def test_backward_indirectly(
fix_seed,
test_input,
test_weight,
test_bias,
test_mode,
expected_weight_grad,
expected_input_grad,
):
binarized_linear(test_input, test_weight, test_bias, test_mode).backward()
assert torch.allclose(
input=test_input.grad,
other=expected_input_grad,
rtol=1e-04,
atol=1e-04,
equal_nan=True,
)
assert torch.allclose(
input=test_weight.grad,
other=expected_weight_grad,
rtol=1e-04,
atol=1e-04,
equal_nan=True,
)
def test_backward_indirectly(
fix_seed,
test_input,
test_weight,
test_bias,
test_mode,
expected_input_grad,
expected_weight_grad,
expected_bias_grad,
):
binarized_linear(test_input, test_weight, test_bias, test_mode).backward()
logger.debug(f"input grad: {test_input.grad}")
logger.debug(f"expected input grad: {expected_input_grad}")
logger.debug(f"weight grad: {test_weight.grad}")
logger.debug(f"expected weight grad: {expected_weight_grad}")
assert torch.allclose(
input=test_input.grad,
other=expected_input_grad,
rtol=1e-04,
atol=1e-04,
equal_nan=True,
)
assert torch.allclose(
input=test_weight.grad,
other=expected_weight_grad,
rtol=1e-04,
atol=1e-04,
equal_nan=True,
)
if expected_bias_grad:
logger.debug(f"bias grad: {test_bias.grad}")
logger.debug(f"expected bias grad: {expected_bias_grad}")
assert torch.allclose(
input=test_bias.grad,
other=expected_bias_grad,
rtol=1e-04,
atol=1e-04,
equal_nan=True,
)
def test_forward(fix_seed, test_input, test_weight, test_bias, test_mode,
expected):
assert torch.allclose(
input=binarized_linear(test_input, test_weight, test_bias, test_mode),
other=expected,
rtol=1e-04,
atol=1e-04,
equal_nan=True,
)
def test_forward(fix_seed, test_input, test_weight, test_bias, test_mode,
expected):
answer = binarized_linear(test_input, test_weight, test_bias, test_mode)
logger.debug(f"answer: {answer}")
logger.debug(f"expected: {expected}")
assert torch.allclose(
input=answer,
other=expected,
rtol=1e-04,
atol=1e-04,
equal_nan=True,
)
def test_supported_mode(fix_seed, test_input, test_weight, test_bias,
test_mode):
with pytest.raises(RuntimeError):
binarized_linear(test_input, test_weight, test_bias, test_mode)
def forward(self, input: torch.Tensor) -> torch.Tensor:
self.clipping()
if self.bias is not None:
return binarized_linear(input, self.weight, self.bias)
return binarized_linear(input, self.weight)