DISCLAIMER: We are working on a clean-up version of the code to be published soon here.

Attribution hackathon

This repository contains code and data for the attribution hackathon.

Getting started

1. Install conda
2. Create the conda environent: conda env create -f attr_hack_env.yml
3. Activate conda environment: conda activate attr_hack
4. Run data processing / simulation (only needed to re-run data processing, you may skip this step).
   • GPP model
     • python Simple_GPP_model/process_predictor-variables.py
     • python Simple_GPP_model/gpp_model.py
   • SM model
     • python simple_sm_model/era_sm_sim.py

Hackathon guide

Relevant files

file	description
hackathon/models/linear.py	A dummy model.
hackathon/base_model.py	The base model class.
hackathon/model_runner.py	The model runner class.
hackathon/data_pipeline.py	The dataloader.
benchmark.py	Evaluate model.

Getting started

1. Make a copy of hackathon/models/linear.py within the same directory and rename it to a meaningful name, e.g., hackathon/models/rnn.py.

2. Replace the class Linear with your model, give it a meaningful name, e.g., class RNN(BaseModel). The model must subclass BaseModel!

3. The model __init__ method must take **kwargs and pass them to the parent class (BaseModel) like so:

   class MyModel(BaseModel):
       def __init__(self, num_features: int, num_targets: int, **kwargs) -> None:
           super(MyModel, self).__init__(**kwargs)
   
           self.linear = torch.nn.Linear(num_features, num_targets)
           self.softplus = torch.nn.Softplus()
   
       def forward(self, x: Tensor) -> Tensor:
           out = self.softplus(self.linear(x))
           return out

4. Define a function model_setup within the same file that returns an initialized model (e.g., hackathon/models/rnn.py). The function must take the argument norm_stats. The model_setup function must follow this pattern:

   def model_setup(norm_stats: dict[str, Tensor]) -> BaseModel:
       """Create a model as subclass of hackathon.base_model.BaseModel.
   
       Parameters
       ----------
       norm_stats: Feature normalization stats with signature {'mean': Tensor, 'std': Tensor},
           both tensors with shape (num_features,).
   
       Returns
       -------
       A model.
       """
       model = MyCustomModel(  # <- Is a subclass of BaseModel
           num_features=8,  # <- A model HP
           num_targets=1,  # <- A model HP
           num_layers=1000000,  # <- A model HP
           ... # more model HPs
           learning_rate=0.01,  # <- BaseModel kwarg
           weight_decay=0.0,  # <- BaseModel kwarg
           norm_stats=norm_stats)  # <- BaseModel kwarg
   
       return model

5. Add your model to the benchmark.py file by importing and adding the model_setup function:

   from hackathon.models.linear import model_setup as linear_model
   
   model_funs = [linear_model]

6. Run your model with python benchmark.py (--quickrun for developer run over one epoch / CV fold with less training and validation data).

Logging

• The log_dir argument sets the base directory of the experiment.
• Cross validation runs are each put into a subdirectory (log_dir/fold_00 etc.)
• The final model (ensemble of all CV models) and its predictions are saved in log_dir/final

0 0 1 1 2 1 1 0 2 1