def __init__(
self,
in_channels,
convnet_out_size,
out_dim,
num_modules,
num_agents,
):
nn.Module.__init__(self)
print('Per task exclusive: all fc')
self.convolutions = nn.Sequential(
SimpleConvNetBlock(in_channels, 32, 3),
SimpleConvNetBlock(32, 32, 3), SimpleConvNetBlock(32, 32, 3),
SimpleConvNetBlock(32, 32, 3), nn.BatchNorm2d(32), Flatten())
# self._loss_layer = Loss(torch.nn.MSELoss(), CorrectClassifiedReward(), discounting=1.)
self._loss_layer = Loss(torch.nn.MSELoss(),
NegLossReward(),
discounting=1.)
self.fc_layers = Sequential(
PerTaskAssignment(),
Selection(*[
LinearWithRelu(convnet_out_size, 48)
for _ in range(num_modules)
]),
Selection(*[LinearWithRelu(48, 48) for _ in range(num_modules)]),
Selection(*[nn.Linear(48, out_dim) for _ in range(num_modules)]),
)
def __init__(self, decision_maker, in_channels, convnet_out_size, out_dim, num_modules, num_agents):
nn.Module.__init__(self)
self.convolutions = nn.Sequential(
SimpleConvNetBlock(in_channels, 32, 3),
SimpleConvNetBlock(32, 32, 3),
SimpleConvNetBlock(32, 32, 3),
SimpleConvNetBlock(32, 32, 3),
nn.BatchNorm2d(32),
Flatten()
)
self._loss_func = Loss(torch.nn.MSELoss(), CorrectClassifiedReward(), discounting=1.)
self._initialization = Initialization()
self._per_task_assignment = PerTaskAssignment()
self._decision_1 = decision_maker(
num_modules, convnet_out_size, num_agents=num_agents, policy_storage_type='tabular',
additional_reward_func=CollaborationReward(reward_ratio=0.3, num_actions=num_modules))
self._decision_2 = decision_maker(
num_modules, convnet_out_size, num_agents=num_agents, policy_storage_type='tabular',
additional_reward_func=CollaborationReward(reward_ratio=0.3, num_actions=num_modules))
self._decision_3 = decision_maker(
num_modules, convnet_out_size, num_agents=num_agents, policy_storage_type='tabular',
additional_reward_func=CollaborationReward(reward_ratio=0.3, num_actions=num_modules))
self._selection_1 = Selection(*[LinearWithRelu(convnet_out_size, 48) for _ in range(num_modules)])
self._selection_2 = Selection(*[LinearWithRelu(48, 48) for _ in range(num_modules)])
self._selection_3 = Selection(*[nn.Linear(48, out_dim) for _ in range(num_modules)])
def __init__(
self,
in_channels,
convnet_out_size,
out_dim,
num_modules,
num_agents,
):
PerTask_all_fc.__init__(
self,
in_channels,
convnet_out_size,
out_dim,
num_modules,
num_agents,
)
print('Per task exclusive: last fc')
self.convolutions = nn.Sequential(self.convolutions, Flatten(),
LinearWithRelu(convnet_out_size, 48),
LinearWithRelu(48, 48))
self._loss_layer = Loss(torch.nn.MSELoss(),
CorrectClassifiedReward(),
discounting=1.)
self.fc_layers = Sequential(
PerTaskAssignment(),
Selection(*[nn.Linear(48, out_dim) for _ in range(num_modules)]),
)
def _create_routing(self, routing_agent, num_agents, exploration, policy_storage_type, detach, approx_hidden_dims,
additional_reward_func, dimensionality_defs):
list_type = nn.ModuleList if not self.recurrent else FakeFlatModuleList
effective_width = self.routing_width if not self.recurrent else self.routing_width + 1 # for termination action
effective_depth = self.routing_depth if not self.recurrent else 1
base_selection = [] if not self.recurrent else [Identity()]
router = list_type([
routing_agent(
num_selections=effective_width,
in_features=dimensionality_defs[i],
num_agents=num_agents,
exploration=exploration,
policy_storage_type=policy_storage_type,
detach=detach,
approx_hidden_dims=approx_hidden_dims,
additional_reward_func=additional_reward_func
) for i in range(effective_depth)
])
selection = list_type([
Selection(*(base_selection + [ # need base selection for termination action
nn.Linear(dimensionality_defs[i], dimensionality_defs[i + 1])
for _ in range(effective_width)
]))
for i in range(effective_depth)
])
return router, selection
def __init__(self, decision_maker, in_channels, convnet_out_size, out_dim,
num_modules, num_agents):
PerTask_all_fc.__init__(self, in_channels, convnet_out_size, out_dim,
num_modules, num_agents)
print('Routing Networks: all fc')
self._initialization = Initialization()
# self._per_task_assignment = PerTaskAssignment()
self._per_task_assignment = decision_maker(
num_agents,
convnet_out_size,
num_agents=1,
policy_storage_type='approx',
additional_reward_func=CollaborationReward(
reward_ratio=0.3, num_actions=num_modules))
self._decision_1 = decision_maker(
num_modules,
convnet_out_size,
num_agents=num_agents,
policy_storage_type='approx',
additional_reward_func=CollaborationReward(
reward_ratio=0.3, num_actions=num_modules))
self._decision_2 = decision_maker(
num_modules,
48,
num_agents=num_agents,
policy_storage_type='approx',
additional_reward_func=CollaborationReward(
reward_ratio=0.3, num_actions=num_modules))
self._decision_3 = decision_maker(
num_modules,
48,
num_agents=num_agents,
policy_storage_type='approx',
additional_reward_func=CollaborationReward(
reward_ratio=0.3, num_actions=num_modules))
self._selection_1 = Selection(
*
[LinearWithRelu(convnet_out_size, 48) for _ in range(num_modules)])
self._selection_2 = Selection(
*[LinearWithRelu(48, 48) for _ in range(num_modules)])
self._selection_3 = Selection(
*[nn.Linear(48, out_dim) for _ in range(num_modules)])