Optuna Pytorch:目标函数返回的值不能强制转换为浮点值

Question

def autotune(trial):
cfg= { 'device' : "cuda" if torch.cuda.is_available() else "cpu",
#   'train_batch_size' : 64,
#   'test_batch_size' : 1000,
#   'n_epochs' : 1,
#   'seed' : 0,
#   'log_interval' : 100,
#   'save_model' : False,
#   'dropout_rate' : trial.suggest_uniform('dropout_rate',0,1.0),
'lr' : trial.suggest_loguniform('lr', 1e-3, 1e-2),
'momentum' : trial.suggest_uniform('momentum', 0.4, 0.99),
'optimizer': trial.suggest_categorical('optimizer',[torch.optim.Adam,torch.optim.SGD, torch.optim.RMSprop, torch.optim.$
'activation': F.tanh}
optimizer = cfg['optimizer'](model.parameters(), lr=cfg['lr'])
#optimizer = torch.optim.Adam(model.parameters(),lr=0.001

正如你在上面看到的，我正在尝试运行Optuna试验，为我的CNN模型搜索最优化的超参数。

# Train the model
# use small epoch for large dataset
# An epoch is 1 run through all the training data
# losses = [] # use this array for plotting losses
for _ in range(epochs):
# using data_loader 
for i, (data, labels) in enumerate(trainloader):
# Forward and get a prediction
# x is the training data which is X_train
if name.lower() == "rnn":
model.hidden = (torch.zeros(1,1,model.hidden_sz),
torch.zeros(1,1,model.hidden_sz))
y_pred = model.forward(data)
# compute loss/error by comparing predicted out vs acutal labels
loss = criterion(y_pred, labels)
#losses.append(loss)
if i%10==0:  # print out loss at every 10 epoch
print(f'epoch {i} and loss is: {loss}')
#Backpropagation
optimizer.zero_grad()
loss.backward()
optimizer.step()
study = optuna.create_study(sampler=optuna.samplers.TPESampler(), direction='minimize',pruner=optuna.pruners.SuccessiveHalvingPrune$
study.optimize(autotune, n_trials=1)

但是，当我运行上面的代码来调整并找出我的最佳参数时，出现了下面的错误，似乎试验失败了，尽管我仍然得到了历元损失和值。请告知，谢谢！

[W 2020-11-11 13:59:48,000] Trial 0 failed, because the returned value from the objective function cannot be cast to float. Returned value is: None
Traceback (most recent call last):
File "autotune2", line 481, in <module>
n_instances, n_features, scores = run_analysis()
File "autotune2", line 350, in run_analysis
print(study.best_params)
File "/home/shar/anaconda3/lib/python3.7/site-packages/optuna/study.py", line 67, in best_params
return self.best_trial.params
File "/home/shar/anaconda3/lib/python3.7/site-packages/optuna/study.py", line 92, in best_trial
return copy.deepcopy(self._storage.get_best_trial(self._study_id))
File "/home/shar/anaconda3/lib/python3.7/site-packages/optuna/storages/_in_memory.py", line 287, in get_best_trial
raise ValueError("No trials are completed yet.")
ValueError: No trials are completed yet.

Answer 1

由于研究的目标函数必须返回浮点值，因此引发此异常。

在你的情况下，问题出在这一行：

study.optimize(autotune, n_trials=1)

您之前定义的自动调谐函数不返回值，也不能用于优化。

如何修复？

对于超参数搜索，自动调谐函数必须返回经过一些训练后可以获得的某个度量，如损失或交叉熵。

对代码的快速修复可以是这样的：

def autotune():
cfg= { 'device' : "cuda" if torch.cuda.is_available() else "cpu"
...etc...
}
best_loss = 1e100;  # or larger
# Train the model
for _ in range(epochs):
for i, (data, labels) in enumerate(trainloader):
... (train the model) ...
# compute loss/error by comparing predicted out vs actual labels
loss = criterion(y_pred, labels)
best_loss = min(loss,best_loss)
return best_loss

Optuna repo中的Pythorch有一个很好的例子，它使用pythoch回调来检索准确性(但如果需要，可以很容易地更改为使用RMSE(。它还使用了一个以上的实验，并取超参数的中值。