Future prediction length

[MartinPerry]

If the output prediction length is longer than input frames, it is solved via this code (in SimVP class and same logic is applied to WaST and TAU as well):

pred_y = []
d = aft_seq_length // pre_seq_length
m = aft_seq_length % pre_seq_length
            
cur_seq = batch_x.clone()
for _ in range(d):
	cur_seq = self.model(cur_seq)
    pred_y.append(cur_seq)

if m != 0:
	cur_seq = self.model(cur_seq)
    pred_y.append(cur_seq[:, :m])
            
pred_y = torch.cat(pred_y, dim=1) 

Doesn't this degrade quality, when input to the model are basically the predicted frames? So the model sees correct past data and later some predicted data affected by error?