def forward(img_a, img_b):
img_a /= 255.
img_b /= 255.
img_ab = generator(img_a, name='atob', reuse=False)
img_ba = generator(img_b, name='btoa', reuse=False)
img_aba = generator(img_ab, name='btoa', reuse=True)
img_bab = generator(img_ba, name='atob', reuse=True)
logit_fake_a = discriminator(img_ba, name='a', reuse=False)
logit_fake_b = discriminator(img_ab, name='b', reuse=False)
score_fake_a = O.sigmoid(logit_fake_a)
score_fake_b = O.sigmoid(logit_fake_b)
for name in ['img_a', 'img_b', 'img_ab', 'img_ba', 'img_aba', 'img_bab', 'score_fake_a', 'score_fake_b']:
dpc.add_output(locals()[name], name=name)
if env.phase is env.Phase.TRAIN:
logit_real_a = discriminator(img_a, name='a', reuse=True)
logit_real_b = discriminator(img_b, name='b', reuse=True)
score_real_a = O.sigmoid(logit_real_a)
score_real_b = O.sigmoid(logit_real_b)
all_g_loss = 0.
all_d_loss = 0.
r_loss_ratio = 0.9
for pair_name, (real, fake), (logit_real, logit_fake), (score_real, score_fake) in zip(
['lossa', 'lossb'],
[(img_a, img_aba), (img_b, img_bab)],
[(logit_real_a, logit_fake_a), (logit_real_b, logit_fake_b)],
[(score_real_a, score_fake_a), (score_real_b, score_fake_b)]):
with env.name_scope(pair_name):
d_loss_real = O.sigmoid_cross_entropy_with_logits(logits=logit_real, labels=O.ones_like(logit_real)).mean(name='d_loss_real')
d_loss_fake = O.sigmoid_cross_entropy_with_logits(logits=logit_fake, labels=O.zeros_like(logit_fake)).mean(name='d_loss_fake')
g_loss = O.sigmoid_cross_entropy_with_logits(logits=logit_fake, labels=O.ones_like(logit_fake)).mean(name='g_loss')
d_acc_real = (score_real > 0.5).astype('float32').mean(name='d_acc_real')
d_acc_fake = (score_fake < 0.5).astype('float32').mean(name='d_acc_fake')
g_accuracy = (score_fake > 0.5).astype('float32').mean(name='g_accuracy')
d_accuracy = O.identity(.5 * (d_acc_real + d_acc_fake), name='d_accuracy')
d_loss = O.identity(.5 * (d_loss_real + d_loss_fake), name='d_loss')
# r_loss = O.raw_l2_loss('raw_r_loss', real, fake).flatten2().sum(axis=1).mean(name='r_loss')
r_loss = O.raw_l2_loss('raw_r_loss', real, fake).mean(name='r_loss')
# r_loss = O.raw_cross_entropy_prob('raw_r_loss', real, fake).flatten2().sum(axis=1).mean(name='r_loss')
# all_g_loss += g_loss + r_loss
all_g_loss += (1 - r_loss_ratio) * g_loss + r_loss_ratio * r_loss
all_d_loss += d_loss
for v in [d_loss_real, d_loss_fake, g_loss, d_acc_real, d_acc_fake, g_accuracy, d_accuracy, d_loss, r_loss]:
dpc.add_output(v, name=re.sub('^tower/\d+/', '', v.name)[:-2], reduce_method='sum')
dpc.add_output(all_g_loss, name='g_loss', reduce_method='sum')
dpc.add_output(all_d_loss, name='d_loss', reduce_method='sum')
def make_network(env):
is_train = env.phase is env.Phase.TRAIN
if is_train:
slave_devices = env.slave_devices
env.set_slave_devices([])
with env.create_network() as net:
h, w, c = get_input_shape()
dpc = env.create_dpcontroller()
with dpc.activate():
def inputs():
state = O.placeholder('state', shape=(None, h, w, c))
return [state]
def forward(x):
_ = x / 255.0
with O.argscope(O.conv2d, nonlin=O.relu):
_ = O.conv2d('conv0', _, 32, 5)
_ = O.max_pooling2d('pool0', _, 2)
_ = O.conv2d('conv1', _, 32, 5)
_ = O.max_pooling2d('pool1', _, 2)
_ = O.conv2d('conv2', _, 64, 4)
_ = O.max_pooling2d('pool2', _, 2)
_ = O.conv2d('conv3', _, 64, 3)
dpc.add_output(_, name='feature')
dpc.set_input_maker(inputs).set_forward_func(forward)
_ = dpc.outputs['feature']
_ = O.fc('fc0', _, 512, nonlin=O.p_relu)
policy = O.fc('fc_policy', _, get_player_nr_actions())
value = O.fc('fc_value', _, 1)
expf = O.scalar('explore_factor', 1, trainable=False)
policy_explore = O.softmax(policy * expf, name='policy_explore')
policy = O.softmax(policy, name='policy')
value = value.remove_axis(1, name='value')
net.add_output(policy_explore, name='policy_explore')
net.add_output(policy, name='policy')
net.add_output(value, name='value')
if is_train:
action = O.placeholder('action', shape=(None, ), dtype='int64')
future_reward = O.placeholder('future_reward', shape=(None, ))
log_policy = O.log(policy + 1e-6)
log_pi_a_given_s = (
log_policy *
O.one_hot(action, get_player_nr_actions())).sum(axis=1)
advantage = (future_reward -
O.zero_grad(value)).rename('advantage')
policy_cost = (log_pi_a_given_s *
advantage).mean(name='policy_cost')
xentropy_cost = (-policy *
log_policy).sum(axis=1).mean(name='xentropy_cost')
value_loss = O.raw_l2_loss('raw_value_loss', future_reward,
value).mean(name='value_loss')
entropy_beta = O.scalar('entropy_beta', 0.01, trainable=False)
loss = O.add_n(
[-policy_cost, -xentropy_cost * entropy_beta, value_loss],
name='loss')
net.set_loss(loss)
for v in [
policy_cost, xentropy_cost, value_loss,
value.mean(name='predict_value'),
advantage.rms(name='rms_advantage'), loss
]:
summary.scalar(v)
if is_train:
env.set_slave_devices(slave_devices)
def make_network(env):
with env.create_network() as net:
net.dist = O.distrib.GaussianDistribution('policy',
size=get_action_shape()[0],
fixed_std=False)
state = O.placeholder('state', shape=(None, ) + get_input_shape())
batch_size = state.shape[0]
# We have to define variable scope here for later optimization.
with env.variable_scope('policy'):
_ = state
_ = O.fc('fc1', _, 64, nonlin=O.relu)
_ = O.fc('fc2', _, 64, nonlin=O.relu)
mu = O.fc('fc_mu', _, net.dist.sample_size, nonlin=O.tanh)
logstd = O.variable('logstd',
O.truncated_normal_initializer(stddev=0.01),
shape=(net.dist.sample_size, ),
trainable=True)
logstd = O.tile(logstd.add_axis(0), [batch_size, 1])
theta = O.concat([mu, logstd], axis=1)
policy = net.dist.sample(batch_size=batch_size,
theta=theta,
process_theta=True)
policy = O.clip_by_value(policy, -1, 1)
net.add_output(theta, name='theta')
net.add_output(policy, name='policy')
if env.phase == env.Phase.TRAIN:
theta_old = O.placeholder('theta_old',
shape=(None, net.dist.param_size))
action = O.placeholder('action',
shape=(None, net.dist.sample_size))
advantage = O.placeholder('advantage', shape=(None, ))
entropy_beta = O.scalar('entropy_beta', g.entropy_beta)
log_prob = net.dist.log_likelihood(action,
theta,
process_theta=True)
log_prob_old = net.dist.log_likelihood(action,
theta_old,
process_theta=True)
ratio = O.exp(log_prob - log_prob_old)
epsilon = get_env('ppo.epsilon')
surr1 = ratio * advantage # surrogate from conservative policy iteration
surr2 = O.clip_by_value(ratio, 1.0 - epsilon,
1.0 + epsilon) * advantage
policy_loss = -O.reduce_mean(O.min(
surr1, surr2)) # PPO's pessimistic surrogate (L^CLIP)
entropy = net.dist.entropy(theta, process_theta=True).mean()
entropy_loss = -entropy_beta * entropy
net.add_output(policy_loss, name='policy_loss')
net.add_output(entropy_loss, name='entropy_loss')
summary.scalar('policy_entropy', entropy)
with env.variable_scope('value'):
_ = state
_ = O.fc('fc1', _, 64, nonlin=O.relu)
_ = O.fc('fc2', _, 64, nonlin=O.relu)
value = O.fc('fcv', _, 1)
value = value.remove_axis(1)
net.add_output(value, name='value')
if env.phase == env.Phase.TRAIN:
value_label = O.placeholder('value_label', shape=(None, ))
value_old = O.placeholder('value_old', shape=(None, ))
value_surr1 = O.raw_l2_loss('raw_value_loss_surr1', value,
value_label)
value_clipped = value_old + O.clip_by_value(
value - value_old, -epsilon, epsilon)
value_surr2 = O.raw_l2_loss('raw_value_loss_surr2', value_clipped,
value_label)
value_loss = O.reduce_mean(O.max(value_surr1, value_surr2))
net.add_output(value_loss, name='value_loss')
if env.phase == env.Phase.TRAIN:
loss = O.identity(policy_loss + entropy_loss + value_loss,
name='total_loss')
net.set_loss(loss)
def make_network(env):
use_linear_vr = get_env('trpo.use_linear_vr')
with env.create_network() as net:
net.dist = O.distrib.GaussianDistribution('policy',
size=get_action_shape()[0],
fixed_std=False)
if use_linear_vr:
from tartist.app.rl.utils.math import LinearValueRegressor
net.value_regressor = LinearValueRegressor()
state = O.placeholder('state', shape=(None, ) + get_input_shape())
# state = O.moving_average(state)
# state = O.clip_by_value(state, -10, 10)
batch_size = state.shape[0]
# We have to define variable scope here for later optimization.
with env.variable_scope('policy'):
_ = state
with O.argscope(O.fc):
_ = O.fc('fc1', _, 64, nonlin=O.relu)
_ = O.fc('fc2', _, 64, nonlin=O.relu)
mu = O.fc('fc_mu', _, net.dist.sample_size, nonlin=O.tanh)
logstd = O.variable(
'logstd',
O.truncated_normal_initializer(stddev=0.01),
shape=(net.dist.sample_size, ),
trainable=True)
logstd = O.tile(logstd.add_axis(0), [batch_size, 1])
theta = O.concat([mu, logstd], axis=1)
policy = net.dist.sample(batch_size=batch_size,
theta=theta,
process_theta=True)
policy = O.clip_by_value(policy, -1, 1)
net.add_output(theta, name='theta')
net.add_output(policy, name='policy')
if env.phase == env.Phase.TRAIN:
theta_old = O.placeholder('theta_old',
shape=(None, net.dist.param_size))
action = O.placeholder('action',
shape=(None, net.dist.sample_size))
advantage = O.placeholder('advantage', shape=(None, ))
log_prob = net.dist.log_likelihood(action,
theta,
process_theta=True)
log_prob_old = net.dist.log_likelihood(action,
theta_old,
process_theta=True)
# Importance sampling of surrogate loss (L in paper).
ratio = O.exp(log_prob - log_prob_old)
policy_loss = -O.reduce_mean(ratio * advantage)
kl = net.dist.kl(theta_p=theta_old,
theta_q=theta,
process_theta=True).mean()
kl_self = net.dist.kl(theta_p=O.zero_grad(theta),
theta_q=theta,
process_theta=True).mean()
entropy = net.dist.entropy(theta, process_theta=True).mean()
net.add_output(policy_loss, name='policy_loss')
net.add_output(kl, name='kl')
net.add_output(kl_self, name='kl_self')
summary.scalar('policy_entropy',
entropy,
collections=[rl.train.ACGraphKeys.POLICY_SUMMARIES])
if not use_linear_vr:
with env.variable_scope('value'):
value = O.fc('fcv', state, 1)
net.add_output(value, name='value')
if env.phase == env.Phase.TRAIN:
value_label = O.placeholder('value_label', shape=(None, ))
value_loss = O.raw_l2_loss('raw_value_loss', value,
value_label).mean(name='value_loss')
net.add_output(value_loss, name='value_loss')