def __init__()

in lama/modules/gpt_connector.py [0:0]

• 28 lines of code
• 4 McCabe index (conditional complexity)

    def __init__(self, args):
        super().__init__()

        if args.gpt_model_dir is not None:
            # load bert model from file
            gpt_model_name = str(args.gpt_model_dir) + "/"
            dict_file = gpt_model_name
            print("loading Open AI GPT model from {}".format(gpt_model_name))
        else:
            # load GPT model from huggingface cache
            gpt_model_name = args.gpt_model_name
            dict_file = gpt_model_name

        # Load pre-trained model tokenizer (vocabulary)
        self.tokenizer = OpenAIGPTTokenizer.from_pretrained(dict_file)

        # GPT uses different way to represent BPE then BERT. Namely, the
        # final suffixes are indicated with </w> suffix, while pieces that must
        # be followed are written as is. In BERT the prefixes are written as is
        # while the parts that must follow (not be followed!) have '##' prefix.
        # There is no one-to-one coversion. But at least we may make pieces that
        # may form a full word look the same.
        # Note that we should be very careful now,
        # tokenizer.convert_tokens_to_ids won't work with our vocabulary.
        def convert_word(word):
            if word == OPENAI_UNK:
                return word
            if word == '\n</w>':
                # Redefine symbol EOS to improve visualization.
                return OPENAI_EOS
            return word[:-4] if word.endswith('</w>') else f'{word}##'

        _, gpt_vocab = zip(*sorted(self.tokenizer.decoder.items()))
        self.vocab = [convert_word(word) for word in gpt_vocab]
        self._init_inverse_vocab()

        # Get UNK symbol as it's written in the origin GPT vocab.
        unk_index = self.inverse_vocab[OPENAI_UNK]
        self.unk_symbol = self.tokenizer.decoder[unk_index]

        # Load pre-trained model (weights)
        self.gpt_model = OpenAIGPTLMHeadModel.from_pretrained(gpt_model_name)
        self.gpt_model.eval()
        print(self.gpt_model.config)

        # Sanity check.
        assert len(self.vocab) == self.gpt_model.config.vocab_size
        assert 0 == self.gpt_model.config.n_special

        self.eos_id = self.inverse_vocab[OPENAI_EOS]
        self.model_vocab = self.vocab