def model(action_axes):
return neon.Sequential([
neon.Affine(
nout=10,
weight_init=neon.GlorotInit(),
bias_init=neon.ConstantInit(),
activation=neon.Tanh(),
),
neon.Affine(weight_init=neon.GlorotInit(),
bias_init=neon.ConstantInit(),
activation=neon.Tanh(),
axes=(action_axes, )),
])
def __init__(self, activation_function=neon.Rectlin(), name="middleware_embedder"):
self.name = name
self.input = None
self.output = None
self.weights_init = neon.GlorotInit()
self.biases_init = neon.ConstantInit()
self.activation_function = activation_function
def baselines_model(action_axes):
return neon.Sequential([
neon.Affine(
nout=64,
weight_init=neon.XavierInit(),
bias_init=neon.ConstantInit(),
activation=neon.Rectlin(),
batch_norm=False,
),
neon.Affine(
axes=action_axes,
weight_init=neon.XavierInit(),
bias_init=neon.ConstantInit(),
activation=None,
),
])
def __init__(self, input_size, batch_size=None, activation_function=neon.Rectlin(), name="embedder"):
self.name = name
self.input_size = input_size
self.batch_size = batch_size
self.activation_function = activation_function
self.weights_init = neon.GlorotInit()
self.biases_init = neon.ConstantInit()
self.input = None
self.output = None
def __init__(self, tuning_parameters, head_idx=0, loss_weight=1., is_local=True):
self.head_idx = head_idx
self.name = "head"
self.output = []
self.loss = []
self.loss_type = []
self.regularizations = []
self.loss_weight = force_list(loss_weight)
self.weights_init = neon.GlorotInit()
self.biases_init = neon.ConstantInit()
self.target = []
self.input = []
self.is_local = is_local
self.batch_size = tuning_parameters.batch_size
def model(action_axes):
"""
Given the expected action axes, return a model mapping from observation to
action axes for use by the dqn agent.
"""
return neon.Sequential([
neon.Preprocess(lambda x: x / 255.0),
neon.Convolution(
(8, 8, 32),
neon.XavierInit(),
strides=4,
activation=neon.Rectlin(),
),
neon.Convolution(
(4, 4, 64),
neon.XavierInit(),
strides=2,
activation=neon.Rectlin(),
),
neon.Convolution(
(3, 3, 64),
neon.XavierInit(),
strides=1,
activation=neon.Rectlin(),
),
neon.Affine(
nout=512,
weight_init=neon.XavierInit(),
bias_init=neon.ConstantInit(),
activation=neon.Rectlin(),
),
neon.Affine(weight_init=neon.XavierInit(),
bias_init=neon.ConstantInit(),
activation=None,
axes=(action_axes, )),
])