Aan de slagGa gratis aan de slag

Adding methods to the MultiHeadAttention class

In this exercise, you'll build the rest of the MultiHeadAttention class from the ground up by defining four methods:

  • .split_heads(): split and transform the input embeddings between the attention heads
  • .compute_attention(): calculate the scaled dot-product attention weights multiplied by the values matrix
  • .combine_heads(): transform the attention weights back into the same shape as the input embeddings, x
  • .forward(): call the other methods to pass the input embeddings through each process

torch.nn has been imported as nn, torch.nn.functional is available as F, and torch is also available.

Deze oefening maakt deel uit van de cursus

Transformer Models with PyTorch

Cursus bekijken

Praktische interactieve oefening

Probeer deze oefening eens door deze voorbeeldcode in te vullen.

class MultiHeadAttention(nn.Module):
    def __init__(self, d_model, num_heads):
        super().__init__()
        self.num_heads = num_heads
        self.d_model = d_model
        self.head_dim = d_model // num_heads
        self.query_linear = nn.Linear(d_model, d_model, bias=False)
        self.key_linear = nn.Linear(d_model, d_model, bias=False)
        self.value_linear = nn.Linear(d_model, d_model, bias=False)
        self.output_linear = nn.Linear(d_model, d_model)

    def split_heads(self, x, batch_size):
        seq_length = x.size(1)
        # Split the input embeddings and permute
        x = x.____
        return x.permute(0, 2, 1, 3)
Code bewerken en uitvoeren