Audio-MLP

MLP-based models for learning audio representations. Submission for HEAR-2021@NeurIPS'21.

Setup

pip install -e 'git+https://github.com/ID56/HEAR-2021-Audio-MLP.git#egg=hearaudiomlp'

Usage

from hearaudiomlp.kwmlp import load_model, get_timestamp_embeddings, get_scene_embeddings # or from hearaudiomlp.audiomlp

model = load_model("checkpoints/kwmlp.pth")

b, ms, sr = 2, 1000, 16000
dummy_input = torch.randn(b, int(sr * ms / 1000))

embeddings, timestamps = get_timestamp_embeddings(dummy_input, model)
scene_embeddings = get_scene_embeddings(dummy_input, model)

Models

Model Name	# Params†	GFLOPS*†	Sampling Rate	Hop Length	Timestamp Embedding	Scene Embedding	Location
kwmlp	424K	0.034	16000	10ms	64	1024	kwmlp(1.7Mb)
audiomae	213K	0.023	16000	10ms	8	1584	audiomae(0.9Mb)

† _{Only considering the encoder, which is used for generating embeddings.}
* _{Although there is no direct way to count FLOPS like parameters, you can use facebookresearch/fvcore. The FLOPS measured are per 1 single input spectrogram (tensor of shape (1, 40, 98)).}

The trained checkpoints are licensed under a Creative Commons Attribution-ShareAlike 4.0 International License, as per HEAR-2021 requirements. You may also download them from drive: [ kwmlp | audiomae ].

Notes

All models were trained on:

• A standard Kaggle environment: a single 16GiB NVIDIA Tesla P100, CUDA 11.0, CuDNN 8.0.5, python 3.7.10.
• KW-MLP was trained on Google Speech Commands V2-35, and the weights are a direct port of its paper [1].
• AudioMAE is an adaptation of KW-MLP that was trained on the training splits from the HEAR2021 Open tasks.
• Both models primarily utilize gated-MLPs [2].

References

@misc{morshed2021attentionfree,
      title   = {Attention-Free Keyword Spotting}, 
      author  = {Mashrur M. Morshed and Ahmad Omar Ahsan},
      year    = {2021},
      eprint  = {2110.07749},
      archivePrefix = {arXiv},
      primaryClass  = {cs.LG}
}

@misc{liu2021pay,
      title  = {Pay Attention to MLPs}, 
      author = {Hanxiao Liu and Zihang Dai and David R. So and Quoc V. Le},
      year   = {2021},
      eprint = {2105.08050},
      archivePrefix = {arXiv},
      primaryClass  = {cs.LG}
}