WordWeights.py

import torch
from torch import Tensor
from torch import nn
from typing import Union, Tuple, List, Iterable, Dict
import os
import json
import logging

class WordWeights(nn.Module):
    """This model can weight word embeddings, for example, with idf-values."""

    def __init__(self, vocab: List[str], word_weights: Dict[str, float], unknown_word_weight: float = 1):
        """

        :param vocab:
            Vocabulary of the tokenizer
        :param word_weights:
            Mapping of tokens to a float weight value. Words embeddings are multiplied by  this float value. Tokens in word_weights must not be equal to the vocab (can contain more or less values)
        :param unknown_word_weight:
            Weight for words in vocab, that do not appear in the word_weights lookup. These can be for example rare words in the vocab, where no weight exists.
        """
        super(WordWeights, self).__init__()
        self.config_keys = ['vocab', 'word_weights', 'unknown_word_weight']
        self.vocab = vocab
        self.word_weights = word_weights
        self.unknown_word_weight = unknown_word_weight

        weights = []
        num_unknown_words = 0
        for word in vocab:
            weight = unknown_word_weight
            if word in word_weights:
                weight = word_weights[word]
            elif word.lower() in word_weights:
                weight = word_weights[word.lower()]
            else:
                num_unknown_words += 1
            weights.append(weight)
        
        logging.info("{} of {} words without a weighting value. Set weight to {}".format(num_unknown_words, len(vocab), unknown_word_weight))

        self.emb_layer = nn.Embedding(len(vocab), 1)
        self.emb_layer.load_state_dict({'weight': torch.FloatTensor(weights).unsqueeze(1)})


    def forward(self, features: Dict[str, Tensor]):
        attention_mask = features['attention_mask']
        token_embeddings = features['token_embeddings']

        #Compute a weight value for each token
        token_weights_raw = self.emb_layer(features['input_ids']).squeeze(-1)
        token_weights = token_weights_raw * attention_mask.float()
        token_weights_sum = torch.sum(token_weights, 1)

        #Multiply embedding by token weight value
        token_weights_expanded = token_weights.unsqueeze(-1).expand(token_embeddings.size())
        token_embeddings = token_embeddings * token_weights_expanded

        features.update({'token_embeddings': token_embeddings, 'token_weights_sum': token_weights_sum})
        return features

    def get_config_dict(self):
        return {key: self.__dict__[key] for key in self.config_keys}

    def save(self, output_path):
        with open(os.path.join(output_path, 'config.json'), 'w') as fOut:
            json.dump(self.get_config_dict(), fOut, indent=2)

    @staticmethod
    def load(input_path):
        with open(os.path.join(input_path, 'config.json')) as fIn:
            config = json.load(fIn)

        return WordWeights(**config)