ?](https://stackoverflow.com/a/57656007/8736575)\n[https://www.tensorflow.org/guide/keras/mixed_precision](https://www.tensorflow.org/guide/keras/mixed_precision)\n[https://arxiv.org/abs/1710.03740](https://arxiv.org/abs/1710.03740)", "803726": "Not useful with P100 GPUs which doe not have mixed precision hardware. It works really well with V100 GPUs but unfortunately those are not the ones available on Kaggle. I have a couple of benchmarks with and without mixed precision.\nCNNs for image recognition: https://www.kaggle.com/c/flower-classification-with-tpus/discussion/137983#784972\nBERT for NLP: https://www.kaggle.com/c/jigsaw-multilingual-toxic-comment-classification/discussion/140022#793524", "804197": "thank you it is really helpful" }, "source": "meta" }