def build_graph(self, assemble_model=False):
input_placeholder = layer.InputPlaceholder(dppo_config.config.input)
policy_head = Network(input_placeholder)
if dppo_config.config.output.continuous:
output = layer.Dense(policy_head, dppo_config.config.output.action_size, init_var=0.01)
actor = ConcatFixedStd(output)
actor_layers = [output, actor]
else:
actor = layer.Dense(policy_head, dppo_config.config.output.action_size,
activation=layer.Activation.Softmax, init_var=0.01)
actor_layers = [actor]
value_head = Network(input_placeholder)
critic = layer.Dense(value_head, 1)
feeds = dict(head=actor, weights=layer.Weights(policy_head, *actor_layers),
ph_state=input_placeholder.ph_state)
if dppo_config.config.use_lstm:
feeds.update(policy_head.lstm_items)
self.actor = Subnet(**feeds)
feeds = dict(head=critic, weights=layer.Weights(value_head, critic),
ph_state=input_placeholder.ph_state)
if dppo_config.config.use_lstm:
feeds.update(value_head.lstm_items)
self.critic = Subnet(**feeds)
if assemble_model:
self.policy = PolicyModel(self.actor)
self.value_func = ValueModel(self.critic)
if dppo_config.config.use_lstm:
self.lstm_zero_state = [self.actor.lstm_zero_state, self.critic.lstm_zero_state]
def build_graph(self):
input = layer.Input(cfg.config.input)
self.ph_action = graph.Placeholder(
np.float32, (None, cfg.config.output.action_size))
sizes = cfg.config.hidden_sizes
assert len(
sizes) > 1, 'You need to provide sizes at least for 2 layers'
dense_1st = layer.Dense(layer.Flatten(input), sizes[0],
layer.Activation.Relu)
dense_2nd = layer.DoubleDense(dense_1st, self.ph_action, sizes[1],
layer.Activation.Relu)
layers = [input, dense_1st, dense_2nd]
net = layer.GenericLayers(dense_2nd, [
dict(type=layer.Dense, size=size, activation=layer.Activation.Relu)
for size in sizes[2:]
])
if len(sizes[2:]) > 0:
layers.append(net)
self.critic = layer.Dense(net, 1, init_var=3e-3)
self.ph_state = input.ph_state
layers.append(self.critic)
self.weights = layer.Weights(*layers)
def build_graph(self, input_placeholder):
input = layer.ConfiguredInput(dppo_config.config.input, input_placeholder=input_placeholder)
layers = [input]
sizes = dppo_config.config.hidden_sizes
activation = layer.get_activation(dppo_config.config)
fc_layers = layer.GenericLayers(layer.Flatten(input),
[dict(type=layer.Dense, size=size, activation=activation)
for size in sizes[:-1]])
layers.append(fc_layers)
last_size = fc_layers.node.shape.as_list()[-1]
if len(sizes) > 0:
last_size = sizes[-1]
if dppo_config.config.use_lstm:
lstm = layer.lstm(dppo_config.config.lstm_type,
graph.Expand(fc_layers, 0), n_units=last_size,
n_cores=dppo_config.config.lstm_num_cores)
head = graph.Reshape(lstm, [-1, last_size])
layers.append(lstm)
self.lstm_items = {"ph_lstm_state": lstm.ph_state,
"lstm_zero_state": lstm.zero_state,
"lstm_state": lstm.state,
"lstm_reset_timestep": lstm.reset_timestep}
else:
head = layer.Dense(fc_layers, last_size, activation=activation)
layers.append(head)
self.ph_state = input.ph_state
self.weight = layer.Weights(*layers)
return head.node
def build_graph(self):
input_placeholder = layer.InputPlaceholder(da3c_config.config.input)
actor_head = Head(input_placeholder)
actor = layer.Actor(actor_head, da3c_config.config.output)
critic_head = Head(input_placeholder)
critic = layer.Dense(critic_head, 1)
self.ph_state = input_placeholder.ph_state
feeds = dict(head=actor, weights=layer.Weights(actor_head, actor))
if da3c_config.config.use_lstm:
feeds.update(actor_head.lstm_items)
self.actor = Subnet(**feeds)
feeds = dict(head=graph.Flatten(critic),
weights=layer.Weights(critic_head, critic))
if da3c_config.config.use_lstm:
feeds.update(critic_head.lstm_items)
self.critic = Subnet(**feeds)
self.weights = layer.Weights(actor_head, actor, critic_head, critic)
if da3c_config.config.use_lstm:
self.lstm_zero_state = (self.actor.lstm_zero_state,
self.critic.lstm_zero_state)
def build_graph(self):
input = layer.Input(trpo_config.config.input)
head = layer.GenericLayers(layer.Flatten(input),
[dict(type=layer.Dense, size=size, activation=layer.Activation.Tanh)
for size in trpo_config.config.hidden_sizes])
if trpo_config.config.output.continuous:
output = layer.Dense(head, trpo_config.config.output.action_size)
actor = ConcatFixedStd(output)
actor_layers = [output, actor]
else:
actor = layer.Dense(head, trpo_config.config.output.action_size,
activation=layer.Activation.Softmax)
actor_layers = [actor]
self.ph_state = input.ph_state
self.actor = actor
self.weights = layer.Weights(*([input, head] + actor_layers))
def build_graph(self):
input = layer.Input(cfg.input)
self.perception =\
layer.Dense(layer.Flatten(input), cfg.d, # d=256
activation=layer.Activation.Relu)
self.weights = layer.Weights(input, self.perception)
self.ph_state = input.ph_state
def build_graph(self):
self.ph_perception =\
graph.Placeholder(np.float32, shape=(None, cfg.d), name="ph_perception")
# tf.placeholder(tf.float32, shape=[None, cfg.d], name="ph_perception")
self.Mspace =\
layer.Dense(self.ph_perception, cfg.d, # d=256
activation=layer.Activation.Relu)
Mspace_expanded = graph.Expand(self.Mspace, 0)
self.lstm = DilatedLSTMCell(cfg.d, num_cores=cfg.d)
# needs wrap as layer to retrieve weights
self.ph_step_size =\
graph.Placeholder(np.float32, shape=(1,), name="ph_m_step_size")
# tf.placeholder(tf.float32, [1], name="ph_m_step_size")
self.ph_initial_lstm_state =\
graph.Placeholder(np.float32, shape=(1, self.lstm.state_size), name="ph_m_lstm_state")
# tf.placeholder(tf.float32, [1, self.lstm.state_size], name="ph_m_lstm_state")
lstm_outputs, self.lstm_state = tf.nn.dynamic_rnn(
self.lstm,
Mspace_expanded,
initial_state=self.ph_initial_lstm_state,
sequence_length=self.ph_step_size,
time_major=False)
lstm_outputs = tf.reshape(lstm_outputs, [-1, cfg.d])
sg_lstm_outputs = graph.TfNode(lstm_outputs)
self.goal = tf.nn.l2_normalize(graph.Flatten(sg_lstm_outputs), dim=1)
critic = layer.Dense(sg_lstm_outputs, 1)
self.value = layer.Flatten(critic)
self.weights = layer.Weights(
self.Mspace, graph.TfNode((self.lstm.matrix, self.lstm.bias)),
critic)
self.lstm_state_out =\
graph.VarAssign(graph.Variable(np.zeros([1, self.lstm.state_size]),
dtype=np.float32, name="lstm_state_out"),
np.zeros([1, self.lstm.state_size]))
def build_graph(self):
input = layer.Input(pg_config.config.input)
dense = layer.GenericLayers(layer.Flatten(input),
[dict(type=layer.Dense, size=size, activation=layer.Activation.Relu)
for size in pg_config.config.hidden_sizes])
actor = layer.Dense(dense, pg_config.config.output.action_size,
activation=layer.Activation.Softmax)
self.state = input.ph_state
self.weights = layer.Weights(input, dense, actor)
return actor.node
def build_graph(self):
input = layer.ConfiguredInput(trpo_config.config.input)
# add one extra feature for timestep
ph_step = graph.Placeholder(np.float32, shape=[None, 1])
state = (input.ph_state, ph_step)
concatenated = graph.Concat([layer.Flatten(input), ph_step], axis=1)
activation = layer.Activation.get_activation(
trpo_config.config.activation)
head = layer.GenericLayers(concatenated, [
dict(type=layer.Dense, size=size, activation=activation)
for size in trpo_config.config.hidden_sizes
])
value = layer.Dense(head, 1)
ph_ytarg_ny = graph.Placeholder(np.float32)
mse = graph.TfNode(
tf.reduce_mean(tf.square(ph_ytarg_ny.node - value.node)))
weights = layer.Weights(input, head, value)
sg_get_weights_flatten = graph.GetVariablesFlatten(weights)
sg_set_weights_flatten = graph.SetVariablesFlatten(weights)
l2 = graph.TfNode(1e-3 * tf.add_n([
tf.reduce_sum(tf.square(v))
for v in utils.Utils.flatten(weights.node)
]))
loss = graph.TfNode(l2.node + mse.node)
sg_gradients = optimizer.Gradients(weights, loss=loss)
sg_gradients_flatten = graph.GetVariablesFlatten(
sg_gradients.calculate)
self.op_value = self.Op(value, state=state)
self.op_get_weights_flatten = self.Op(sg_get_weights_flatten)
self.op_set_weights_flatten = self.Op(
sg_set_weights_flatten, value=sg_set_weights_flatten.ph_value)
self.op_compute_loss_and_gradient = self.Ops(loss,
sg_gradients_flatten,
state=state,
ytarg_ny=ph_ytarg_ny)
self.op_losses = self.Ops(loss,
mse,
l2,
state=state,
ytarg_ny=ph_ytarg_ny)
def build_graph(self):
input = layer.ConfiguredInput(config.input)
hidden = layer.GenericLayers(layer.Flatten(input), [
dict(type=layer.Dense, size=size, activation=layer.Activation.Tanh)
for size in config.hidden_sizes
])
weights = [input, hidden]
if config.dueling_dqn:
if config.hidden_sizes:
v_input, a_input = tf.split(hidden.node, [
config.hidden_sizes[-1] // 2, config.hidden_sizes[-1] // 2
],
axis=1)
v_input = graph.TfNode(v_input)
a_input = graph.TfNode(a_input)
else:
v_input, a_input = hidden, hidden
v_output = layer.Dense(v_input, 1)
a_output = layer.Dense(a_input, config.output.action_size)
output = v_output.node + a_output.node - tf.reduce_mean(
a_output.node, axis=1, keep_dims=True)
output = graph.TfNode(output)
weights.extend([v_output, a_output])
else:
output = layer.Dense(hidden, config.output.action_size)
weights.append(output)
self.ph_state = input.ph_state
self.output = output
self.weights = layer.Weights(*weights)
def build_graph(self):
conv_layer = dict(type=layer.Convolution, activation=layer.Activation.Elu,
n_filters=32, filter_size=[3, 3], stride=[2, 2],
border=layer.Border.Same)
input = layer.Input(cfg.config.input, descs=[dict(conv_layer)] * 4)
shape = [None] + [cfg.config.output.action_size]
self.ph_probs = graph.Placeholder(np.float32, shape=shape, name='act_probs')
self.ph_taken = graph.Placeholder(np.int32, shape=(None,), name='act_taken')
flattened_input = layer.Flatten(input)
last_size = flattened_input.node.shape.as_list()[-1]
inverse_inp = graph.Reshape(input, [-1, last_size*2])
get_first = graph.TfNode(inverse_inp.node[:, :last_size])
get_second = graph.TfNode(inverse_inp.node[:, last_size:])
forward_inp = graph.Concat([get_first, self.ph_probs], axis=1)
fc_size = cfg.config.hidden_sizes[-1]
inv_fc1 = layer.Dense(inverse_inp, fc_size, layer.Activation.Relu)
inv_fc2 = layer.Dense(inv_fc1, shape[-1]) # layer.Activation.Softmax
fwd_fc1 = layer.Dense(forward_inp, fc_size, layer.Activation.Relu)
fwd_fc2 = layer.Dense(fwd_fc1, last_size)
inv_loss = graph.SparseSoftmaxCrossEntropyWithLogits(inv_fc2, self.ph_taken).op
fwd_loss = graph.L2loss(fwd_fc2.node - get_second.node).op
self.ph_state = input.ph_state # should be even wrt to batch_size for now
self.rew_out = graph.TfNode(cfg.config.icm.nu * fwd_loss)
self.loss = graph.TfNode(cfg.config.icm.beta * fwd_loss + (1 - cfg.config.icm.beta) * inv_loss)
layers = [input, inv_fc1, inv_fc2, fwd_fc1, fwd_fc2]
self.weights = layer.Weights(*layers)
def build_graph(self, input_placeholder):
conv_layer = dict(type=layer.Convolution,
activation=layer.Activation.Elu,
n_filters=32,
filter_size=[3, 3],
stride=[2, 2],
border=layer.Border.Same)
input_layers = [dict(conv_layer)
] * 4 if da3c_config.config.input.universe else None
input = layer.Input(da3c_config.config.input,
descs=input_layers,
input_placeholder=input_placeholder)
sizes = da3c_config.config.hidden_sizes
layers = [input]
flattened_input = layer.Flatten(input)
fc_layers = layer.GenericLayers(flattened_input, [
dict(
type=layer.Dense, size=size, activation=layer.Activation.Relu6)
for size in sizes[:-1]
])
layers.append(fc_layers)
last_size = fc_layers.node.shape.as_list()[-1]
if len(sizes) > 0:
last_size = sizes[-1]
if da3c_config.config.use_lstm:
lstm = layer.LSTM(graph.Expand(fc_layers, 0), n_units=last_size)
head = graph.Reshape(lstm, [-1, last_size])
layers.append(lstm)
self.ph_lstm_state = lstm.ph_state
self.lstm_zero_state = lstm.zero_state
self.lstm_state = lstm.state
else:
head = layer.Dense(fc_layers,
last_size,
activation=layer.Activation.Relu6)
layers.append(head)
self.ph_state = input.ph_state
self.weight = layer.Weights(*layers)
return head.node
def build_graph(self):
input = layer.ConfiguredInput(da3c_config.config.input)
sizes = da3c_config.config.hidden_sizes
layers = [input]
flattened_input = layer.Flatten(input)
last_size = flattened_input.node.shape.as_list()[-1]
if len(sizes) > 0:
last_size = sizes[-1]
if da3c_config.config.use_lstm:
lstm = layer.lstm(da3c_config.config.lstm_type,
graph.Expand(flattened_input, 0),
n_units=last_size,
n_cores=da3c_config.config.lstm_num_cores)
head = graph.Reshape(lstm, [-1, last_size])
layers.append(lstm)
self.ph_lstm_state = lstm.ph_state
self.lstm_zero_state = lstm.zero_state
self.lstm_state = lstm.state
self.lstm_reset_timestep = lstm.reset_timestep
else:
activation = layer.get_activation(da3c_config.config)
head = layer.GenericLayers(flattened_input, [
dict(type=layer.Dense, size=size, activation=activation)
for size in sizes
])
layers.append(head)
actor = layer.Actor(head, da3c_config.config.output)
critic = layer.Dense(head, 1)
layers.extend((actor, critic))
self.ph_state = input.ph_state
self.actor = actor
self.critic = graph.Flatten(critic)
self.weights = layer.Weights(*layers)
self.actor_weights = layer.Weights(actor)
def build_graph(self):
conv_layer = dict(type=layer.Convolution, activation=layer.Activation.Elu,
n_filters=32, filter_size=[3, 3], stride=[2, 2],
border=layer.Border.Same)
input_layers = [dict(conv_layer)] * 4 if da3c_config.config.input.universe else None
input = layer.Input(da3c_config.config.input, descs=input_layers)
sizes = da3c_config.config.hidden_sizes
layers = [input]
flattened_input = layer.Flatten(input)
last_size = flattened_input.node.shape.as_list()[-1]
if len(sizes) > 0:
last_size = sizes[-1]
if da3c_config.config.use_lstm:
lstm = layer.LSTM(graph.Expand(flattened_input, 0), n_units=last_size)
head = graph.Reshape(lstm, [-1, last_size])
layers.append(lstm)
self.ph_lstm_state = lstm.ph_state
self.lstm_zero_state = lstm.zero_state
self.lstm_state = lstm.state
else:
head = layer.GenericLayers(flattened_input,
[dict(type=layer.Dense, size=size,
activation=layer.Activation.Relu) for size in sizes])
layers.append(head)
actor = layer.Actor(head, da3c_config.config.output)
critic = layer.Dense(head, 1)
layers.extend((actor, critic))
self.ph_state = input.ph_state
self.actor = actor
self.critic = graph.Flatten(critic)
self.weights = layer.Weights(*layers)
self.actor_weights = layer.Weights(actor)