CSE5002 Course Project by Zheng Dong
Training code is adapted from my repo Torch-MTS, which I developed and maintained for my personal research.
The required packages are:
pytorch>=1.11
numpy
scipy
scikit-learn
pandas
matplotlib
pyyaml
torchinfo
-
mkdir data/and runscripts/gen_dataset.pyThis will generate four
.npzfiles indata/:adj.npz, data.npz, label.npz, indices.npz.You can skip this step because I already uploaded the four files.
-
Run models: ADFGCN, LapeGCN, MLP
Here we use ADFGCN as an example.
- Modify the hyper-parameters in
configs/ADFGCN.yaml - Run model train-test by
python scripts/train.py -m adfgcn -g <your_gpu_id>
- Training logs will be saved in
logs/ADFGCN/ - Trained models will be saved in
saved_models/ADFGCN/, with.ptformat
Other models are similar.
Example training log:
<logs/ADFGCN-2023-05-20-20-03-08.log> x-(5298, 6) y-(5298,) --------- ADFGCN --------- { "lr": 0.01, "weight_decay": 0.001, "milestones": [ 10 ], "clip_grad": 5, "max_epochs": 1000, "early_stop": 50, "pass_device": true, "model_args": { "num_nodes": 5298, "adj_path": "../data/adj.npz", "input_classes_list": [ 6, 3, 43, 44, 64, 2506 ], "embedding_dim": 8, "output_dim": 11, "hidden_dim": 32, "node_embedding_dim": 16, "cheb_k": 1, "num_layers": 2, "dropout": 0.1, "device": "cuda:0" } } ========================================================================================== Layer (type:depth-idx) Output Shape Param # ========================================================================================== ADFGCN [5298, 11] 169,536 ├─ModuleList: 1-1 -- -- │ └─Embedding: 2-1 [5298, 8] 48 │ └─Embedding: 2-2 [5298, 8] 24 │ └─Embedding: 2-3 [5298, 8] 344 │ └─Embedding: 2-4 [5298, 8] 352 │ └─Embedding: 2-5 [5298, 8] 512 │ └─Embedding: 2-6 [5298, 8] 20,048 ├─Linear: 1-2 [5298, 32] 1,568 ├─ModuleList: 1-3 -- -- │ └─DFGCN: 2-7 [5298, 32] 5,152 │ │ └─Dropout: 3-1 [5298, 32] -- │ │ └─Dropout: 3-2 [5298, 32] -- │ │ └─Dropout: 3-3 [5298, 32] -- │ │ └─Dropout: 3-4 [5298, 32] -- │ │ └─Dropout: 3-5 [5298, 32] -- │ └─DFGCN: 2-8 [5298, 32] 5,152 │ │ └─Dropout: 3-6 [5298, 32] -- │ │ └─Dropout: 3-7 [5298, 32] -- │ │ └─Dropout: 3-8 [5298, 32] -- │ │ └─Dropout: 3-9 [5298, 32] -- │ │ └─Dropout: 3-10 [5298, 32] -- ├─Sequential: 1-4 [5298, 11] -- │ └─Linear: 2-9 [5298, 32] 2,080 │ └─ReLU: 2-10 [5298, 32] -- │ └─Dropout: 2-11 [5298, 32] -- │ └─Linear: 2-12 [5298, 11] 363 ========================================================================================== Total params: 205,179 Trainable params: 205,179 Non-trainable params: 0 Total mult-adds (M): 134.25 ========================================================================================== Input size (MB): 0.13 Forward/backward pass size (MB): 5.21 Params size (MB): 0.10 Estimated Total Size (MB): 5.44 ========================================================================================== Loss: CrossEntropyLoss 2023-05-20 20:03:09.446458 Epoch 1 Train Loss = 2.48147 Train acc = 0.02736 Val acc = 0.18200 2023-05-20 20:03:09.672309 Epoch 2 Train Loss = 2.17014 Train acc = 0.20219 Val acc = 0.15600 2023-05-20 20:03:09.897267 Epoch 3 Train Loss = 2.00025 Train acc = 0.18836 Val acc = 0.27000 2023-05-20 20:03:10.121297 Epoch 4 Train Loss = 1.89027 Train acc = 0.25605 Val acc = 0.33800 2023-05-20 20:03:10.345895 Epoch 5 Train Loss = 1.79081 Train acc = 0.30674 Val acc = 0.40000 ... 2023-05-20 20:04:38.384733 Epoch 400 Train Loss = 0.61484 Train acc = 0.75634 Val acc = 0.73200 2023-05-20 20:04:38.597126 Epoch 401 Train Loss = 0.61650 Train acc = 0.75634 Val acc = 0.73200 2023-05-20 20:04:38.823033 Epoch 402 Train Loss = 0.61868 Train acc = 0.75317 Val acc = 0.73800 2023-05-20 20:04:39.044909 Epoch 403 Train Loss = 0.61037 Train acc = 0.76872 Val acc = 0.73200 2023-05-20 20:04:39.266680 Epoch 404 Train Loss = 0.60906 Train acc = 0.76181 Val acc = 0.73400 Early stopping at epoch: 404 Best at epoch 354: Train Loss = 0.64359 Train acc = 0.74827 Val acc = 0.74600 --------- Test --------- Test acc = 0.70801 Inference time: 0.02 s - Modify the hyper-parameters in