def step(self, optimizer, arch_optimizer):
self.train()
losses = []
train_bandit = utils.get_data_queue_bandit(self.args, self.contexts,
self.pos_weights)
cnt = 0
time_data = 0
time_forward = 0
time_update = 0
end = -1
for step, features in enumerate(train_bandit):
if end != -1:
time_data += time.time() - end
begin = time.time()
optimizer.zero_grad()
arch_optimizer.zero_grad()
output, error_loss, loss_alpha = self.forward(features)
time_forward += time.time() - begin
losses.append(error_loss.cpu().detach().item())
begin2 = time.time()
optimizer.step()
arch_optimizer.step()
time_update += time.time() - begin2
cnt += 1
end = time.time()
print("time_data: ", time_data)
print("time_forward: ", time_forward)
print("time_update: ", time_update)
print("cnt: ", cnt)
return np.mean(losses)
def step(self, optimizer, arch_optimizer):
self.train()
losses = []
train_bandit = utils.get_data_queue_bandit(self.args, self.contexts, self.pos_weights)
cnt = 0
for step, features in enumerate(train_bandit):
optimizer.zero_grad()
output, error_loss, loss_alpha = self.forward(features)
losses.append(error_loss.cpu().detach().item())
optimizer.step()
cnt += 1
print("cnt: ", cnt)
return np.mean(losses)
def step(self, optimizer, arch_optimizer, epoch):
self.train()
losses = []
train_bandit = utils.get_data_queue_bandit(self.args, self.contexts,
self.pos_weights)
if epoch < self.args.search_epoch:
train_epoch = (epoch + 1) * 5
else:
train_epoch = 1
for k in range(train_epoch):
for step, features in enumerate(train_bandit):
optimizer.zero_grad()
output, error_loss, loss_alpha = self.forward(features)
losses.append(error_loss.cpu().detach().item())
optimizer.step()
return np.mean(losses)
