Thank you everyone for participating in this competition, thank you for your time and effort - I admire your dedication to submit hundreds of solutions to this challenge. Thanks everyone for your discussion posts and code notebook and thanks to all of you who posted their solutions. We'll carefully read all of them and take notes regarding things that worked and things that didn't, feedback and things we could do better next time.
\nWe had the largest field of participants of all BirdCLEF editions this year, which is a great success. Now, we need to transform this participation into conservation impact. And here is how you can help to achieve that:
\nWe want to put your best single model into the hands of ecologists who do conservation work in Kenya.
\nSubmitting a solution, getting a score and rank is fine, but how well does your model actually perform in practice? There is a way to find out. As you may know, we're collaborating with Natural State this year. Natural State has ongoing conservation efforts in Kenya, and their ecologists need machine learning tools to advance their efforts.
\nSo how can we help?
\nRaven Pro is an audio processing software tool and is widely used by ecologists to analyze audio data. Raven Pro also has a machine learning component that allows ecologists to run deep learning models within Raven without having to write any code. This allows automated analyses of audio monitoring data, greatly improving the process of identifying bird species in large audio collections.
\nIf ecologists can run your model in Raven, you can help to advance on-the-ground monitoring efforts.
\nHow does it work?
\nIs it complicated?
\nYes and no…Raven uses TensorFlow for Java to run inference and thus requires a TensorFlow saved model. If you built your model in TensorFlow, that should be straight forward. Yet, if you used a different framework, it might be a bit more complicated, and you might have to use conversion tools like ONNX to get to a TF saved model. Also, Raven models need to accept raw audio as input, not spectrograms - which might be another constraint.
\nCan we use the power of the Kaggle community to help Natural State?
\nWe would really appreciate your effort to try to convert your best single model into a Raven model. If you need a Raven license to test your model and the conversion, please email @holgerklinck at holger.klinck@cornell.edu.
\nPlease use this thread to let us know if you are having problems with the conversion, and also please let us know if you have gotten your model to work in Raven and share your model.
\nThanks again for all your efforts - now make them count!
", "messages": [ { "id": "2274992", "postDate": "05/26/2023 12:20:28", "content": "Thank you everyone for participating in this competition, thank you for your time and effort - I admire your dedication to submit hundreds of solutions to this challenge. Thanks everyone for your discussion posts and code notebook and thanks to all of you who posted their solutions. We'll carefully read all of them and take notes regarding things that worked and things that didn't, feedback and things we could do better next time.
\nWe had the largest field of participants of all BirdCLEF editions this year, which is a great success. Now, we need to transform this participation into conservation impact. And here is how you can help to achieve that:
\nWe want to put your best single model into the hands of ecologists who do conservation work in Kenya.
\nSubmitting a solution, getting a score and rank is fine, but how well does your model actually perform in practice? There is a way to find out. As you may know, we're collaborating with Natural State this year. Natural State has ongoing conservation efforts in Kenya, and their ecologists need machine learning tools to advance their efforts.
\nSo how can we help?
\nRaven Pro is an audio processing software tool and is widely used by ecologists to analyze audio data. Raven Pro also has a machine learning component that allows ecologists to run deep learning models within Raven without having to write any code. This allows automated analyses of audio monitoring data, greatly improving the process of identifying bird species in large audio collections.
\nIf ecologists can run your model in Raven, you can help to advance on-the-ground monitoring efforts.
\nHow does it work?
\nIs it complicated?
\nYes and no…Raven uses TensorFlow for Java to run inference and thus requires a TensorFlow saved model. If you built your model in TensorFlow, that should be straight forward. Yet, if you used a different framework, it might be a bit more complicated, and you might have to use conversion tools like ONNX to get to a TF saved model. Also, Raven models need to accept raw audio as input, not spectrograms - which might be another constraint.
\nCan we use the power of the Kaggle community to help Natural State?
\nWe would really appreciate your effort to try to convert your best single model into a Raven model. If you need a Raven license to test your model and the conversion, please email @holgerklinck at holger.klinck@cornell.edu.
\nPlease use this thread to let us know if you are having problems with the conversion, and also please let us know if you have gotten your model to work in Raven and share your model.
\nThanks again for all your efforts - now make them count!
", "rawMarkdown": "Thank you everyone for participating in this competition, thank you for your time and effort - I admire your dedication to submit hundreds of solutions to this challenge. Thanks everyone for your discussion posts and code notebook and thanks to all of you who posted their solutions. We'll carefully read all of them and take notes regarding things that worked and things that didn't, feedback and things we could do better next time.\n\nWe had the largest field of participants of all BirdCLEF editions this year, which is a great success. Now, we need to transform this participation into conservation impact. And here is how you can help to achieve that:\n\n**We want to put your best single model into the hands of ecologists who do conservation work in Kenya.**\n\nSubmitting a solution, getting a score and rank is fine, but how well does your model actually perform in practice? There is a way to find out. As you may know, we're collaborating with [Natural State](https://www.naturalstate.org) this year. Natural State has ongoing conservation efforts in Kenya, and their ecologists need machine learning tools to advance their efforts.\n\nSo how can we help? \n\n[Raven Pro](https://ravensoundsoftware.com/software/raven-pro/) is an audio processing software tool and is widely used by ecologists to analyze audio data. Raven Pro also has a machine learning component that allows ecologists to run deep learning models within Raven without having to write any code. This allows automated analyses of audio monitoring data, greatly improving the process of identifying bird species in large audio collections.\n\n**If ecologists can run your model in Raven, you can help to advance on-the-ground monitoring efforts.**\n\nHow does it work?\n\n- Raven has a well-documented [knowledge base](https://ravensoundsoftware.com/article-categories/learning-detector/) which has all the resources\n- Models need to be converted to the Raven model format, [here is how that works](https://ravensoundsoftware.com/wp-content/uploads/2023/05/MarchineLearningDetectorInRavenPro_v4.pdf)\n- Models need to be distributed so that ecologists can use them\n\nIs it complicated?\n\nYes and no...Raven uses TensorFlow for Java to run inference and thus requires a TensorFlow saved model. If you built your model in TensorFlow, that should be straight forward. Yet, if you used a different framework, it might be a bit more complicated, and you might have to use conversion tools like ONNX to get to a TF saved model. Also, Raven models need to accept raw audio as input, not spectrograms - which might be another constraint.\n\n**Can we use the power of the Kaggle community to help Natural State?**\n\nWe would really appreciate your effort to try to convert your best single model into a Raven model. If you need a Raven license to test your model and the conversion, please email @holgerklinck at [holger.klinck@cornell.edu](mailto:holger.klinck@cornell.edu).\n\nPlease use this thread to let us know if you are having problems with the conversion, and also please let us know if you have gotten your model to work in Raven and share your model.\n\nThanks again for all your efforts - now make them count!", "votes": null }, { "id": "2315111", "postDate": "06/23/2023 20:17:17", "content": "Hi Stefan,
\nThank you for the competition. Will it be possible to release the test data as public, or somehow allow evaluation of models that do not conform to the inference requirements. I had a model that was very successful on the train set, but I couldn't get it to run fast enough. I would like to be able to test it and compare it to other people's models on Kaggle.
", "rawMarkdown": "Hi Stefan,\n\nThank you for the competition. Will it be possible to release the test data as public, or somehow allow evaluation of models that do not conform to the inference requirements. I had a model that was very successful on the train set, but I couldn't get it to run fast enough. I would like to be able to test it and compare it to other people's models on Kaggle.", "votes": null } ], "comments": [ { "id": 2315111, "author_name": "pelegshilo", "author_url": "", "post_date": "06/23/2023 20:17:17", "content": "Hi Stefan,
\nThank you for the competition. Will it be possible to release the test data as public, or somehow allow evaluation of models that do not conform to the inference requirements. I had a model that was very successful on the train set, but I couldn't get it to run fast enough. I would like to be able to test it and compare it to other people's models on Kaggle.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2274992": "Thank you everyone for participating in this competition, thank you for your time and effort - I admire your dedication to submit hundreds of solutions to this challenge. Thanks everyone for your discussion posts and code notebook and thanks to all of you who posted their solutions. We'll carefully read all of them and take notes regarding things that worked and things that didn't, feedback and things we could do better next time.\n\nWe had the largest field of participants of all BirdCLEF editions this year, which is a great success. Now, we need to transform this participation into conservation impact. And here is how you can help to achieve that:\n\n**We want to put your best single model into the hands of ecologists who do conservation work in Kenya.**\n\nSubmitting a solution, getting a score and rank is fine, but how well does your model actually perform in practice? There is a way to find out. As you may know, we're collaborating with [Natural State](https://www.naturalstate.org) this year. Natural State has ongoing conservation efforts in Kenya, and their ecologists need machine learning tools to advance their efforts.\n\nSo how can we help? \n\n[Raven Pro](https://ravensoundsoftware.com/software/raven-pro/) is an audio processing software tool and is widely used by ecologists to analyze audio data. Raven Pro also has a machine learning component that allows ecologists to run deep learning models within Raven without having to write any code. This allows automated analyses of audio monitoring data, greatly improving the process of identifying bird species in large audio collections.\n\n**If ecologists can run your model in Raven, you can help to advance on-the-ground monitoring efforts.**\n\nHow does it work?\n\n- Raven has a well-documented [knowledge base](https://ravensoundsoftware.com/article-categories/learning-detector/) which has all the resources\n- Models need to be converted to the Raven model format, [here is how that works](https://ravensoundsoftware.com/wp-content/uploads/2023/05/MarchineLearningDetectorInRavenPro_v4.pdf)\n- Models need to be distributed so that ecologists can use them\n\nIs it complicated?\n\nYes and no...Raven uses TensorFlow for Java to run inference and thus requires a TensorFlow saved model. If you built your model in TensorFlow, that should be straight forward. Yet, if you used a different framework, it might be a bit more complicated, and you might have to use conversion tools like ONNX to get to a TF saved model. Also, Raven models need to accept raw audio as input, not spectrograms - which might be another constraint.\n\n**Can we use the power of the Kaggle community to help Natural State?**\n\nWe would really appreciate your effort to try to convert your best single model into a Raven model. If you need a Raven license to test your model and the conversion, please email @holgerklinck at [holger.klinck@cornell.edu](mailto:holger.klinck@cornell.edu).\n\nPlease use this thread to let us know if you are having problems with the conversion, and also please let us know if you have gotten your model to work in Raven and share your model.\n\nThanks again for all your efforts - now make them count!", "2315111": "Hi Stefan,\n\nThank you for the competition. Will it be possible to release the test data as public, or somehow allow evaluation of models that do not conform to the inference requirements. I had a model that was very successful on the train set, but I couldn't get it to run fast enough. I would like to be able to test it and compare it to other people's models on Kaggle." }, "source": "meta" }