Pytorch version of ST-ResNet
The code implementation refers to [ST-ResNet-Pytorch](https://github.com/BruceBinBoxing/ST-ResNet-Pytorch)
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 = 500Four Way Network
C-Way, P-Way, T-Way are stack by ResUnit
E-Way is stack by nn.linear() and nn.ReLU()
| 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;