def __init__(self, seq2reward_network: Seq2RewardNetwork,
params: Seq2RewardTrainerParameters):
self.seq2reward_network = seq2reward_network
self.params = params
self.optimizer = torch.optim.Adam(self.seq2reward_network.parameters(),
lr=params.learning_rate)
self.minibatch_size = self.params.batch_size
self.loss_reporter = NoOpLossReporter()
# PageHandler must use this to activate evaluator:
self.calc_cpe_in_training = True
# Turning off Q value output during training:
self.view_q_value = params.view_q_value
# permutations used to do planning
device = get_device(self.seq2reward_network)
self.all_permut = gen_permutations(
params.multi_steps, len(self.params.action_names)).to(device)
def __init__(
self,
compress_model_network: FullyConnectedNetwork,
seq2reward_network: Seq2RewardNetwork,
params: Seq2RewardTrainerParameters,
):
self.compress_model_network = compress_model_network
self.seq2reward_network = seq2reward_network
self.params = params
self.optimizer = torch.optim.Adam(
self.compress_model_network.parameters(),
lr=params.compress_model_learning_rate,
)
self.minibatch_size = self.params.compress_model_batch_size
self.loss_reporter = NoOpLossReporter()
# PageHandler must use this to activate evaluator:
self.calc_cpe_in_training = True
# permutations used to do planning
device = get_device(self.compress_model_network)
self.all_permut = gen_permutations(
params.multi_steps, len(self.params.action_names)).to(device)