If I want to change the length of the forecast, should I change the value of horizon?

[enze5088]

if set horizon ={3, 6, 12, 24}, change outputTensor and TrainYTenosr to (batchsize,horizon,Feature dimension)? But no matter how I change the value of horizon, the output is [batchsize,Feature dimension]. Should I adjust the output layer?

[MarcDjandji]

I encountered the same problem, and after some digging, I saw that the source of the issue is coming from the _batchify method of the Data_utility class, which is located In the utils.py. The _batchify method takes as input the horizon but never uses it resulting in no effect on the size of the input/output size.

[yajunDai]

I

the horizon is the h of y(t+h+1) . _batchify use 'self.h' is equal to horizon. And the output dim is dependent on h, I think.

[AlainOUYANG]

I encountered the same problem, and after some digging, I saw that the source of the issue is coming from the _batchify method of the Data_utility class, which is located In the utils.py. The _batchify method takes as input the horizon but never uses it resulting in no effect on the size of the input/output size.

Agreed. See the 70th line of utils.py

def _batchify(self, idx_set, horizon):

    n = len(idx_set);
    X = torch.zeros((n,self.P,self.m));
    Y = torch.zeros((n,self.m));

    for i in range(n):
        end = idx_set[i] - self.h + 1;
        start = end - self.P;
        X[i,:,:] = torch.from_numpy(self.dat[start:end, :]);
        Y[i,:] = torch.from_numpy(self.dat[idx_set[i], :]);  # <- This line.

It seems the code always performs one-step forecasting no matter how you change the horizon.

Am I right?

[YuboCoco]

I encountered the same problem, and after some digging, I saw that the source of the issue is coming from the _batchify method of the Data_utility class, which is located In the utils.py. The _batchify method takes as input the horizon but never uses it resulting in no effect on the size of the input/output size.

Agreed. See the 70th line of utils.py

def _batchify(self, idx_set, horizon):

    n = len(idx_set);
    X = torch.zeros((n,self.P,self.m));
    Y = torch.zeros((n,self.m));

    for i in range(n):
        end = idx_set[i] - self.h + 1;
        start = end - self.P;
        X[i,:,:] = torch.from_numpy(self.dat[start:end, :]);
        Y[i,:] = torch.from_numpy(self.dat[idx_set[i], :]);  # <- This line.

It seems the code always performs one-step forecasting no matter how you change the horizon.

Am I right?

You are right，this code only work for one step forecasting，it seems like the output result of One-step method have better performance.However, this is not effective in practical applications. I think change the shape of Y[ i , : ] to Y[i , output_step , : ] will be worked. : )