def grad_variance(self):
grad_var_ops = []
tensor_to_avg = []
for t, g in zip(self._tvars, self._grads):
if isinstance(g, ops.IndexedSlices):
tensor_to_avg.append(
tf.reshape(tf.unsorted_segment_sum(g.values, g.indices,
g.dense_shape[0]),
shape=t.get_shape()))
else:
tensor_to_avg.append(g)
avg_op = self._moving_averager.apply(tensor_to_avg)
grad_var_ops.append(avg_op)
with tf.control_dependencies([avg_op]):
self._grad_avg = [
self._moving_averager.average(val) for val in tensor_to_avg
]
self._grad_avg_squared = [tf.square(val) for val in self._grad_avg]
self._grad_var = tf.maximum(
tf.constant(EPS, dtype=self._grad_norm_squared_avg.dtype),
self._grad_norm_squared_avg -
tf.add_n([tf.reduce_sum(val) for val in self._grad_avg_squared]))
if self._sparsity_debias:
self._grad_var *= self._sparsity_avg
return grad_var_ops
def extract_feature(waveforms, params):
'''extract fbank with delta-delta and do cmvn
waveforms: [batch, samples]
'''
p = params
with tf.variable_scope('feature_extractor'):
mel_fbanks = extract_logfbank_with_delta(waveforms, params)
# shape: [1, nframes, nbins, nchannels]
fbank_size = utils.shape_list(mel_fbanks)
#assert fbank_size[0] == 1
# This replaces CMVN estimation on data
if not p.audio_global_cmvn:
mean = tf.reduce_mean(mel_fbanks, keepdims=True, axis=1)
variance = tf.reduce_mean(tf.square(mel_fbanks - mean),
keepdims=True,
axis=1)
else:
assert p.audio_cmvn_path, p.audio_cmvn_path
mean, variance = utils.load_cmvn(p.audio_cmvn_path)
var_epsilon = 1e-09
mel_fbanks = utils.apply_cmvn(mel_fbanks, mean, variance, var_epsilon)
# Later models like to flatten the two spatial dims. Instead, we add a
# unit spatial dim and flatten the frequencies and channels.
batch_size = fbank_size[0]
feats = tf.concat([
tf.reshape(
mel_fbanks,
[batch_size, fbank_size[1], fbank_size[2], fbank_size[3]]),
tf.zeros((batch_size, p.num_zeropad_frames, fbank_size[2],
fbank_size[3]))
], 1)
return feats # shape [batch_size, nframes, featue_size, chnanels]
def call(self,
logits=None,
input_length=None,
labels=None,
label_length=None,
**kwargs):
assert "soft_lables" in kwargs
soft_labels = kwargs["soft_labels"]
loss_standard = cross_entropy(
logits=logits,
input_length=input_length,
labels=labels,
label_length=label_length,
smoothing=self.smoothing)
loss_soft = cross_entropy(
logits=logits / self.T,
input_length=input_length,
labels=soft_labels,
label_length=label_length,
smoothing=self.smoothing)
# Since the magnitudes of the gradients produced by the soft targets
# scale as 1/T2 , it is important to multiply them by T2 when using
# both hard and soft targets
total_loss = self.alpha * tf.square(
self.T) * loss_soft + (1 - self.alpha) * loss_standard
return total_loss
def before_apply(self):
self._moving_averager = tf.train.ExponentialMovingAverage(
decay=self._beta, zero_debias=self._zero_debias)
assert self._grads is not None and len(self._grads) > 0
before_apply_ops = []
# get per var g**2 and norm**2
self._grad_squared = []
self._grad_norm_squared = []
for v, g in zip(self._tvars, self._grads):
if g is None:
continue
with ops.colocate_with(v):
self._grad_squared.append(tf.square(g))
self._grad_norm_squared = [
tf.reduce_sum(grad_squared) for grad_squared in self._grad_squared
]
if self._sparsity_debias:
avg_op_sparsity = self.grad_sparsity()
before_apply_ops.append(avg_op_sparsity)
# the following running average on squared norm of gradient is shared
# by `grad_variance` and `dist_to_opt`
avg_op = self._moving_averager.apply(self._grad_norm_squared)
with tf.control_dependencies([avg_op]):
self._grad_norm_squared_avg = [
self._moving_averager.average(val)
for val in self._grad_norm_squared
]
self._grad_norm_squared = tf.add_n(self._grad_norm_squared)
self._grad_norm_squared_avg = tf.add_n(self._grad_norm_squared_avg)
before_apply_ops.append(avg_op)
with tf.control_dependencies([avg_op]):
curv_range_ops = self.curvature_range()
before_apply_ops += curv_range_ops
grad_var_ops = self.grad_variance()
before_apply_ops += grad_var_ops
dist_to_opt_ops = self.dist_to_opt()
before_apply_ops += dist_to_opt_ops
return tf.group(*before_apply_ops)
def arcface_loss(embedding,
labels,
out_num,
weights=None,
s=64.,
m=0.5,
limit_to_pi=True):
'''
https://github.com/auroua/InsightFace_TF/blob/master/losses/face_losses.py
:param embedding: the input embedding vectors
:param labels: the input labels, the shape should be eg: (batch_size, 1)
:param s: scalar value default is 64
:param out_num: output class num
:param weights: a tf.variable with shape (embedding.shape[-1], out_num)
or None to make a new one internally. default = None
:param m: the margin value, default is 0.5
:return: the final cacualted output, this output is send into the tf.nn.softmax directly
'''
cos_m = math.cos(m)
sin_m = math.sin(m)
mm = sin_m * m # issue 1
threshold = math.cos(math.pi - m)
with tf.variable_scope('arcface_loss'):
# inputs and weights norm
embedding_norm = tf.norm(embedding, axis=1, keep_dims=True)
embedding = tf.div(embedding, embedding_norm, name='norm_embedding')
if weights is None:
weights = tf.get_variable(
name='weights',
shape=[embedding.shape[-1].value, out_num],
initializer=tf.initializer.glorot_unifrom())
weights_norm = tf.norm(weights, axis=0, keep_dims=True)
weights = tf.div(weights, weights_norm, name='norm_weights')
# cos(theta+m)
cos_t = tf.matmul(embedding, weights, name='cos_t')
cos_t2 = tf.square(cos_t, name='cos_2')
sin_t2 = tf.subtract(1., cos_t2, name='sin_2')
sin_t = tf.sqrt(sin_t2, name='sin_t')
cos_mt = s * tf.subtract(tf.multiply(cos_t, cos_m),
tf.multiply(sin_t, sin_m),
name='cos_mt')
if limit_to_pi:
# this condition controls the theta+m should in range [0, pi]
# 0<=theta+m<=pi
# -m<=theta<=pi-m
cond_v = cos_t - threshold
cond = tf.cast(tf.nn.relu(cond_v, name='if_else'), dtype=tf.bool)
keep_val = s * (cos_t - mm)
cos_mt_temp = tf.where(cond, cos_mt, keep_val)
else:
cos_mt_temp = cos_mt
mask = tf.one_hot(labels, depth=out_num, name='one_hot_mask')
# mask = tf.squeeze(mask, 1)
inv_mask = tf.subtract(1., mask, name='inverse_mask')
s_cos_t = tf.multiply(s, cos_t, name='scalar_cos_t')
output = tf.add(tf.multiply(s_cos_t, inv_mask),
tf.multiply(cos_mt_temp, mask),
name='arcface_loss_output')
return output