Turn out too well

[TiAmo888]

Why are your results so good? Even if the method you proposed is removed, the prediction results are still very good using only the full connection layer. I suspect that there may be problems with the method you proposed

[jackyue1994]

Yeah, it seems that FreMLP is not functioning as intended when it operates on the embed_size dimension. It fails to learn any inter-series or intra-series dependencies. See issue 12.

[vincent05r]

did you keep the dimension extension layer? if so that might be very interesting since it suggests that the linear layer in it does some very heavy lifting.