def __init__(self, hp):
super().__init__()
self._hp = hp
if self._hp.low_dim:
self.encoder = Linear(in_dim=self._hp.state_size * 2,
out_dim=128,
builder=self._hp.builder)
out_size = 128
else:
self.encoder = ConvEncoder(self._hp)
out_size = self.encoder.get_output_size()[0] * 5 * 5
self.mlp = torch.nn.Sequential()
self.mlp.add_module(
'linear_1',
Linear(in_dim=out_size, out_dim=128, builder=self._hp.builder))
for i in range(10):
self.mlp.add_module(
f'linear_{i+2}',
Linear(in_dim=128, out_dim=128, builder=self._hp.builder))
self.mlp.add_module(f'relu_{i+2}', torch.nn.ReLU())
self.mlp.add_module(
'linear_final',
Linear(in_dim=128,
out_dim=self._hp.action_size,
builder=self._hp.builder))
self.mlp.add_module('tanh', torch.nn.Tanh())
def __init__(self, hp, inp_dim, feature_size):
super().__init__()
self._hp = hp
self.inp_dim = inp_dim
self.feature_size = feature_size
self.linear = Linear(in_dim=inp_dim, out_dim=2*feature_size, builder=self._hp.builder)
def __init__(self, hp):
super().__init__()
self._hp = hp
if self._hp.low_dim:
self.linear1 = Linear(in_dim=4,
out_dim=128,
builder=self._hp.builder)
else:
self.encoder = ConvEncoder(self._hp)
out_size = self.encoder.get_output_size()
self.linear1 = Linear(in_dim=64 * 5 * 5,
out_dim=128,
builder=self._hp.builder)
self.linear2 = Linear(in_dim=128 + self._hp.action_size,
out_dim=1,
builder=self._hp.builder)
def build_network(self, build_encoder=True):
self.encoder = ConvEncoder(self._hp)
out_size = self.encoder.get_output_size()
self.spatial_softmax = SpatialSoftmax(
out_size[1], out_size[2], out_size[0]) # height, width, channel
self.linear = Linear(in_dim=out_size[0] * 2,
out_dim=1,
builder=self._hp.builder)
def __init__(self, hp):
super().__init__()
self._hp = hp
self.encoder = ConvEncoder(self._hp)
out_size = self.encoder.get_output_size()
self.linear1 = Linear(in_dim=64 * 5 * 5,
out_dim=128,
builder=self._hp.builder)
self.linear2 = Linear(in_dim=128,
out_dim=self._hp.hidden_size * 2,
builder=self._hp.builder)
self.linear3 = Linear(in_dim=self._hp.hidden_size,
out_dim=128,
builder=self._hp.builder)
self.linear4 = Linear(in_dim=128,
out_dim=64 * 5 * 5,
builder=self._hp.builder)
self.decoder = ConvDecoder(self._hp)
def __init__(self, hp, num_outputs):
super().__init__()
self._hp = hp
self.num_outputs = num_outputs
self.ll_size = ll_size = self._hp.linear_layer_size
self.fc_layers = torch.nn.ModuleList()
if self._hp.low_dim:
self.linear1 = Linear(in_dim=2 * self._hp.state_size,
out_dim=ll_size,
builder=self._hp.builder)
self.linear2 = Linear(in_dim=ll_size + self._hp.action_size,
out_dim=ll_size,
builder=self._hp.builder)
else:
self.encoder = ConvEncoder(self._hp)
out_size = self.encoder.get_output_size()
self.linear1 = Linear(in_dim=out_size[0] * out_size[1] *
out_size[2],
out_dim=ll_size,
builder=self._hp.builder)
self.linear2 = Linear(in_dim=ll_size + self._hp.action_size,
out_dim=ll_size,
builder=self._hp.builder)
for i in range(3):
self.fc_layers.append(
Linear(in_dim=ll_size,
out_dim=ll_size,
builder=self._hp.builder))
self.fc_layers.append(
Linear(in_dim=ll_size,
out_dim=self.num_outputs,
builder=self._hp.builder))
def build_network(self, build_encoder=True):
self.encoder = ConvEncoder(self._hp)
out_size = self.encoder.get_output_size()
if self._hp.spatial_softmax:
self.spatial_softmax = SpatialSoftmax(
out_size[1], out_size[2],
out_size[0]) # height, width, channel
self.linear = Linear(in_dim=out_size[0] * 2,
out_dim=1,
builder=self._hp.builder)
else:
self.linear = Linear(in_dim=256,
out_dim=1,
builder=self._hp.builder)
self.fc_layers = []
self.fc_layers.append(
Linear(in_dim=out_size[0] * out_size[1] * out_size[2],
out_dim=256,
builder=self._hp.builder))
for i in range(3):
self.fc_layers.append(
Linear(in_dim=256, out_dim=256, builder=self._hp.builder))
def __init__(self, hp):
super().__init__()
self._hp = hp
self.linear1 = Linear(in_dim=self._hp.hidden_size,
out_dim=self._hp.hidden_size,
builder=self._hp.builder)