mirror of
https://github.com/storytold/artcraft.git
synced 2026-11-19 08:05:20 +00:00
theoretically TTS HTTP works with inference-job
This commit is contained in:
1 parent
3d4a870190
commit
cc1d7f6de1
8 files changed
+78
-45
No files matched your search
Generated
+1
@@ -47,6 +47,7 @@
|
||||
<w>vocoder</w>
|
||||
<w>vocoders</w>
|
||||
<w>vocodes</w>
|
||||
<w>waveglow</w>
|
||||
</words>
|
||||
</dictionary>
|
||||
</component>
|
||||
@@ -81,7 +81,7 @@ SO_VITS_SVC_INFERENCE_MAYBE_DEFAULT_CONFIG_PATH="/models/voice_conversion/so-vit
|
||||
|
||||
TT2_LEGACY_ROOT_DIRECTORY="/models/tts"
|
||||
TT2_LEGACY_MAYBE_PYTHON_INTERPRETER="python3.6"
|
||||
TT2_LEGACY_MAYBE_DOCKER_IMAGE_SHA="9f1945388110"
|
||||
TT2_LEGACY_MAYBE_DOCKER_IMAGE_SHA="0d29da13e2df"
|
||||
|
||||
# ================================= #
|
||||
# #
|
||||
|
||||
+1
@@ -6,3 +6,4 @@ pub mod tacotron2_dependencies;
|
||||
pub mod tacotron2_inference_command;
|
||||
pub mod tacotron2_inference_sidecar_client;
|
||||
pub mod tacotron2_sidecar_health_check_client;
|
||||
pub mod vocoder_option;
|
||||
+32
-13
@@ -25,7 +25,8 @@ use crate::job::job_loop::job_success_result::{JobSuccessResult, ResultEntity};
|
||||
use crate::job::job_loop::process_single_job_error::ProcessSingleJobError;
|
||||
use crate::job::job_types::tts::tacotron2_v2_early_fakeyou::download_static_dependencies::download_static_dependencies;
|
||||
use crate::job::job_types::tts::tacotron2_v2_early_fakeyou::seconds_to_decoder_steps::seconds_to_decoder_steps;
|
||||
use crate::job::job_types::tts::tacotron2_v2_early_fakeyou::tacotron2_inference_command::{InferenceArgs, MelMultiplyFactor, VocoderForInferenceOption};
|
||||
use crate::job::job_types::tts::tacotron2_v2_early_fakeyou::tacotron2_inference_command::{InferenceArgs, MelMultiplyFactor};
|
||||
use crate::job::job_types::tts::tacotron2_v2_early_fakeyou::vocoder_option::VocoderForInferenceOption;
|
||||
use crate::job_dependencies::JobDependencies;
|
||||
use crate::util::maybe_download_file_from_bucket::maybe_download_file_from_bucket;
|
||||
|
||||
@@ -198,7 +199,6 @@ pub async fn process_job(args: ProcessJobArgs<'_>) -> Result<JobSuccessResult, P
|
||||
|
||||
// ==================== RUN INFERENCE SCRIPT ==================== //
|
||||
|
||||
|
||||
let mut maybe_mel_multiply_factor = None;
|
||||
|
||||
if let Some(factor) = tts_model.maybe_custom_mel_multiply_factor {
|
||||
@@ -207,19 +207,38 @@ pub async fn process_job(args: ProcessJobArgs<'_>) -> Result<JobSuccessResult, P
|
||||
maybe_mel_multiply_factor = Some(MelMultiplyFactor::DefaultMultiplyFactor);
|
||||
}
|
||||
|
||||
// TODO(bt,2023-11-27): Handle LRU cache on python side
|
||||
let maybe_unload_model_path = None;
|
||||
|
||||
let inference_start_time = Instant::now();
|
||||
|
||||
let _r = model_dependencies.inference_command.execute_inference(InferenceArgs {
|
||||
synthesizer_checkpoint_path: &tts_synthesizer_fs_path,
|
||||
text_pipeline_type: text_pipeline_type_or_guess.to_str(),
|
||||
vocoder: vocoder_option,
|
||||
maybe_mel_multiply_factor,
|
||||
max_decoder_steps,
|
||||
input_text_filename: &text_input_fs_path,
|
||||
output_audio_filename: &output_audio_fs_path,
|
||||
output_spectrogram_filename: &output_spectrogram_fs_path,
|
||||
output_metadata_filename: &output_metadata_fs_path,
|
||||
});
|
||||
if model_dependencies.sidecar.use_sidecar_instead_of_shell {
|
||||
let _r = model_dependencies.sidecar.inference_client.request_inference(
|
||||
&cleaned_inference_text,
|
||||
max_decoder_steps,
|
||||
&tts_synthesizer_fs_path,
|
||||
&text_pipeline_type_or_guess.to_str(),
|
||||
vocoder_option,
|
||||
&output_audio_fs_path,
|
||||
&output_spectrogram_fs_path,
|
||||
&output_metadata_fs_path,
|
||||
maybe_unload_model_path,
|
||||
tts_model.use_default_mel_multiply_factor,
|
||||
tts_model.maybe_custom_mel_multiply_factor,
|
||||
).await.map_err(|e| ProcessSingleJobError::Other(e))?;
|
||||
} else {
|
||||
let _r = model_dependencies.inference_command.execute_inference(InferenceArgs {
|
||||
synthesizer_checkpoint_path: &tts_synthesizer_fs_path,
|
||||
text_pipeline_type: text_pipeline_type_or_guess.to_str(),
|
||||
vocoder: vocoder_option,
|
||||
maybe_mel_multiply_factor,
|
||||
max_decoder_steps,
|
||||
input_text_filename: &text_input_fs_path,
|
||||
output_audio_filename: &output_audio_fs_path,
|
||||
output_spectrogram_filename: &output_spectrogram_fs_path,
|
||||
output_metadata_filename: &output_metadata_fs_path,
|
||||
});
|
||||
}
|
||||
|
||||
let inference_duration = Instant::now().duration_since(inference_start_time);
|
||||
|
||||
|
||||
+2
@@ -21,6 +21,7 @@ pub struct Tacotron2Dependencies {
|
||||
}
|
||||
|
||||
pub struct SidecarDeps {
|
||||
pub use_sidecar_instead_of_shell: bool,
|
||||
pub inference_client: Tacotron2InferenceSidecarClient,
|
||||
pub health_check_client: Tacotron2SidecarHealthCheckClient,
|
||||
}
|
||||
@@ -56,6 +57,7 @@ impl Tacotron2Dependencies {
|
||||
hifigan_vocoder_model_filename,
|
||||
hifigan_superres_vocoder_model_filename,
|
||||
sidecar: SidecarDeps {
|
||||
use_sidecar_instead_of_shell: true, // TODO: ENV VAR
|
||||
inference_client,
|
||||
health_check_client,
|
||||
}
|
||||
|
||||
+2
-12
@@ -3,10 +3,11 @@ use std::path::{Path, PathBuf};
|
||||
use log::info;
|
||||
use subprocess::{Popen, PopenConfig};
|
||||
|
||||
use errors::AnyhowResult;
|
||||
use filesys::path_to_string::path_to_string;
|
||||
use subprocess_common::docker_options::{DockerFilesystemMount, DockerGpu, DockerOptions};
|
||||
|
||||
use crate::AnyhowResult;
|
||||
use crate::job::job_types::tts::tacotron2_v2_early_fakeyou::vocoder_option::VocoderForInferenceOption;
|
||||
|
||||
/// This command is used to run tacotron2 (v1 "early fakeyou") inference
|
||||
#[derive(Clone)]
|
||||
@@ -27,17 +28,6 @@ pub struct Tacotron2InferenceCommand {
|
||||
maybe_docker_options: Option<DockerOptions>,
|
||||
}
|
||||
|
||||
#[derive(Clone)]
|
||||
pub enum VocoderForInferenceOption<P: AsRef<Path>> {
|
||||
Waveglow {
|
||||
waveglow_vocoder_checkpoint_path: P,
|
||||
},
|
||||
HifiganSuperres {
|
||||
hifigan_vocoder_checkpoint_path: P,
|
||||
hifigan_superres_vocoder_checkpoint_path: P,
|
||||
}
|
||||
}
|
||||
|
||||
pub enum MelMultiplyFactor {
|
||||
// NB: Default is typically "1.4"
|
||||
DefaultMultiplyFactor,
|
||||
|
||||
+26
-19
@@ -6,8 +6,11 @@ use hyper::client::Client;
|
||||
use log::info;
|
||||
|
||||
use container_common::anyhow_result::AnyhowResult;
|
||||
use filesys::path_to_string::path_to_string;
|
||||
use mysql_queries::column_types::vocoder_type::VocoderType;
|
||||
|
||||
use crate::job::job_types::tts::tacotron2_v2_early_fakeyou::vocoder_option::VocoderForInferenceOption;
|
||||
|
||||
/// This is used to talk to the TT2 Python HTTP sidecar (rather than shelling out over bash).
|
||||
/// This keeps models loaded in memory and results in faster inference time.
|
||||
|
||||
@@ -67,11 +70,8 @@ impl Tacotron2InferenceSidecarClient {
|
||||
raw_text: &str,
|
||||
max_decoder_steps: u32,
|
||||
synthesizer_checkpoint_path: P,
|
||||
vocoder_type: VocoderType,
|
||||
text_pipeline_type: &str,
|
||||
hifigan_vocoder_checkpoint_path: P,
|
||||
hifigan_superres_vocoder_checkpoint_path: P,
|
||||
waveglow_vocoder_checkpoint_path: P,
|
||||
vocoder_option: VocoderForInferenceOption<P>,
|
||||
output_audio_filename: P,
|
||||
output_spectrogram_filename: P,
|
||||
output_metadata_filename: P,
|
||||
@@ -80,23 +80,30 @@ impl Tacotron2InferenceSidecarClient {
|
||||
maybe_custom_mel_multiply_factor: Option<f64>,
|
||||
) -> AnyhowResult<()> {
|
||||
|
||||
let waveglow_vocoder_checkpoint_path = waveglow_vocoder_checkpoint_path
|
||||
.as_ref()
|
||||
.to_str()
|
||||
.map(|s| s.to_string())
|
||||
.ok_or(anyhow!("bad waveglow vocoder path"))?;
|
||||
// TODO(bt,2023-11-26): Clean this mutability / multi-configuration up
|
||||
let vocoder_type;
|
||||
|
||||
let hifigan_vocoder_checkpoint_path = hifigan_vocoder_checkpoint_path
|
||||
.as_ref()
|
||||
.to_str()
|
||||
.map(|s| s.to_string())
|
||||
.ok_or(anyhow!("bad hifigan vocoder path"))?;
|
||||
let waveglow_vocoder_checkpoint_path;
|
||||
let hifigan_vocoder_checkpoint_path;
|
||||
let hifigan_superres_vocoder_checkpoint_path;
|
||||
|
||||
let hifigan_superres_vocoder_checkpoint_path = hifigan_superres_vocoder_checkpoint_path
|
||||
.as_ref()
|
||||
.to_str()
|
||||
.map(|s| s.to_string())
|
||||
.ok_or(anyhow!("bad hifigan super resolution vocoder path"))?;
|
||||
match vocoder_option {
|
||||
VocoderForInferenceOption::Waveglow { waveglow_vocoder_checkpoint_path : path } => {
|
||||
vocoder_type = VocoderType::WaveGlow;
|
||||
waveglow_vocoder_checkpoint_path = path_to_string(path.as_ref());
|
||||
hifigan_vocoder_checkpoint_path = "BOGUS_NOT_USED".to_string();
|
||||
hifigan_superres_vocoder_checkpoint_path = "BOGUS_NOT_USED".to_string();
|
||||
}
|
||||
VocoderForInferenceOption::HifiganSuperres {
|
||||
hifigan_vocoder_checkpoint_path: path,
|
||||
hifigan_superres_vocoder_checkpoint_path : super_res_path,
|
||||
} => {
|
||||
vocoder_type = VocoderType::HifiGanSuperResolution;
|
||||
waveglow_vocoder_checkpoint_path = "BOGUS_NOT_USED".to_string();
|
||||
hifigan_vocoder_checkpoint_path = path_to_string(path.as_ref());
|
||||
hifigan_superres_vocoder_checkpoint_path = path_to_string(super_res_path.as_ref());
|
||||
}
|
||||
}
|
||||
|
||||
let synthesizer_checkpoint_path = synthesizer_checkpoint_path
|
||||
.as_ref()
|
||||
|
||||
+13
@@ -0,0 +1,13 @@
|
||||
use std::path::Path;
|
||||
|
||||
#[derive(Clone)]
|
||||
pub enum VocoderForInferenceOption<P: AsRef<Path>> {
|
||||
Waveglow {
|
||||
waveglow_vocoder_checkpoint_path: P,
|
||||
},
|
||||
HifiganSuperres {
|
||||
hifigan_vocoder_checkpoint_path: P,
|
||||
hifigan_superres_vocoder_checkpoint_path: P,
|
||||
}
|
||||
}
|
||||
|
||||
Reference in new issue
Block a user