def make_network(env, h=None, w=None):
with env.create_network() as net:
if h is None:
img = O.placeholder('img', shape=(1, None, None, 3))
else:
img = O.variable('img', np.zeros([1, h, w, 3]))
net.add_output(img, name='img')
_ = img
_ = _ - get_env('neural_style.image_mean').reshape(1, 1, 1, 3)
_ = O.pad_rb_multiple_of(_, 32)
def stacked_conv(prefix, nr_convs, in_, channel, kernel=(3, 3), padding='SAME', nonlin=O.relu):
for i in range(1, nr_convs + 1):
in_ = O.conv2d('{}_{}'.format(prefix, i), in_, channel, kernel, padding=padding, nonlin=nonlin)
return in_
_ = stacked_conv('conv1', 2, _, 64)
_ = O.pooling2d('pool1', _, (2, 2))
_ = stacked_conv('conv2', 2, _, 128)
_ = O.pooling2d('pool2', _, (2, 2))
_ = stacked_conv('conv3', 3, _, 256)
_ = O.pooling2d('pool3', _, (2, 2))
_ = stacked_conv('conv4', 3, _, 512)
_ = O.pooling2d('pool4', _, (2, 2))
_ = stacked_conv('conv5', 3, _, 512)
_ = O.pooling2d('pool5', _, (2, 2))
for l in get_env('neural_style.content_layers'):
net.add_output(net.find_var_by_name(l[0] + '/bias'), name=l[0])
for l in get_env('neural_style.style_layers'):
net.add_output(net.find_var_by_name(l[0] + '/bias'), name=l[0])
def make_param_gs(env, var_list, name_scope):
var_shapes = [as_tftensor(v).get_shape().as_list() for v in var_list]
for vs, v in zip(var_shapes, var_list):
assert None not in vs, 'Could not determine the shape for optimizable variable: {}.'.format(
v)
var_nr_elems = [
as_tftensor(v).get_shape().num_elements() for v in var_list
]
nr_total_elems = sum(var_nr_elems)
param_nr_elems = nr_total_elems
with env.name_scope(name_scope):
# Parameter getter
param_getter = vectorize_var_list(var_list)
# Parameter setter
flat_variables_tensor = O.placeholder('flat_variable_tensor',
shape=(nr_total_elems, ))
var_assigns = []
index = 0
for v, vs, vn in zip(var_list, var_shapes, var_nr_elems):
value = flat_variables_tensor[index:index + vn].reshape(vs)
# Use tf.assign because tf.group use non-3rdparty-compatible codes.
var_assigns.append(
tf.assign(v, value, name='assign_{}'.format(escape_name(v))))
index += vn
param_setter = tf.group(*var_assigns)
param_provider = as_tftensor(flat_variables_tensor)
return param_nr_elems, param_getter, param_setter, param_provider
def testPaddingCenter(self):
a = O.placeholder('a', shape=(16, 15, 15, 3))
b = O.pad_center(a, [17, 17])
self.assertTupleEqual(b.static_shape, (16, 17, 17, 3))
avar = np.random.normal(size=(16, 15, 15, 3))
bvar = np.pad(avar, [[0, 0], [1, 1], [1, 1], [0, 0]], mode='constant')
self.assertTensorClose(b.eval(a=avar), bvar)
def testPaddingRBMultiple(self):
a = O.placeholder('a', shape=(16, 15, 15, 3))
b = O.pad_rb_multiple_of(a, 8)
self.assertTupleEqual(b.static_shape, (16, 16, 16, 3))
avar = np.random.normal(size=(16, 15, 15, 3))
bvar = np.pad(avar, [[0, 0], [0, 1], [0, 1], [0, 0]], mode='constant')
self.assertTensorClose(b.eval(a=avar), bvar)
def testCropLU(self):
a = O.placeholder('a', shape=(16, 17, 17, 3))
b = O.crop_lu(a, [15, 15])
self.assertTupleEqual(b.static_shape, (16, 15, 15, 3))
avar = np.random.normal(size=(16, 17, 17, 3))
bvar = avar[:, :-2, :-2, :]
self.assertTensorClose(b.eval(a=avar), bvar)
def make_network(env):
with env.create_network() as net:
dpc = env.create_dpcontroller()
with dpc.activate():
def inputs():
h, w, c = 28, 28, 1
img = O.placeholder('img', shape=(None, h, w, c))
return [img]
def forward(img):
_ = img
_ = O.conv2d('conv1',
_,
16, (3, 3),
padding='SAME',
nonlin=O.identity)
_ = O.batch_norm('bn1', _)
_ = O.relu(_)
_ = O.pooling2d('pool1', _, kernel=2)
_ = O.conv2d('conv2',
_,
32, (3, 3),
padding='SAME',
nonlin=O.identity)
_ = O.batch_norm('bn2', _)
_ = O.relu(_)
_ = O.pooling2d('pool2', _, kernel=2)
dpc.add_output(_, name='feature')
dpc.set_input_maker(inputs).set_forward_func(forward)
_ = dpc.outputs['feature']
_ = O.fc('fc1', _, 64)
_ = O.fc('fc2', _, 10)
prob = O.softmax(_, name='prob')
pred = _.argmax(axis=1).astype('int32', name='pred')
net.add_output(prob)
net.add_output(pred)
if env.phase is env.Phase.TRAIN:
label = O.placeholder('label', shape=(None, ), dtype='int32')
loss = O.sparse_softmax_cross_entropy_with_logits(
logits=_, labels=label).mean()
loss = O.identity(loss, name='loss')
net.set_loss(loss)
accuracy = O.eq(label, pred).astype('float32').mean()
error = 1. - accuracy
summary.scalar('accuracy', accuracy)
summary.scalar('error', error)
summary.inference.scalar('loss', loss)
summary.inference.scalar('accuracy', accuracy)
summary.inference.scalar('error', error)
def make_network(env):
with env.create_network() as net:
nr_classes = get_env('dataset.nr_classes')
conv_bn_relu = functools.partial(O.conv2d, nonlin=O.bn_relu)
conv2d = conv_bn_relu
dpc = env.create_dpcontroller()
with dpc.activate():
def inputs():
h, w, c = 32, 32, 3
img = O.placeholder('img', shape=(None, h, w, c))
return [img]
def forward(img):
_ = img
_ = conv2d('conv1.1', _, 16, (3, 3), padding='SAME')
_ = conv2d('conv1.2', _, 16, (3, 3), padding='SAME')
_ = O.pooling2d('pool1', _, kernel=3, stride=2)
_ = conv2d('conv2.1', _, 32, (3, 3), padding='SAME')
_ = conv2d('conv2.2', _, 32, (3, 3), padding='SAME')
_ = O.pooling2d('pool2', _, kernel=3, stride=2)
_ = conv2d('conv3.1', _, 64, (3, 3), padding='VALID')
_ = conv2d('conv3.2', _, 64, (3, 3), padding='VALID')
_ = conv2d('conv3.3', _, 64, (3, 3), padding='VALID')
dpc.add_output(_, name='feature')
dpc.set_input_maker(inputs).set_forward_func(forward)
_ = dpc.outputs['feature']
_ = O.fc('fc1', _, 128, nonlin=O.relu)
_ = O.fc('fc2', _, 64, nonlin=O.relu)
_ = O.fc('linear', _, nr_classes)
prob = O.softmax(_, name='prob')
pred = _.argmax(axis=1).astype('int32', name='pred')
net.add_output(prob)
net.add_output(pred)
if env.phase is env.Phase.TRAIN:
label = O.placeholder('label', shape=(None, ), dtype='int32')
loss = O.sparse_softmax_cross_entropy_with_logits(logits=_, labels=label).mean()
loss = O.identity(loss, name='loss')
net.set_loss(loss)
accuracy = O.eq(label, pred).astype('float32').mean()
error = 1. - accuracy
summary.scalar('accuracy', accuracy)
summary.scalar('error', error)
summary.inference.scalar('loss', loss)
summary.inference.scalar('accuracy', accuracy)
summary.inference.scalar('error', error)
def testAdvancedIndexing(self):
a = O.placeholder('a', shape=(5, 5))
a_val = np.arange(25).reshape((5, 5)).astype('float32')
feed_dict = {a.name: a_val}
self.assertTensorClose(a[0:3].eval(feed_dict=feed_dict), a_val[0:3])
self.assertTensorClose(a[0:3, 0:3].eval(feed_dict=feed_dict),
a_val[0:3, 0:3])
with self.assertRaises(NotImplementedError):
self.assertTensorClose(a.set_sub[0:3](1).eval(feed_dict=feed_dict),
np.array([1, 1, 1, 3, 4]))
if True:
self.assertTensorClose(a.ai[[0, 3]].eval(feed_dict=feed_dict),
a_val[[0, 3]])
self.assertTensorClose(
a.ai[[0, 3], [0, 3]].eval(feed_dict=feed_dict), a_val[[0, 3],
[0, 3]])
with self.assertRaises(NotImplementedError):
self.assertTensorClose(
a.set_ai[[0, 3]](1).eval(feed_dict=feed_dict),
np.array([1, 1, 1, 3, 4]))
def make_network(env):
is_train = env.phase is env.Phase.TRAIN
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))
next_state = O.placeholder('next_state', shape=(None, h, w, c))
return [state, next_state]
@O.auto_reuse
def phi(x):
_ = x / 255.0
# Nature structure
with O.argscope(O.conv2d, nonlin=O.relu):
_ = O.conv2d('conv1', _, 32, 8, stride=4)
_ = O.conv2d('conv2', _, 64, 4, stride=2)
_ = O.conv2d('conv3', _, 64, 3, stride=1)
return _
def forward(state, next_state):
dpc.add_output(phi(state), name='feature')
dpc.add_output(phi(next_state), name='next_feature')
dpc.set_input_maker(inputs).set_forward_func(forward)
@O.auto_reuse
def phi_fc(feature):
_ = feature
_ = O.fc('fc0',
_,
512,
nonlin=functools.partial(O.leaky_relu, alpha=0.01))
q_pred = O.fc('fcq', _, get_player_nr_actions())
q_max = q_pred.max(axis=1)
q_argmax = q_pred.argmax(axis=1)
return q_pred, q_max, q_argmax
_ = dpc.outputs['feature']
q_pred, q_max, q_argmax = phi_fc(_)
_ = dpc.outputs['next_feature']
next_q_pred, next_q_max, _ = phi_fc(_)
net.add_output(q_pred, name='q_pred')
net.add_output(q_max, name='q_max')
net.add_output(q_argmax, name='q_argmax')
if is_train:
reward = O.placeholder('reward', shape=(None, ), dtype='float32')
action = O.placeholder('action', shape=(None, ), dtype='int64')
is_over = O.placeholder('is_over', shape=(None, ), dtype='bool')
assert get_env('dqn.nr_td_steps') == 1
this_q_pred = (q_pred *
O.one_hot(action, get_player_nr_actions())).sum(
axis=1)
this_q_label = reward + get_env('dqn.gamma') * (
1 - is_over.astype('float32')) * O.zero_grad(next_q_max)
summary.scalar('this_q_pred', this_q_pred.mean())
summary.scalar('this_q_label', this_q_label.mean())
summary.scalar('reward', reward.mean())
summary.scalar('is_over', is_over.astype('float32').mean())
q_loss = O.raw_smooth_l1_loss('raw_q_loss', this_q_pred,
this_q_label).mean(name='q_loss')
net.set_loss(q_loss)
def inputs():
img = O.placeholder('img', shape=(None, h, w, c))
# only for demo-time
zc = O.placeholder('zc', shape=(1, net.zc_distrib.sample_size))
return [img, zc]
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):
is_train = env.phase is env.Phase.TRAIN
# device control: always use master device only for training session
if is_train:
slave_devices = env.slave_devices
env.set_slave_devices([])
with env.create_network() as net:
input_length, = get_input_shape()
action_length, = get_action_shape()
dpc = env.create_dpcontroller()
with dpc.activate():
def inputs():
state = O.placeholder('state', shape=(None, input_length))
return [state]
# forward policy network and value network separately (actor-critic)
def forward(x):
_ = x
_ = O.fc('fcp1', _, 512, nonlin=O.relu)
_ = O.fc('fcp2', _, 256, nonlin=O.relu)
dpc.add_output(_, name='feature_p')
_ = x
_ = O.fc('fcv1', _, 512, nonlin=O.relu)
_ = O.fc('fcv2', _, 256, nonlin=O.relu)
dpc.add_output(_, name='feature_v')
dpc.set_input_maker(inputs).set_forward_func(forward)
_ = dpc.outputs['feature_p']
# mu and std, assuming spherical covariance
policy_mu = O.fc('fc_policy_mu', _, action_length)
# In this example, we do not use variance. instead, we use fixed value.
# policy_var = O.fc('fc_policy_var', _, 1, nonlin=O.softplus)
# policy_var = O.tile(policy_var, [1, action_length], name='policy_var')
# policy_std = O.sqrt(policy_var, name='policy_std')
actor_space = get_env('a3c.actor_space')
nr_bins = actor_space.shape[1]
# Instead of using normal distribution, we use Laplacian distribution for policy.
# And also, we are sampling from a truncated Laplacian distribution (only care the value in the
# action space). To simplify the computation, we discretize the action space.
actor_space = O.constant(actor_space)
actor_space = O.tile(actor_space.add_axis(0), [policy_mu.shape[0], 1, 1])
policy_mu3 = O.tile(policy_mu.add_axis(2), [1, 1, nr_bins])
# policy_std3 = O.tile(policy_std.add_axis(2), [1, 1, nr_bins])
# logits = O.abs(actor_space - policy_mu3) / (policy_std3 + 1e-2)
# Here, we force the std of the policy to be 1.
logits_explore = -O.abs(actor_space - policy_mu3)
policy_explore = O.softmax(logits_explore)
# Clip the policy for output
action_range = get_action_range()
action_range = tuple(map(O.constant, action_range))
action_range = tuple(map(lambda x: O.tile(x.add_axis(0), [policy_mu.shape[0], 1]), action_range))
policy_output = O.clip_by_value(policy_mu, *action_range)
_ = dpc.outputs['feature_v']
value = O.fc('fc_value', _, 1)
value = value.remove_axis(1, name='value')
# Note that, here the policy_explore is a discrete policy,
# and policy is actually the continuous one.
net.add_output(policy_explore, name='policy_explore')
net.add_output(policy_output, name='policy')
net.add_output(value, name='value')
if is_train:
action = O.placeholder('action', shape=(None, action_length), dtype='int64')
future_reward = O.placeholder('future_reward', shape=(None, ))
entropy_beta = O.scalar('entropy_beta', 0.1, trainable=False)
# Since we discretized the action space, use cross entropy here.
log_policy = O.log(policy_explore + 1e-4)
log_pi_a_given_s = (log_policy * O.one_hot(action, nr_bins)).sum(axis=2).sum(axis=1)
advantage = (future_reward - O.zero_grad(value)).rename('advantage')
# Important trick: using only positive advantage to perform gradient assent. This stabilizes the training.
advantage = advantage * O.zero_grad((advantage > 0.).astype('float32'))
policy_loss = O.identity(-(log_pi_a_given_s * advantage).mean(), name='policy_loss')
# As mentioned, there is no trainable variance.
# entropy_loss = O.identity(-entropy_beta * (policy_std ** 2.).sum(axis=1).mean(), name='entropy_loss')
value_loss = O.raw_smooth_l1_loss('raw_value_loss', future_reward, value).mean(name='value_loss')
loss = O.add_n([policy_cost, value_loss], name='loss')
net.set_loss(loss)
for v in [policy_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 inputs():
state = O.placeholder('state', shape=(None, input_length))
return [state]
def make_network(env):
with env.create_network() as net:
n = 2
nr_classes = get_env('dataset.nr_classes')
conv2d = functools.partial(O.conv2d,
kernel=3,
use_bias=False,
padding='SAME')
conv_bn_relu = functools.partial(conv2d, nonlin=O.bn_relu)
dpc = env.create_dpcontroller()
with dpc.activate():
def inputs():
h, w, c = 32, 32, 3
img = O.placeholder('img', shape=(None, h, w, c))
return [img]
def residual(name, x, first=False, inc_dim=False):
in_channel = x.static_shape[3]
out_channel = in_channel
stride = 1
if inc_dim:
out_channel = in_channel * 2
stride = 2
with env.variable_scope(name):
_ = x if first else O.bn_relu(x)
_ = conv_bn_relu('conv1', _, out_channel, stride=stride)
_ = conv2d('conv2', _, out_channel)
if inc_dim:
x = O.pooling2d('pool', x, kernel=2)
x = O.pad(x, [[0, 0], [0, 0], [0, 0],
[in_channel // 2, in_channel // 2]])
print(name, x.static_shape)
_ = _ + x
return _
def forward(img):
_ = img / 128.0 - 1.0
_ = conv_bn_relu('conv0', _, 16)
_ = residual('res1.0', _, first=True)
for i in range(1, n):
_ = residual('res1.{}'.format(i), _)
_ = residual('res2.0', _, inc_dim=True)
for i in range(1, n):
_ = residual('res2.{}'.format(i), _)
_ = residual('res3.0', _, inc_dim=True)
for i in range(1, n):
_ = residual('res3.{}'.format(i), _)
_ = O.batch_norm('bn_last', _)
_ = O.relu(_)
_ = _.mean(axis=[1, 2]) # global avg pool
dpc.add_output(_, name='feature')
dpc.set_input_maker(inputs).set_forward_func(forward)
_ = dpc.outputs['feature']
_ = O.fc('linear', _, nr_classes)
prob = O.softmax(_, name='prob')
pred = _.argmax(axis=1).astype('int32', name='pred')
net.add_output(prob)
net.add_output(pred)
if env.phase is env.Phase.TRAIN:
label = O.placeholder('label', shape=(None, ), dtype='int32')
loss = O.sparse_softmax_cross_entropy_with_logits(
logits=_, labels=label).mean()
loss = O.identity(loss, name='loss')
net.set_loss(loss)
accuracy = O.eq(label, pred).astype('float32').mean()
error = 1. - accuracy
summary.scalar('accuracy', accuracy)
summary.scalar('error', error)
summary.inference.scalar('loss', loss)
summary.inference.scalar('accuracy', accuracy)
summary.inference.scalar('error', error)
def inputs():
img = O.placeholder('img', shape=(None, h, w, c))
if env.phase is env.Phase.TRAIN:
return [img]
else:
return []
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 inputs():
img = O.placeholder('img', shape=(None, h, w, c))
return [img]
def make_network(env):
with env.create_network() as net:
state = O.placeholder('state', shape=(None, ) + get_input_shape())
logits = O.fc('fc', state, get_action_shape())
net.add_output(logits, name='policy')
def make_rpredictor_network(env):
is_train = env.phase is env.Phase.TRAIN
with env.create_network() as net:
h, w, c = get_input_shape()
# Hack(MJY):: forced RGB input (instead of combination of history frames)
c = 3
dpc = env.create_dpcontroller()
with dpc.activate():
def inputs():
state = O.placeholder('state', shape=(None, h, w, c))
t1_state = O.placeholder('t1_state', shape=(None, h, w, c))
t2_state = O.placeholder('t2_state', shape=(None, h, w, c))
return [state, t1_state, t2_state]
@O.auto_reuse
def forward_conv(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)
return _
def forward(x, t1, t2):
dpc.add_output(forward_conv(x), name='feature')
dpc.add_output(forward_conv(t1), name='t1_feature')
dpc.add_output(forward_conv(t2), name='t2_feature')
dpc.set_input_maker(inputs).set_forward_func(forward)
@O.auto_reuse
def forward_fc(feature, action):
action = O.one_hot(action, get_player_nr_actions())
_ = O.concat([feature.flatten2(), action], axis=1)
_ = O.fc('fc0', _, 512, nonlin=O.p_relu)
reward = O.fc('fc_reward', _, 1)
return reward
action = O.placeholder('action', shape=(None, ), dtype='int64')
net.add_output(forward_fc(dpc.outputs['feature'], action),
name='reward')
if is_train:
t1_action = O.placeholder('t1_action',
shape=(None, ),
dtype='int64')
t1_reward_exp = O.exp(
forward_fc(dpc.outputs['t1_feature'], t1_action).sum())
t2_action = O.placeholder('t2_action',
shape=(None, ),
dtype='int64')
t2_reward_exp = O.exp(
forward_fc(dpc.outputs['t2_feature'], t2_action).sum())
pref = O.placeholder('pref')
pref = O.callback_injector(pref)
p1, p2 = 1 - pref, pref
p_greater = t1_reward_exp / (t1_reward_exp + t2_reward_exp)
loss = -p1 * O.log(p_greater) - p2 * O.log(1 - p_greater)
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')
def inputs():
state = O.placeholder('state', shape=(None, h, w, c))
return [state]
def inputs():
img_a = O.placeholder('img_a', shape=(None, h, w, c))
img_b = O.placeholder('img_b', shape=(None, h, w, c))
return [img_a, img_b]
