Test multihead attention

tests/test_transformer.py

@@ -0,0 +1,52 @@
+import numpy as np
+import pytest
+import torch
+
+from neuralogic.core import Template, Settings, V, R
+from neuralogic.dataset import Dataset
+from neuralogic.nn.module import MultiheadAttention
+
+
+@pytest.mark.parametrize(
+    "qdim, kdim, vdim, num_heads, sequence_len",
+    [
+        (10, 12, 14, 1, 3),  # One head
+        (10, 12, 14, 2, 3),  # Two heads
+    ],
+)
+def test_multiheadattention(qdim: int, kdim: int, vdim: int, num_heads: int, sequence_len: int):
+    keys = torch.rand((sequence_len, kdim))
+    queries = torch.rand((sequence_len, qdim))
+    values = torch.rand((sequence_len, vdim))
+
+    mha = torch.nn.MultiheadAttention(qdim, num_heads, bias=False, kdim=kdim, vdim=vdim)
+
+    template = Template()
+    template += MultiheadAttention(qdim, num_heads, "out", "q", "k", "v", vdim=vdim, kdim=kdim)
+
+    model = template.build(Settings(iso_value_compression=False, chain_pruning=False))
+
+    params = model.parameters()
+    torch_params = list(mha.parameters())
+
+    params["weights"][1] = torch_params[0].detach().numpy().tolist()
+    params["weights"][2] = torch_params[2].detach().numpy().tolist()
+    params["weights"][3] = torch_params[1].detach().numpy().tolist()
+    params["weights"][4] = torch_params[3].detach().numpy().tolist()
+
+    model.load_state_dict(params)
+
+    example = []
+    for i in range(sequence_len):
+        example.append(R.v(i)[values[i]])
+        example.append(R.k(i)[keys[i]])
+        example.append(R.q(i)[queries[i]])
+
+    dataset = Dataset([example], [R.out(i) for i in range(sequence_len)])
+    built_dataset = model.build_dataset(dataset)
+
+    torch_results = mha(queries, keys, values)
+    pyneuralogic_results = model(built_dataset, train=False)
+
+    for t_res, pnl_res in zip(torch_results[0], pyneuralogic_results):
+        assert np.allclose(t_res.detach().numpy(), np.array(pnl_res))