opus/dnn/torch/osce/models/lavoce.py
Jan Buethe 00580a63aa
bugfix
2023-09-22 11:39:22 +02:00

274 lines
No EOL
12 KiB
Python

"""
/* Copyright (c) 2023 Amazon
Written by Jan Buethe */
/*
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
- Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
- Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
"""
import torch
from torch import nn
import torch.nn.functional as F
import numpy as np
from utils.layers.limited_adaptive_comb1d import LimitedAdaptiveComb1d
from utils.layers.limited_adaptive_conv1d import LimitedAdaptiveConv1d
from utils.layers.td_shaper import TDShaper
from utils.layers.noise_shaper import NoiseShaper
from utils.complexity import _conv1d_flop_count
from utils.endoscopy import write_data
from models.nns_base import NNSBase
from models.lpcnet_feature_net import LPCNetFeatureNet
from .scale_embedding import ScaleEmbedding
def print_channels(y, prefix="", name="", rate=16000):
num_channels = y.size(1)
for i in range(num_channels):
channel_name = f"{prefix}_c{i:02d}"
if len(name) > 0: channel_name += "_" + name
ch = y[0,i,:].detach().cpu().numpy()
ch = ((2**14) * ch / np.max(ch)).astype(np.int16)
write_data(channel_name, ch, rate)
class LaVoce(nn.Module):
""" Linear-Adaptive VOCodEr """
FEATURE_FRAME_SIZE=160
FRAME_SIZE=80
def __init__(self,
num_features=20,
pitch_embedding_dim=64,
cond_dim=256,
pitch_max=300,
kernel_size=15,
preemph=0.85,
comb_gain_limit_db=-6,
global_gain_limits_db=[-6, 6],
conv_gain_limits_db=[-6, 6],
norm_p=2,
avg_pool_k=4,
pulses=False,
innovate1=True,
innovate2=False,
innovate3=False,
ftrans_k=2):
super().__init__()
self.num_features = num_features
self.cond_dim = cond_dim
self.pitch_max = pitch_max
self.pitch_embedding_dim = pitch_embedding_dim
self.kernel_size = kernel_size
self.preemph = preemph
self.pulses = pulses
self.ftrans_k = ftrans_k
assert self.FEATURE_FRAME_SIZE % self.FRAME_SIZE == 0
self.upsamp_factor = self.FEATURE_FRAME_SIZE // self.FRAME_SIZE
# pitch embedding
self.pitch_embedding = nn.Embedding(pitch_max + 1, pitch_embedding_dim)
# feature net
self.feature_net = LPCNetFeatureNet(num_features + pitch_embedding_dim, cond_dim, self.upsamp_factor)
# noise shaper
self.noise_shaper = NoiseShaper(cond_dim, self.FRAME_SIZE)
# comb filters
left_pad = self.kernel_size // 2
right_pad = self.kernel_size - 1 - left_pad
self.cf1 = LimitedAdaptiveComb1d(self.kernel_size, cond_dim, frame_size=self.FRAME_SIZE, overlap_size=40, use_bias=False, padding=[left_pad, right_pad], max_lag=pitch_max + 1, gain_limit_db=comb_gain_limit_db, global_gain_limits_db=global_gain_limits_db, norm_p=norm_p)
self.cf2 = LimitedAdaptiveComb1d(self.kernel_size, cond_dim, frame_size=self.FRAME_SIZE, overlap_size=40, use_bias=False, padding=[left_pad, right_pad], max_lag=pitch_max + 1, gain_limit_db=comb_gain_limit_db, global_gain_limits_db=global_gain_limits_db, norm_p=norm_p)
self.af_prescale = LimitedAdaptiveConv1d(2, 1, self.kernel_size, cond_dim, frame_size=self.FRAME_SIZE, use_bias=False, padding=[self.kernel_size - 1, 0], gain_limits_db=conv_gain_limits_db, norm_p=norm_p)
self.af_mix = LimitedAdaptiveConv1d(2, 2, self.kernel_size, cond_dim, frame_size=self.FRAME_SIZE, use_bias=False, padding=[self.kernel_size - 1, 0], gain_limits_db=conv_gain_limits_db, norm_p=norm_p)
# spectral shaping
self.af1 = LimitedAdaptiveConv1d(1, 2, self.kernel_size, cond_dim, frame_size=self.FRAME_SIZE, use_bias=False, padding=[self.kernel_size - 1, 0], gain_limits_db=conv_gain_limits_db, norm_p=norm_p)
# non-linear transforms
self.tdshape1 = TDShaper(cond_dim, frame_size=self.FRAME_SIZE, avg_pool_k=avg_pool_k, innovate=innovate1)
self.tdshape2 = TDShaper(cond_dim, frame_size=self.FRAME_SIZE, avg_pool_k=avg_pool_k, innovate=innovate2)
self.tdshape3 = TDShaper(cond_dim, frame_size=self.FRAME_SIZE, avg_pool_k=avg_pool_k, innovate=innovate3)
# combinators
self.af2 = LimitedAdaptiveConv1d(2, 2, self.kernel_size, cond_dim, frame_size=self.FRAME_SIZE, use_bias=False, padding=[self.kernel_size - 1, 0], gain_limits_db=conv_gain_limits_db, norm_p=norm_p)
self.af3 = LimitedAdaptiveConv1d(2, 1, self.kernel_size, cond_dim, frame_size=self.FRAME_SIZE, use_bias=False, padding=[self.kernel_size - 1, 0], gain_limits_db=conv_gain_limits_db, norm_p=norm_p)
self.af4 = LimitedAdaptiveConv1d(2, 1, self.kernel_size, cond_dim, frame_size=self.FRAME_SIZE, use_bias=False, padding=[self.kernel_size - 1, 0], gain_limits_db=conv_gain_limits_db, norm_p=norm_p)
# feature transforms
self.post_cf1 = nn.Conv1d(cond_dim, cond_dim, ftrans_k)
self.post_cf2 = nn.Conv1d(cond_dim, cond_dim, ftrans_k)
self.post_af1 = nn.Conv1d(cond_dim, cond_dim, ftrans_k)
self.post_af2 = nn.Conv1d(cond_dim, cond_dim, ftrans_k)
self.post_af3 = nn.Conv1d(cond_dim, cond_dim, ftrans_k)
def create_phase_signals(self, periods):
batch_size = periods.size(0)
progression = torch.arange(1, self.FRAME_SIZE + 1, dtype=periods.dtype, device=periods.device).view((1, -1))
progression = torch.repeat_interleave(progression, batch_size, 0)
phase0 = torch.zeros(batch_size, dtype=periods.dtype, device=periods.device).unsqueeze(-1)
chunks = []
for sframe in range(periods.size(1)):
f = (2.0 * torch.pi / periods[:, sframe]).unsqueeze(-1)
if self.pulses:
alpha = torch.cos(f).view(batch_size, 1, 1)
chunk_sin = torch.sin(f * progression + phase0).view(batch_size, 1, self.FRAME_SIZE)
pulse_a = torch.relu(chunk_sin - alpha) / (1 - alpha)
pulse_b = torch.relu(-chunk_sin - alpha) / (1 - alpha)
chunk = torch.cat((pulse_a, pulse_b), dim = 1)
else:
chunk_sin = torch.sin(f * progression + phase0).view(batch_size, 1, self.FRAME_SIZE)
chunk_cos = torch.cos(f * progression + phase0).view(batch_size, 1, self.FRAME_SIZE)
chunk = torch.cat((chunk_sin, chunk_cos), dim = 1)
phase0 = phase0 + self.FRAME_SIZE * f
chunks.append(chunk)
phase_signals = torch.cat(chunks, dim=-1)
return phase_signals
def flop_count(self, rate=16000, verbose=False):
frame_rate = rate / self.FRAME_SIZE
# feature net
feature_net_flops = self.feature_net.flop_count(frame_rate)
comb_flops = self.cf1.flop_count(rate) + self.cf2.flop_count(rate)
af_flops = self.af1.flop_count(rate) + self.af2.flop_count(rate) + self.af3.flop_count(rate) + self.af4.flop_count(rate) + self.af_prescale.flop_count(rate) + self.af_mix.flop_count(rate)
feature_flops = (_conv1d_flop_count(self.post_cf1, frame_rate) + _conv1d_flop_count(self.post_cf2, frame_rate)
+ _conv1d_flop_count(self.post_af1, frame_rate) + _conv1d_flop_count(self.post_af2, frame_rate) + _conv1d_flop_count(self.post_af3, frame_rate))
if verbose:
print(f"feature net: {feature_net_flops / 1e6} MFLOPS")
print(f"comb filters: {comb_flops / 1e6} MFLOPS")
print(f"adaptive conv: {af_flops / 1e6} MFLOPS")
print(f"feature transforms: {feature_flops / 1e6} MFLOPS")
return feature_net_flops + comb_flops + af_flops + feature_flops
def feature_transform(self, f, layer):
f = f.permute(0, 2, 1)
f = F.pad(f, [self.ftrans_k - 1, 0])
f = torch.tanh(layer(f))
return f.permute(0, 2, 1)
def forward(self, features, periods, debug=False):
periods = periods.squeeze(-1)
pitch_embedding = self.pitch_embedding(periods)
full_features = torch.cat((features, pitch_embedding), dim=-1)
cf = self.feature_net(full_features)
# upsample periods
periods = torch.repeat_interleave(periods, self.upsamp_factor, 1)
# pre-net
ref_phase = torch.tanh(self.create_phase_signals(periods))
if debug: print_channels(ref_phase, prefix="lavoce_01", name="pulse")
x = self.af_prescale(ref_phase, cf)
noise = self.noise_shaper(cf)
if debug: print_channels(torch.cat((x, noise), dim=1), prefix="lavoce_02", name="inputs")
y = self.af_mix(torch.cat((x, noise), dim=1), cf)
if debug: print_channels(y, prefix="lavoce_03", name="postselect1")
# temporal shaping + innovating
y1 = y[:, 0:1, :]
y2 = self.tdshape1(y[:, 1:2, :], cf)
if debug: print_channels(y2, prefix="lavoce_04", name="postshape1")
y = torch.cat((y1, y2), dim=1)
y = self.af2(y, cf, debug=debug)
if debug: print_channels(y, prefix="lavoce_05", name="postselect2")
cf = self.feature_transform(cf, self.post_af2)
y1 = y[:, 0:1, :]
y2 = self.tdshape2(y[:, 1:2, :], cf)
if debug: print_channels(y2, prefix="lavoce_06", name="postshape2")
y = torch.cat((y1, y2), dim=1)
y = self.af3(y, cf, debug=debug)
if debug: print_channels(y, prefix="lavoce_07", name="postmix1")
cf = self.feature_transform(cf, self.post_af3)
# spectral shaping
y = self.cf1(y, cf, periods, debug=debug)
if debug: print_channels(y, prefix="lavoce_08", name="postcomb1")
cf = self.feature_transform(cf, self.post_cf1)
y = self.cf2(y, cf, periods, debug=debug)
if debug: print_channels(y, prefix="lavoce_09", name="postcomb2")
cf = self.feature_transform(cf, self.post_cf2)
y = self.af1(y, cf, debug=debug)
if debug: print_channels(y, prefix="lavoce_10", name="postselect3")
cf = self.feature_transform(cf, self.post_af1)
# final temporal env adjustment
y1 = y[:, 0:1, :]
y2 = self.tdshape3(y[:, 1:2, :], cf)
if debug: print_channels(y2, prefix="lavoce_11", name="postshape3")
y = torch.cat((y1, y2), dim=1)
y = self.af4(y, cf, debug=debug)
if debug: print_channels(y, prefix="lavoce_12", name="postmix2")
return y
def process(self, features, periods, debug=False):
self.eval()
device = next(iter(self.parameters())).device
with torch.no_grad():
# run model
f = features.unsqueeze(0).to(device)
p = periods.unsqueeze(0).to(device)
y = self.forward(f, p, debug=debug).squeeze()
# deemphasis
if self.preemph > 0:
for i in range(len(y) - 1):
y[i + 1] += self.preemph * y[i]
# clip to valid range
out = torch.clip((2**15) * y, -2**15, 2**15 - 1).short()
return out