Repository structure

This repository has a modular structure, where you can have multiple implementations for each module and switch between them by changing the configuration file.

We use OmegaConf to manage the configuration files. Please, see OmegaConf documentation for more details.

The configuration files are stored in the configs/ directory. Each of them has configurations for four main modules of the pipeline:

• dataset is a module for data loading and preprocessing. dataset modules are stored in datasets/ directory. Each dataset module defines a Dataset class.
• model is a module where a GNN model is defined. model modules are stored in the models/ directory. Each model module defines a Model class which inherits from torch.nn.Module.
  • The Model objects create an input graph, pass it through the GNN and returns the predicted accelerations.
  • The GNN architecture is defined in the core_model module. These are stored in the models/core_models/ directory. Each core_model module defines an EncodeProcessDecode class.
• runner module defines the training and validation loops. runner modules are stored in the runners/ directory. Each runner module defines
  • Runner class which defines:
    • valid_rollout() method used to generate a sequence of garment geometries for validation
    • forward() method used in training. It runs a forward pass of the model and performs an optimization step.
  • create_optimizer() function that builds an optimizer and a scheduler for the training process
  • run_epoch() which iterates over the dataloader and calls runner.forward() for each training sample
• criterions are modules that define the objective terms. One configuration file may have an unlimited number of criterions, their values are summed up. criterion modules are stored in the criterions/ directory. Each criterion module defines a Criterion class which inherits from torch.nn.Module.

Each module also defines a Config class which declares the parameters of the module. Their value may be overridden in the configuration file on in the command line (see OmegaConf documentation).

Pytorch Geometric for handling graphs

We use Pytorch Geometric to handle the graph data and run message-passing steps.

For details, please, see Pytorch Geometric documentation and datasets/README.md.