def nn1(t: ti.i32):
for i in range(n_hidden):
actuation = 0.0
for j in ti.static(range(n_sin_waves)):
actuation += weights1[i, j] * ti.sin(spring_omega * t * dt +
2 * math.pi / n_sin_waves * j)
for j in ti.static(range(n_objects)):
offset = x[t, j] - x[t, head_id]
# use a smaller weight since there are too many of them
actuation += weights1[i, j * 6 + n_sin_waves] * offset[0] * 0.05
actuation += weights1[i,
j * 6 + n_sin_waves + 1] * offset[1] * 0.05
actuation += weights1[i, j * 6 + n_sin_waves + 2] * v[t,
i][0] * 0.05
actuation += weights1[i, j * 6 + n_sin_waves + 3] * v[t,
i][1] * 0.05
actuation += weights1[i, j * 6 + n_sin_waves +
4] * rotation[t, i] * 0.05
actuation += weights1[i, j * 6 + n_sin_waves + 5] * omega[t,
i] * 0.05
actuation += weights1[i, n_objects * 6 + n_sin_waves] * goal[None][0]
actuation += weights1[i,
n_objects * 6 + n_sin_waves + 1] * goal[None][1]
actuation += bias1[i]
actuation = ti.tanh(actuation)
hidden[t, i] = actuation
def nn2(t: ti.i32):
for i in range(n_gravitation):
act = 0.0
for j in ti.static(range(n_hidden)):
act += hidden[t, j] * weight2[j, i]
act += bias2[i]
gravitation[t, i] = ti.tanh(act)
def nn1(t: ti.i32):
for i in range(n_hidden):
actuation = 0.0
for j in ti.static(range(n_sin_waves)):
actuation += weights1[i, j] * ti.sin(spring_omega * t * dt +
2 * math.pi / n_sin_waves * j)
for j in ti.static(range(n_objects)):
offset = x[t, j] - center[t]
# use a smaller weight since there are too many of them
actuation += weights1[i, j * 4 + n_sin_waves] * offset[0] * 0.05
actuation += weights1[i,
j * 4 + n_sin_waves + 1] * offset[1] * 0.05
actuation += weights1[i, j * 4 + n_sin_waves + 2] * v[t,
j][0] * 0.05
actuation += weights1[i, j * 4 + n_sin_waves + 3] * v[t,
j][1] * 0.05
if ti.static(duplicate_v > 0):
for j in ti.static(range(duplicate_v)):
actuation += weights1[i, n_objects * 4 + n_sin_waves + j * 2] * target_v[t][0]
actuation += weights1[i, n_objects * 4 + n_sin_waves + j * 2 + 1] * target_v[t][1]
if ti.static(duplicate_h > 0):
for j in ti.static(range(duplicate_h)):
actuation += weights1[i, n_objects * 4 + n_sin_waves + duplicate_v * 2 + j] * target_h[None]
actuation += bias1[i]
actuation = ti.tanh(actuation)
hidden[t, i] = actuation
def compute_actuation(t: ti.i32):
for i in range(n_actuators):
act = 0.0
for j in ti.static(range(n_sin_waves)):
act += weights[i, j] * ti.sin(actuation_omega * t * dt +
2 * math.pi / n_sin_waves * j)
act += bias[i]
actuation[t, i] = ti.tanh(act)
def nn2(t: ti.i32):
for i in range(n_springs):
actuation = 0.0
for j in ti.static(range(n_hidden)):
actuation += weights2[i, j] * hidden[t, j]
actuation += bias2[i]
actuation = ti.tanh(actuation)
act[t, i] = actuation
def bias(t: ti.i32):
for i in range(d1):
act = h1_prev[t, i] + b[i]
if ti.static(activation == 'relu'):
act = ti.max(act, 0.)
if ti.static(activation == 'tanh'):
act = ti.tanh(act)
h1[t, i] = act
def thinc(wl, wc, wr, beta):
w0 = wc
w1 = wc
if (wr - wc) * (wc - wl) > 0.0:
# use thinc reconstruction
eps = 1.0e-15
wmin = min(wr, wl)
wmax = max(wr, wl)
wdelta = wmax - wmin
theta = sign(wr - wl)
C = (wc - wmin + eps) / (wdelta + eps)
B = ti.exp(theta * beta * (2 * C - 1))
A = (B / cosh(beta) - 1) / ti.tanh(beta)
# reconstructed value on right side of left face
w0 = wmin + wdelta / 2.0 * (1.0 + theta * A)
# reconstructed value on left side of right face
w1 = wmin + wdelta / 2.0 * (1.0 + theta * (ti.tanh(beta) + A) /
(1.0 + A * ti.tanh(beta)))
return w0, w1
def nn1(t: ti.i32):
for i in range(n_hidden):
act = 0.0
act += (x[t][0] - 0.5) * weight1[0, i]
act += (x[t][1] - 0.5) * weight1[1, i]
act += v[t][0] * weight1[2, i]
act += v[t][1] * weight1[3, i]
act += (goal[t][0] - 0.5) * weight1[4, i]
act += (goal[t][1] - 0.5) * weight1[5, i]
act += (goal_v[t][0] - 0.5) * weight1[6, i]
act += (goal_v[t][1] - 0.5) * weight1[7, i]
act += bias1[i]
hidden[t, i] = ti.tanh(act)
def _forward(self, t: ti.i32, nn_input: ti.template()):
for model_id, k, i, j in ti.ndrange(self.n_models, self.batch_size,
self.n_hidden, self.n_input):
self.hidden[model_id, t, k,
i] += self.weights1[model_id, i,
j] * nn_input[model_id, t, k, j]
if ti.static(self.activation):
for model_id, k, i in ti.ndrange(self.n_models, self.batch_size,
self.n_hidden):
self.output[model_id, t, k,
i] = ti.tanh(self.hidden[model_id, t, k, i] +
self.bias1[model_id, i])
else:
for model_id, k, i in ti.ndrange(self.n_models, self.batch_size,
self.n_hidden):
self.output[model_id, t, k,
i] = self.hidden[model_id, t, k,
i] + self.bias1[model_id, i]
def func():
xi[0] = -yi[None]
xi[1] = ~yi[None]
xi[2] = ti.logical_not(yi[None])
xi[3] = ti.abs(yi[None])
xf[0] = -yf[None]
xf[1] = ti.abs(yf[None])
xf[2] = ti.sqrt(yf[None])
xf[3] = ti.sin(yf[None])
xf[4] = ti.cos(yf[None])
xf[5] = ti.tan(yf[None])
xf[6] = ti.asin(yf[None])
xf[7] = ti.acos(yf[None])
xf[8] = ti.tanh(yf[None])
xf[9] = ti.floor(yf[None])
xf[10] = ti.ceil(yf[None])
xf[11] = ti.exp(yf[None])
xf[12] = ti.log(yf[None])
def test_trigonometric(): grad_test(lambda x: ti.tanh(x), lambda x: np.tanh(x)) grad_test(lambda x: ti.sin(x), lambda x: np.sin(x)) grad_test(lambda x: ti.cos(x), lambda x: np.cos(x)) grad_test(lambda x: ti.acos(x), lambda x: np.arccos(x)) grad_test(lambda x: ti.asin(x), lambda x: np.arcsin(x))
def nonlinear1():
for i in range(n_hidden):
output1_nonlinear[i] = ti.tanh(output1[i])
lambda x: x * x,
lambda x: x**2,
lambda x: x * x * x,
lambda x: x * x * x * x,
lambda x: 0.4 * x * x - 3,
lambda x: (x - 3) * (x - 1),
lambda x: (x - 3) * (x - 1) + x * x,
])
@if_has_autograd
@test_utils.test()
def test_poly(tifunc):
grad_test(tifunc)
@pytest.mark.parametrize('tifunc,npfunc', [
(lambda x: ti.tanh(x), lambda x: np.tanh(x)),
(lambda x: ti.sin(x), lambda x: np.sin(x)),
(lambda x: ti.cos(x), lambda x: np.cos(x)),
(lambda x: ti.acos(x), lambda x: np.arccos(x)),
(lambda x: ti.asin(x), lambda x: np.arcsin(x)),
])
@if_has_autograd
@test_utils.test(exclude=[ti.vulkan])
def test_trigonometric(tifunc, npfunc):
grad_test(tifunc, npfunc)
@pytest.mark.parametrize('tifunc', [
lambda x: 1 / x,
lambda x: (x + 1) / (x - 1),
lambda x: (x + 1) * (x + 2) / ((x - 1) * (x + 3)),
