mirror of
https://github.com/storytold/storyteller-ml.git
synced 2026-10-16 00:13:07 +00:00
Brandon Thomas
508 lines
18 KiB
Python
508 lines
18 KiB
Python
# -*- coding: utf-8 -*-
|
|
"""SoftVCTraceAndOnnx.ipynb
|
|
|
|
Automatically generated by Colaboratory.
|
|
|
|
Original file is located at
|
|
https://colab.research.google.com/drive/1g0YsQQY8fzXHmZUWAMzQO60c6OUlyKBl
|
|
"""
|
|
|
|
!pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu116
|
|
!pip3 install sklearn
|
|
!pip3 install soundfile
|
|
!pip3 install matplotlib
|
|
|
|
!git clone https://github.com/bshall/hubert
|
|
!git clone https://github.com/bshall/acoustic-model
|
|
!git clone https://github.com/bshall/hifigan
|
|
|
|
!nvidia-smi
|
|
|
|
from google.colab import drive
|
|
drive.mount('/content/drive')
|
|
|
|
# Torch Script Inference Hubert
|
|
import copy
|
|
from typing import Optional, Tuple
|
|
import random
|
|
|
|
from sklearn.cluster import KMeans
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
from torch.nn.modules.utils import consume_prefix_in_state_dict_if_present
|
|
import torch.nn.utils as utils
|
|
|
|
from torch.nn.utils.weight_norm import WeightNorm
|
|
from typing import Dict, Any
|
|
from collections import OrderedDict
|
|
|
|
def remove_weight_norm(module):
|
|
module_list = [mod for mod in module.children()]
|
|
if len(module_list) == 0:
|
|
for k, hook in module._forward_pre_hooks.items():
|
|
if isinstance(hook, WeightNorm):
|
|
hook.remove(module)
|
|
del module._forward_pre_hooks[k]
|
|
else:
|
|
for mod in module_list:
|
|
remove_weight_norm(mod)
|
|
|
|
|
|
def modify_model(state_dict:Dict[str,Any]):
|
|
new_state_dict = OrderedDict()
|
|
for k in state_dict['acoustic-model'].keys():
|
|
value = state_dict['acoustic-model'][k]
|
|
name = k[7:] # remove `module.` because they used ddp training to save
|
|
new_state_dict[name] = value
|
|
return new_state_dict
|
|
|
|
# @torch.jit.export
|
|
# def generate(self, x: torch.Tensor) -> torch.Tensor:
|
|
# x = self.encoder(x)
|
|
# return self.decoder.generate(x)
|
|
class AcousticModelScripted(nn.Module):
|
|
def __init__(self, discrete: bool = False, upsample: bool = True):
|
|
super().__init__()
|
|
self.encoder = Encoder(discrete, upsample)
|
|
self.decoder = Decoder()
|
|
|
|
def forward(self, x: torch.Tensor, mels: torch.Tensor) -> torch.Tensor:
|
|
x = self.encoder(x)
|
|
return self.decoder(x, mels)
|
|
|
|
@torch.jit.export
|
|
def generate(self, x: torch.Tensor) -> torch.Tensor:
|
|
x = self.encoder(x)
|
|
return self.decoder.generate(x)
|
|
|
|
|
|
class Encoder(nn.Module):
|
|
def __init__(self, discrete: bool = False, upsample: bool = True):
|
|
super().__init__()
|
|
self.embedding = nn.Embedding(100 + 1, 256) if discrete else None
|
|
self.prenet = PreNet(256, 256, 256)
|
|
self.convs = nn.Sequential(
|
|
nn.Conv1d(256, 512, 5, 1, 2),
|
|
nn.ReLU(),
|
|
nn.InstanceNorm1d(512),
|
|
nn.ConvTranspose1d(512, 512, 4, 2, 1) if upsample else nn.Identity(),
|
|
nn.Conv1d(512, 512, 5, 1, 2),
|
|
nn.ReLU(),
|
|
nn.InstanceNorm1d(512),
|
|
nn.Conv1d(512, 512, 5, 1, 2),
|
|
nn.ReLU(),
|
|
nn.InstanceNorm1d(512),
|
|
)
|
|
|
|
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
|
if self.embedding is not None:
|
|
x = self.embedding(x)
|
|
x = self.prenet(x)
|
|
x = self.convs(x.transpose(1, 2))
|
|
return x.transpose(1, 2)
|
|
|
|
|
|
class Decoder(nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.prenet = PreNet(128, 256, 256)
|
|
self.lstm1 = nn.LSTM(512 + 256, 768, batch_first=True)
|
|
self.lstm2 = nn.LSTM(768, 768, batch_first=True)
|
|
self.lstm3 = nn.LSTM(768, 768, batch_first=True)
|
|
self.proj = nn.Linear(768, 128, bias=False)
|
|
|
|
def forward(self, x: torch.Tensor, mels: torch.Tensor) -> torch.Tensor:
|
|
mels = self.prenet(mels)
|
|
x, _ = self.lstm1(torch.cat((x, mels), dim=-1))
|
|
res = x
|
|
x, _ = self.lstm2(x)
|
|
x = res + x
|
|
res = x
|
|
x, _ = self.lstm3(x)
|
|
x = res + x
|
|
return self.proj(x)
|
|
|
|
@torch.inference_mode()
|
|
def generate(self, xs: torch.Tensor) -> torch.Tensor:
|
|
m = torch.zeros(xs.size(0), 128, device=xs.device)
|
|
h1 = torch.zeros(1, xs.size(0), 768, device=xs.device)
|
|
c1 = torch.zeros(1, xs.size(0), 768, device=xs.device)
|
|
h2 = torch.zeros(1, xs.size(0), 768, device=xs.device)
|
|
c2 = torch.zeros(1, xs.size(0), 768, device=xs.device)
|
|
h3 = torch.zeros(1, xs.size(0), 768, device=xs.device)
|
|
c3 = torch.zeros(1, xs.size(0), 768, device=xs.device)
|
|
|
|
mel = []
|
|
for x in torch.unbind(xs, dim=1):
|
|
m = self.prenet(m)
|
|
x = torch.cat((x, m), dim=1).unsqueeze(1)
|
|
x1, (h1, c1) = self.lstm1(x, (h1, c1))
|
|
x2, (h2, c2) = self.lstm2(x1, (h2, c2))
|
|
x = x1 + x2
|
|
x3, (h3, c3) = self.lstm3(x, (h3, c3))
|
|
x = x + x3
|
|
m = self.proj(x).squeeze(1)
|
|
mel.append(m)
|
|
return torch.stack(mel, dim=1)
|
|
|
|
|
|
class PreNet(nn.Module):
|
|
def __init__(
|
|
self,
|
|
input_size: int,
|
|
hidden_size: int,
|
|
output_size: int,
|
|
dropout: float = 0.5,
|
|
):
|
|
super().__init__()
|
|
self.net = nn.Sequential(
|
|
nn.Linear(input_size, hidden_size),
|
|
nn.ReLU(),
|
|
nn.Dropout(dropout),
|
|
nn.Linear(hidden_size, output_size),
|
|
nn.ReLU(),
|
|
nn.Dropout(dropout),
|
|
)
|
|
|
|
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
|
return self.net(x)
|
|
|
|
import IPython.display as display
|
|
import torch, torchaudio
|
|
import torch.nn as nn
|
|
import sys
|
|
import soundfile as sf
|
|
|
|
from time import time
|
|
import torch.jit
|
|
sys.path.insert(0, './hubert/hubert')
|
|
from model import HubertSoft
|
|
|
|
sys.path.insert(0, './acoustic-model/')
|
|
from acoustic.model import AcousticModel
|
|
|
|
#@title select input wav
|
|
file_name = "/content/drive/MyDrive/VCModels/Source/brandon1.wav"#@param { type: "string"}
|
|
|
|
source, sr = torchaudio.load(f"{file_name}")
|
|
source = torchaudio.functional.resample(source, sr, 16000)
|
|
source = source.unsqueeze(0).cuda()
|
|
print(f"Wav Shape and Type {source.shape} {source.dtype}")
|
|
if source.shape[1] == 2: # remove channel
|
|
source = torch.mean(source, dim=0).unsqueeze(0)
|
|
print(f"Removing Channel to mono Wav Shape and Type {source.shape} {source.dtype}")
|
|
|
|
display.Audio(source.squeeze().cpu(), rate=16000)
|
|
|
|
import datetime
|
|
|
|
hubert_soft = torch.hub.load("bshall/hubert:main", "hubert_soft").cuda()
|
|
hubert_soft.eval()
|
|
|
|
script_acoustic = True #@param {type: "boolean"}
|
|
print(script_acoustic)
|
|
|
|
# Load the acoustic model (either hubert_soft or hubert_discrete)
|
|
input_acoustic_model_file = "/content/drive/MyDrive/VCModels/Padme/model-24000.pt" #@param {type: "string"}
|
|
acoustic_dict = torch.load(input_acoustic_model_file) # this used ddp ## REPLACE THIS! <-- !
|
|
if script_acoustic == True:
|
|
acoustic = AcousticModelScripted().cuda()
|
|
else:
|
|
acoustic = AcousticModel().cuda()
|
|
|
|
acoustic.eval()
|
|
|
|
new_state_dict = modify_model(state_dict=acoustic_dict)
|
|
acoustic.load_state_dict(new_state_dict)
|
|
|
|
import importlib.util
|
|
import sys
|
|
|
|
spec = importlib.util.spec_from_file_location("hifigan.utils", "/content/hifigan/hifigan/utils.py")
|
|
u = importlib.util.module_from_spec(spec)
|
|
sys.modules["hifigan.utils"] = u
|
|
spec.loader.exec_module(u)
|
|
|
|
spec = importlib.util.spec_from_file_location("hifigan.generator", "/content/hifigan/hifigan/generator.py")
|
|
g = importlib.util.module_from_spec(spec)
|
|
sys.modules["hifigan.generator"] = g
|
|
spec.loader.exec_module(g)
|
|
|
|
new_state_dict = modify_model(state_dict=acoustic_dict)
|
|
acoustic.load_state_dict(new_state_dict)
|
|
input_hifigan_checkpoint_path = "/content/drive/MyDrive/VCModels/Hifigan/Padme.pt" #@param {type: "string"}
|
|
if input_hifigan_checkpoint_path:
|
|
checkpoint = torch.load(input_hifigan_checkpoint_path)["generator"]["model"]
|
|
hifigan = g.HifiganGenerator()
|
|
consume_prefix_in_state_dict_if_present(checkpoint, "module.")
|
|
hifigan.load_state_dict(checkpoint)
|
|
hifigan.remove_weight_norm()
|
|
else:
|
|
hifigan = torch.hub.load("bshall/hifigan:main", "hifigan_hubert_soft")
|
|
|
|
hifigan.cuda()
|
|
hifigan.eval()
|
|
|
|
#utils.remove_weight_norm(hubert_soft.positional_embedding.conv)
|
|
# Convert to the target speaker
|
|
units = None
|
|
mel = None
|
|
|
|
unit_start_time = 0.0
|
|
unit_end_time = 0.0
|
|
|
|
acoustic_start_time = 0.0
|
|
acoustic_end_time = 0.0
|
|
|
|
hifi_start_time = 0.0
|
|
hifi_end_time = 0.0
|
|
|
|
with torch.inference_mode():
|
|
unit_start_time = datetime.datetime.now()
|
|
units = hubert_soft.units(source)
|
|
print(f"Units and Type {units.shape} {units.dtype}")
|
|
unit_end_time = datetime.datetime.now()
|
|
|
|
hubert_soft_script = torch.jit.trace_module(hubert_soft,{"units":source})
|
|
torch.jit.save(hubert_soft_script, 'hubert_soft.jit')
|
|
|
|
print(f"Unit Generation Time {(unit_end_time - unit_start_time).total_seconds() * 1000}")
|
|
|
|
print("Hubert Soft Exported")
|
|
|
|
acoustic_start_time = datetime.datetime.now()
|
|
mel = acoustic.generate(units).transpose(1, 2)
|
|
print(f"Mel and Type {mel.shape} {mel.dtype}")
|
|
acoustic_end_time = datetime.datetime.now()
|
|
print(f"Acoustic Time {(acoustic_end_time - acoustic_start_time).total_seconds() * 1000}")
|
|
if script_acoustic == True:
|
|
output_acoustic_scripted = '/content/drive/MyDrive/VCModels/FinishedModels/Padme.recast' #@param {type: "string"}
|
|
print("Exporting Scripted")
|
|
acoustic_model_script = torch.jit.script(acoustic)
|
|
torch.jit.save(acoustic_model_script, output_acoustic_scripted)
|
|
else:
|
|
print("Exporting Traced")
|
|
acoustic_model_script = torch.jit.trace_module(acoustic,{"generate":units})
|
|
output_acoustic_traced = '' #@param {type: "string"}
|
|
torch.jit.save(acoustic_model_script, output_acoustic_traced)
|
|
|
|
print("Acoustic Model Exported")
|
|
hifi_start_time = datetime.datetime.now()
|
|
|
|
target = hifigan(mel)
|
|
print(f"Wavform and Type {target.shape} {target.dtype}")
|
|
hifi_end_time = datetime.datetime.now()
|
|
print(f"MEL {target.shape}")
|
|
print(f"Hifi Gan Time {(hifi_end_time - hifi_start_time).total_seconds() * 1000}")
|
|
|
|
hifi_gan_script = torch.jit.trace_module(hifigan,{"forward":mel})
|
|
output_hifigan_jit = '/content/drive/MyDrive/VCModels/FinishedModels/Padme-hifigan-jit.pt' #@param {type: "string"}
|
|
torch.jit.save(hifi_gan_script, output_hifigan_jit)
|
|
print("Hifi Gan Model Exported")
|
|
|
|
import datetime
|
|
|
|
jit_unit_start_time = 0.0
|
|
jit_unit_end_time = 0.0
|
|
jit_unit_final = 0.0
|
|
|
|
jit_acoustic_start_time = 0.0
|
|
jit_acoustic_end_time = 0.0
|
|
jit_acoustic_final = 0.0
|
|
|
|
jit_hifi_start_time = 0.0
|
|
jit_hifi_end_time = 0.0
|
|
jit_hifi_final = 0.0
|
|
|
|
jit_model = torch.jit.load('hubert_soft.jit').cuda()
|
|
jit_model.eval()
|
|
with torch.no_grad():
|
|
jit_unit_start_time = datetime.datetime.now()
|
|
jit_tensor = jit_model.units(source)
|
|
jit_unit_end_time = datetime.datetime.now()
|
|
jit_unit_final = jit_unit_end_time - jit_unit_start_time
|
|
print(jit_unit_final.total_seconds() * 1000)
|
|
print(jit_tensor.shape)
|
|
assert torch.allclose(jit_tensor,units)
|
|
|
|
print(script_acoustic)
|
|
if script_acoustic == True:
|
|
acoustic_jit_model = torch.jit.load(output_acoustic_scripted).cuda()
|
|
print("Using Scripted Model")
|
|
else:
|
|
acoustic_jit_model = torch.jit.load(output_acoustic_traced).cuda()
|
|
acoustic_jit_model.eval()
|
|
with torch.no_grad():
|
|
if script_acoustic == True:
|
|
print("Using Scripted Model")
|
|
jit_acoustic_start_time = datetime.datetime.now()
|
|
jit_mel = acoustic_jit_model.generate(jit_tensor)
|
|
else:
|
|
jit_acoustic_start_time = datetime.datetime.now()
|
|
jit_mel = acoustic_jit_model.generate(jit_tensor)
|
|
jit_acoustic_end_time = datetime.datetime.now()
|
|
jit_acoustic_final = jit_acoustic_end_time - jit_acoustic_start_time
|
|
print(jit_acoustic_final.total_seconds() * 1000)
|
|
|
|
jit_mel = jit_mel.transpose(1,2)
|
|
print(jit_mel.shape)
|
|
assert torch.allclose(mel,jit_mel)
|
|
|
|
hifi_gan = torch.jit.load(output_hifigan_jit).cuda()
|
|
hifi_gan.eval()
|
|
with torch.no_grad():
|
|
jit_hifi_start_time = datetime.datetime.now()
|
|
jit_target = hifi_gan(jit_mel)
|
|
jit_hifi_end_time = datetime.datetime.now()
|
|
jit_hifi_final = jit_hifi_end_time - jit_hifi_start_time
|
|
print(jit_hifi_final.total_seconds() * 1000)
|
|
assert torch.allclose(target,jit_target)
|
|
|
|
sf.write(f'./sample_jit.wav', jit_target.squeeze().detach().cpu().numpy(), samplerate=16000)
|
|
sf.write(f'./sample.wav', target.squeeze().cpu().detach().numpy(), samplerate=16000)
|
|
|
|
#uploaded = files.upload()
|
|
import datetime
|
|
|
|
jit_unit_start_time = 0.0
|
|
jit_unit_end_time = 0.0
|
|
jit_unit_final = 0.0
|
|
|
|
jit_acoustic_start_time = 0.0
|
|
jit_acoustic_end_time = 0.0
|
|
jit_acoustic_final = 0.0
|
|
|
|
jit_hifi_start_time = 0.0
|
|
jit_hifi_end_time = 0.0
|
|
jit_hifi_final = 0.0
|
|
|
|
|
|
source, sr = torchaudio.load(f"{file_name}")
|
|
source = torchaudio.functional.resample(source, sr, 16000)
|
|
source = source.unsqueeze(0).cuda()
|
|
print(f"Wav Shape and Type {source.shape} {source.dtype}")
|
|
if source.shape[1] == 2: # remove channel
|
|
source = torch.mean(source, dim=1).unsqueeze(0)
|
|
print(f"Removing Channel to mono Wav Shape and Type {source.shape} {source.dtype}")
|
|
|
|
display.Audio(source.squeeze().cpu(), rate=16000)
|
|
|
|
jit_model = torch.jit.load('hubert_soft.jit').cuda()
|
|
jit_model.eval()
|
|
|
|
if script_acoustic == True:
|
|
acoustic_jit_model = torch.jit.load(output_acoustic_scripted).cuda()
|
|
print("Loading Scripted Model")
|
|
else:
|
|
acoustic_jit_model = torch.jit.load(output_acoustic_traced).cuda()
|
|
print("Loading Tracing Model")
|
|
|
|
acoustic_jit_model.eval()
|
|
|
|
hifi_gan = torch.jit.load(output_hifigan_jit).cuda()
|
|
hifi_gan.eval()
|
|
print("Loaded Models")
|
|
|
|
# Warm up the model to get the real time
|
|
for i in range(0,5):
|
|
with torch.no_grad():
|
|
jit_unit_start_time = datetime.datetime.now()
|
|
jit_tensor = jit_model.units(source)
|
|
jit_unit_end_time = datetime.datetime.now()
|
|
|
|
if script_acoustic == True:
|
|
print("Using Scripted Model TT2")
|
|
jit_acoustic_start_time = datetime.datetime.now()
|
|
jit_mel = acoustic_jit_model.generate(jit_tensor)
|
|
else:
|
|
print("Using Traced Model TT2")
|
|
jit_acoustic_start_time = datetime.datetime.now()
|
|
jit_mel = acoustic_jit_model.generate(jit_tensor)
|
|
jit_mel = jit_mel.transpose(1, 2)
|
|
|
|
jit_acoustic_end_time = datetime.datetime.now()
|
|
|
|
jit_hifi_start_time = datetime.datetime.now()
|
|
jit_target = hifi_gan(jit_mel)
|
|
jit_hifi_end_time = datetime.datetime.now()
|
|
|
|
|
|
jit_unit_final = jit_unit_end_time - jit_unit_start_time
|
|
jit_acoustic_final = jit_acoustic_end_time - jit_acoustic_start_time
|
|
jit_hifi_final = jit_hifi_end_time - jit_hifi_start_time
|
|
|
|
print(f"16 Khz Wav Shape {source.shape}")
|
|
|
|
print(f"Jit Tensor Shape for Unit Inputs: {jit_tensor.shape}")
|
|
print(f"Hubert Runtime : {jit_unit_final.total_seconds() * 1000}")
|
|
|
|
print(f"Tactron 2 Runtime Mel Output Shape: {jit_acoustic_final.total_seconds() * 1000}")
|
|
print(f"Jit Tensor Shape for Mel output: {jit_mel.shape}")
|
|
|
|
print(f"Hifi Runtime Mel Output Shape: {jit_hifi_final.total_seconds() * 1000}")
|
|
print(f"Jit Tensor Shape for Mel output: {jit_target.shape}")
|
|
|
|
sf.write(f'./{file_name}_jit.wav', jit_target.squeeze().detach().cpu().numpy(), samplerate=16000)
|
|
display.Audio(jit_target.squeeze().cpu(), rate=16000)
|
|
|
|
# silero vad
|
|
import datetime
|
|
|
|
|
|
source, sr = torchaudio.load(f"{file_name}")
|
|
source = torchaudio.functional.resample(source, sr, 16000)
|
|
|
|
print(source.shape)
|
|
|
|
input_vad_file = "/content/drive/MyDrive/tts/softvc-checkpoints/silero_vad.jit" #@param {type: "string"}
|
|
|
|
class VADIterator:
|
|
# You will only need to verify the 100 ms chunks to see this checks a minute of audio
|
|
# sound_sample_size run vad in 30 ms, 60 ms or 100 ms
|
|
def __init__(self,threshold: float = 0.5, sampling_rate: int = 16000,sound_sample_size:float = 100):
|
|
self.silero_jit = torch.jit.load(input_vad_file).cuda()
|
|
self.silero_jit.eval()
|
|
self.sampling_rate = sampling_rate
|
|
self.sound_sample_size = sound_sample_size
|
|
|
|
jit_silero_start_time = 0.0
|
|
jit_silero_end_time = 0.0
|
|
jit_silero_final = 0.0
|
|
|
|
def operating_frame_size(self,sampling_rate:int,ms_of_audio:float):
|
|
return int((ms_of_audio * sampling_rate) / 1000) # number of frames
|
|
|
|
def check_voice_activity(self,source_sample) -> float:
|
|
probs_over_time = []
|
|
# formula we know that we can get 1600 operating frames / 16000 sr * 1000 = 100 ms of audio from 1600 frames.
|
|
# this gives you the operating frames given the audio size you want.
|
|
op_frame_size = self.operating_frame_size(sampling_rate=self.sampling_rate,ms_of_audio=self.sound_sample_size)
|
|
print(op_frame_size)
|
|
iters = int(source_sample.shape[-1] / op_frame_size) # get the chunks
|
|
#print(iters)
|
|
for i in range(0,iters): # get sound_sample_size ms chunks
|
|
|
|
#print(window_size)
|
|
chunk = source_sample[i * op_frame_size:(i + 1) * op_frame_size]
|
|
with torch.no_grad():# slide window across using VAD 100ms
|
|
jit_silero_start_time = datetime.datetime.now()
|
|
try:
|
|
print(f"Current {i}")
|
|
prob = self.silero_jit(chunk.cuda(),sr=self.sampling_rate).item() # input shape is [1*] just an array of samples. 1 = false
|
|
jit_silero_end_time = datetime.datetime.now()
|
|
print(f"Probability: {prob}")
|
|
jit_silero_final = jit_silero_end_time - jit_silero_start_time
|
|
#print(f"Silero VAD Time: {jit_silero_final.total_seconds() * 1000}")
|
|
probs_over_time.append(prob)
|
|
except Exception as e:
|
|
print(e)
|
|
|
|
return probs_over_time
|
|
|
|
iterator = VADIterator(sound_sample_size=100)
|
|
source_sample = source_sample=source.squeeze().squeeze()
|
|
print(source_sample.shape)
|
|
probs = iterator.check_voice_activity(source_sample=source_sample)
|
|
#print(probs) |