search

cure-lab / SCINet

The GitHub repository for the paper: “Time Series is a Special Sequence: Forecasting with Sample Convolution and Interaction“. (NeurIPS 2022)
Apache License 2.0
633 stars 128 forks source link

Questions about training time #20

Closed MatthewYZhang closed 2 years ago

MatthewYZhang commented 2 years ago

Hi all,

Appreciate your work.

I tried to train using the following command, and I noticed that it may take about 10 minutes to train only one epoch. I am using V100 to train.

python run_financial.py --dataset_name electricity --window_size 168 --horizon 24 --hidden-size 8 --single_step 1 --stacks 2 --levels 3 --lr 9e-3 --dropout 0 --batch_size 32 --model_name ele_I168_o24_lr9e-3_bs32_dp0_h8_s2l3_w0.5 --groups 321

I wonder if this is normal, because as far as I am concerned, your network mainly consists of conv1d and sampling, and they may not took that much time if If you can kindly offer some training specs and training time, I would be very thankful.

Matthew

ailingzengzzz commented 2 years ago

Hi @MatthewYZhang ,

It's normal. In the "electricity" dataset, we regard it as a univariate dataset. Thus, we use "--groups 321" (321 is the number of variates) in the arguments to split the spatial influence on different varieties. This operation is also known as a “depthwise convolution”, and it will take a longer time to be processed.