def transform(data, network, time_factor=1):
x = data.placeholder
from returnn.tf.compat import v1 as tf
# summary("features", x)
step = network.global_train_step
step1 = tf.where(tf.greater_equal(step, 1000), 1, 0)
step2 = tf.where(tf.greater_equal(step, 2000), 1, 0)
def get_masked():
x_masked = x
x_masked = random_mask(
x_masked,
batch_axis=data.batch_dim_axis,
axis=data.time_dim_axis,
min_num=step1 + step2,
max_num=tf.maximum(tf.shape(x)[data.time_dim_axis] // 100, 2) *
(1 + step1 + step2 * 2),
max_dims=20 // time_factor)
x_masked = random_mask(x_masked,
batch_axis=data.batch_dim_axis,
axis=data.feature_dim_axis,
min_num=step1 + step2,
max_num=2 + step1 + step2 * 2,
max_dims=data.dim // 5)
#summary("features_mask", x_masked)
return x_masked
x = network.cond_on_train(get_masked, lambda: x)
return x
def transform(data, network, time_factor=1):
x = data.placeholder
from returnn.tf.compat import v1 as tf
# summary("features", x)
step = network.global_train_step
step1 = tf.where(tf.greater_equal(step, 1000), 1, 0)
step2 = tf.where(tf.greater_equal(step, 2000), 1, 0)
def get_masked():
x_masked = x
x_masked = random_mask(
x_masked,
batch_axis=data.batch_dim_axis,
axis=data.time_dim_axis,
min_num=step1 + step2,
max_num=tf.maximum(tf.shape(x)[data.time_dim_axis] // 100, 2) *
(1 + step1 + step2 * 2),
max_dims=20 // time_factor)
x_masked = random_mask(x_masked,
batch_axis=data.batch_dim_axis,
axis=data.feature_dim_axis,
min_num=step1 + step2,
max_num=2 + step1 + step2 * 2,
max_dims=data.dim // 5)
#summary("features_mask", x_masked)
return x_masked
cond1 = network.train_flag
cond2 = tf.greater_equal(
tf.shape(x)[data.time_dim_axis],
20) # ignore specaug for utterances less than 20 frames
x = tf.cond(tf.logical_and(cond1, cond2), get_masked, lambda: x)
return x
def transform(source, **kwargs):
from returnn.tf.compat import v1 as tf
data = source(0, as_data=True)
time_factor = 1 # for switchout == 6
x = data.placeholder
network = kwargs["self"].network
from returnn.tf.compat import v1 as tf
step = network.global_train_step
step1 = tf.where(tf.greater_equal(step, 1000), 1, 0)
step2 = tf.where(tf.greater_equal(step, 2000), 1, 0)
def get_masked():
x_masked = x
x_masked = random_mask(
x_masked,
batch_axis=data.batch_dim_axis,
axis=data.time_dim_axis,
min_num=step1 + step2,
max_num=tf.maximum(tf.shape(x)[data.time_dim_axis] // 100, 2) *
(1 + step1 + step2 * 2),
max_dims=20 // time_factor)
x_masked = random_mask(x_masked,
batch_axis=data.batch_dim_axis,
axis=data.feature_dim_axis,
min_num=step1 + step2,
max_num=2 + step1 + step2 * 2,
max_dims=data.dim // 5)
return x_masked
x = network.cond_on_train(get_masked, lambda: x)
return x
def _mask(x, batch_axis, axis, pos, max_amount):
"""
:param tf.Tensor x: (batch,time,feature)
:param int batch_axis:
:param int axis:
:param tf.Tensor pos: (batch,)
:param int|tf.Tensor max_amount: inclusive
"""
from returnn.tf.compat import v1 as tf
ndim = x.get_shape().ndims
n_batch = tf.shape(x)[batch_axis]
dim = tf.shape(x)[axis]
amount = tf.random_uniform(shape=(n_batch, ),
minval=1,
maxval=max_amount + 1,
dtype=tf.int32)
pos2 = tf.minimum(pos + amount, dim)
idxs = tf.expand_dims(tf.range(0, dim), 0) # (1,dim)
pos_bc = tf.expand_dims(pos, 1) # (batch,1)
pos2_bc = tf.expand_dims(pos2, 1) # (batch,1)
cond = tf.logical_and(tf.greater_equal(idxs, pos_bc),
tf.less(idxs, pos2_bc)) # (batch,dim)
if batch_axis > axis:
cond = tf.transpose(cond) # (dim,batch)
cond = tf.reshape(cond, [
tf.shape(x)[i] if i in (batch_axis, axis) else 1 for i in range(ndim)
])
from TFUtil import where_bc
x = where_bc(cond, 0.0, x)
return x
def _get_mask(x, axis, pos, max_amount):
"""
:param tf.Tensor x: (batch,time,feature)
:param int axis:
:param tf.Tensor pos: (batch,)
:param int max_amount: inclusive
"""
from returnn.tf.compat import v1 as tf
n_batch = tf.shape(x)[0]
dim = tf.shape(x)[axis]
amount = tf.random_uniform(shape=(n_batch, ),
minval=1,
maxval=max_amount + 1,
dtype=tf.int32)
pos2 = tf.minimum(pos + amount, dim)
idxs = tf.expand_dims(tf.range(0, dim), 0) # (1,dim)
pos_bc = tf.expand_dims(pos, 1) # (batch,1)
pos2_bc = tf.expand_dims(pos2, 1) # (batch,1)
cond = tf.logical_and(tf.greater_equal(idxs, pos_bc),
tf.less(idxs, pos2_bc)) # (batch,dim)
return cond