Commit 369e8e2e authored by Phil Wang's avatar Phil Wang
Browse files

allow for the structured quantized dropout from encodec paper

parent f54e19d5

README.md

+10 −0
Original line number Diff line number Diff line
@@ -454,3 +454,13 @@ $ accelerate launch train.py 
    status = {to be published - one attention stabilization technique is circulating within Google Brain, being used by multiple teams}

}

```



```bibtex

@article{Defossez2022HighFN,

    title   = {High Fidelity Neural Audio Compression},

    author  = {Alexandre D'efossez and Jade Copet and Gabriel Synnaeve and Yossi Adi},

    journal = {ArXiv},

    year    = {2022},

    volume  = {abs/2210.13438}

}

```

audiolm_pytorch/soundstream.py

+2 −0
Original line number Diff line number Diff line
@@ -362,6 +362,7 @@ class SoundStream(nn.Module): 
        rq_num_quantizers = 8,

        rq_commitment_weight = 1.,

        rq_ema_decay = 0.95,

        rq_quantize_dropout_multiple_of = 1,

        input_channels = 1,

        discr_multi_scales = (1, 0.5, 0.25),

        stft_normalized = False,
@@ -437,6 +438,7 @@ class SoundStream(nn.Module): 
            codebook_size = codebook_size,

            decay = rq_ema_decay,

            commitment_weight = rq_commitment_weight,

            quantize_dropout_multiple_of = rq_quantize_dropout_multiple_of,

            kmeans_init = True,

            threshold_ema_dead_code = 2,

            quantize_dropout = True,

audiolm_pytorch/version.py

+1 −1
Original line number Diff line number Diff line 
__version__ = '0.21.0'
 
__version__ = '0.21.2'

setup.py

+1 −1
Original line number Diff line number Diff line
@@ -33,7 +33,7 @@ setup( 
    'torchaudio',

    'transformers',

    'tqdm',

    'vector-quantize-pytorch>=0.10.15'

    'vector-quantize-pytorch>=1.0.6'

  ],

  classifiers=[

    'Development Status :: 4 - Beta',