def test_gradients_against_finite_difference(self, temp=1.0, j=0):
def num_grad(f, x, eps=1e-4):
shape = x.shape
x_flat = x.ravel()
grad_flat = np.zeros_like(x_flat)
for j in range(len(x_flat)):
e = np.zeros_like(x_flat)
e[j] = 1
call1 = f((x_flat + eps * e).reshape(shape))
call2 = f((x_flat - eps * e).reshape(shape))
grad_flat[j] = (call1 - call2) / (2 * eps)
return grad_flat.reshape(shape)
toy_sub = self._toy_sub.numpy()
toy_gap_open = self._toy_gap_open.numpy()
toy_gap_extend = self._toy_gap_extend.numpy()
_, g_sim_mat, g_gap_open, g_gap_extend = (
npy_ops._soft_sw_affine(sim_mat=toy_sub[j], gap_open=toy_gap_open[j],
gap_extend=toy_gap_extend[j], temperature=temp,
ret_grads=True))
# Gradient w.r.t. sim_mat.
def f_sim_mat(x):
return npy_ops._soft_sw_affine(sim_mat=x, gap_open=toy_gap_open[j],
gap_extend=toy_gap_extend[j],
temperature=temp, ret_grads=False)
g_sim_mat_num = num_grad(f_sim_mat, toy_sub[j].astype(np.float64))
self.assertAllClose(g_sim_mat, g_sim_mat_num)
# Gradient w.r.t. gap_open.
def f_gap_open(x):
return npy_ops._soft_sw_affine(sim_mat=toy_sub[j], gap_open=x,
gap_extend=toy_gap_extend[j],
temperature=temp, ret_grads=False)
g_gap_open_num = num_grad(f_gap_open, toy_gap_open[j].astype(np.float64))
self.assertAllClose(g_gap_open, g_gap_open_num)
# Gradient w.r.t. gap_extend.
def f_gap_ext(x):
return npy_ops._soft_sw_affine(sim_mat=toy_sub[j],
gap_open=toy_gap_open[j], gap_extend=x,
temperature=temp, ret_grads=False)
g_gap_extend_num = num_grad(f_gap_ext, toy_gap_extend[j].astype(np.float64))
self.assertAllClose(g_gap_extend, g_gap_extend_num)
def f_gap_ext(x):
return npy_ops._soft_sw_affine(sim_mat=toy_sub[j],
gap_open=toy_gap_open[j], gap_extend=x,
temperature=temp, ret_grads=False)