Hello everyone! I have published a dataset with a trained ResNet-34 model to tackle this task with CNNs and Mel spectrograms.
\n\nNotebook: https://www.kaggle.com/tarunpaparaju/birdcall-identification-spectrogram-resnet\nDataset: https://www.kaggle.com/tarunpaparaju/pretrained-resnet34-for-birdcall-classification
\n\nThis dataset contains a ResNet-34 model trained on Mel spectrograms from the Cornell Birdcall Identification dataset. It can be used to identify bird species from audio clips with high accuracy (around 55% on unseen clips) spanning 264 different species mentioned on https://www.xeno-canto.org/.
To use this pre-trained model with PyTorch, you first need to convert your audio clip to a Mel spectrogram image to feed into the model. Refer to my kernel on birdcall classification to understand how to generate these Mel spectrograms. Make sure your audio signal is of length 1000000 and set the spectrogram features to 256. Then finally convert the image to a 3-channel version (repetition) and apply classic ImageNet normalization with albumentations. Once you convert the audio clip/s to Mel spectrograms, define the ResNet model and the load the pre-trained weights from this dataset. The code snippet below demonstrates how to set up the model with pre-trained weights. Now, you can use this model to classify bird species!
", "messages": [ { "id": "900922", "postDate": "06/25/2020 06:30:10", "content": "
Hello everyone! I have published a dataset with a trained ResNet-34 model to tackle this task with CNNs and Mel spectrograms.
\n\nNotebook: https://www.kaggle.com/tarunpaparaju/birdcall-identification-spectrogram-resnet\nDataset: https://www.kaggle.com/tarunpaparaju/pretrained-resnet34-for-birdcall-classification
\n\nThis dataset contains a ResNet-34 model trained on Mel spectrograms from the Cornell Birdcall Identification dataset. It can be used to identify bird species from audio clips with high accuracy (around 55% on unseen clips) spanning 264 different species mentioned on https://www.xeno-canto.org/.
To use this pre-trained model with PyTorch, you first need to convert your audio clip to a Mel spectrogram image to feed into the model. Refer to my kernel on birdcall classification to understand how to generate these Mel spectrograms. Make sure your audio signal is of length 1000000 and set the spectrogram features to 256. Then finally convert the image to a 3-channel version (repetition) and apply classic ImageNet normalization with albumentations. Once you convert the audio clip/s to Mel spectrograms, define the ResNet model and the load the pre-trained weights from this dataset. The code snippet below demonstrates how to set up the model with pre-trained weights. Now, you can use this model to classify bird species!
", "rawMarkdown": "Hello everyone! I have published a dataset with a trained ResNet-34 model to tackle this task with CNNs and Mel spectrograms.\n\n________________________________________________________________________________________________________________________________________________________________________________\n\n### TLDR\n\nNotebook: https://www.kaggle.com/tarunpaparaju/birdcall-identification-spectrogram-resnet\nDataset: https://www.kaggle.com/tarunpaparaju/pretrained-resnet34-for-birdcall-classification\n\n________________________________________________________________________________________________________________________________________________________________________________\n\n### Content\n\nThis dataset contains a
ResNet-34 model trained on Mel spectrograms from the [Cornell Birdcall Identification](https://www.kaggle.com/c/birdsong-recognition) dataset. It can be used to identify bird species from audio clips with high accuracy (around 55% on unseen clips) spanning 264 different species mentioned on https://www.xeno-canto.org/.\n\n________________________________________________________________________________________________________________________________________________________________________________\n\n### Usage\n\nTo use this pre-trained model with PyTorch, you first need to convert your audio clip to a Mel spectrogram image to feed into the model. Refer to [my kernel on birdcall classification](https://www.kaggle.com/tarunpaparaju/birdcall-identification-spectrogram-resnet) to understand how to generate these Mel spectrograms. Make sure your audio signal is of length 1000000 and set the spectrogram features to 256. Then finally convert the image to a 3-channel version (repetition) and apply classic ImageNet normalization with albumentations. Once you convert the audio clip/s to Mel spectrograms, define the ResNet model and the load the pre-trained weights from this dataset. The code snippet below demonstrates how to set up the model with pre-trained weights. Now, you can use this model to classify bird species!\n\n________________________________________________________________________________________________________________________________________________________________________________\n\n ![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1351163%2Fd1aed4b170f94ec95d97fff9163330b6%2Fcarbon.png?generation=1593065716014933&alt=media\")
",
"votes": null
},
{
"id": "900954",
"postDate": "06/25/2020 06:59:53",
"content": "P.S.: I forgot to add torch.load in the final line. Call torch.load before setting the state dict.
torch.load in the final line. Call torch.load before setting the state dict.",
"votes": null
}
],
"comments": [
{
"id": 900954,
"author_name": "tarunpaparaju",
"author_url": "",
"post_date": "06/25/2020 06:59:53",
"content": "P.S.: I forgot to add torch.load in the final line. Call torch.load before setting the state dict.
ResNet-34 model trained on Mel spectrograms from the [Cornell Birdcall Identification](https://www.kaggle.com/c/birdsong-recognition) dataset. It can be used to identify bird species from audio clips with high accuracy (around 55% on unseen clips) spanning 264 different species mentioned on https://www.xeno-canto.org/.\n\n________________________________________________________________________________________________________________________________________________________________________________\n\n### Usage\n\nTo use this pre-trained model with PyTorch, you first need to convert your audio clip to a Mel spectrogram image to feed into the model. Refer to [my kernel on birdcall classification](https://www.kaggle.com/tarunpaparaju/birdcall-identification-spectrogram-resnet) to understand how to generate these Mel spectrograms. Make sure your audio signal is of length 1000000 and set the spectrogram features to 256. Then finally convert the image to a 3-channel version (repetition) and apply classic ImageNet normalization with albumentations. Once you convert the audio clip/s to Mel spectrograms, define the ResNet model and the load the pre-trained weights from this dataset. The code snippet below demonstrates how to set up the model with pre-trained weights. Now, you can use this model to classify bird species!\n\n________________________________________________________________________________________________________________________________________________________________________________\n\n ",
"900954": "P.S.: I forgot to add torch.load in the final line. Call torch.load before setting the state dict."
},
"source": "meta"
}