def artanh(x):
if x.dtype == tf.float32:
result = tf.atanh(tf.clip_by_value(x, clip_value_min=tf.constant([-1], dtype=tf.float32)+1e-7, clip_value_max=tf.constant([1],dtype=tf.float32)-1e-7))
elif x.dtype == tf.float64:
result = tf.atanh(tf.clip_by_value(x, clip_value_min=tf.constant([-1], dtype=tf.float64)+1e-16, clip_value_max=tf.constant([1],dtype=tf.float64)-1e-16))
else:
raise ValueError('invalid dtype!')
return result
def pair_dy(self):
"""
Rapidity difference between all pairs of particles.
"""
# dy = y1 - y2 = atanh(beta1) - atanh(beta2)
beta = tf.clip_by_value(self.beta(), self.epsilon, 1 - self.epsilon)
dy = tf.atanh(tf.expand_dims(beta, axis=-1)) - tf.atanh(tf.expand_dims(beta, axis=-2))
# return only upper triangle without diagonal
return tf.gather(tf.reshape(dy, [-1, self.n**2]), self.triu_indices, axis=1)
def soft_round_inverse(y, alpha, eps=1e-3):
"""Inverse of soft_round().
This is described in Sec. 4.1. in the paper
> "Universally Quantized Neural Compression"<br />
> Eirikur Agustsson & Lucas Theis<br />
> https://arxiv.org/abs/2006.09952
Args:
y: tf.Tensor. Inputs to this function.
alpha: Float or tf.Tensor. Controls smoothness of the approximation.
eps: Float. Threshold below which soft_round() is assumed to equal the
identity function.
Returns:
tf.Tensor
"""
# This guards the gradient of tf.where below against NaNs, while maintaining
# correctness, as for alpha < eps the result is ignored.
alpha_bounded = tf.maximum(alpha, eps)
m = tf.floor(y) + .5
s = (y - m) * (tf.tanh(alpha_bounded / 2.) * 2.)
r = tf.atanh(s) / alpha_bounded
# `r` must be between -.5 and .5 by definition. In case atanh becomes +-inf
# due to numerical instability, this prevents the forward pass from yielding
# infinite values. Note that it doesn't prevent the backward pass from
# returning non-finite values.
r = tf.clip_by_value(r, -.5, .5)
# For very low alphas, soft_round behaves like identity.
return tf.where(alpha < eps, y, m + r, name="soft_round_inverse")
def nlogp(self, dist, action):
''' negative logp of unnormalized action '''
before_squahed_action = tf.atanh(
tf.clip_by_value(action, -1 + EPS, 1 - EPS))
log_likelihood = dist.log_prob(before_squahed_action)
log_likelihood -= tf.reduce_sum(tf.log(1 - action**2 + EPS), axis=1)
return -tf.reduce_mean(log_likelihood)
def testSampleFromDiscretizedMixLogistic(self):
batch = 2
height = 4
width = 4
num_mixtures = 5
seed = 42
logits = tf.concat( # assign all probability mass to first component
[tf.ones([batch, height, width, 1]) * 1e8,
tf.zeros([batch, height, width, num_mixtures - 1])],
axis=-1)
locs = tf.random_uniform([batch, height, width, num_mixtures * 3],
minval=-.9, maxval=.9)
log_scales = tf.ones([batch, height, width, num_mixtures * 3]) * -1e8
coeffs = tf.atanh(tf.zeros([batch, height, width, num_mixtures * 3]))
pred = tf.concat([logits, locs, log_scales, coeffs], axis=-1)
locs_0 = locs[..., :3]
expected_sample = tf.clip_by_value(locs_0, -1., 1.)
actual_sample = common_layers.sample_from_discretized_mix_logistic(
pred, seed=seed)
actual_sample_val, expected_sample_val = self.evaluate(
[actual_sample, expected_sample])
# Use a low tolerance: samples numerically differ, as the actual
# implementation clips log-scales so they always contribute to sampling.
self.assertAllClose(actual_sample_val, expected_sample_val, atol=1e-2)
def _graph_fn_unsquash(self, values):
if get_backend() == "tf":
return tf.atanh((values - self.low) /
(self.high - self.low) * 2.0 - 1.0)
elif get_backend() == "tf":
return torch.atanh((values - self.low) /
(self.high - self.low) * 2.0 - 1.0)
def neglogp(self, x):
neglogp_likelihood = 0.5 * tf.reduce_sum(tf.square((tf.atanh(x)-self.mean)/(self.std+self.EPS)), axis=-1) \
+ 0.5 * np.log(2.0 * np.pi) * tf.cast(tf.shape(x)[-1], tf.float32) \
+ tf.reduce_sum(self.logstd, axis=-1)
policy = self.sample()
return neglogp_likelihood + tf.reduce_sum(
tf.log(1 - policy**2 + self.EPS), axis=-1)
def mlp_actor_critic(x, a, hidden_sizes=(400,300), activation=tf.nn.relu,
output_activation=None, policy=mlp_gaussian_policy, action_space=None):
action_scale = action_space.high[0]
a_unsqueeze = a / action_scale
a_unsqueeze = tf.atanh(a_unsqueeze)
# policy
with tf.variable_scope('pi'):
mu, pi, logp_pi, logp_a = policy(x, a_unsqueeze, hidden_sizes, activation, output_activation)
mu, pi, logp_pi, logp_a = apply_squashing_func(mu, pi, logp_pi, a_unsqueeze, logp_a)
# make sure actions are in correct range
mu *= action_scale
pi *= action_scale
# vfs
vf_mlp = lambda x: tf.squeeze(mlp(x, list(hidden_sizes) + [1], activation, None), axis=1)
with tf.variable_scope('q1'):
q1 = vf_mlp(tf.concat([x, a], axis=-1))
with tf.variable_scope('q1', reuse=True):
q1_pi = vf_mlp(tf.concat([x, pi], axis=-1))
with tf.variable_scope('q2'):
q2 = vf_mlp(tf.concat([x, a], axis=-1))
with tf.variable_scope('q2', reuse=True):
q2_pi = vf_mlp(tf.concat([x, pi], axis=-1))
with tf.variable_scope('v'):
v = vf_mlp(x)
with tf.variable_scope('Q'):
Q = vf_mlp(tf.concat([x, a], axis=-1))
with tf.variable_scope('Q', reuse=True):
Q_pi = vf_mlp(tf.concat([x, pi], axis=-1))
with tf.variable_scope('R'):
R = vf_mlp(x)
return mu, pi, logp_pi, q1, q2, q1_pi, q2_pi, v, Q, Q_pi, R
def testDiscretizedMixLogisticLoss(self):
batch = 2
height = 4
width = 4
channels = 3
num_mixtures = 5
logits = tf.concat( # assign all probability mass to first component
[tf.ones([batch, height, width, 1]) * 1e8,
tf.zeros([batch, height, width, num_mixtures - 1])],
axis=-1)
locs = tf.random_uniform([batch, height, width, num_mixtures * 3],
minval=-.9, maxval=.9)
log_scales = tf.random_uniform([batch, height, width, num_mixtures * 3],
minval=-1., maxval=1.)
coeffs = tf.atanh(tf.zeros([batch, height, width, num_mixtures * 3]))
pred = tf.concat([logits, locs, log_scales, coeffs], axis=-1)
# Test labels that don't satisfy edge cases where 8-bit value is 0 or 255.
labels = tf.random_uniform([batch, height, width, channels],
minval=-.9, maxval=.9)
locs_0 = locs[..., :3]
log_scales_0 = log_scales[..., :3]
centered_labels = labels - locs_0
inv_stdv = tf.exp(-log_scales_0)
plus_in = inv_stdv * (centered_labels + 1. / 255.)
min_in = inv_stdv * (centered_labels - 1. / 255.)
cdf_plus = tf.nn.sigmoid(plus_in)
cdf_min = tf.nn.sigmoid(min_in)
expected_loss = -tf.reduce_sum(tf.log(cdf_plus - cdf_min), axis=-1)
actual_loss = common_layers.discretized_mix_logistic_loss(
pred=pred, labels=labels)
actual_loss_val, expected_loss_val = self.evaluate(
[actual_loss, expected_loss])
self.assertAllClose(actual_loss_val, expected_loss_val, rtol=1e-5)
def testSampleFromDiscretizedMixLogistic(self):
batch = 2
height = 4
width = 4
num_mixtures = 5
seed = 42
logits = tf.concat( # assign all probability mass to first component
[tf.ones([batch, height, width, 1]) * 1e8,
tf.zeros([batch, height, width, num_mixtures - 1])],
axis=-1)
locs = tf.random_uniform([batch, height, width, num_mixtures * 3],
minval=-.9, maxval=.9)
log_scales = tf.ones([batch, height, width, num_mixtures * 3]) * -1e8
coeffs = tf.atanh(tf.zeros([batch, height, width, num_mixtures * 3]))
pred = tf.concat([logits, locs, log_scales, coeffs], axis=-1)
locs_0 = locs[..., :3]
expected_sample = tf.clip_by_value(locs_0, -1., 1.)
actual_sample = common_layers.sample_from_discretized_mix_logistic(
pred, seed=seed)
actual_sample_val, expected_sample_val = self.evaluate(
[actual_sample, expected_sample])
# Use a low tolerance: samples numerically differ, as the actual
# implementation clips log-scales so they always contribute to sampling.
self.assertAllClose(actual_sample_val, expected_sample_val, atol=1e-2)
def soft_round_inverse(y, alpha, eps=1e-12):
"""Inverse of soft_round().
This is described in Sec. 4.1. in the paper
> "Universally Quantized Neural Compression"<br />
> Eirikur Agustsson & Lucas Theis<br />
> https://arxiv.org/abs/2006.09952
Args:
y: tf.Tensor. Inputs to this function.
alpha: Float or tf.Tensor. Controls smoothness of the approximation.
eps: Float. Threshold below which soft_round() is assumed to equal the
identity function.
Returns:
tf.Tensor
"""
if isinstance(alpha, (float, int)) and alpha < eps:
return tf.identity(y, name="soft_round_inverse")
m = tf.floor(y) + 0.5
s = (y - m) * (tf.tanh(alpha / 2.0) * 2.0)
# We have -0.5 <= (y-m) <= 0.5 and -1 < tanh < 1, so
# -1 <= s <= 1. However tf.atanh is only stable for inputs
# in the range [-1+1e-7, 1-1e-7], so we (safely) clip s to this range.
# In the rare case where `1-|s| < 1e-7`, we use straight-through for the
# gradient.
s = _clip_st(s)
r = tf.atanh(s) / tf.maximum(alpha, eps)
# For very low alphas, soft_round behaves like identity
return tf.where(alpha < eps, y, m + r, name="soft_round_inverse")
def eta(self):
"""
Pseudorapidity.
"""
return tf.atanh(
tf.clip_by_value(self.pz() / self.p(), self.epsilon - 1,
1 - self.epsilon))
def cont_bern_mean(lam, l_lim=0.49, u_lim=0.51):
# continuous Bernoulli mean funtion in tensorflow
# just like the normalizing constant, it is computed in a numerically stable way around 0.5
cut_lam = tf.where(tf.logical_or(tf.less(lam, l_lim), tf.greater(lam, u_lim)), lam, l_lim * tf.ones_like(lam))
mu = cut_lam / (2.0 * cut_lam - 1.0) + 1.0 / (2.0 * tf.atanh(1.0 - 2.0 * cut_lam))
taylor = 0.5 + (lam - 0.5) / 3.0 + 16.0 / 45.0 * tf.pow(lam - 0.5, 3)
return tf.where(tf.logical_or(tf.less(lam, l_lim), tf.greater(lam, u_lim)), mu, taylor)
def h_log(c, x, input):
xpy = mobius_add(c, -x, input)
# print('xpy')
# print(xpy.shape)
# output = tf.identity(xpy, name = 'output')
# print(c)
# output = tf.identity(xpy, name = 'output')
xpy_norm = safe_norm(xpy, axis=-1, keepdims=True)
# output = tf.identity(xpy_norm, name = 'output')
# print('clip')
# print(xpy_norm.shape)
# xpy_norm = tf.clip_by_value(
# xpy_norm,
# clip_value_min = -1. / np.sqrt(c) * (1. - 1e-3),
# clip_value_max = 1. / np.sqrt(c) * (1. - 1e-3)
# )
# print(xpy_norm.shape)
output = (2. / (np.sqrt(c) * h_lambda(c, x)) *
tf.atanh(np.sqrt(c) * xpy_norm) * xpy / xpy_norm)
# output = tf.atanh(np.sqrt(c) * xpy_norm )
# print('norm')
# print(xpy.shape)
# print(safe_norm(xpy).shape)
return output
def apply_harmonic_bias(channels, num_layers):
"""Offset network outputs to ensure harmonic distribution of initial alpha.
The first num_layers-1 channels are the ones that will become the alpha
channels for layers [1, N-1]. (There is no channel corresponding to the alpha
of the back layer because is it always 1.0, i.e. fully opaque.)
We adjust these first num_layers-1 channels so that instead of all layer
alphas having an initial mean of 0.5, the Nth layer from the back has an
initial mean of 1/N. This harmonic distribution allows each layer to
contribute equal weight when the layers are composed.
Args:
channels: [..., N] Network output before final tanh activation.
num_layers: How many layers we are predicting an MPI for.
Returns:
[..., N] Adjusted output.
"""
# The range below begins at 2 because the back layer is not predicted, as it's
# always fully opaque.
alpha = 1.0 / tf.range(2, num_layers + 1, dtype=tf.float32)
# Convert to desired offset before activation and scaling:
shift = tf.atanh(2.0 * alpha - 1.0)
# Remaining channels are left as is.
no_shift = tf.zeros([tf.shape(channels)[-1] - (num_layers - 1)])
shift = tf.concat([shift, no_shift], axis=-1)
return channels + shift
def _inverse(self, y):
# 0.99999997 is the maximum value such that atanh(x) is valid for both
# tf.float32 and tf.float64
y = tf.where(tf.less_equal(tf.abs(y), 1.),
tf.clip_by_value(y, -0.99999997, 0.99999997),
y)
return tf.atanh(y)
def testDiscretizedMixLogisticLoss(self):
batch = 2
height = 4
width = 4
channels = 3
num_mixtures = 5
logits = tf.concat( # assign all probability mass to first component
[tf.ones([batch, height, width, 1]) * 1e8,
tf.zeros([batch, height, width, num_mixtures - 1])],
axis=-1)
locs = tf.random_uniform([batch, height, width, num_mixtures * 3],
minval=-.9, maxval=.9)
log_scales = tf.random_uniform([batch, height, width, num_mixtures * 3],
minval=-1., maxval=1.)
coeffs = tf.atanh(tf.zeros([batch, height, width, num_mixtures * 3]))
pred = tf.concat([logits, locs, log_scales, coeffs], axis=-1)
# Test labels that don't satisfy edge cases where 8-bit value is 0 or 255.
labels = tf.random_uniform([batch, height, width, channels],
minval=-.9, maxval=.9)
locs_0 = locs[..., :3]
log_scales_0 = log_scales[..., :3]
centered_labels = labels - locs_0
inv_stdv = tf.exp(-log_scales_0)
plus_in = inv_stdv * (centered_labels + 1. / 255.)
min_in = inv_stdv * (centered_labels - 1. / 255.)
cdf_plus = tf.nn.sigmoid(plus_in)
cdf_min = tf.nn.sigmoid(min_in)
expected_loss = -tf.reduce_sum(tf.log(cdf_plus - cdf_min), axis=-1)
actual_loss = common_layers.discretized_mix_logistic_loss(
pred=pred, labels=labels)
actual_loss_val, expected_loss_val = self.evaluate(
[actual_loss, expected_loss])
self.assertAllClose(actual_loss_val, expected_loss_val, rtol=1e-5)
def log_pis_for(self, actions):
if self._squash:
raw_actions = tf.atanh(actions)
log_pis = self._distribution.log_prob(raw_actions)
log_pis -= self._squash_correction(raw_actions)
return log_pis
return self._distribution.log_prob(raw_actions)
def h_matmul(c, M, x):
Mx = tf.matmul(M, x)
output = (1. / np.sqrt(c) * tf.tanh(
safe_norm(Mx) / safe_norm(x) * tf.atanh(np.sqrt(c) * safe_norm(x))) *
Mx / safe_norm(Mx))
return output
def _inverse(self, y):
dtype = y.dtype
y = tf.cast(y, tf.float32)
y = tf.where(tf.less_equal(tf.abs(y), 1.),
tf.clip_by_value(y, -0.99999997, 0.99999997), y)
y = tf.atanh(y)
y = tf.cast(y, dtype)
return y
def logpac(self, action):
from stable_baselines.sac.policies import gaussian_likelihood, EPS
act_mu = self.policy_tf.act_mu
log_std = tf.log(self.policy_tf.std)
# Potentially we need to clip atanh and pass gradient
log_u = gaussian_likelihood(
tf.atanh(tf.clip_by_value(action, -0.99, 0.99)), act_mu, log_std)
log_ac = log_u - tf.reduce_sum(tf.log(1 - action**2 + EPS), axis=1)
return log_ac
def neglogp(self, x):
if self.squash:
return 0.5 * tf.reduce_sum(tf.square((tf.atanh(x) - self.mean) / self.std), axis=-1) \
+ 0.5 * np.log(2.0 * np.pi) * tf.cast(tf.shape(x)[-1], tf.float32) \
+ tf.reduce_sum(self.logstd, axis=-1) + tf.reduce_sum(tf.log(1-x**2+1e-6), axis=-1)
else:
return 0.5 * tf.reduce_sum(tf.square((x - self.mean) / self.std), axis=-1) \
+ 0.5 * np.log(2.0 * np.pi) * tf.cast(tf.shape(x)[-1], tf.float32) \
+ tf.reduce_sum(self.logstd, axis=-1)
def unsquash_action(mu, pi, log_std):
"""
desquash action from [-1, 1] to [-inf, inf]
"""
_pi = tf.atanh(pi)
log_pi = Policy.gaussian_likelihood(_pi, mu, log_std)
sub = tf.reduce_sum(tf.math.log(Policy.clip_but_pass_gradient(1 - pi**2, l=0, h=1) + 1e-6), axis=1, keepdims=True)
log_pi -= sub
return log_pi
def cont_bern_log_norm(lam, l_lim=0.49, u_lim=0.51):
# computes the log normalizing constant of a continuous Bernoulli distribution in a numerically stable way.
# returns the log normalizing constant for lam in (0, l_lim) U (u_lim, 1) and a Taylor approximation in
# [l_lim, u_lim].
# cut_y below might appear useless, but it is important to not evaluate log_norm near 0.5 as tf.where evaluates
# both options, regardless of the value of the condition.
cut_lam = tf.where(tf.logical_or(tf.less(lam, l_lim), tf.greater(lam, u_lim)), lam, l_lim * tf.ones_like(lam))
log_norm = tf.log(tf.abs(2.0 * tf.atanh(1 - 2.0 * cut_lam))) - tf.log(tf.abs(1 - 2.0 * cut_lam))
taylor = tf.log(2.0) + 4.0 / 3.0 * tf.pow(lam - 0.5, 2) + 104.0 / 45.0 * tf.pow(lam - 0.5, 4)
return tf.where(tf.logical_or(tf.less(lam, l_lim), tf.greater(lam, u_lim)), log_norm, taylor)
def tf_my_mob_mat_distance(mat_x, mat_y):
# input shape: [features, nodes]
mat = tf_my_mob_mat_addition(-mat_x, mat_y)
# mat = mat + EPS
mat_norm = tf.norm(mat, axis=2)
mat_norm = tf.clip_by_value(mat_norm,
clip_value_min=1e-8,
clip_value_max=clip_value)
res = 2. * tf.atanh(mat_norm)
return res
def log_prob(self, value, **kwargs):
if self.squash:
# from SAC paper: https://arxiv.org/pdf/1801.01290.pdf
u = tf.atanh(value)
correction = tf.reduce_sum(tf.log(1 - value ** 2 + EPSILON), axis=1)
# correction = tf.reduce_sum(tf.log1p(-tf.square(value) + EPSILON), axis=1)
log_prob = super().log_prob(u, **kwargs) - correction
else:
log_prob = super().log_prob(value, **kwargs)
log_prob = tf.reduce_sum(log_prob, axis=-1)
return log_prob
def AdaIN_adv_tanh(content, epsilon=1e-5):
meanC, varC = tf.nn.moments(content, [1, 2], keep_dims=True)
bs = settings.config["BATCH_SIZE"]
content_shape = content.shape.as_list()
new_shape = [bs, 1, 1, content_shape[3]]
with tf.variable_scope("scale"):
sigmaS = tf.get_variable("sigma_S", shape=new_shape,
initializer=tf.zeros_initializer())
meanS = tf.get_variable("mean_S", shape=new_shape,
initializer=tf.zeros_initializer())
sigmaC = tf.sqrt(tf.add(varC, epsilon))
p=tf.sqrt(1.5)
def get_mid_range(l,r):
_mid=(l+r)/2.0
_range=(r-l)/2.0
return _mid,_range
sign=tf.sign(meanC)
abs_meanC=tf.abs(meanC)
_sigma_mid, _sigma_range = get_mid_range(sigmaC/p, sigmaC*p)
_mean_mid, _mean_range = get_mid_range(abs_meanC/p, abs_meanC*p)
sigmaSp = _sigma_range*tf.nn.tanh(sigmaS)+_sigma_mid
meanSp = sign * (_mean_range*tf.nn.tanh(meanS)+_mean_mid)
ops_bound = []
ops_asgn = [tf.assign(sigmaS, tf.atanh((sigmaC-_sigma_mid)/ (_sigma_range +1e-4) )),
tf.assign(meanS, tf.atanh((abs_meanC-_mean_mid)/(_mean_range + 1e-4) ))]
#ops_asgn = [sigmaS.initializer, meanS.initializer]#
#ops_asgn = [tf.assign(sigmaS, sigmaC-_sigma_mid),
# tf.assign(meanS, meanC-_mean_mid)]
return (content - meanC) * sigmaSp / sigmaC + meanSp , ops_asgn, ops_bound, sigmaSp, meanSp, meanS, sigmaS
def _build_baseline_policy_and_kl(self, target_dist, obs_input, action_input):
EPS = 1e-6
self.behavior_policy = Actor(self.action_dim, self.max_action, hidden_dim=self.hidden_dim)
_, behavior_action_logp, behavior_dist = self.behavior_policy([obs_input])
before_squahed_action = tf.atanh(tf.clip_by_value(action_input, -1 + EPS, 1 - EPS))
log_likelihood = behavior_dist.log_prob(before_squahed_action)
log_likelihood -= tf.reduce_sum(tf.log(1 - action_input ** 2 + EPS), axis=1)
behavior_loss = -tf.reduce_mean(log_likelihood)
behavior_optimizer = tf.train.AdamOptimizer(self.learning_rate)
behavior_train_op = behavior_optimizer.minimize(behavior_loss, var_list=self.behavior_policy.trainable_variables)
self.sess.run(tf.variables_initializer(behavior_optimizer.variables()))
return target_dist.kl_divergence(behavior_dist)[:, None], behavior_train_op, behavior_loss
def _define_ops(self):
super()._define_ops()
# Loss to be optimized by attacker.
self.loss: tf.Tensor = None
# A single step of the attack. Will be run in order.
self.step: List[tf.Tensor] = None
# The output perturbed image.
self.output: tf.Tensor = None
# >>> Your code here <<<
w = tf.Variable(tf.zeros(self.batch_shape), name="w")
Xadv = 0.5 * (tf.tanh(w) + 1)
logits = self.model.logits(Xadv)
self.init_inputs.append(
tf.assign(w, tf.atanh(1.9*(self.X_var-0.5)))
)
term1 = tf.reduce_sum(
tf.square(Xadv - self.X_var),
[1, 2, 3]
)
if self.target is None:
others_score = tf.reduce_max((1 - self.Yi) * logits, axis=1)
target_score = tf.reduce_sum(self.Yi * logits, axis=1)
term2 = tf.maximum(target_score - others_score, -self.k)
else:
target_onehot = tf.one_hot(
np.repeat(self.target, self.batch_size),
self.model.num_classes
)
others_score = tf.reduce_max((1 - target_onehot) * logits, axis=1)
target_score = tf.reduce_sum(target_onehot * logits, axis=1)
term2 = tf.maximum(others_score - target_score, -self.k)
self.loss = term1 + self.c * term2
grad = tf.gradients(ys=self.loss, xs=w)[0]
w_updated = tf.assign(w, w - self.lr * grad)
self.step = [w_updated,
tf.assign(
self.lr,
self.lr * self.learning_rate_decay
)]
self.output = tf.clip_by_value(Xadv, 0, 1)
def tf_my_prod_mat_log_map_zero(M, c):
sqrt_c = tf.sqrt(c)
# M = tf.transpose(M)
M = M + EPS
M = tf.clip_by_norm(M, clip_norm=clip_value, axes=0)
m_norm = tf.norm(M, axis=0)
atan_norm = tf.atanh(
tf.clip_by_value(m_norm * sqrt_c,
clip_value_min=-0.9,
clip_value_max=0.9))
M_cof = atan_norm / m_norm / sqrt_c
res = M * M_cof
return res
def _create_q_update(self):
"""Create a minimization operation for Q-function update."""
opponent_actions, opponent_actions_log_pis = self.opponent_policy.actions_for(
observations=self._next_observations_ph,
reuse=tf.AUTO_REUSE,
with_log_pis=True)
assert_shape(opponent_actions, [None, self._opponent_action_dim])
prior = self._get_opponent_prior(self._next_observations_ph)
raw_actions = tf.atanh(opponent_actions)
prior_log_pis = prior.dist.log_prob(raw_actions)
prior_log_pis = prior_log_pis - squash_correction(raw_actions)
actions, actions_log_pis = self.policy.actions_for(
observations=self._next_observations_ph,
reuse=tf.AUTO_REUSE,
with_log_pis=True,
opponent_actions=opponent_actions)
with tf.variable_scope('target_joint_q_agent_{}'.format(
self._agent_id),
reuse=tf.AUTO_REUSE):
q_value_targets = self.target_joint_qf.output_for(
observations=self._next_observations_ph,
actions=actions,
opponent_actions=opponent_actions)
q_value_targets = q_value_targets - self._annealing_pl * actions_log_pis - opponent_actions_log_pis + prior_log_pis
assert_shape(q_value_targets, [None])
self._q_values = self.joint_qf.output_for(self._observations_ph,
self._actions_pl,
self._opponent_actions_pl,
reuse=True)
assert_shape(self._q_values, [None])
ys = tf.stop_gradient(self._reward_scale * self._rewards_pl +
(1 - self._terminals_pl) * self._discount *
q_value_targets)
assert_shape(ys, [None])
bellman_residual = 0.5 * tf.reduce_mean((ys - self._q_values)**2)
with tf.variable_scope('target_joint_qf_opt_agent_{}'.format(
self._agent_id),
reuse=tf.AUTO_REUSE):
if self._train_qf:
td_train_op = tf.train.AdamOptimizer(self._qf_lr).minimize(
loss=bellman_residual,
var_list=self.joint_qf.get_params_internal())
self._training_ops.append(td_train_op)
self._bellman_residual = bellman_residual
def __init__(self,
grid_height,
grid_width,
target_control_points,
input_shape,
bounded=True,
**kwargs):
'''
tps_localizer will generate the source_control_point in the input images.
Input
--------
grid_height -- The y dimension of the target_control_points
grid_width -- The x dimension of the target_control_points
target_control_points -- [x,y] of shape (N,2)
input_shape -- The image 2D size of shape (H,W)
bounded -- If the grid extent is bounded from -1 to 1 or not
'''
super(tps_localizer, self).__init__(**kwargs)
assert tf.shape(target_control_points)[0] == grid_width * grid_height
self.output_dim = tf.shape(target_control_points)[0]
self.layer1 = tf.keras.layers.Conv2D(filters=10,
kernel_size=5,
input_shape=input_dims)
self.layer2 = tf.keras.layers.MaxPool2D(pool_size=2)
self.layer3 = tf.keras.layers.Activation.ReLU()
self.layer4 = tf.keras.layers.Conv2D(filters=20, kernel_size=5)
self.layer5 = tf.keras.layers.SpatialDropout2D(rate=0.5)
self.layer6 = tf.keras.layers.MaxPool2D(pool_size=2)
self.layer7 = tf.keras.layers.Activation.ReLU()
self.layer8 = tf.keras.layers.Flatten()
self.layer9 = tf.keras.layers.Dense(units=50, activation='relu')
self.layer10 = tf.keras.layers.Dropout(rate=0.5)
if bounded:
self.layer11 = tf.keras.layers.Dense(
units=self.output_dim,
activation='tanh',
kernel_initializer=tf.keras.initializers.Zeros(),
bias_initializer=tf.keras.initializers.Constant(
tf.atanh(target_control_points), dtype="float32"))
else:
self.layer11 = tf.keras.layers.Dense(
units=self.output_dim,
activation='linear',
kernel_initializer=tf.keras.initializers.Zeros(),
bias_initializer=tf.keras.initializers.Constant(
target_control_points, dtype="float32"))
def _inverse(self, y): return tf.atanh(y)
def adv_train_arctan_net(input_images, clip_norm=1.5):
with tf.variable_scope('adv_encoder') as scope:
width = 32
height = 32
batch_size = 128
# code_length = 6000
input_images = input_images/255
arctan_images = tf.atanh(((input_images*2)-1)*0.999999)
mean, var = tf.nn.moments(arctan_images, axes=tuple(range(1,len(input_images.shape))), keep_dims=True)
normed_input_images = (arctan_images-mean)/var
# Convolutional layer 1
conv1 = tf.layers.conv2d(inputs=normed_input_images,
filters=64,
kernel_size=(5, 5),
# kernel_initializer=tf.contrib.layers.xavier_initializer(),
activation=tf.nn.leaky_relu,
padding='SAME',
name='adv_conv1')
# maxpool layer1
maxpool1 = tf.layers.max_pooling2d(conv1, (3,3), (2,2), 'SAME')
# Convolutional layer 2
conv2 = tf.layers.conv2d(inputs=maxpool1,
filters=128,
kernel_size=(5, 5),
# kernel_initializer=tf.contrib.layers.xavier_initializer(),
activation=tf.nn.leaky_relu,
padding='SAME',
name='adv_conv2')
# maxpool layer2
maxpool2 = tf.layers.max_pooling2d(conv2, (3,3), (2,2), 'SAME')
deconv1 = tf.layers.conv2d_transpose(maxpool2, 64, (5,5), (2,2), 'SAME',
activation=tf.nn.leaky_relu,
name='adv_deconv1')
adv_mask = tf.layers.conv2d_transpose(deconv1, 3, (5,5), (2,2), 'SAME',
# activation=tf.nn.tanh,
name='adv_deconv2')
arctan_adv_images = adv_mask + normed_input_images
unscaled_adv_images = tf.tanh(arctan_adv_images)
unscaled_diff = unscaled_adv_images - input_images
# clip_norm = 1.5
scaled_dif = tf.clip_by_norm(unscaled_diff, clip_norm)
adv_images = tf.clip_by_value(scaled_dif+input_images,0,1)
output_images = tf.reshape(adv_images, (batch_size, height, width, 3)) * 255.0
dif = adv_images - input_images
# Display the training images in the visualizer.
tf.summary.image('adv_images', output_images)
# Reconstruction L2 loss
mean_square_error = tf.reduce_mean(tf.square(dif), axis=list(range(1,len(dif.shape))))
loss = tf.reduce_mean(mean_square_error, name='dis_loss')
return loss, output_images