Coverage for hypergan/ops/tensorflow/activations.py : 27%

import numpy as np 

import tensorflow as tf 

def lrelu(x, leak=0.2, name="lrelu"): 

with tf.variable_scope(name): 

f1 = 0.5 * (1 + leak) 

f2 = 0.5 * (1 - leak) 

return f1 * x + f2 * abs(x) 

# http://stackoverflow.com/questions/39975676/how-to-implement-prelu-activation-in-tensorflow 

def prelu(prefix, i, _x): 

name = (prefix+"prelu_"+str(i)) 

orig_shape = _x.get_shape() 

_x = tf.reshape(_x, [config['batch_size'], -1]) 

#print("prelu for", _x.get_shape()[-1]) 

alphas = tf.get_variable(name, 

_x.get_shape()[-1], 

initializer=tf.random_normal_initializer(mean=0.0,stddev=0.01), 

dtype=tf.float32) 

pos = tf.nn.relu(_x) 

neg = alphas * (_x - abs(_x)) * 0.5 

return tf.reshape(pos + neg, orig_shape) 

def selu(x): 

alpha = 1.6732632423543772848170429916717 

scale = 1.0507009873554804934193349852946 

return scale*tf.where(x>=0.0, x, alpha*tf.nn.elu(x)) 

def sin_and_cos(x, name="ignored"): 

return tf.concat(axis=len(x.get_shape()) - 1, values=[tf.sin(x), tf.cos(x)]) 

def maxout(x, k = 2): 

shape = [int(e) for e in x.get_shape()] 

ax = len(shape) 

ch = shape[-1] 

assert ch % k == 0 

shape[-1] = ch // k 

shape.append(k) 

x = tf.reshape(x, shape) 

return tf.reduce_max(x, ax) 

rng = np.random.RandomState([2016, 6, 1]) 

def offset_maxout(x, k = 2): 

shape = [int(e) for e in x.get_shape()] 

ax = len(shape) 

ch = shape[-1] 

assert ch % k == 0 

shape[-1] = ch // k 

shape.append(k) 

x = tf.reshape(x, shape) 

ofs = rng.randn(1000, k).max(axis=1).mean() 

return tf.reduce_max(x, ax) - ofs 

def lrelu_sq(x): 

""" 

Concatenates lrelu and square 

""" 

dim = len(x.get_shape()) - 1 

return tf.concat(axis=dim, values=[lrelu(x), tf.minimum(tf.abs(x), tf.square(x))]) 

def decayer(x, name="decayer"): 

with tf.variable_scope(name): 

scale = tf.get_variable("scale", [1], initializer=tf.constant_initializer(1.,dtype=config['dtype']),dtype=config['dtype']) 

decay_scale = tf.get_variable("decay_scale", [1], initializer=tf.constant_initializer(1.,dtype=config['dtype']),dtype=config['dtype']) 

relu = tf.nn.relu(x) 

return scale * relu / (1. + tf.abs(decay_scale) * tf.square(decay_scale)) 

def decayer2(x, name="decayer"): 

with tf.variable_scope(name): 

scale = tf.get_variable("scale", [int(x.get_shape()[-1])], initializer=tf.constant_initializer(1.,dtype=config['dtype']),dtype=config['dtype']) 

decay_scale = tf.get_variable("decay_scale", [int(x.get_shape()[-1])], initializer=tf.constant_initializer(1.,dtype=config['dtype']), dtype=config['dtype']) 

relu = tf.nn.relu(x) 

return scale * relu / (1. + tf.abs(decay_scale) * tf.square(decay_scale)) 

def masked_relu(x, name="ignored"): 

shape = [int(e) for e in x.get_shape()] 

prefix = [0] * (len(shape) - 1) 

most = shape[:-1] 

assert shape[-1] % 2 == 0 

half = shape[-1] // 2 

first_half = tf.slice(x, prefix + [0], most + [half]) 

second_half = tf.slice(x, prefix + [half], most + [half]) 

return tf.nn.relu(first_half) * tf.nn.sigmoid(second_half) 

def _phase_shift(I, r): 

# Helper function with main phase shift operation 

bsize, a, b, c = I.get_shape().as_list() 

X = tf.reshape(I, (bsize, a, b, r, r)) 

X = tf.transpose(X, (0, 1, 2, 4, 3)) # bsize, a, b, 1, 1 

X = tf.split(axis=1, num_or_size_splits=a, value=X) # a, [bsize, b, r, r] 

X = tf.concat(axis=2, values=[tf.squeeze(x) for x in X]) # bsize, b, a*r, r 

X = tf.split(axis=1, num_or_size_splits=b, value=X) # b, [bsize, a*r, r] 

X = tf.concat(axis=2, values=[tf.squeeze(x) for x in X]) # 

bsize, a*r, b*r 

return tf.reshape(X, (bsize, a*r, b*r, 1)) 

def phase_shift(X, r, color=False): 

# Main OP that you can arbitrarily use in you tensorflow code 

if color: 

Xc = tf.split(axis=3, num_or_size_splits=3, value=X) 

X = tf.concat(axis=3, values=[_phase_shift(x, r) for x in Xc]) 

else: 

X = _phase_shift(X, r) 

return X 

def minmax(net): 

net = tf.minimum(net, 1) 

net = tf.maximum(net, -1) 

return net 

def minmaxzero(net): 

net = tf.minimum(net, 1) 

net = tf.maximum(net, 0) 

return net