def build_model(self):
"""Creates and initializes the shared and controller models."""
if self.args.network_type == 'Net':
self.shared = models.Network(self.args)
else:
raise NotImplementedError(f'Network type '
f'`{self.args.network_type}` is not '
f'defined')
self.controller = models.Controller(self.args)
def build_model(self):
"""Creates and initializes the shared and controller models."""
self.shared = models.RNN(self.args, self.dataset)
self.controller = models.Controller(self.args)
if self.args.num_gpu == 1:
self.shared.cuda()
self.controller.cuda()
elif self.args.num_gpu > 1:
raise NotImplementedError('`num_gpu > 1` is in progress')
def build_model(self):
"""Creates and initializes the shared and controller models."""
if self.args.network_type == 'unet':
self.shared = models.Unet(self.args)
else:
raise NotImplementedError(f'Network type '
f'`{self.args.network_type}` is not '
f'defined')
self.controller = models.Controller(self.args)
if self.args.num_gpu == 1:
self.shared.cuda()
self.controller.cuda()
elif self.args.num_gpu > 1:
raise NotImplementedError('`num_gpu > 1` is in progress')
def build_model(self):
"""Creates and initializes the shared and controller models."""
if self.args.network_type == 'rnn':
self.shared = models.RNN(self.args, self.dataset)
elif self.args.network_type == 'cnn':
self.shared = models.CNN(self.args, self.dataset)
else:
raise NotImplementedError(
'Network type `{0}` is not defined'.format(
self.args.network_type))
self.controller = models.Controller(
self.args
) # 构建了一个orward：Embedding（130,100）->lstm（100,100）->decoder的列表，对应25个decoder
if self.args.num_gpu == 1:
self.shared.cuda()
self.controller.cuda()
elif self.args.num_gpu > 1:
raise NotImplementedError('`num_gpu > 1` is in progress')
def build_model(self):
"""Creates and initializes the shared and controller models."""
if self.args.network_type == 'rnn':
self.shared = models.RNN(self.args, self.dataset)
self.controller = models.Controller(self.args)
elif self.args.network_type == 'micro_cnn':
self.shared = models.CNN(self.args, self.dataset)
self.controller = models.CNNMicroController(self.args)
else:
raise NotImplementedError(f'Network type '
f'`{self.args.network_type}` is not '
f'defined')
if self.args.num_gpu == 1:
if torch.__version__ == '0.3.1':
self.shared.cuda()
self.controller.cuda()
else:
self.shared.to(self.device)
self.controller.to(self.device)
elif self.args.num_gpu > 1:
raise NotImplementedError('`num_gpu > 1` is in progress')
def build_model(self):
"""Creates and initializes the shared and controller models."""
if self.args.network_type == 'rnn':
self.shared = models.RNN(self.args, self.dataset)
elif self.args.network_type == 'cnn':
print("----- begin to init cnn------")
self.shared = models.CNN(self.args, self.dataset)
# self.shared = self.shared.cuda()
else:
raise NotImplementedError(f'Network type '
f'`{self.args.network_type}` is not '
f'defined')
print("---- begin to init controller-----")
self.controller = models.Controller(self.args)
#self.controller = self.controller.cuda()
print("===begin to cuda")
if True:
print("cuda")
self.shared.cuda()
self.controller.cuda()
print("finish cuda")
elif self.args.num_gpu > 1:
raise NotImplementedError('`num_gpu > 1` is in process')
def train_controller(self):
avg_reward_base = None
baseline = None
adv_history = []
entropy_history = []
reward_history = []
controller = models.Controller(self.n_tranformers, self.n_scalers,
self.n_constructers, self.n_selecters,
self.n_models, self.func_names,
self.lstm_size, self.temperature,
self.tanh_constant, self.save_dir)
controller_optimizer = _get_optimizer(self.optimizer)
controller_optim = controller_optimizer(controller.parameters(),
lr=self.controller_lr)
controller.train()
total_loss = 0
results_dag = []
results_acc = []
random_history = []
acc_history = []
for step in range(self.controller_max_step):
# sample models
dags, actions, sample_entropy, sample_log_probs = controller()
sample_entropy = torch.sum(sample_entropy)
sample_log_probs = torch.sum(sample_log_probs)
# print(sample_log_probs)
print(actions)
random_actions = self.random_actions()
with torch.no_grad():
acc = self.get_reward(actions)
random_acc = self.get_reward(torch.LongTensor(random_actions))
random_history.append(random_acc)
results_acc.append(acc)
results_dag.append(dags)
acc_history.append(acc)
rewards = torch.tensor(acc)
if self.entropy_weight is not None:
rewards += self.entropy_weight * sample_entropy
reward_history.append(rewards)
entropy_history.append(sample_entropy)
# moving average baseline
if baseline is None:
baseline = rewards
else:
decay = self.ema_baseline_decay
baseline = decay * baseline + (1 - decay) * rewards
adv = rewards - baseline
adv_history.append(adv)
# policy loss
loss = sample_log_probs * adv
# update
controller_optim.zero_grad()
loss.backward()
if self.controller_grad_clip > 0:
torch.nn.utils.clip_grad_norm(controller.parameters(),
self.controller_grad_clip)
controller_optim.step()
total_loss += loss.item()
if ((step % self.log_step) == 0) and (step > 0):
self._summarize_controller_train(total_loss, adv_history,
entropy_history,
reward_history, acc_history,
random_history,
avg_reward_base, dags)
reward_history, adv_history, entropy_history,acc_history,random_history = [], [], [],[],[]
total_loss = 0
self.controller_step += 1
max_acc = np.max(results_acc)
max_dag = results_dag[np.argmax(results_acc)]
path = os.path.join(self.model_dir, 'networks', 'best.png')
utils.draw_network(max_dag[0], path)
# np.sort(results_acc)[-10:]
return np.sort(list(set(results_acc)))[-10:]