yyjeremy / gpd Goto Github PK

The official implementation of the ICLR 2024 paper entitled "Spatio-Temporal Few-Shot Learning via Diffusive Neural Network Generation".

In this project, we propose a novel framework, GPD, which performs generative pre-training on a collection of model parameters optimized with data from source cities. Our proposed approach recasts spatio-temporal graph transfer learning as pre-training a generative hypernetwork, which generates tailored model parameters guided by prompts. Our framework has the potential to revolutionize smart city applications in data-scarce environments and contribute to more sustainable and efficient urban development.

• Tested OS: Linux
• Python >= 3.8
• torch == 1.12.0
• torch_geometric == 2.2.0
• Tensorboard

1. Install Pytorch with the correct CUDA version.
2. Use the pip install -r requirements.txt command to install all of the Python modules and packages used in this project.

The data used for training and evaluation can be found in Time-Series data. After downloading the data, move them to ./Data.

For each city, we provide the following data:

• Graph data: It records the adjacency matrix of the spatiotemporal graph.
• Time series data: It records the temporal sequential data for each node.

We provide two time-series datasets: crowd flow (including DC, BM, man) and traffic speed (including metr-la, pems-bay, shenzhen, chengdu_m).

The details of these two data sets are as follows:

To train node-level models with the traffic dataset, run:

cd Pretrain

CUDA_VISIBLE_DEVICES=0 python main.py --taskmode task4 --model v_GWN --test_data metr-la --ifnewname 1 --aftername TrafficData

After full-trained, run Pretrain\PrepareParams\model2tensor.py to extract parameters from the trained model. And put the params-dataset in ./Data.

To train diffusion model and generate the parameters of the target city:

cd GPD

CUDA_VISIBLE_DEVICES=0 python 1Dmain.py --expIndex 140 --targetDataset metr-la --modeldim 512 --epochs 80000 --diffusionstep 500 --basemodel v_GWN --denoise Trans1

• expIndex assigns a special number to the experiment.
• targetDataset specifies the target dataset, which can be selected from ['DC', 'BM', 'man', 'metr-la', 'pemes-bay', 'shenzhen', 'chengdu_m'].
• modeldim specifies the hidden dim of the Transformer.
• epochs specifies the number of iterations.
• diffusionstep specifies the total steps of the diffusion process.
• basemodel specifies the spatio-temporal graph model, which can be selected from ['v_STGCN5', 'v_GWN'].
• denoise model specifies the conditioning strategies, which can be selected from ['Trans1', 'Trans2', 'Trans3', 'Trans4', 'Trans5'].
  • Trans1: Pre-conditioning with inductive bias.
  • Trans2: Pre-conditioning.
  • Trans3: Pre-adaptive conditioning.
  • Trans4: Post-adaptive Conditioning.
  • Trans5: Adaptive norm conditioning.

The sample result is in GPD/Output/expXX/.

To finetune the generated parameters of the target city and evaluate, run:

cd Pretrain

CUDA_VISIBLE_DEVICES=0 python main.py --taskmode task7 --model v_GWN --test_data metr-la --ifnewname 1 --aftername finetune_7days --epochs 600 --target_days 7

• taskmode 'task7' means finetune after diffusion sampling.
• model specifies the spatio-temporal graph model, which can be selected from ['v_STGCN5', 'v_GWN'].
• test_data specifies the dataset, which can be selected from ['DC', 'BM', 'man', 'metr-la', 'pemes-bay', 'shenzhen', 'chengdu_m'].
• ifnewname assign 1 to better distinguish the results of the current experiment.
• aftername Use with --ifnewname 1 to give an identification name to the log file and results folder of the current experiment.
• epochs specifies the number of iterations.
• target_days specifies the amount of data used in finetune stage.

Let me give an example of the overall instructions. If you want to set 'metr-la' as target city:

• In pretrain: set the test_data as 'PMS-Bay', 'Didi-Chengdu', and 'Didi-Shenzhen' respectively to pretrain the models of other three source cities.
• In Diffusion: set the targetDataset as 'metr-la'.
• In finetune: set the test_dataset as 'metr-la'.

Since finetune and pretraining share the same code framework and use the same set of parameter names, this can be a little confusing and I will try to make the distinction between them in later versions of the code.