facebook/mms-tts-eng
Original36.3Mparameters
Fixture16.3Mparameters
Layer stacks5compared
Findings0to review
✓ Same architecture, smaller
Every module class, layer kind and checked property of the original is in the fixture.
Layer stacks
One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.
text_encoder.encoder.layers
Original · 6
Fixture · 2
| Kind | Original | Fixture | |
|---|---|---|---|
| EncoderLayer | 6 | 2 | ✓ kept |
flow.flows
Original · 4
Fixture · 4
| Kind | Original | Fixture | |
|---|---|---|---|
| ResidualCouplingLayer | 4 | 4 | ✓ kept |
decoder.resblocks
Original · 12
Fixture · 12
| Kind | Original | Fixture | |
|---|---|---|---|
| HifiGanResidualBlock | 12 | 12 | ✓ kept |
posterior_encoder.wavenet.in_layers
Original · 16
Fixture · 16
| Kind | Original | Fixture | |
|---|---|---|---|
| ModuleDict | 16 | 16 | ✓ kept |
posterior_encoder.wavenet.res_skip_layers
Original · 16
Fixture · 16
| Kind | Original | Fixture | |
|---|---|---|---|
| ModuleDict | 16 | 16 | ✓ kept |
Where the parameters are
Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.
Original
Fixture
| Component | Original | Share | Fixture | Share |
|---|---|---|---|---|
| decoder.resblocks | 11.0M | 30.2% | 11.0M | 67.2% |
| flow.flows | 7.1M | 19.6% | 827.5K | 5.1% |
| text_encoder.encoder.layers | 6.2M | 17.1% | 133.1K | 0.8% |
| posterior_encoder.wavenet.in_layers | 5.9M | 16.3% | 659.5K | 4.0% |
| decoder.upsampler | 2.7M | 7.3% | 2.7M | 16.3% |
| posterior_encoder.wavenet.res_skip_layers | 1.2M | 3.2% | 130.9K | 0.8% |
| decoder.conv_pre | 688.6K | 1.9% | 688.6K | 4.2% |
| 13 others | 1.6M | 4.3% | 254.4K | 1.6% |
Preserved properties
| Config | Property | Original | Fixture | |
|---|---|---|---|---|
| model | model_type | vits | vits | ✓ kept |
| model | vocab_size | 38 | 38 | ✓ kept |
| model | hidden_act | relu | relu | ✓ kept |
| model | sampling_rate | 16000 | 16000 | ✓ kept |
Module classes
Instances of each class in the built model (prefix Vits omitted).
| Class | Original | Fixture |
|---|---|---|
| Conv1d | 177 | 169 |
| ModuleList | 80 | 80 |
| LayerNorm | 72 | 64 |
| ModuleDict | 64 | 64 |
| ParametrizationList | 64 | 64 |
| ParametrizedConv1d | 64 | 64 |
| _WeightNorm | 64 | 64 |
| Dropout | 27 | 19 |
| Linear | 24 | 8 |
| HifiGanResidualBlock | 12 | 12 |
| DilatedDepthSeparableConv | 10 | 10 |
| ConvFlow | 8 | 8 |
| ReLU | 6 | 2 |
| Attention | 6 | 2 |
| EncoderLayer | 6 | 2 |
| FeedForward | 6 | 2 |
| WaveNet | 5 | 5 |
| ConvTranspose1d | 4 | 4 |
| ResidualCouplingLayer | 4 | 4 |
| ElementwiseAffine | 2 | 2 |
| Embedding | 1 | 1 |
| Encoder | 1 | 1 |
| HifiGan | 1 | 1 |
| Model | 1 | 1 |
| PosteriorEncoder | 1 | 1 |
| ResidualCouplingBlock | 1 | 1 |
| StochasticDurationPredictor | 1 | 1 |
| TextEncoder | 1 | 1 |
Config changes
4 fields changed, 50 unchanged.
| Field | Original | Fixture |
|---|---|---|
| ffn_dim | 768 | 128 |
| hidden_size | 192 | 64 |
| num_attention_heads | 2 | 4 |
| num_hidden_layers | 6 | 2 |