opus/dnn/training_tf2/lpcnet.py
Jean-Marc Valin d45ab6fcb6 Move back to tanh for frame rate network
Swish has lower loss, but doesn't seem to improve quality
2022-09-24 03:22:57 -04:00

339 lines
15 KiB
Python

#!/usr/bin/python3
'''Copyright (c) 2018 Mozilla
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 FOUNDATION 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 math
import tensorflow as tf
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input, GRU, Dense, Embedding, Reshape, Concatenate, Lambda, Conv1D, Multiply, Add, Bidirectional, MaxPooling1D, Activation, GaussianNoise
from tensorflow.compat.v1.keras.layers import CuDNNGRU
from tensorflow.keras import backend as K
from tensorflow.keras.constraints import Constraint
from tensorflow.keras.initializers import Initializer
from tensorflow.keras.callbacks import Callback
from mdense import MDense
import numpy as np
import h5py
import sys
from tf_funcs import *
from diffembed import diff_Embed
from parameters import set_parameter
frame_size = 160
pcm_bits = 8
embed_size = 128
pcm_levels = 2**pcm_bits
def interleave(p, samples):
p2=tf.expand_dims(p, 3)
nb_repeats = pcm_levels//(2*p.shape[2])
p3 = tf.reshape(tf.repeat(tf.concat([1-p2, p2], 3), nb_repeats), (-1, samples, pcm_levels))
return p3
def tree_to_pdf(p, samples):
return interleave(p[:,:,1:2], samples) * interleave(p[:,:,2:4], samples) * interleave(p[:,:,4:8], samples) * interleave(p[:,:,8:16], samples) \
* interleave(p[:,:,16:32], samples) * interleave(p[:,:,32:64], samples) * interleave(p[:,:,64:128], samples) * interleave(p[:,:,128:256], samples)
def tree_to_pdf_train(p):
#FIXME: try not to hardcode the 2400 samples (15 frames * 160 samples/frame)
return tree_to_pdf(p, 2400)
def tree_to_pdf_infer(p):
return tree_to_pdf(p, 1)
def quant_regularizer(x):
Q = 128
Q_1 = 1./Q
#return .01 * tf.reduce_mean(1 - tf.math.cos(2*3.1415926535897931*(Q*x-tf.round(Q*x))))
return .01 * tf.reduce_mean(K.sqrt(K.sqrt(1.0001 - tf.math.cos(2*3.1415926535897931*(Q*x-tf.round(Q*x))))))
class Sparsify(Callback):
def __init__(self, t_start, t_end, interval, density, quantize=False):
super(Sparsify, self).__init__()
self.batch = 0
self.t_start = t_start
self.t_end = t_end
self.interval = interval
self.final_density = density
self.quantize = quantize
def on_batch_end(self, batch, logs=None):
#print("batch number", self.batch)
self.batch += 1
if self.quantize or (self.batch > self.t_start and (self.batch-self.t_start) % self.interval == 0) or self.batch >= self.t_end:
#print("constrain");
layer = self.model.get_layer('gru_a')
w = layer.get_weights()
p = w[1]
nb = p.shape[1]//p.shape[0]
N = p.shape[0]
#print("nb = ", nb, ", N = ", N);
#print(p.shape)
#print ("density = ", density)
for k in range(nb):
density = self.final_density[k]
if self.batch < self.t_end and not self.quantize:
r = 1 - (self.batch-self.t_start)/(self.t_end - self.t_start)
density = 1 - (1-self.final_density[k])*(1 - r*r*r)
A = p[:, k*N:(k+1)*N]
A = A - np.diag(np.diag(A))
#This is needed because of the CuDNNGRU strange weight ordering
A = np.transpose(A, (1, 0))
L=np.reshape(A, (N//4, 4, N//8, 8))
S=np.sum(L*L, axis=-1)
S=np.sum(S, axis=1)
SS=np.sort(np.reshape(S, (-1,)))
thresh = SS[round(N*N//32*(1-density))]
mask = (S>=thresh).astype('float32')
mask = np.repeat(mask, 4, axis=0)
mask = np.repeat(mask, 8, axis=1)
mask = np.minimum(1, mask + np.diag(np.ones((N,))))
#This is needed because of the CuDNNGRU strange weight ordering
mask = np.transpose(mask, (1, 0))
p[:, k*N:(k+1)*N] = p[:, k*N:(k+1)*N]*mask
#print(thresh, np.mean(mask))
if self.quantize and ((self.batch > self.t_start and (self.batch-self.t_start) % self.interval == 0) or self.batch >= self.t_end):
if self.batch < self.t_end:
threshold = .5*(self.batch - self.t_start)/(self.t_end - self.t_start)
else:
threshold = .5
quant = np.round(p*128.)
res = p*128.-quant
mask = (np.abs(res) <= threshold).astype('float32')
p = mask/128.*quant + (1-mask)*p
w[1] = p
layer.set_weights(w)
class SparsifyGRUB(Callback):
def __init__(self, t_start, t_end, interval, grua_units, density, quantize=False):
super(SparsifyGRUB, self).__init__()
self.batch = 0
self.t_start = t_start
self.t_end = t_end
self.interval = interval
self.final_density = density
self.grua_units = grua_units
self.quantize = quantize
def on_batch_end(self, batch, logs=None):
#print("batch number", self.batch)
self.batch += 1
if self.quantize or (self.batch > self.t_start and (self.batch-self.t_start) % self.interval == 0) or self.batch >= self.t_end:
#print("constrain");
layer = self.model.get_layer('gru_b')
w = layer.get_weights()
p = w[0]
N = p.shape[0]
M = p.shape[1]//3
for k in range(3):
density = self.final_density[k]
if self.batch < self.t_end and not self.quantize:
r = 1 - (self.batch-self.t_start)/(self.t_end - self.t_start)
density = 1 - (1-self.final_density[k])*(1 - r*r*r)
A = p[:, k*M:(k+1)*M]
#This is needed because of the CuDNNGRU strange weight ordering
A = np.reshape(A, (M, N))
A = np.transpose(A, (1, 0))
N2 = self.grua_units
A2 = A[:N2, :]
L=np.reshape(A2, (N2//4, 4, M//8, 8))
S=np.sum(L*L, axis=-1)
S=np.sum(S, axis=1)
SS=np.sort(np.reshape(S, (-1,)))
thresh = SS[round(M*N2//32*(1-density))]
mask = (S>=thresh).astype('float32')
mask = np.repeat(mask, 4, axis=0)
mask = np.repeat(mask, 8, axis=1)
A = np.concatenate([A2*mask, A[N2:,:]], axis=0)
#This is needed because of the CuDNNGRU strange weight ordering
A = np.transpose(A, (1, 0))
A = np.reshape(A, (N, M))
p[:, k*M:(k+1)*M] = A
#print(thresh, np.mean(mask))
if self.quantize and ((self.batch > self.t_start and (self.batch-self.t_start) % self.interval == 0) or self.batch >= self.t_end):
if self.batch < self.t_end:
threshold = .5*(self.batch - self.t_start)/(self.t_end - self.t_start)
else:
threshold = .5
quant = np.round(p*128.)
res = p*128.-quant
mask = (np.abs(res) <= threshold).astype('float32')
p = mask/128.*quant + (1-mask)*p
w[0] = p
layer.set_weights(w)
class PCMInit(Initializer):
def __init__(self, gain=.1, seed=None):
self.gain = gain
self.seed = seed
def __call__(self, shape, dtype=None):
num_rows = 1
for dim in shape[:-1]:
num_rows *= dim
num_cols = shape[-1]
flat_shape = (num_rows, num_cols)
if self.seed is not None:
np.random.seed(self.seed)
a = np.random.uniform(-1.7321, 1.7321, flat_shape)
#a[:,0] = math.sqrt(12)*np.arange(-.5*num_rows+.5,.5*num_rows-.4)/num_rows
#a[:,1] = .5*a[:,0]*a[:,0]*a[:,0]
a = a + np.reshape(math.sqrt(12)*np.arange(-.5*num_rows+.5,.5*num_rows-.4)/num_rows, (num_rows, 1))
return self.gain * a.astype("float32")
def get_config(self):
return {
'gain': self.gain,
'seed': self.seed
}
class WeightClip(Constraint):
'''Clips the weights incident to each hidden unit to be inside a range
'''
def __init__(self, c=2):
self.c = c
def __call__(self, p):
# Ensure that abs of adjacent weights don't sum to more than 127. Otherwise there's a risk of
# saturation when implementing dot products with SSSE3 or AVX2.
return self.c*p/tf.maximum(self.c, tf.repeat(tf.abs(p[:, 1::2])+tf.abs(p[:, 0::2]), 2, axis=1))
#return K.clip(p, -self.c, self.c)
def get_config(self):
return {'name': self.__class__.__name__,
'c': self.c}
constraint = WeightClip(0.992)
def new_lpcnet_model(rnn_units1=384, rnn_units2=16, nb_used_features=20, batch_size=128, training=False, adaptation=False, quantize=False, flag_e2e = False, cond_size=128, lpc_order=16, lpc_gamma=1., lookahead=2):
pcm = Input(shape=(None, 1), batch_size=batch_size)
dpcm = Input(shape=(None, 3), batch_size=batch_size)
feat = Input(shape=(None, nb_used_features), batch_size=batch_size)
pitch = Input(shape=(None, 1), batch_size=batch_size)
dec_feat = Input(shape=(None, cond_size))
dec_state1 = Input(shape=(rnn_units1,))
dec_state2 = Input(shape=(rnn_units2,))
padding = 'valid' if training else 'same'
fconv1 = Conv1D(cond_size, 3, padding=padding, activation='tanh', name='feature_conv1')
fconv2 = Conv1D(cond_size, 3, padding=padding, activation='tanh', name='feature_conv2')
pembed = Embedding(256, 64, name='embed_pitch')
cat_feat = Concatenate()([feat, Reshape((-1, 64))(pembed(pitch))])
cfeat = fconv2(fconv1(cat_feat))
fdense1 = Dense(cond_size, activation='tanh', name='feature_dense1')
fdense2 = Dense(cond_size, activation='tanh', name='feature_dense2')
if flag_e2e and quantize:
fconv1.trainable = False
fconv2.trainable = False
fdense1.trainable = False
fdense2.trainable = False
cfeat = fdense2(fdense1(cfeat))
error_calc = Lambda(lambda x: tf_l2u(x[0] - tf.roll(x[1],1,axis = 1)))
if flag_e2e:
lpcoeffs = diff_rc2lpc(name = "rc2lpc")(cfeat)
else:
lpcoeffs = Input(shape=(None, lpc_order), batch_size=batch_size)
real_preds = diff_pred(name = "real_lpc2preds")([pcm,lpcoeffs])
weighting = lpc_gamma ** np.arange(1, 17).astype('float32')
weighted_lpcoeffs = Lambda(lambda x: x[0]*x[1])([lpcoeffs, weighting])
tensor_preds = diff_pred(name = "lpc2preds")([pcm,weighted_lpcoeffs])
past_errors = error_calc([pcm,tensor_preds])
embed = diff_Embed(name='embed_sig',initializer = PCMInit())
cpcm = Concatenate()([tf_l2u(pcm),tf_l2u(tensor_preds),past_errors])
cpcm = GaussianNoise(.3)(cpcm)
cpcm = Reshape((-1, embed_size*3))(embed(cpcm))
cpcm_decoder = Reshape((-1, embed_size*3))(embed(dpcm))
rep = Lambda(lambda x: K.repeat_elements(x, frame_size, 1))
quant = quant_regularizer if quantize else None
if training:
rnn = CuDNNGRU(rnn_units1, return_sequences=True, return_state=True, name='gru_a', stateful=True,
recurrent_constraint = constraint, recurrent_regularizer=quant)
rnn2 = CuDNNGRU(rnn_units2, return_sequences=True, return_state=True, name='gru_b', stateful=True,
kernel_constraint=constraint, recurrent_constraint = constraint, kernel_regularizer=quant, recurrent_regularizer=quant)
else:
rnn = GRU(rnn_units1, return_sequences=True, return_state=True, recurrent_activation="sigmoid", reset_after='true', name='gru_a', stateful=True,
recurrent_constraint = constraint, recurrent_regularizer=quant)
rnn2 = GRU(rnn_units2, return_sequences=True, return_state=True, recurrent_activation="sigmoid", reset_after='true', name='gru_b', stateful=True,
kernel_constraint=constraint, recurrent_constraint = constraint, kernel_regularizer=quant, recurrent_regularizer=quant)
rnn_in = Concatenate()([cpcm, rep(cfeat)])
md = MDense(pcm_levels, activation='sigmoid', name='dual_fc')
gru_out1, _ = rnn(rnn_in)
gru_out1 = GaussianNoise(.005)(gru_out1)
gru_out2, _ = rnn2(Concatenate()([gru_out1, rep(cfeat)]))
ulaw_prob = Lambda(tree_to_pdf_train)(md(gru_out2))
if adaptation:
rnn.trainable=False
rnn2.trainable=False
md.trainable=False
embed.Trainable=False
m_out = Concatenate(name='pdf')([tensor_preds,real_preds,ulaw_prob])
if not flag_e2e:
model = Model([pcm, feat, pitch, lpcoeffs], m_out)
else:
model = Model([pcm, feat, pitch], [m_out, cfeat])
model.rnn_units1 = rnn_units1
model.rnn_units2 = rnn_units2
model.nb_used_features = nb_used_features
model.frame_size = frame_size
if not flag_e2e:
encoder = Model([feat, pitch], cfeat)
dec_rnn_in = Concatenate()([cpcm_decoder, dec_feat])
else:
encoder = Model([feat, pitch], [cfeat,lpcoeffs])
dec_rnn_in = Concatenate()([cpcm_decoder, dec_feat])
dec_gru_out1, state1 = rnn(dec_rnn_in, initial_state=dec_state1)
dec_gru_out2, state2 = rnn2(Concatenate()([dec_gru_out1, dec_feat]), initial_state=dec_state2)
dec_ulaw_prob = Lambda(tree_to_pdf_infer)(md(dec_gru_out2))
if flag_e2e:
decoder = Model([dpcm, dec_feat, dec_state1, dec_state2], [dec_ulaw_prob, state1, state2])
else:
decoder = Model([dpcm, dec_feat, dec_state1, dec_state2], [dec_ulaw_prob, state1, state2])
# add parameters to model
set_parameter(model, 'lpc_gamma', lpc_gamma, dtype='float64')
set_parameter(model, 'flag_e2e', flag_e2e, dtype='bool')
set_parameter(model, 'lookahead', lookahead, dtype='int32')
return model, encoder, decoder