def build_model(self, reuse, dev, ntype):
with tf.variable_scope(self.name) and tf.device(dev):
if reuse:
tf.get_variable_scope().reuse_variables()
assert tf.get_variable_scope().reuse
# Set inputs of networks
self.minimap = tf.placeholder(tf.float32, [None, U.minimap_channel(), self.msize, self.msize], name='minimap')
self.screen = tf.placeholder(tf.float32, [None, U.screen_channel(), self.ssize, self.ssize], name='screen')
self.info = tf.placeholder(tf.float32, [None, self.isize], name='info')
# Build networks
net = build_net(self.minimap, self.screen, self.info, self.msize, self.ssize, len(actions.FUNCTIONS), ntype)
self.spatial_action, self.non_spatial_action, self.value = net
# Set targets and masks
self.valid_spatial_action = tf.placeholder(tf.float32, [None], name='valid_spatial_action')
self.spatial_action_selected = tf.placeholder(tf.float32, [None, self.ssize**2], name='spatial_action_selected')
self.valid_non_spatial_action = tf.placeholder(tf.float32, [None, len(actions.FUNCTIONS)], name='valid_non_spatial_action')
self.non_spatial_action_selected = tf.placeholder(tf.float32, [None, len(actions.FUNCTIONS)], name='non_spatial_action_selected')
self.value_target = tf.placeholder(tf.float32, [None], name='value_target')
# Compute log probability
spatial_action_prob = tf.reduce_sum(self.spatial_action * self.spatial_action_selected, axis=1)
spatial_action_log_prob = tf.log(tf.clip_by_value(spatial_action_prob, 1e-10, 1.))
non_spatial_action_prob = tf.reduce_sum(self.non_spatial_action * self.non_spatial_action_selected, axis=1)
valid_non_spatial_action_prob = tf.reduce_sum(self.non_spatial_action * self.valid_non_spatial_action, axis=1)
valid_non_spatial_action_prob = tf.clip_by_value(valid_non_spatial_action_prob, 1e-10, 1.)
non_spatial_action_prob = non_spatial_action_prob / valid_non_spatial_action_prob
non_spatial_action_log_prob = tf.log(tf.clip_by_value(non_spatial_action_prob, 1e-10, 1.))
self.summary.append(tf.summary.histogram('spatial_action_prob', spatial_action_prob))
self.summary.append(tf.summary.histogram('non_spatial_action_prob', non_spatial_action_prob))
# Compute losses, more details in https://arxiv.org/abs/1602.01783
# Policy loss and value loss
action_log_prob = self.valid_spatial_action * spatial_action_log_prob + non_spatial_action_log_prob
advantage = tf.stop_gradient(self.value_target - self.value)
policy_loss = - tf.reduce_mean(action_log_prob * advantage)
value_loss = - tf.reduce_mean(self.value * advantage)
self.summary.append(tf.summary.scalar('policy_loss', policy_loss))
self.summary.append(tf.summary.scalar('value_loss', value_loss))
# TODO: policy penalty
loss = policy_loss + value_loss
# Build the optimizer
self.learning_rate = tf.placeholder(tf.float32, None, name='learning_rate')
opt = tf.train.RMSPropOptimizer(self.learning_rate, decay=0.99, epsilon=1e-10)
grads = opt.compute_gradients(loss)
cliped_grad = []
for grad, var in grads:
self.summary.append(tf.summary.histogram(var.op.name, var))
self.summary.append(tf.summary.histogram(var.op.name+'/grad', grad))
grad = tf.clip_by_norm(grad, 10.0)
cliped_grad.append([grad, var])
self.train_op = opt.apply_gradients(cliped_grad)
self.summary_op = tf.summary.merge(self.summary)
self.saver = tf.train.Saver(max_to_keep=100)
def build_model(self, reuse, dev, ntype):
with tf.variable_scope(self.name) and tf.device(dev):
if reuse:
tf.get_variable_scope().reuse_variables()
assert tf.get_variable_scope().reuse
# Set inputs of networks
self.minimap = tf.placeholder(tf.float32, [None, U.minimap_channel(), self.msize, self.msize], name='minimap')
self.screen = tf.placeholder(tf.float32, [None, U.screen_channel(), self.ssize, self.ssize], name='screen')
self.info = tf.placeholder(tf.float32, [None, self.isize], name='info')
# Build networks
net = build_net(self.minimap, self.screen, self.info, self.msize, self.ssize, len(actions.FUNCTIONS), ntype)
self.spatial_action, self.non_spatial_action, self.value = net
# Set targets and masks
self.valid_spatial_action = tf.placeholder(tf.float32, [None], name='valid_spatial_action')
self.spatial_action_selected = tf.placeholder(tf.float32, [None, self.ssize**2], name='spatial_action_selected')
self.valid_non_spatial_action = tf.placeholder(tf.float32, [None, len(actions.FUNCTIONS)], name='valid_non_spatial_action')
self.non_spatial_action_selected = tf.placeholder(tf.float32, [None, len(actions.FUNCTIONS)], name='non_spatial_action_selected')
self.value_target = tf.placeholder(tf.float32, [None], name='value_target')
# Compute log probability
spatial_action_prob = tf.reduce_sum(self.spatial_action * self.spatial_action_selected, axis=1)
spatial_action_log_prob = tf.log(tf.clip_by_value(spatial_action_prob, 1e-10, 1.))
non_spatial_action_prob = tf.reduce_sum(self.non_spatial_action * self.non_spatial_action_selected, axis=1)
valid_non_spatial_action_prob = tf.reduce_sum(self.non_spatial_action * self.valid_non_spatial_action, axis=1)
valid_non_spatial_action_prob = tf.clip_by_value(valid_non_spatial_action_prob, 1e-10, 1.)
non_spatial_action_prob = non_spatial_action_prob / valid_non_spatial_action_prob
non_spatial_action_log_prob = tf.log(tf.clip_by_value(non_spatial_action_prob, 1e-10, 1.))
self.summary.append(tf.summary.histogram('spatial_action_prob', spatial_action_prob))
self.summary.append(tf.summary.histogram('non_spatial_action_prob', non_spatial_action_prob))
# Compute losses, more details in https://arxiv.org/abs/1602.01783
# Policy loss and value loss
action_log_prob = self.valid_spatial_action * spatial_action_log_prob + non_spatial_action_log_prob
advantage = tf.stop_gradient(self.value_target - self.value)
policy_loss = - tf.reduce_mean(action_log_prob * advantage)
value_loss = - tf.reduce_mean(self.value * advantage)
self.summary.append(tf.summary.scalar('policy_loss', policy_loss))
self.summary.append(tf.summary.scalar('value_loss', value_loss))
# TODO: policy penalty
loss = policy_loss + value_loss
# Build the optimizer
self.learning_rate = tf.placeholder(tf.float32, None, name='learning_rate')
opt = tf.train.RMSPropOptimizer(self.learning_rate, decay=0.99, epsilon=1e-10)
grads = opt.compute_gradients(loss)
cliped_grad = []
for grad, var in grads:
self.summary.append(tf.summary.histogram(var.op.name, var))
self.summary.append(tf.summary.histogram(var.op.name+'/grad', grad))
grad = tf.clip_by_norm(grad, 10.0)
cliped_grad.append([grad, var])
self.train_op = opt.apply_gradients(cliped_grad)
self.summary_op = tf.summary.merge(self.summary)
self.saver = tf.train.Saver(max_to_keep=100)
def build_model(self, reuse, dev, ntype):
with tf.variable_scope(self.name) and tf.device(dev):
if reuse:
tf.get_variable_scope().reuse_variables()
assert tf.get_variable_scope().reuse
# Set inputs of networks
self.minimap = tf.placeholder(tf.float32, [None, U.minimap_channel(), self.msize, self.msize], name='minimap')
self.screen = tf.placeholder(tf.float32, [None, U.screen_channel(), self.ssize, self.ssize], name='screen')
self.info = tf.placeholder(tf.float32, [None, self.isize], name='info')
# Build networks
net = build_net(self.minimap, self.screen, self.info, self.msize, self.ssize, len(actions.FUNCTIONS), ntype)
self.spatial_action, self.non_spatial_action, self.value = net
# Set targets and masks
self.valid_spatial_action = tf.placeholder(tf.float32, [None], name='valid_spatial_action')
self.spatial_action_selected = tf.placeholder(tf.float32, [None, self.ssize**2], name='spatial_action_selected')
self.valid_non_spatial_action = tf.placeholder(tf.float32, [None, len(actions.FUNCTIONS)], name='valid_non_spatial_action')
self.non_spatial_action_selected = tf.placeholder(tf.float32, [None, len(actions.FUNCTIONS)], name='non_spatial_action_selected')
self.value_target = tf.placeholder(tf.float32, [None], name='value_target')
# Compute log probability
spatial_action_prob = tf.clip_by_value(tf.reduce_sum(self.spatial_action * self.spatial_action_selected, axis=1), 1e-10, 1.)
non_spatial_action_prob = tf.clip_by_value(tf.reduce_sum(self.non_spatial_action * self.non_spatial_action_selected * self.valid_non_spatial_action, axis=1), 1e-10, 1.)
q_value = spatial_action_prob * self.valid_spatial_action * self.ispatial + non_spatial_action_prob
self.delta = self.value_target - q_value
#self.clipped_error = tf.where(tf.abs(self.delta) < 1.0, 0.5 * tf.square(self.delta), tf.abs(self.delta) - 0.5, name='clipped_error')
#value_loss = tf.reduce_mean(self.clipped_error, name='value_loss')
value_loss = tf.reduce_mean(tf.square(self.delta))
self.summary.append(tf.summary.histogram('spatial_action_prob', spatial_action_prob))
self.summary.append(tf.summary.histogram('non_spatial_action_prob', non_spatial_action_prob))
self.summary.append(tf.summary.scalar('value_loss', value_loss))
# Build the optimizer
self.learning_rate = tf.placeholder(tf.float32, None, name='learning_rate')
opt = tf.train.RMSPropOptimizer(self.learning_rate, decay=0.99, epsilon=1e-10)
grads = opt.compute_gradients(value_loss)
cliped_grad = []
for grad, var in grads:
self.summary.append(tf.summary.histogram(var.op.name, var))
grad = grad if grad is not None else tf.zeros_like(var)
self.summary.append(tf.summary.histogram(var.op.name+'/grad', grad))
grad = tf.clip_by_norm(grad, 10.0)
cliped_grad.append([grad, var])
self.train_op = opt.apply_gradients(cliped_grad)
self.summary_op = tf.summary.merge(self.summary)
self.saver = tf.train.Saver(max_to_keep=100)
def build_model(self, reuse, dev, ntype):
with tf.variable_scope(self.name) and tf.device(dev):
if reuse:
tf.get_variable_scope().reuse_variables()
assert tf.get_variable_scope().reuse
# Set inputs of networks
self.minimap = tf.placeholder(
tf.float32,
[None, U.minimap_channel(), self.msize, self.msize],
name='minimap')
self.screen = tf.placeholder(
tf.float32,
[None, U.screen_channel(), self.ssize, self.ssize],
name='screen')
self.info = tf.placeholder(tf.float32, [None, self.isize],
name='info')
# Build a3c base networks
net = build_net(self.minimap,
self.screen,
self.info,
self.msize,
self.ssize,
len(actions.FUNCTIONS),
ntype,
reuse=False)
self.spatial_action, self.non_spatial_action, self.value = net
# Set targets and masks
self.valid_spatial_action = tf.placeholder(
tf.float32, [None], name='valid_spatial_action')
self.spatial_action_selected = tf.placeholder(
tf.float32, [None, self.ssize**2],
name='spatial_action_selected')
self.valid_non_spatial_action = tf.placeholder(
tf.float32, [None, len(actions.FUNCTIONS)],
name='valid_non_spatial_action')
self.non_spatial_action_selected = tf.placeholder(
tf.float32, [None, len(actions.FUNCTIONS)],
name='non_spatial_action_selected')
self.value_target = tf.placeholder(tf.float32, [None],
name='value_target')
# Compute log probability
spatial_action_prob = tf.reduce_sum(self.spatial_action *
self.spatial_action_selected,
axis=1)
spatial_action_log_prob = tf.log(
tf.clip_by_value(spatial_action_prob, 1e-10, 1.))
non_spatial_action_prob = tf.reduce_sum(
self.non_spatial_action * self.non_spatial_action_selected,
axis=1)
valid_non_spatial_action_prob = tf.reduce_sum(
self.non_spatial_action * self.valid_non_spatial_action,
axis=1)
valid_non_spatial_action_prob = tf.clip_by_value(
valid_non_spatial_action_prob, 1e-10, 1.)
non_spatial_action_prob = non_spatial_action_prob / valid_non_spatial_action_prob
non_spatial_action_log_prob = tf.log(
tf.clip_by_value(non_spatial_action_prob, 1e-10, 1.))
self.summary.append(
tf.summary.histogram('spatial_action_prob',
spatial_action_prob))
self.summary.append(
tf.summary.histogram('non_spatial_action_prob',
non_spatial_action_prob))
# Compute a3closses, more details in https://arxiv.org/abs/1602.01783
# Policy loss and value loss
action_log_prob = self.valid_spatial_action * spatial_action_log_prob + non_spatial_action_log_prob
advantage = tf.stop_gradient(self.value_target - self.value)
policy_loss = -tf.reduce_mean(action_log_prob * advantage)
value_loss = -tf.reduce_mean(self.value * advantage)
self.summary.append(tf.summary.scalar('policy_loss', policy_loss))
self.summary.append(tf.summary.scalar('value_loss', value_loss))
# TODO: policy penalty
a3c_loss = policy_loss + value_loss
#pc_part_start
self.pc_minimap = tf.placeholder(
tf.float32,
[None, U.minimap_channel(), self.msize, self.msize],
name='pc_minimap')
self.pc_screen = tf.placeholder(
tf.float32,
[None, U.screen_channel(), self.ssize, self.ssize],
name='pc_screen')
self.pc_info = tf.placeholder(tf.float32, [None, self.isize],
name='info')
self.pc_valid_non_spatial_action = tf.placeholder(
tf.float32, [None, len(actions.FUNCTIONS)],
name='pc_valid_non_spatial_action')
pc_net = build_pc_net(self.pc_minimap, self.pc_screen,
self.pc_info, self.msize, self.ssize,
len(actions.FUNCTIONS),
self.pc_valid_non_spatial_action)
pc_q, pc_q_max = pc_net
pc_a = tf.placeholder("float", [None, len(actions.FUNCTIONS)])
pc_a_reshaped = tf.reshape(
self.pc_a, [-1, 1, 1, len(actions.FUNCTIONS)])
# Extract Q for taken action
pc_qa_ = tf.multiply(self.pc_q, pc_a_reshaped)
pc_qa = tf.reduce_sum(pc_qa_, reduction_indices=3, keep_dims=False)
# (-1, 20, 20)
# TD target for Q
self.pc_r = tf.placeholder("float", [None, 20, 20])
pc_loss = self._pixel_change_lambda * tf.nn.l2_loss(self.pc_r -
pc_qa)
# Build the optimizer
loss = pc_loss + a3c_loss
self.learning_rate = tf.placeholder(tf.float32,
None,
name='learning_rate')
opt = tf.train.RMSPropOptimizer(self.learning_rate,
decay=0.99,
epsilon=1e-10)
grads = opt.compute_gradients(loss)
cliped_grad = []
for grad, var in grads:
self.summary.append(tf.summary.histogram(var.op.name, var))
self.summary.append(
tf.summary.histogram(var.op.name + '/grad', grad))
grad = tf.clip_by_norm(grad, 10.0)
cliped_grad.append([grad, var])
self.train_op = opt.apply_gradients(cliped_grad)
self.summary_op = tf.summary.merge(self.summary)
self.saver = tf.train.Saver(max_to_keep=100)
def build_model(self, reuse, dev, ntype):
with tf.variable_scope(self.name) and tf.device(dev):
if reuse:
tf.get_variable_scope().reuse_variables()
assert tf.get_variable_scope().reuse
# Set inputs of networks
self.minimap = tf.placeholder(
tf.float32,
[None, U.minimap_channel(), self.msize, self.msize],
name='minimap')
self.screen = tf.placeholder(
tf.float32,
[None, U.screen_channel(), self.ssize, self.ssize],
name='screen')
self.info = tf.placeholder(tf.float32, [None, self.isize],
name='info')
# create master and subpolicies
self.subpolicy_Q = build_net(self.minimap, self.screen, self.info,
self.msize, self.ssize, num_units + 2,
'master_policy')
# Set targets and masks for master policy update
self.learning_rate = tf.placeholder(tf.float32,
None,
name='learning_rate')
self.action_input = tf.placeholder("float", [None, num_units + 2])
self.y_input = tf.placeholder("float", [None])
self.Q_action = tf.reduce_sum(tf.multiply(self.subpolicy_Q,
self.action_input),
reduction_indices=1)
self.cost = tf.reduce_mean(tf.square(self.y_input - self.Q_action))
self.master_train_op = tf.train.AdamOptimizer(
self.learning_rate).minimize(self.cost)
# Set targets and masks for subpolicies update
self.valid_spatial_action = tf.placeholder(
tf.float32, [None], name='valid_spatial_action_')
self.spatial_action_selected = tf.placeholder(
tf.float32, [None, self.ssize**2],
name='spatial_action_selected')
self.valid_non_spatial_action = tf.placeholder(
tf.float32, [None, len(actions.FUNCTIONS)],
name='valid_non_spatial_action_')
self.non_spatial_action_selected = tf.placeholder(
tf.float32, [None, len(actions.FUNCTIONS)],
name='non_spatial_action_selected_')
self.value_target = tf.placeholder(tf.float32, [None],
name='value_target_')
# Build the optimizer
opt = tf.train.AdamOptimizer(self.learning_rate)
self.subpolicy = build_net(self.minimap, self.screen, self.info,
self.msize, self.ssize,
len(actions.FUNCTIONS), 'fcn')
self.spatial_action, self.non_spatial_action, self.value = self.subpolicy
# Compute log probability
spatial_action_prob = tf.reduce_sum(self.spatial_action *
self.spatial_action_selected,
axis=1)
spatial_action_log_prob = tf.log(
tf.clip_by_value(spatial_action_prob, 1e-10, 1.))
non_spatial_action_prob = tf.reduce_sum(
self.non_spatial_action * self.non_spatial_action_selected,
axis=1)
valid_non_spatial_action_prob = tf.reduce_sum(
self.non_spatial_action * self.valid_non_spatial_action,
axis=1)
valid_non_spatial_action_prob = tf.clip_by_value(
valid_non_spatial_action_prob, 1e-10, 1.)
non_spatial_action_prob = non_spatial_action_prob / valid_non_spatial_action_prob
non_spatial_action_log_prob = tf.log(
tf.clip_by_value(non_spatial_action_prob, 1e-10, 1.))
self.summary.append(
tf.summary.histogram('spatial_action_prob_',
spatial_action_prob))
self.summary.append(
tf.summary.histogram('non_spatial_action_prob_',
non_spatial_action_prob))
# Compute losses, more details in https://arxiv.org/abs/1602.01783
# Policy loss and value loss
action_log_prob = self.valid_spatial_action * spatial_action_log_prob + non_spatial_action_log_prob
advantage = tf.stop_gradient(self.value_target - self.value)
policy_loss = -tf.reduce_mean(action_log_prob * advantage)
value_loss = -tf.reduce_mean(self.value * advantage)
self.summary.append(tf.summary.scalar('policy_loss_', policy_loss))
self.summary.append(tf.summary.scalar('value_loss_', value_loss))
# TODO: policy penalty
loss = policy_loss + value_loss
grads = opt.compute_gradients(loss)
cliped_grad = []
for grad, var in grads:
# get around of master policy gradients
if grad is None:
continue
self.summary.append(tf.summary.histogram(var.op.name, var))
self.summary.append(
tf.summary.histogram(var.op.name + '/grad', grad))
grad = tf.clip_by_norm(grad, 10.0)
cliped_grad.append([grad, var])
self.train_op = opt.apply_gradients(cliped_grad)
self.summary_op = tf.summary.merge(self.summary)
self.saver = tf.train.Saver(max_to_keep=100)