Method __init__

Language_Model/GPT1.py:271–310

(
            self,
            vocabulary_size=40479,
            embedding_size=768,
            sequence_length=256,
            num_heads=12,
            num_layers=4,
            learn_embeddings=False,
            _tokens_embedding_weight=None,
            _positional_embedding_weight=None,
    )

class MiniGPT1(nn.Module):
    def __init__(
            self,
            vocabulary_size=40479,
            embedding_size=768,
            sequence_length=256,
            num_heads=12,
            num_layers=4,
            learn_embeddings=False,
            _tokens_embedding_weight=None,
            _positional_embedding_weight=None,
    ):
        super(MiniGPT1, self).__init__()
        self.vocabulary_size = vocabulary_size
        self.embedding_size = embedding_size
        self.sequence_length = sequence_length
        self.num_heads = num_heads
        self.num_layers = num_layers
        self.learn_embeddings = learn_embeddings
        self.head_size = embedding_size // num_heads

        self.embedding = GPT1Embedding(
            vocabulary_size,
            embedding_size,
            sequence_length,
            _tokens_embedding_weight=_tokens_embedding_weight,
            _positional_embedding_weight=_positional_embedding_weight,
        )
        self.layers = nn.ModuleList(
            [
                Block(self.head_size, 4 * embedding_size, num_heads, sequence_length)
                for _ in range(num_layers)
            ]
        )
        self.classifier = nn.Linear(embedding_size, vocabulary_size, bias=False)

        # Tying classifier and embedding weights
        self.classifier.weight = self.embedding.tokens.weight

        # Freeze the embedding weights, depending on learn_embeddings
        self.embedding.requires_grad_(learn_embeddings)

    def get_embeddings(self, inputs):
        """Get the embeddings for some input sequence.