The repo for temporal and spatial speed models, targets two scenarios:

• Speed imputation: increase the coverage of speed derived from GPS observations.
• Speed forecasting: predict future speed in the next 15 minute or 1 hour.

• Graph Convolutional Networks + LSTMs/CNNs
• (unfinished) Tensor Factorization

• PyTorch version >= 1.2.0
• Python version >= 3.6
• For training new models, you'll also need an NVIDIA GPU

• preprocess.py does feature extraction
• trainer.py train TGCN model
  • python trainer.py --train: train the data
  • python trainer.py --test: predict data and save in parquet format

For each subgrid and each dataset (i.e., test, train, validation) a parquet file with predicted values is created. With the following schema:

message schema {
  optional double 1553983200;
  optional double 1553986800;
  optional double 1553990400;
  optional double 1553994000;
  optional double 1553997600;
  optional double 1554001200;
  optional double 1554004800;
  optional double 1554008400;
  optional double 1554012000;
  ...
  optional double 1554883200;
  optional int64 from_node;
  optional int64 to_node;
}

We can claim GPU(s) when creating a pod using the k8s-trainer.yml script. Currently, only one GPU can be claimed for one pod.

To speed up and optimize CPU / GPU loads, we move all the data tranformation logics from Pytorch dataloader to preprocessing component. In specific:

• The preprocess.py script expects data with numpy tensors (each tensor corresponds to a grid), in the shape of [num_nodes, 1, num_timesteps]. Currently, all the downstream transformations are done using numpy / pandas on a single machine.

  Then for every tensor, we enrich with features extracted from JURBEY and some basic time-senstive features, window-slicing it to new tensor of shape [num_nodes, num_features, num_look_back_step, num_timesteps] for data, and [num_nodes, num_features, num_look_ahead_step, num_timesteps] for target, similarly for mask.

• NOTE 1: the grid-based clustering is done separately and the mapping is in the cluster-mapping.csv file.

• NOTE 2: The current memory botteneck is from the window-slicing of the feature tensors, there are 2 alternatives: (1) do it sequentially for every chunk of time steps and (2) use distributed framework i.e., Spark.

Apache 2.0