![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F12870466%2F2c5ec9bfb0bd7983c1b5e7cf7d35bb7c%2FScreenshot_95.jpg?generation=1684832109956677&alt=media\")
\n:)
i have two preprocessing function one used torchaudio other used librosaa . But i find librosa mel spectograms slightly more better than torchaudio ( not much difference generates same featured mel specs but slight noise pickup in ta ) . I was training a model no matter what model i use resnet18 - efficientnetb0 training time for one epoch with batch size 32 takes around 45 - 50 mins ,when use librosa . I stuck finding why it takes so long .After some struggling , i found that torchaudio is faster than librosa ,now training time reduced to 18-20min per epoch with same config .
\nAny tips for reducing further more ,i'm using efficientnetv2 small model training with pytorch lightningg ⚡.
\nthank you 😄
", "messages": [ { "id": 2273132, "postDate": "2023-05-25T03:10:42.943Z", "content": "Nice one. thanks
", "rawMarkdown": "Nice one. thanks" } ] }, { "id": 2270321, "postDate": "2023-05-23T05:52:53.990Z", "content": "i have two preprocessing function one used torchaudio other used librosaa . But i find librosa mel spectograms slightly more better than torchaudio ( not much difference generates same featured mel specs but slight noise pickup in ta ) . I was training a model no matter what model i use resnet18 - efficientnetb0 training time for one epoch with batch size 32 takes around 45 - 50 mins ,when use librosa . I stuck finding why it takes so long .After some struggling , i found that torchaudio is faster than librosa ,now training time reduced to 18-20min per epoch with same config .
\nAny tips for reducing further more ,i'm using efficientnetv2 small model training with pytorch lightningg ⚡.
\nthank you 😄
", "rawMarkdown": "i have two preprocessing function one used torchaudio other used librosaa . But i find librosa mel spectograms slightly more better than torchaudio ( not much difference generates same featured mel specs but slight noise pickup in ta ) . I was training a model no matter what model i use resnet18 - efficientnetb0 training time for one epoch with batch size 32 takes around 45 - 50 mins ,when use librosa . I stuck finding why it takes so long .After some struggling , i found that torchaudio is faster than librosa ,now training time reduced to 18-20min per epoch with same config .\n\nAny tips for reducing further more ,i'm using efficientnetv2 small model training with pytorch lightningg ⚡.\n\nthank you 😄", "votes": 1 }, { "id": 2270401, "postDate": "2023-05-23T06:58:36.667Z", "rawMarkdown": "", "votes": -4, "isDeleted": true, "replies": [ { "id": 2270590, "postDate": "2023-05-23T08:56:56.737Z", "content": "
\n:)
On another note tho, if you want to minimise the differences between librosa and torchaudio, check out this notebook made at the end of the previous BirdClef comp here that does exactely that !
", "rawMarkdown": "\n:)\n\nOn another note tho, if you want to minimise the differences between librosa and torchaudio, check out this notebook made at the end of the previous BirdClef comp [here ](https://www.kaggle.com/code/nomorevotch/create-the-same-mel-from-librosa-and-torchaudio/notebook)that does exactely that !", "votes": 2, "replies": [ { "id": 2270638, "postDate": "2023-05-23T09:43:07.840Z", "content": "Thanks for sharing , i was asking about how to make same mels in both torchaudio and librosa , the notebook u shared really helped me . Now i can update my training pipeline . 😄
", "rawMarkdown": "Thanks for sharing , i was asking about how to make same mels in both torchaudio and librosa , the notebook u shared really helped me . Now i can update my training pipeline . 😄", "votes": 1 } ] } ] } ], "comments": [ { "id": 2273132, "author_name": "Soheil Tehranipour", "author_url": "", "post_date": "2023-05-25T03:10:42.943000", "content": "Nice one. thanks
", "votes": 0, "replies": [] } ] }, { "id": 2270401, "author_name": "", "author_url": "", "post_date": "2023-05-23T06:58:36.667000", "content": "", "votes": -4, "replies": [ { "id": 2270590, "author_name": "JEANMPIA", "author_url": "", "post_date": "2023-05-23T08:56:56.737000", "content": "
\n:)
On another note tho, if you want to minimise the differences between librosa and torchaudio, check out this notebook made at the end of the previous BirdClef comp here that does exactely that !
", "votes": 2, "replies": [ { "id": 2270638, "author_name": "Altair Farooque", "author_url": "", "post_date": "2023-05-23T09:43:07.840000", "content": "Thanks for sharing , i was asking about how to make same mels in both torchaudio and librosa , the notebook u shared really helped me . Now i can update my training pipeline . 😄
", "votes": 1, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2273132": "1. Integration with PyTorch\n2. GPU acceleration\n3. Differentiable operations\n4. Compatibility with Torch ecosystem\n5. Community support", "2270321": "i have two preprocessing function one used torchaudio other used librosaa . But i find librosa mel spectograms slightly more better than torchaudio ( not much difference generates same featured mel specs but slight noise pickup in ta ) . I was training a model no matter what model i use resnet18 - efficientnetb0 training time for one epoch with batch size 32 takes around 45 - 50 mins ,when use librosa . I stuck finding why it takes so long .After some struggling , i found that torchaudio is faster than librosa ,now training time reduced to 18-20min per epoch with same config .\n\nAny tips for reducing further more ,i'm using efficientnetv2 small model training with pytorch lightningg ⚡.\n\nthank you 😄", "2270401": "" } }