ST-ResNet

Pytorch version of ST-ResNet

The code implementation refers to [ST-ResNet-Pytorch](https://github.com/BruceBinBoxing/ST-ResNet-Pytorch)

Dataset

Four types of input data:

• Closeness: (bs, 2 * l_c, h, w)
• Period: (bs, 2 * l_p, h, w)
• Trend: (bs, 2 * l_t, h, w)
• External-Feature: (bs, 28)

TaxiBJ:

l_c, l_p, l_t = 3, 1, 1
h, w = 32, 32
bs = 500

Network

Four Way Network

C-Way, P-Way, T-Way are stack by ResUnit

E-Way is stack by nn.linear() and nn.ReLU()

Experiment

ID	resume	structure	training mode	TaxiBJ(rmse)	BikeNYC(rmse)
20220725A	random(42)	L4-woBN	training	33.1854	-
20220725B	20220725A	L4-woBN	finetuning	-	-

Note:

• training: train set, 500epoch;
• finetuning: train+val set, 100epoch;
• woBN: without BN layers;