Implement neural network and training pipeline prototype

[williamstark01]

Potentially using PyTorch Lightning if it helps.

[yangtcai]

I think the issue can be divided into three parts

• The DETR model This part will produce the prediction results, the input of the model is the subsequence, and the output is the boundaries and the classes of the repeat sequences, the prototype of the DETR model is the following

  class DETR(nn.Module):
   def forward(sample):
       """
        Parameters:
            -- sample: batched sequences, of shape [batch_size x seq_len x embedding_len ]
                 eg, when choosing one hot encoding, embedding_len will be 5, [A, T, C, G, N]
            -- pred_logits: the classification logits (including no-object) for all queries.
                              Shape= [batch_size x num_queries x (num_classes + 1)]
            -- pred_boundaries: The normalized boundaries coordinates for all queries, represented as
                             (center, width). These values are normalized in [0, 1].
       """
        pass

• Hungarian matcher As the DETR model only produce num_queries prediction results, the order of prediction sets and target set has still not been decided, so the purpose of the part is that find the minimal matching between the prediction set and the target set.
• The Loss function It includes the classification loss and the segment loss.

[williamstark01]

@yangtcai that sounds like a good idea, feel free to open new issues for these subtasks and add them to the project board.

[williamstark01]

Implemented in #24