def inv_bias_act(x,
act='linear',
alpha=0.2,
gain=None,
lrmul=1,
bias_var='bias',
reverse=False):
assert act in ['linear', 'lrelu']
x = tf.transpose(x, [0, 3, 1, 2])
b = tf.get_variable(bias_var,
shape=[x.shape[1]],
initializer=tf.initializers.zeros()) * lrmul
if reverse:
if act == 'lrelu':
if gain is None:
gain = np.sqrt(2)
x = x / gain
mask = tf.cast(x < 0, x.dtype) * (1.0 / alpha - 1.0) + 1.0
x = x * mask
x = fused_bias_act(x, -b)
return tf.transpose(x, [0, 2, 3, 1])
return tf.transpose(
fused_bias_act(x,
b=tf.cast(b, x.dtype),
act=act,
alpha=alpha,
gain=gain), [0, 2, 3, 1])
def apply_bias_act(x, act="linear", alpha=None, gain=None, lrmul=1, bias_var="bias"):
b = (
tf.get_variable(
bias_var, shape=[x.shape[1]], initializer=tf.initializers.zeros()
)
* lrmul
)
return fused_bias_act(x, b=tf.cast(b, x.dtype), act=act, alpha=alpha, gain=gain)
def apply_bias_act(x,
act='linear',
gain=None,
lrmul=1,
clamp=None,
bias_var='bias',
trainable=True):
b = tf.get_variable(bias_var,
shape=[x.shape[1]],
initializer=tf.initializers.zeros(),
trainable=trainable) * lrmul
return fused_bias_act(x,
b=tf.cast(b, x.dtype),
act=act,
gain=gain,
clamp=clamp)
def apply_bias_act(x,
act='linear',
alpha=None,
gain=None,
lrmul=1,
bias_var='bias',
impl='cuda'):
b = tf.get_variable(bias_var,
shape=[x.shape[1]],
initializer=tf.initializers.zeros()) * lrmul
return fused_bias_act(x,
b=tf.cast(b, x.dtype),
act=act,
alpha=alpha,
gain=gain,
impl=impl)