def hard_sigmoid(x):
"""
An approximation of sigmoid.
More approximate and faster than ultra_fast_sigmoid.
Approx in 3 parts: 0, scaled linear, 1.
Removing the slope and shift does not make it faster.
"""
# Use the same dtype as determined by "upgrade_to_float",
# and perform computation in that dtype.
out_dtype = aes.upgrade_to_float(aes.Scalar(dtype=x.dtype))[0].dtype
slope = constant(0.2, dtype=out_dtype)
shift = constant(0.5, dtype=out_dtype)
x = (x * slope) + shift
x = clip(x, 0, 1)
return x
def test_jax_basic():
rng = np.random.default_rng(28494)
x = matrix("x")
y = matrix("y")
b = vector("b")
# `ScalarOp`
z = cosh(x**2 + y / 3.0)
# `[Inc]Subtensor`
out = aet_subtensor.set_subtensor(z[0], -10.0)
out = aet_subtensor.inc_subtensor(out[0, 1], 2.0)
out = out[:5, :3]
out_fg = FunctionGraph([x, y], [out])
test_input_vals = [
np.tile(np.arange(10), (10, 1)).astype(config.floatX),
np.tile(np.arange(10, 20), (10, 1)).astype(config.floatX),
]
(jax_res, ) = compare_jax_and_py(out_fg, test_input_vals)
# Confirm that the `Subtensor` slice operations are correct
assert jax_res.shape == (5, 3)
# Confirm that the `IncSubtensor` operations are correct
assert jax_res[0, 0] == -10.0
assert jax_res[0, 1] == -8.0
out = clip(x, y, 5)
out_fg = FunctionGraph([x, y], [out])
compare_jax_and_py(out_fg, test_input_vals)
out = aet.diagonal(x, 0)
out_fg = FunctionGraph([x], [out])
compare_jax_and_py(
out_fg, [np.arange(10 * 10).reshape((10, 10)).astype(config.floatX)])
out = aet_slinalg.cholesky(x)
out_fg = FunctionGraph([x], [out])
compare_jax_and_py(
out_fg,
[(np.eye(10) + rng.standard_normal(size=(10, 10)) * 0.01).astype(
config.floatX)],
)
# not sure why this isn't working yet with lower=False
out = aet_slinalg.Cholesky(lower=False)(x)
out_fg = FunctionGraph([x], [out])
compare_jax_and_py(
out_fg,
[(np.eye(10) + rng.standard_normal(size=(10, 10)) * 0.01).astype(
config.floatX)],
)
out = aet_slinalg.solve(x, b)
out_fg = FunctionGraph([x, b], [out])
compare_jax_and_py(
out_fg,
[
np.eye(10).astype(config.floatX),
np.arange(10).astype(config.floatX),
],
)
out = aet.diag(b)
out_fg = FunctionGraph([b], [out])
compare_jax_and_py(out_fg, [np.arange(10).astype(config.floatX)])
out = aet_nlinalg.det(x)
out_fg = FunctionGraph([x], [out])
compare_jax_and_py(
out_fg, [np.arange(10 * 10).reshape((10, 10)).astype(config.floatX)])
out = aet_nlinalg.matrix_inverse(x)
out_fg = FunctionGraph([x], [out])
compare_jax_and_py(
out_fg,
[(np.eye(10) + rng.standard_normal(size=(10, 10)) * 0.01).astype(
config.floatX)],
)