Structure

• Additional Data/ contains the peptide data split for cross validation

• csvs/ contains the model configurations required to reproduce classical and quantum results in the paper; additional parameters can be added by examining the arg_type_map variable in run.py

• model/ contains the TorchQuantum and PyTorch quantum and classical models, as well as utility files

• utils/ contains other miscellaneous functions used in processing

• run.py and run_classical.py run the specified model and save data using Tensorboard

• average_folds.py averages test and validation F1 data from executed runs

• get_attributions.py and get_attributions_classical.py compute the Captum IG/SVS attributions

Modifications to TorchQuantum

A small number of modifications have to be made to TorchQuantum to support the recurrent methods used here. As of TQ 0.1.7, these are:

1. Replace line 90 in the GeneralEncoder class in encoding/encodings.py with the following four lines:

   50  class GeneralEncoder(Encoder, metaclass=ABCMeta):
   
       ...
   
   90      def forward(self, qdev: tq.QuantumDevice, x, reset=True):
   91          self.q_device = qdev
   92          if reset:
   93             self.q_device.reset_states(x.shape[0])

i.e. add reset as a default parameter and add the if statement resetting the state.

2. Create the MeasureOne class in measurement/measurements.py:

   class MeasureOne(tq.QuantumModule):
       """Obtain the expectation value of all the qubits."""
   
       def __init__(self, obs, v_c_reg_mapping=None):
           super().__init__()
           self.obs = obs
           self.v_c_reg_mapping = v_c_reg_mapping
   
       def forward(self, qdev: tq.QuantumDevice):
           x = expval(qdev, [0], [self.obs()])
   
           if self.v_c_reg_mapping is not None:
               c2v_mapping = self.v_c_reg_mapping["c2v"]
               """
               the measurement is not normal order, need permutation
               """
               perm = []
               for k in range(x.shape[-1]):
                   if k in c2v_mapping.keys():
                       perm.append(c2v_mapping[k])
               x = x[:, perm]
   
           if self.noise_model_tq is not None and self.noise_model_tq.is_add_noise:
               return self.noise_model_tq.apply_readout_error(x)
           else:
               return x
   
       def set_v_c_reg_mapping(self, mapping):
           self.v_c_reg_mapping = mapping

This is identical to the pre-existing MeasureAll class except for the first line of the forward method, which here returns only a single observable.

Then add MeasureAll to the __all__ list defined on line 17 in the same file.

Usage

1. Install the requirements in a venv by using pip install -r requirements.txt.
2. Then install TorchQuantum in editable mode following the instructions on their homepage, and the modifications required above.
3. Create a .csv file detailing the structure of the model you wish to run (csvs/ contain some examples of this).
4. Attributions can be calculated and the best F1 score over folds found by using get_attributions.py and average_folds.py.