File size: 10,741 Bytes
aed64b5 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 |
import logging
import os
import numpy as np
import tensorflow as tf
# pylint: disable=E0611,E0401
import tensorflow.keras.backend as K
# pylint: disable=E0611,E0401
from tensorflow.keras import layers, regularizers
# pylint: disable=E0611,E0401
from tensorflow.keras.layers import (
BatchNormalization,
Conv2D,
Dense,
Dropout,
Input,
Lambda,
Reshape,
)
# pylint: disable=E0611,E0401
from tensorflow.keras.models import Model
# pylint: disable=E0611,E0401
from tensorflow.keras.optimizers import Adam
from deep_speaker.constants import NUM_FBANKS, SAMPLE_RATE, NUM_FRAMES
from deep_speaker.triplet_loss import deep_speaker_loss
logger = logging.getLogger(__name__)
@tf.function
def tf_normalize(data, ndims, eps=0, adjusted=False):
data = tf.convert_to_tensor(data, name='data')
reduce_dims = [-i - 1 for i in range(ndims)]
# pylint: disable=E1123,E1120
data = tf.cast(data, dtype=tf.dtypes.float32)
data_num = tf.reduce_prod(data.shape[-ndims:])
data_mean = tf.reduce_mean(data, axis=reduce_dims, keepdims=True)
# Apply a minimum normalization that protects us against uniform images.
stddev = tf.math.reduce_std(data, axis=reduce_dims, keepdims=True)
adjusted_stddev = stddev
if adjusted:
min_stddev = tf.math.rsqrt(tf.cast(data_num, tf.dtypes.float32))
eps = tf.maximum(eps, min_stddev)
if eps > 0:
adjusted_stddev = tf.maximum(adjusted_stddev, eps)
return (data - data_mean) / adjusted_stddev
@tf.function
def tf_fbank(samples):
"""
Compute Mel-filterbank energy features from an audio signal.
See python_speech_features.fbank
"""
frame_length = int(0.025 * SAMPLE_RATE)
frame_step = int(0.01 * SAMPLE_RATE)
fft_length = 512
fft_bins = fft_length // 2 + 1
pre_emphasis = samples[:, 1:] - 0.97 * samples[:, :-1]
# Original implementation from python_speech_features
# frames = tf.expand_dims(sigproc.framesig(preemphasis[0], frame_length,
# frame_step, winfunc=lambda x: np.ones((x,))), 0)
# powspec = sigproc.powspec(frames, fft_length)
# Tensorflow impl #1, using manually-split frames and rfft
# spec = tf.abs(tf.signal.rfft(frames, [fft_length]))
# powspec = tf.square(spec) / fft_length
# Tensorflow impl #2, using stft to handle framing automatically
# (There is a one-off mismatch on the number of frames on the resulting tensor, but I guess this is ok)
spec = tf.abs(tf.signal.stft(pre_emphasis, frame_length, frame_step, fft_length, window_fn=tf.ones))
powspec = tf.square(spec) / fft_length
# Matrix to transform spectrum to mel-frequencies
# Original implementation from python_speech_features
# linear_to_mel_weight_matrix = get_filterbanks(NUM_FBANKS, fft_length,
# SAMPLE_RATE, 0, SAMPLE_RATE/2).astype(np.float32).T
linear_to_mel_weight_matrix = tf.signal.linear_to_mel_weight_matrix(
num_mel_bins=NUM_FBANKS,
num_spectrogram_bins=fft_bins,
sample_rate=SAMPLE_RATE,
lower_edge_hertz=0,
upper_edge_hertz=SAMPLE_RATE / 2,
)
feat = tf.matmul(powspec, linear_to_mel_weight_matrix)
# feat = tf.where(feat == 0, np.finfo(np.float32).eps, feat)
return feat
class DeepSpeakerModel:
# I thought it was 3 but maybe energy is added at a 4th dimension.
# would be better to have 4 dimensions:
# MFCC, DIFF(MFCC), DIFF(DIFF(MFCC)), ENERGIES (probably tiled across the frequency domain).
# this seems to help match the parameter counts.
def __init__(
self,
batch_input_shape=(None, NUM_FRAMES, NUM_FBANKS, 1),
include_softmax=False,
num_speakers_softmax=None,
pcm_input=False
):
if pcm_input:
batch_input_shape = None
self.include_softmax = include_softmax
if self.include_softmax:
assert num_speakers_softmax > 0
self.clipped_relu_count = 0
# http://cs231n.github.io/convolutional-networks/
# conv weights
# #params = ks * ks * nb_filters * num_channels_input
# Conv128-s
# 5*5*128*128/2+128
# ks*ks*nb_filters*channels/strides+bias(=nb_filters)
# take 100 ms -> 4 frames.
# if signal is 3 seconds, then take 100ms per 100ms and average out this network.
# 8*8 = 64 features.
# used to share all the layers across the inputs
# num_frames = K.shape() - do it dynamically after.
if pcm_input:
batch_input_shape = batch_input_shape or (None, None) # Batch-size, num-samples
inputs = Input(batch_shape=batch_input_shape, name='raw_inputs')
x = inputs
x = Lambda(tf_fbank)(x)
x = Lambda(lambda x_: tf_normalize(x_, 1, 1e-12))(x)
x = Lambda(lambda x_: tf.expand_dims(x_, axis=-1))(x)
else:
batch_input_shape = batch_input_shape or (None, None, NUM_FBANKS, 1)
inputs = Input(batch_shape=batch_input_shape, name='input')
x = inputs
x = self.cnn_component(x)
x = Reshape((-1, 2048))(x)
# Temporal average layer. axis=1 is time.
x = Lambda(lambda y: K.mean(y, axis=1), name='average')(x)
if include_softmax:
logger.info('Including a Dropout layer to reduce overfitting.')
# used for softmax because the dataset we pre-train on might be too small. easy to overfit.
x = Dropout(0.5)(x)
x = Dense(512, name='affine')(x)
if include_softmax:
# Those weights are just when we train on softmax.
x = Dense(num_speakers_softmax, activation='softmax')(x)
else:
# Does not contain any weights.
x = Lambda(lambda y: K.l2_normalize(y, axis=1), name='ln')(x)
self.m = Model(inputs, x, name='ResCNN')
def keras_model(self):
return self.m
def get_weights(self):
w = self.m.get_weights()
if self.include_softmax:
w.pop() # last 2 are the W_softmax and b_softmax.
w.pop()
return w
def clipped_relu(self, inputs):
relu = Lambda(lambda y: K.minimum(K.maximum(y, 0), 20), name=f'clipped_relu_{self.clipped_relu_count}')(inputs)
self.clipped_relu_count += 1
return relu
def identity_block(self, input_tensor, kernel_size, filters, stage, block):
conv_name_base = f'res{stage}_{block}_branch'
x = Conv2D(filters,
kernel_size=kernel_size,
strides=1,
activation=None,
padding='same',
kernel_initializer='glorot_uniform',
kernel_regularizer=regularizers.l2(l=0.0001),
name=conv_name_base + '_2a')(input_tensor)
x = BatchNormalization(name=conv_name_base + '_2a_bn')(x)
x = self.clipped_relu(x)
x = Conv2D(
filters,
kernel_size=kernel_size,
strides=1,
activation=None,
padding='same',
kernel_initializer='glorot_uniform',
kernel_regularizer=regularizers.l2(l=0.0001),
name=conv_name_base + '_2b',
)(x)
x = BatchNormalization(name=conv_name_base + '_2b_bn')(x)
x = self.clipped_relu(x)
x = layers.add([x, input_tensor])
x = self.clipped_relu(x)
return x
def conv_and_res_block(self, inp, filters, stage):
conv_name = 'conv{}-s'.format(filters)
# TODO: why kernel_regularizer?
o = Conv2D(filters,
kernel_size=5,
strides=2,
activation=None,
padding='same',
kernel_initializer='glorot_uniform',
kernel_regularizer=regularizers.l2(l=0.0001), name=conv_name)(inp)
o = BatchNormalization(name=conv_name + '_bn')(o)
o = self.clipped_relu(o)
for i in range(3):
o = self.identity_block(o, kernel_size=3, filters=filters, stage=stage, block=i)
return o
def cnn_component(self, inp):
x = self.conv_and_res_block(inp, 64, stage=1)
x = self.conv_and_res_block(x, 128, stage=2)
x = self.conv_and_res_block(x, 256, stage=3)
x = self.conv_and_res_block(x, 512, stage=4)
return x
def set_weights(self, w):
for layer, layer_w in zip(self.m.layers, w):
layer.set_weights(layer_w)
logger.info(f'Setting weights for [{layer.name}]...')
def main():
# Looks correct to me.
# I have 37K but paper reports 41K. which is not too far.
dsm = DeepSpeakerModel()
dsm.m.summary()
# I suspect num frames to be 32.
# Then fbank=64, then total would be 32*64 = 2048.
# plot_model(dsm.m, to_file='model.png', dpi=300, show_shapes=True, expand_nested=True)
def _train():
# x = np.random.uniform(size=(6, 32, 64, 4)) # 6 is multiple of 3.
# y_softmax = np.random.uniform(size=(6, 100))
# dsm = DeepSpeakerModel(batch_input_shape=(None, 32, 64, 4), include_softmax=True, num_speakers_softmax=100)
# dsm.m.compile(optimizer=Adam(lr=0.01), loss='categorical_crossentropy')
# print(dsm.m.predict(x).shape)
# print(dsm.m.evaluate(x, y_softmax))
# w = dsm.get_weights()
dsm = DeepSpeakerModel(batch_input_shape=(None, 32, 64, 4), include_softmax=False)
# dsm.m.set_weights(w)
dsm.m.compile(optimizer=Adam(lr=0.01), loss=deep_speaker_loss)
# it works!!!!!!!!!!!!!!!!!!!!
# unit_batch_size = 20
# anchor = np.ones(shape=(unit_batch_size, 32, 64, 4))
# positive = np.array(anchor)
# negative = np.ones(shape=(unit_batch_size, 32, 64, 4)) * (-1)
# batch = np.vstack((anchor, positive, negative))
# x = batch
# y = np.zeros(shape=(len(batch), 512)) # not important.
# print('Starting to fit...')
# while True:
# print(dsm.m.train_on_batch(x, y))
# should not work... and it does not work!
unit_batch_size = 20
negative = np.ones(shape=(unit_batch_size, 32, 64, 4)) * (-1)
batch = np.vstack((negative, negative, negative))
x = batch
y = np.zeros(shape=(len(batch), 512)) # not important.
print('Starting to fit...')
while True:
print(dsm.m.train_on_batch(x, y))
def _test_checkpoint_compatibility():
dsm = DeepSpeakerModel(batch_input_shape=(None, 32, 64, 4), include_softmax=True, num_speakers_softmax=10)
dsm.m.save_weights('test.h5')
dsm = DeepSpeakerModel(batch_input_shape=(None, 32, 64, 4), include_softmax=False)
dsm.m.load_weights('test.h5', by_name=True)
os.remove('test.h5')
if __name__ == '__main__':
_test_checkpoint_compatibility()
|