How to build my own model for RT prediction with meta embedding?

[kanzy]

Hello,

I am a newbie of AlphaPeptDeep and Torch in general, and I am trying to learn through your tutorial https://alphapeptdeep.readthedocs.io/en/latest/nbs/tutorial_building_rt_model.html in Jupyter Notebook.

After running the tutorial code without issue, I want to add the meta information "peptide charge" as the additional feature in the model. I found there is the existing class Encoder_26AA_Mod_Charge_CNN_LSTM_AttnSum in building_block.py, so I just tried modifying the class RT_LSTM_Module in the tutorial code:

    self.rt_encoder = building_block.Encoder_26AA_Mod_Charge_CNN_LSTM_AttnSum(
        hidden
    )

  def forward(self,
      aa_indices,
      mod_x,
      charges,
  ):
      x = self.rt_encoder(aa_indices, mod_x, charges)
      x = self.dropout(x)
      return self.rt_decoder(x).squeeze(1)

I also added a line in the class RT_ModelInterface function _get_features_from_batch_df:

   charges = torch.Tensor(batch_df['charge'].values)

Then, when I try to train the new model, it gives the below error:

File ~\alphapeptdeep\peptdeep\model\model_interface.py:493, in ModelInterface.train(self, precursor_df, batch_size, epoch, warmup_epoch, lr, verbose, verbose_each_epoch, **kwargs)
    491 for epoch in range(epoch):
    492     if self.fixed_sequence_len == 0:
--> 493         batch_cost = self._train_one_epoch(
    494             precursor_df, epoch,
    495             batch_size, verbose_each_epoch,
    496             **kwargs
    497         )
    498     else:
    499         batch_cost = self._train_one_epoch_by_padding_zeros(
    500             precursor_df, epoch,
    501             batch_size, verbose_each_epoch,
    502             **kwargs
    503         )

File ~\alphapeptdeep\peptdeep\model\model_interface.py:822, in ModelInterface._train_one_epoch(self, precursor_df, epoch, batch_size, verbose_each_epoch, **kwargs)
    817 features = self._get_features_from_batch_df(
    818     batch_df, **kwargs
    819 )
    820 if isinstance(features, tuple):
    821     batch_cost.append(
--> 822         self._train_one_batch(targets, *features)
    823     )
    824 else:
    825     batch_cost.append(
    826         self._train_one_batch(targets, features)
    827     )

File ~\alphapeptdeep\peptdeep\model\model_interface.py:842, in ModelInterface._train_one_batch(self, targets, *features)
    840 """Training for a mini batch"""
    841 self.optimizer.zero_grad()
--> 842 predicts = self.model(*features)
    843 cost = self.loss_func(predicts, targets)
    844 cost.backward()

File C:\ProgramData\anaconda3\envs\peptdeep\lib\site-packages\torch\nn\modules\module.py:1553, in Module._wrapped_call_impl(self, *args, **kwargs)
   1551     return self._compiled_call_impl(*args, **kwargs)  # type: ignore[misc]
   1552 else:
-> 1553     return self._call_impl(*args, **kwargs)

File C:\ProgramData\anaconda3\envs\peptdeep\lib\site-packages\torch\nn\modules\module.py:1562, in Module._call_impl(self, *args, **kwargs)
   1557 # If we don't have any hooks, we want to skip the rest of the logic in
   1558 # this function, and just call forward.
   1559 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
   1560         or _global_backward_pre_hooks or _global_backward_hooks
   1561         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1562     return forward_call(*args, **kwargs)
   1564 try:
   1565     result = None

TypeError: forward() missing 1 required positional argument: 'charges'

I thought I already put in the 'charges' argument and couldn't figure out how to fix this problem. Hope you would please help take a look and teach me more about meta embedding in AlphaPeptDeep. Many thanks!