import torch
from torch import nn
from transformers import BertPreTrainedModel

class ParagramSPModel(BertPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.config = config
        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
        # Initialize weights and apply final processing
        self.post_init()

    def filter_input_ids(self, input_ids):
      output = []
      length = input_ids.shape[1]
      for i in range(input_ids.shape[0]):
        ids = input_ids[i]
        filtered_ids = []
        for j in ids:
          if j > 0:
            filtered_ids.append(j)
        if len(filtered_ids) == 0:
          filtered_ids = [0]
        output.append(filtered_ids + [config.pad_token_id] * (length - length(filtered_ids)))
      return torch.tensor(output)
  
    def forward(self, input_ids, attention_mask):
        print(input_ids)
        print(attention_mask)
        input_ids = self.filter_input_ids(input_ids)
        attention_mask = input_ids > 0
        embeddings = self.word_embeddings(input_ids)
        masked_embeddings = embeddings * attention_mask[:, :, None]
        mean_pooled_embeddings = masked_embeddings.sum(dim=1) / attention_mask[:, :, None].sum(dim=1)
        return (embeddings, mean_pooled_embeddings, embeddings)