class TransformerEncoderDecoderConnectionBlock(nn.Module):
def __init__(self, hidden_dim: int, key_query_value_dim: int = 64, num_attention_heads=8, with_hard_concrete_gate=False):
super(TransformerEncoderDecoderConnectionBlock, self).__init__()
self.multihead_attention = MultiHeadAttention(hidden_dim, key_and_query_dim=key_query_value_dim, value_dim=key_query_value_dim, num_heads=num_attention_heads, with_hard_concrete_gate=with_hard_concrete_gate)
self.norm = nn.LayerNorm(hidden_dim)
return
def forward(self, encoder_outputs, decoder_hidden, mask=None):
attention_outputs = self.multihead_attention.forward(q_hidden_inputs=decoder_hidden, k_hidden_inputs=encoder_outputs, v_hidden_inputs=encoder_outputs, mask=mask)
return self.norm(decoder_hidden + attention_outputs)
def test_multihead_attention():
with torch.no_grad():
kq_dim = 4
v_dim = 8
num_heads = 16
hidden_dim = 64
batch_size = 3
seq_len = 7
attention_input = torch.rand((batch_size, seq_len, hidden_dim))
mha = MultiHeadAttention(hidden_dim,
key_and_query_dim=kq_dim,
value_dim=v_dim,
num_heads=num_heads)
mha_ouptut = mha.forward(q_hidden_inputs=attention_input,
k_hidden_inputs=attention_input,
v_hidden_inputs=attention_input,
mask=None)
assert mha_ouptut.size() == attention_input.size()
