def test_optimized_plated_einsum_adjoint(equation, plates, backend):
inputs, outputs, sizes, operands, funsor_operands = make_einsum_example(
equation)
with interpretation(reflect):
fwd_expr = einsum(equation,
*funsor_operands,
plates=plates,
backend=backend)
actuals = adjoint(fwd_expr, funsor_operands)
for operand in operands:
pyro_require_backward(operand)
expected_out = pyro_einsum(equation,
*operands,
modulo_total=False,
plates=plates,
backend=backend)[0]
expected_out._pyro_backward()
for i, (inp, tv, fv) in enumerate(zip(inputs, operands, funsor_operands)):
actual = actuals[fv]
expected = tv._pyro_backward_result
if inp:
actual = actual.align(tuple(inp))
assert isinstance(actual, funsor.Tensor)
assert expected.shape == actual.data.shape
assert torch.allclose(expected, actual.data, atol=1e-7)
def test_optimized_plated_einsum(equation, plates, backend):
inputs, outputs, sizes, operands, funsor_operands = make_einsum_example(
equation)
expected = pyro_einsum.einsum(equation,
*operands,
plates=plates,
backend=backend)[0]
actual = einsum(equation, *funsor_operands, plates=plates, backend=backend)
if len(equation) < 10:
actual_naive = naive_plated_einsum(equation,
*funsor_operands,
plates=plates,
backend=backend)
assert_close(actual, actual_naive)
assert isinstance(actual, funsor.Tensor) and len(outputs) == 1
if len(outputs[0]) > 0:
actual = actual.align(tuple(outputs[0]))
assert expected.shape == actual.data.shape
assert torch.allclose(expected, actual.data)
for output in outputs:
for i, output_dim in enumerate(output):
assert output_dim in actual.inputs
assert actual.inputs[output_dim].dtype == sizes[output_dim]