{ "id": 492135, "title": "Complete Workflow: Spectrograms->ImageNet->.56 LB", "url": "/competitions/birdclef-2024/discussion/492135", "author_name": "", "post_date": "2024-04-08T18:42:20.725028500Z", "votes": 3, "comment_count": 1, "views": 0, "content": "

Short version - just made my first successful submission at .56 on LB - sharing the notebooks:

\n

https://www.kaggle.com/code/richolson/birdclef2024-simple-mel-spectrogram-generator/
\nhttps://www.kaggle.com/code/richolson/birdclef-2024-spectrograms-imagenet-train
\nhttps://www.kaggle.com/code/richolson/birdclef-2024-spectrograms-imagenet-run

\n

My goal here was just to quickly build something from-scratch that scores better-than-chance. There are easy things that I think will boost the score.

\n

Here's the breakdown:

\n

Generate Mel Spectrograms:
\nhttps://www.kaggle.com/code/richolson/birdclef2024-simple-mel-spectrogram-generator/
\nhttps://www.kaggle.com/datasets/richolson/birdclef-2024-mel-spectrograms

\n

14,522 224x224 images - representing 5-sec clips. This is just a subset of the full data. Took about 45-minutes to generate.

\n

Generating a more complete dataset would just be re-running (I'll probably do that at some point).

\n

Fine-tune an ImageNet (MobileNetV2) on Spectrograms:
\nhttps://www.kaggle.com/code/richolson/birdclef-2024-spectrograms-imagenet-train

\n

Trained for 10-epochs - takes only an hour on CPU. Model performance seemed to level-off about there (0.18 validation accuracy). I haven't played around with hyperparameters much - but I suspect adding more training data is next step.

\n

I chose MobileNetV2 since I knew runtime would be a bottle-neck with the CPU-only limitation.

\n

Run the Model:
\nhttps://www.kaggle.com/code/richolson/birdclef-2024-spectrograms-imagenet-run

\n

So - it turns out it needs about 7 seconds to generate the spectrograms for each 4-minute soundscape - and then another 7 seconds to run the model. So - 14-seconds total per file.

\n

To handle 1100 files in the 2-hour time limit - we need to be closer to 6.5 / seconds per 4-minute file.

\n

Wanted to get things working - so I made a compromise.. The notebook just predicts 1 out of every 3 5-second segment - and then copies those predictions for the next 2 segments. That gets it down to about 4.5 seconds per 4-minute file. I'm sure this hurts me on the LB - but I was happy to get things running.

\n

The current run notebook doesn't utilize all the CPU cores well - so I'm optimistic I can get the model running fully (or more fully) with some further tweaks.

\n

Hope this is useful to someone - please feel free to build on this!

", "messages": [ { "id": "2742120", "postDate": "04/08/2024 18:42:20", "content": "

Short version - just made my first successful submission at .56 on LB - sharing the notebooks:

\n

https://www.kaggle.com/code/richolson/birdclef2024-simple-mel-spectrogram-generator/
\nhttps://www.kaggle.com/code/richolson/birdclef-2024-spectrograms-imagenet-train
\nhttps://www.kaggle.com/code/richolson/birdclef-2024-spectrograms-imagenet-run

\n

My goal here was just to quickly build something from-scratch that scores better-than-chance. There are easy things that I think will boost the score.

\n

Here's the breakdown:

\n

Generate Mel Spectrograms:
\nhttps://www.kaggle.com/code/richolson/birdclef2024-simple-mel-spectrogram-generator/
\nhttps://www.kaggle.com/datasets/richolson/birdclef-2024-mel-spectrograms

\n

14,522 224x224 images - representing 5-sec clips. This is just a subset of the full data. Took about 45-minutes to generate.

\n

Generating a more complete dataset would just be re-running (I'll probably do that at some point).

\n

Fine-tune an ImageNet (MobileNetV2) on Spectrograms:
\nhttps://www.kaggle.com/code/richolson/birdclef-2024-spectrograms-imagenet-train

\n

Trained for 10-epochs - takes only an hour on CPU. Model performance seemed to level-off about there (0.18 validation accuracy). I haven't played around with hyperparameters much - but I suspect adding more training data is next step.

\n

I chose MobileNetV2 since I knew runtime would be a bottle-neck with the CPU-only limitation.

\n

Run the Model:
\nhttps://www.kaggle.com/code/richolson/birdclef-2024-spectrograms-imagenet-run

\n

So - it turns out it needs about 7 seconds to generate the spectrograms for each 4-minute soundscape - and then another 7 seconds to run the model. So - 14-seconds total per file.

\n

To handle 1100 files in the 2-hour time limit - we need to be closer to 6.5 / seconds per 4-minute file.

\n

Wanted to get things working - so I made a compromise.. The notebook just predicts 1 out of every 3 5-second segment - and then copies those predictions for the next 2 segments. That gets it down to about 4.5 seconds per 4-minute file. I'm sure this hurts me on the LB - but I was happy to get things running.

\n

The current run notebook doesn't utilize all the CPU cores well - so I'm optimistic I can get the model running fully (or more fully) with some further tweaks.

\n

Hope this is useful to someone - please feel free to build on this!

", "rawMarkdown": "Short version - just made my first successful submission at .56 on LB - sharing the notebooks:\n\nhttps://www.kaggle.com/code/richolson/birdclef2024-simple-mel-spectrogram-generator/\nhttps://www.kaggle.com/code/richolson/birdclef-2024-spectrograms-imagenet-train\nhttps://www.kaggle.com/code/richolson/birdclef-2024-spectrograms-imagenet-run\n\nMy goal here was just to quickly build something from-scratch that scores better-than-chance. There are easy things that I think will boost the score.\n\nHere's the breakdown:\n\n**Generate Mel Spectrograms:**\nhttps://www.kaggle.com/code/richolson/birdclef2024-simple-mel-spectrogram-generator/\nhttps://www.kaggle.com/datasets/richolson/birdclef-2024-mel-spectrograms\n\n14,522 224x224 images - representing 5-sec clips. This is just a subset of the full data. Took about 45-minutes to generate.\n\nGenerating a more complete dataset would just be re-running (I'll probably do that at some point).\n\n**Fine-tune an ImageNet (MobileNetV2) on Spectrograms:**\nhttps://www.kaggle.com/code/richolson/birdclef-2024-spectrograms-imagenet-train\n\nTrained for 10-epochs - takes only an hour on CPU. Model performance seemed to level-off about there (0.18 validation accuracy). I haven't played around with hyperparameters much - but I suspect adding more training data is next step.\n\nI chose MobileNetV2 since I knew runtime would be a bottle-neck with the CPU-only limitation.\n\n**Run the Model:**\nhttps://www.kaggle.com/code/richolson/birdclef-2024-spectrograms-imagenet-run\n\nSo - it turns out it needs about 7 seconds to generate the spectrograms for each 4-minute soundscape - and then another 7 seconds to run the model. So - 14-seconds total per file.\n\nTo handle 1100 files in the 2-hour time limit - we need to be closer to 6.5 / seconds per 4-minute file.\n\nWanted to get things working - so I made a compromise.. The notebook just predicts 1 out of every 3 5-second segment - and then copies those predictions for the next 2 segments. That gets it down to about 4.5 seconds per 4-minute file. I'm sure this hurts me on the LB - but I was happy to get things running.\n\nThe current run notebook doesn't utilize all the CPU cores well - so I'm optimistic I can get the model running fully (or more fully) with some further tweaks.\n\nHope this is useful to someone - please feel free to build on this!", "votes": null }, { "id": "2742609", "postDate": "04/09/2024 01:41:42", "content": "

I made some adjustments - and managed to get things running fast enough to predict on each 5-second interval.

\n

Unfortunately - this dropped my score from .56 down to .55 on the LB. Go figure…

", "rawMarkdown": "I made some adjustments - and managed to get things running fast enough to predict on each 5-second interval.\n\nUnfortunately - this dropped my score from .56 down to .55 on the LB. Go figure...", "votes": null } ], "comments": [ { "id": 2742609, "author_name": "richolson", "author_url": "", "post_date": "04/09/2024 01:41:42", "content": "

I made some adjustments - and managed to get things running fast enough to predict on each 5-second interval.

\n

Unfortunately - this dropped my score from .56 down to .55 on the LB. Go figure…

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2742120": "Short version - just made my first successful submission at .56 on LB - sharing the notebooks:\n\nhttps://www.kaggle.com/code/richolson/birdclef2024-simple-mel-spectrogram-generator/\nhttps://www.kaggle.com/code/richolson/birdclef-2024-spectrograms-imagenet-train\nhttps://www.kaggle.com/code/richolson/birdclef-2024-spectrograms-imagenet-run\n\nMy goal here was just to quickly build something from-scratch that scores better-than-chance. There are easy things that I think will boost the score.\n\nHere's the breakdown:\n\n**Generate Mel Spectrograms:**\nhttps://www.kaggle.com/code/richolson/birdclef2024-simple-mel-spectrogram-generator/\nhttps://www.kaggle.com/datasets/richolson/birdclef-2024-mel-spectrograms\n\n14,522 224x224 images - representing 5-sec clips. This is just a subset of the full data. Took about 45-minutes to generate.\n\nGenerating a more complete dataset would just be re-running (I'll probably do that at some point).\n\n**Fine-tune an ImageNet (MobileNetV2) on Spectrograms:**\nhttps://www.kaggle.com/code/richolson/birdclef-2024-spectrograms-imagenet-train\n\nTrained for 10-epochs - takes only an hour on CPU. Model performance seemed to level-off about there (0.18 validation accuracy). I haven't played around with hyperparameters much - but I suspect adding more training data is next step.\n\nI chose MobileNetV2 since I knew runtime would be a bottle-neck with the CPU-only limitation.\n\n**Run the Model:**\nhttps://www.kaggle.com/code/richolson/birdclef-2024-spectrograms-imagenet-run\n\nSo - it turns out it needs about 7 seconds to generate the spectrograms for each 4-minute soundscape - and then another 7 seconds to run the model. So - 14-seconds total per file.\n\nTo handle 1100 files in the 2-hour time limit - we need to be closer to 6.5 / seconds per 4-minute file.\n\nWanted to get things working - so I made a compromise.. The notebook just predicts 1 out of every 3 5-second segment - and then copies those predictions for the next 2 segments. That gets it down to about 4.5 seconds per 4-minute file. I'm sure this hurts me on the LB - but I was happy to get things running.\n\nThe current run notebook doesn't utilize all the CPU cores well - so I'm optimistic I can get the model running fully (or more fully) with some further tweaks.\n\nHope this is useful to someone - please feel free to build on this!", "2742609": "I made some adjustments - and managed to get things running fast enough to predict on each 5-second interval.\n\nUnfortunately - this dropped my score from .56 down to .55 on the LB. Go figure..." }, "source": "meta" }