Description
The FFTSModel in models/ffts_model.py does not accept a pred_len parameter for specifying the prediction sequence length. The prediction head is hardcoded to use max_len (input sequence length), making it an equal-length prediction model only.
Current behavior
class FFTSModel(nn.Module):
def __init__(self, feat_dim, max_len, d_model, n_heads, num_layers,
patch_len=16, stride=8, dropout=0.1, activation='gelu',
num_experts=4, top_k=2):
...
# Prediction head hardcodes target_window = max_len
self.head = FlattenHead(feat_dim, self.head_nf, max_len, head_dropout=dropout)This means the model can only predict sequences of the same length as the input, which is inconsistent with other SOTA models in the time series forecasting (PiXTime, TimeXer, PatchTST, iTransformer, etc.) that support flexible prediction lengths.
Files affected
- models/ffts_model.py: FFTSModel.init, FFTSModel.forward
Description
The
FFTSModelinmodels/ffts_model.pydoes not accept apred_lenparameter for specifying the prediction sequence length. The prediction head is hardcoded to usemax_len(input sequence length), making it an equal-length prediction model only.Current behavior
This means the model can only predict sequences of the same length as the input, which is inconsistent with other SOTA models in the time series forecasting (PiXTime, TimeXer, PatchTST, iTransformer, etc.) that support flexible prediction lengths.
Files affected