rustyverse/rustytorch
2
0
Fork 0
Code Issues Pull Requests Actions 6 Packages Projects Releases Wiki Activity

style: cargo fmt --workspace (whitespace/wrapping only, no semantic change)

Browse Source 
Whole-workspace rustfmt pass picked up while iterating on Mamba GPU
backward work. Verified formatting-only via diff sampling; no logic
changed.

Co-Authored-By: Claude Sonnet 5 <[email protected]>
This commit is contained in:
osobh
2026-08-10 07:09:36 -07:00
co-authored by Claude Sonnet 5
parent ad6405663f
commit 4aaa36a57a
305 changed files with 25537 additions and 18337 deletions
Download Patch File Download Diff File Expand all files Collapse all files
crates/models/rtx-tts/src/vocoder/hifigan.rs
+66 -40
View File
@@ -7,10 +7,10 @@
//! Reference: "HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity
//! Speech Synthesis" (Kong et al., 2020)
use super::Vocoder;
use crate::error::{Result, TtsError};
use rtx_tensor::{Device, Tensor};
use serde::{Deserialize, Serialize};
use crate::error::{Result, TtsError};
use super::Vocoder;
/// Configuration for HiFi-GAN vocoder
#[derive(Debug, Clone, Serialize, Deserialize)]
@@ -37,11 +37,7 @@ impl Default for HiFiGANConfig {
upsample_rates: vec![8, 8, 2, 2],
upsample_kernel_sizes: vec![16, 16, 4, 4],
resblock_kernel_sizes: vec![3, 7, 11],
resblock_dilation_sizes: vec![
vec![1, 3, 5],
vec![1, 3, 5],
vec![1, 3, 5],
],
resblock_dilation_sizes: vec![vec![1, 3, 5], vec![1, 3, 5], vec![1, 3, 5]],
initial_channel: 512,
mel_channels: 80,
sample_rate: 22050,
@@ -53,29 +49,39 @@ impl HiFiGANConfig {
/// Validate configuration
pub fn validate(&self) -> Result<()> {
if self.upsample_rates.is_empty() {
return Err(TtsError::InvalidConfig("Upsample rates cannot be empty".to_string()));
return Err(TtsError::InvalidConfig(
"Upsample rates cannot be empty".to_string(),
));
}
if self.upsample_rates.len() != self.upsample_kernel_sizes.len() {
return Err(TtsError::InvalidConfig(
"Upsample rates and kernel sizes must have same length".to_string()
"Upsample rates and kernel sizes must have same length".to_string(),
));
}
if self.resblock_kernel_sizes.is_empty() {
return Err(TtsError::InvalidConfig("Resblock kernel sizes cannot be empty".to_string()));
return Err(TtsError::InvalidConfig(
"Resblock kernel sizes cannot be empty".to_string(),
));
}
if self.resblock_kernel_sizes.len() != self.resblock_dilation_sizes.len() {
return Err(TtsError::InvalidConfig(
"Resblock kernel sizes and dilation sizes must have same length".to_string()
"Resblock kernel sizes and dilation sizes must have same length".to_string(),
));
}
if self.initial_channel == 0 {
return Err(TtsError::InvalidConfig("Initial channel must be > 0".to_string()));
return Err(TtsError::InvalidConfig(
"Initial channel must be > 0".to_string(),
));
}
if self.mel_channels == 0 {
return Err(TtsError::InvalidConfig("Mel channels must be > 0".to_string()));
return Err(TtsError::InvalidConfig(
"Mel channels must be > 0".to_string(),
));
}
if self.sample_rate == 0 {
return Err(TtsError::InvalidConfig("Sample rate must be > 0".to_string()));
return Err(TtsError::InvalidConfig(
"Sample rate must be > 0".to_string(),
));
}
Ok(())
}
@@ -90,9 +96,16 @@ struct ResBlock1 {
}
impl ResBlock1 {
fn new(_channels: usize, kernel_size: usize, dilations: Vec<usize>, device: &Device) -> Result<Self> {
fn new(
_channels: usize,
kernel_size: usize,
dilations: Vec<usize>,
device: &Device,
) -> Result<Self> {
if dilations.is_empty() {
return Err(TtsError::InvalidConfig("Dilations cannot be empty".to_string()));
return Err(TtsError::InvalidConfig(
"Dilations cannot be empty".to_string(),
));
}
Ok(Self {
@@ -113,14 +126,20 @@ impl ResBlock1 {
let residual = self.apply_conv_block(&output, channels, dilation)?;
// Residual connection
output = output.add(&residual)
output = output
.add(&residual)
.map_err(|e| TtsError::TensorError(e.to_string()))?;
}
Ok(output)
}
fn apply_conv_block(&self, input: &Tensor, _channels: usize, _dilation: usize) -> Result<Tensor> {
fn apply_conv_block(
&self,
input: &Tensor,
_channels: usize,
_dilation: usize,
) -> Result<Tensor> {
// Simplified convolution block
// In a full implementation, this would use proper Conv1d with LeakyReLU
@@ -129,10 +148,12 @@ impl ResBlock1 {
let dims = input.dims();
// Apply LeakyReLU-like transformation (y = max(0.2x, x))
let data = input.to_cpu()
let data = input
.to_cpu()
.map_err(|e| TtsError::TensorError(e.to_string()))?;
let transformed: Vec<f32> = data.iter()
let transformed: Vec<f32> = data
.iter()
.map(|&x| if x > 0.0 { x } else { 0.2 * x })
.collect();
@@ -166,7 +187,7 @@ impl HiFiGANGenerator {
let dims = mel.dims();
if dims.len() != 2 {
return Err(TtsError::InvalidInput(
"Mel spectrogram must be 2D [mel_channels, time_frames]".to_string()
"Mel spectrogram must be 2D [mel_channels, time_frames]".to_string(),
));
}
@@ -178,7 +199,8 @@ impl HiFiGANGenerator {
}
// Reshape to [1, mel_channels, time_frames] for convolution
let mut x = mel.unsqueeze(0)
let mut x = mel
.unsqueeze(0)
.map_err(|e| TtsError::TensorError(e.to_string()))?;
// Pre-convolution: expand mel channels to initial_channel
@@ -201,9 +223,11 @@ impl HiFiGANGenerator {
x = self.post_conv(&x)?;
// Remove batch dimension and return [samples]
x = x.squeeze(Some(0))
x = x
.squeeze(Some(0))
.map_err(|e| TtsError::TensorError(e.to_string()))?;
x = x.squeeze(Some(0))
x = x
.squeeze(Some(0))
.map_err(|e| TtsError::TensorError(e.to_string()))?;
Ok(x)
@@ -224,12 +248,18 @@ impl HiFiGANGenerator {
Tensor::from_data(
data,
vec![batch_size, self.config.initial_channel, time_frames],
&self.device
&self.device,
)
.map_err(|e| TtsError::TensorError(e.to_string()))
}
fn upsample_block(&self, input: &Tensor, channels: usize, rate: usize, _idx: usize) -> Result<Tensor> {
fn upsample_block(
&self,
input: &Tensor,
channels: usize,
rate: usize,
_idx: usize,
) -> Result<Tensor> {
// Transposed convolution for upsampling
// In full implementation: ConvTranspose1d with proper upsampling
@@ -246,7 +276,7 @@ impl HiFiGANGenerator {
Tensor::from_data(
data,
vec![batch_size, out_channels, upsampled_frames],
&self.device
&self.device,
)
.map_err(|e| TtsError::TensorError(e.to_string()))
}
@@ -269,13 +299,15 @@ impl HiFiGANGenerator {
let mut result = outputs[0].clone();
for output in outputs.iter().skip(1) {
result = result.add(output)
result = result
.add(output)
.map_err(|e| TtsError::TensorError(e.to_string()))?;
}
// Divide by number of outputs for averaging
let scale = 1.0 / outputs.len() as f32;
result.mul_scalar(scale)
result
.mul_scalar(scale)
.map_err(|e| TtsError::TensorError(e.to_string()))
}
@@ -289,15 +321,12 @@ impl HiFiGANGenerator {
// Create output waveform
let data = vec![0.0f32; batch_size * time_frames];
let mut output = Tensor::from_data(
data,
vec![batch_size, 1, time_frames],
&self.device
)
.map_err(|e| TtsError::TensorError(e.to_string()))?;
let mut output = Tensor::from_data(data, vec![batch_size, 1, time_frames], &self.device)
.map_err(|e| TtsError::TensorError(e.to_string()))?;
// Apply tanh activation for audio range [-1, 1]
output = output.tanh()
output = output
.tanh()
.map_err(|e| TtsError::TensorError(e.to_string()))?;
Ok(output)
@@ -330,10 +359,7 @@ impl HiFiGAN {
pub fn new(config: HiFiGANConfig, device: &Device) -> Result<Self> {
let generator = HiFiGANGenerator::new(config.clone(), device)?;
Ok(Self {
generator,
config,
})
Ok(Self { generator, config })
}
/// Get the generator