two questions about timer

[zxl696]

1.            if self.args.use_ims:
                # output used to calculate loss misaligned patch_len compared to input
                loss = criterion(outputs[:, -self.args.seq_len:, :], batch_y)
            else:
                # only use the forecast window to calculate loss
                loss = criterion(outputs[:, -self.args.pred_len:, :], batch_y[:, -self.args.pred_len:, :])
   

代码这里的计算损失为什么是这样的，使用ims为啥这样计算损失，训练函数里面也没看出来使用ims他的训练逻辑
2. if self.data_type == 'custom':
data_len = len(df_raw)
num_train = int(data_len * 0.75)
num_test = int(data_len * 0.15)
num_vali = data_len - num_train - num_test
border1s = [0, num_train-self.input_len, data_len - num_test-self.input_len]
border2s = [num_train, num_train + num_vali, data_len]
这样分配起始点会不会造成测试集和验证集提前知道数据了

[WenWeiTHU]

Hi, please refer to the two relevant issue:

• Question about finetune loss #24
• Operation on Dataset Autoformer#53
• Train, Val, Test dataset leakage Autoformer#20