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):
super(_WorkerNetwork, self).__init__()
self.lstm = CustomBasicLSTMCell(cfg.d) # d=256
# needs wrap as layer to retrieve weights
self.ph_goal =\
graph.Placeholder(np.float32, shape=(None, cfg.d), name="ph_goal")
# self.ph_goal = tf.placeholder(tf.float32, [None, cfg.d], name="ph_goal")
perception_expanded = graph.Expand(self.perception.node, 0)
self.ph_step_size = \
graph.Placeholder(np.float32, shape=(1,), name="ph_w_step_size")
# tf.placeholder(tf.float32, [1], name="ph_w_step_size")
self.ph_initial_lstm_state = \
graph.Placeholder(np.float32, shape=(1, self.lstm.state_size), name="ph_w_lstm_state")
# tf.placeholder(tf.float32, [1, self.lstm.state_size], name="ph_w_lstm_state")
lstm_outputs, self.lstm_state = tf.nn.dynamic_rnn(
self.lstm,
perception_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)
U = layer.LinearLayer(sg_lstm_outputs,
shape=(cfg.d, cfg.action_size * cfg.k),
transformation=tf.matmul)
U_embedding = tf.transpose(tf.reshape(U, [cfg.action_size, cfg.k, -1]))
w = layer.LinearLayer(self.ph_goal,
shape=(cfg.d, cfg.k),
transformation=tf.matmul,
bias=False)
w_reshaped = tf.reshape(w.node, [-1, 1, cfg.k])
self.pi = layer.MatmulLayer(w_reshaped,
U_embedding,
activation=layer.Activation.Softmax)
self.vi = layer.LinearLayer(sg_lstm_outputs,
shape=(cfg.d, 1),
transformation=tf.matmul)
self.weights = layer.Weights(
self.weights, graph.TfNode((self.lstm.matrix, self.lstm.bias)), U,
w, self.vi)
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_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):
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.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)