Informer must except pred_len > 1

[Hadar Sharvit]

when using the Informermodel and a pred_len == 1, we observe an index error

IndexError: max(): Expected reduction dim 2 to have non-zero size

the reason, we believe, is this:

• the forward method takes in x_dec that has shape (b,1,f) and called in the self.decoder module (a DecoderLayer layer)
• inside, a self-attention mechanism is performed with x_dec (denoted with x ), which in turn called the self.inner_attention module, which is the forward call of the ProbAttention class.
• in this step, keys=x=x_dec, hence the line _, L_K, _, _ = keys.shape sets L_K == 1, which in turns sets U_part to 0 as there is a log taken over L_K in the definition U_part = self.factor * np.ceil(np.log(L_K)).astype('int').item() # c*ln(L_k)
• the last call is the _prob_QK method of the ProbAttention class, that performs M = Q_K_sample.max(-1)[0] - torch.div(Q_K_sample.sum(-1), L_K), resulting in a IndexError: max(): Expected reduction dim 2 to have non-zero size. (Q_K_sample is an empty tensor)

---

the command we used is:

python run_longExp.py --is_training 1 --root_path ./dataset/ --data_path ETTh1.csv --model_id ETTh1_96 --model Informer --data ETTh1 --features M --seq_len 128 --pred_len 1 --label_len 0 --enc_in 7 --des Exp --itr 1 --batch_size 32 --learning_rate 0.005