def he_normal(seed=None, scale=2.):
"""He normal initializer.
It draws samples from a truncated normal distribution centered on 0
with standard deviation (after truncation) given by
`stddev = sqrt(2 / fan_in)` where `fan_in` is the number of
input units in the weight tensor.
Parameters
----------
seed: A Python integer. Used to seed the random generator.
Returns
-------
An initializer.
References
----------
[He et al., 2015]
(https://www.cv-foundation.org/openaccess/content_iccv_2015/html/He_Delving_Deep_into_ICCV_2015_paper.html)
# pylint: disable=line-too-long
([pdf](https://www.cv-foundation.org/openaccess/content_iccv_2015/papers/He_Delving_Deep_into_ICCV_2015_paper.pdf))
"""
return VarianceScaling(scale=scale,
mode="fan_in",
distribution="truncated_normal",
seed=seed)
def he_uniform(seed=None, mode='fan_in', dtype=tf.float32):
'''
mode: a string. 'fan_in', 'fan_out' or 'fan_avg'
'''
return VarianceScaling(scale=2.0,
mode=mode,
distribution='uniform',
seed=seed,
dtype=tf.float32)
def he_normal_fanout(seed=None):
"""He normal initializer.
It draws samples from a truncated normal distribution centered on 0
with `stddev = sqrt(2 / fan_out)`
where `fan_in` is the number of input units in the weight tensor.
To keep aligned with official implementation
"""
return VarianceScaling(
scale=2., mode="fan_out", distribution="truncated_normal", seed=seed)
def create(self):
X_img = tf.placeholder(tf.float32, shape=(None, 84, 84, 4))
Y = tf.placeholder(tf.float32, shape=(None, ))
a = tf.placeholder(tf.int32, shape=(None, 2))
initializer = VarianceScaling(1.0)
initializer_bias = tf.zeros_initializer()
W_conv1 = tf.Variable(initializer([8, 8, 4, 32]))
b_conv1 = tf.Variable(initializer_bias(shape=(32, )))
A_conv1 = tf.nn.relu(
tf.nn.conv2d(X_img, W_conv1, strides=[1, 4, 4, 1], padding='SAME')
+ b_conv1)
W_conv2 = tf.Variable(initializer([4, 4, 32, 64]))
b_conv2 = tf.Variable(initializer_bias(shape=(64, )))
A_conv2 = tf.nn.relu(
tf.nn.conv2d(
A_conv1, W_conv2, strides=[1, 2, 2, 1], padding='SAME') +
b_conv2)
W_conv3 = tf.Variable(initializer([3, 3, 64, 64]))
b_conv3 = tf.Variable(initializer_bias(shape=(64, )))
A_conv3 = tf.nn.relu(
tf.nn.conv2d(
A_conv2, W_conv3, strides=[1, 1, 1, 1], padding='SAME') +
b_conv3)
flatten = tf.reshape(A_conv3, [-1, 11 * 11 * 64])
n_neurons_1 = 512
n_neurons_2 = self.n_actions
W_fc1 = tf.Variable(initializer([11 * 11 * 64, n_neurons_1]))
b_fc1 = tf.Variable(initializer_bias(n_neurons_1))
Z_fc1 = tf.nn.relu(tf.matmul(flatten, W_fc1) + b_fc1)
W_fc2 = tf.Variable(initializer([n_neurons_1, n_neurons_2]))
b_fc2 = tf.Variable(initializer_bias(n_neurons_2))
Qs = tf.matmul(Z_fc1, W_fc2) + b_fc2
action = tf.argmax(Qs, axis=1)
max_q = tf.reduce_max(Qs, axis=1)
loss = tf.reduce_mean(
tf.square(tf.stop_gradient(Y) - tf.gather_nd(Qs, a)))
optimizer = tf.train.RMSPropOptimizer(self.lr, self.decay)
train = optimizer.minimize(loss)
saver = tf.train.Saver()
return (X_img, Y, train, loss, saver, W_conv2, action, Qs, a)
def he_uniform(seed=None):
"""He uniform variance scaling initializer.
It draws samples from a uniform distribution within [-limit, limit]
where `limit` is `sqrt(6 / fan_in)`
where `fan_in` is the number of input units in the weight tensor.
Arguments:
seed: A Python integer. Used to seed the random generator.
Returns:
An initializer.
References:
He et al., http://arxiv.org/abs/1502.01852
"""
return VarianceScaling(
scale=2., mode='fan_in', distribution='uniform', seed=seed)
def he_normal(seed=None):
"""He normal initializer.
It draws samples from a truncated normal distribution centered on 0
with `stddev = sqrt(2 / fan_in)`
where `fan_in` is the number of input units in the weight tensor.
Arguments:
seed: A Python integer. Used to seed the random generator.
Returns:
An initializer.
References:
He et al., http://arxiv.org/abs/1502.01852
"""
return VarianceScaling(
scale=2., mode='fan_in', distribution='normal', seed=seed)
def lecun_uniform(seed=None):
"""LeCun uniform initializer.
It draws samples from a uniform distribution within [-limit, limit]
where `limit` is `sqrt(3 / fan_in)`
where `fan_in` is the number of input units in the weight tensor.
Arguments:
seed: A Python integer. Used to seed the random generator.
Returns:
An initializer.
References:
LeCun 98, Efficient Backprop,
http://yann.lecun.com/exdb/publis/pdf/lecun-98b.pdf
"""
return VarianceScaling(
scale=1., mode='fan_in', distribution='uniform', seed=seed)
def glorot_normal(seed=None):
"""Glorot normal initializer, also called Xavier normal initializer.
It draws samples from a truncated normal distribution centered on 0
with `stddev = sqrt(2 / (fan_in + fan_out))`
where `fan_in` is the number of input units in the weight tensor
and `fan_out` is the number of output units in the weight tensor.
Arguments:
seed: A Python integer. Used to seed the random generator.
Returns:
An initializer.
References:
Glorot & Bengio, AISTATS 2010
http://jmlr.org/proceedings/papers/v9/glorot10a/glorot10a.pdf
"""
return VarianceScaling(
scale=1., mode='fan_avg', distribution='normal', seed=seed)
def glorot_uniform(seed=None):
"""Glorot uniform initializer, also called Xavier uniform initializer.
It draws samples from a uniform distribution within [-limit, limit]
where `limit` is `sqrt(6 / (fan_in + fan_out))`
where `fan_in` is the number of input units in the weight tensor
and `fan_out` is the number of output units in the weight tensor.
Arguments:
seed: A Python integer. Used to seed the random generator.
Returns:
An initializer.
References:
Glorot & Bengio, AISTATS 2010
http://jmlr.org/proceedings/papers/v9/glorot10a/glorot10a.pdf
"""
return VarianceScaling(
scale=1., mode='fan_avg', distribution='uniform', seed=seed)
def lecun_normal(seed=None):
"""LeCun normal initializer.
It draws samples from a truncated normal distribution centered on 0
with `stddev = sqrt(1 / fan_in)`
where `fan_in` is the number of input units in the weight tensor.
Arguments:
seed: A Python integer. Used to seed the random generator.
Returns:
An initializer.
References:
- [Self-Normalizing Neural Networks](https://arxiv.org/abs/1706.02515)
- [Efficient
Backprop](http://yann.lecun.com/exdb/publis/pdf/lecun-98b.pdf)
"""
return VarianceScaling(
scale=1., mode='fan_in', distribution='normal', seed=seed)
def initializer(mode="FAN_IN",
distribution="normal",
is_relu=False,
seed=None):
""" create xavier initializer, include he_initializer for relu
args:
mode: "FAN_IN"(default) or "FAN_OUT" or "FAN_AVG"
distribution: "normal"(default) or "uniform"
is_relu: True, norm variance by 2
in case we output to a ReLU unit
Reference: http://jmlr.org/proceedings/papers/v9/glorot10a/glorot10a.pdf
https://arxiv.org/abs/1502.01852 by Kaiming He
return:
initializer op
"""
mode = mode.lower()
scale = 2. if is_relu else 1.
return VarianceScaling(scale=scale,
mode=mode,
distribution=distribution,
seed=seed)
def selu_normal(seed=None):
return VarianceScaling(scale=1.,
mode='fan_in',
distribution='normal',
seed=seed)
cnn3 = tf.nn.relu(conv_layer3)
flat = tf.contrib.layers.flatten(cnn3)
fc1 = tf.matmul(flat, weights["fc1"]) + bias["fc1"]
fc1 = tf.nn.relu(fc1)
Z = tf.matmul(fc1, weights["fc2"]) + bias["fc2"]
return Z # n x 4
weights_online = {
"cnn1":
tf.get_variable("o_cnn1",
shape=[8, 8, 4, 32],
initializer=VarianceScaling(),
trainable=True,
dtype=tf.float32),
"cnn2":
tf.get_variable("o_cnn2",
shape=[4, 4, 32, 64],
initializer=VarianceScaling(),
trainable=True,
dtype=tf.float32),
"cnn3":
tf.get_variable("o_cnn3",
shape=[3, 3, 64, 64],
initializer=VarianceScaling(),
trainable=True,
dtype=tf.float32),
"fc1":
def lecun_uniform(seed=None):
return VarianceScaling(scale=1.,
mode='fan_in',
distribution='uniform',
seed=seed)
def lecun_normal(seed=None):
return VarianceScaling(scale=1.,
mode='fan_in',
distribution='truncated_normal',
seed=seed)
def glorot_normal(seed=None, dtype=tf.float32):
return VarianceScaling(scale=1.0,
mode='fan_avg',
distribution='normal',
seed=seed,
dtype=tf.float32)