def run_irl(world, car, reward, theta, data):
def gen():
for point in data:
for c, x0, u in zip(world.cars, point['x0'], point['u']):
c.traj.x0.set_value(x0)
for cu, uu in zip(c.traj.u, u):
cu.set_value(uu)
yield
r = car.traj.reward(reward)
g = utils.grad(r, car.traj.u)
H = utils.hessian(r, car.traj.u)
I = tt.eye(utils.shape(H)[0])
reg = utils.vector(1)
reg.set_value([1e-1])
H = H - reg[0] * I
L = tt.dot(g, tt.dot(tn.MatrixInverse()(H), g)) + tt.log(tn.Det()(-H))
for _ in gen():
pass
optimizer = utils.Maximizer(L, [theta],
gen=gen,
method='gd',
eps=0.1,
debug=True,
iters=1000,
inf_ignore=10)
optimizer.maximize()
print theta.get_value()
def test_hess():
x0 = torch.tensor([-1.0, 1.0], requires_grad=True)
fx = rosenbrock(*x0)
auto_ddx = U.hessian(fx, x0)
analytical_ddx = hess_rosenbrock(*x0)
assert (torch.norm(analytical_ddx - auto_ddx) == 0)
print("hess test success")
def image_ridge_line(image):
rms = rms_image(image)
im = image.image.copy()
im_shape = np.shape(im)
im[im < 5. * rms] = 0
# (2, imsize, imsize) array with gradient (in x & y -directions)
grad = np.gradient(im, edge_order=2)
from utils import hessian
# (2, 2, imsize, imsize) array with second derivates
hess = hessian(im)
det_ar = np.zeros(im_shape)
for i in range(im_shape[0]):
for j in range(im_shape[1]):
det_ar[i, j] = np.linalg.det(hess[:, :, i, j])
def train_normal(num_epochs,
model,
dset_train,
batch_size,
grad_clip,
logging_freq,
optim="sgd",
**kwargs):
train_loader = torch.utils.data.DataLoader(dset_train,
batch_size=batch_size,
shuffle=True)
model.train()
for epoch in range(num_epochs):
epoch_loss = AverageMeter()
for batch_idx, batch in enumerate(train_loader):
x, y = batch
w_optimizer.zero_grad()
y_pred = model(x)
loss = criterion(y_pred, y)
loss.backward(retain_graph=True)
epoch_loss.update(loss.item())
if optim == "newton":
linear_weight = list(model.weight_params())[0]
hessian_newton = torch.inverse(
hessian(loss * 1, linear_weight,
linear_weight).reshape(linear_weight.size()[1],
linear_weight.size()[1]))
with torch.no_grad():
for w in model.weight_params():
w = w.subtract_(torch.matmul(w.grad, hessian_newton))
elif optim == "sgd":
torch.nn.utils.clip_grad_norm_(model.weight_params(), 1)
w_optimizer.step()
else:
raise NotImplementedError
wandb.log({
"Train loss": epoch_loss.avg,
"Epoch": epoch,
"Batch": batch_idx
})
if batch_idx % logging_freq == 0:
print("Epoch: {}, Batch: {}, Loss: {}, Alphas: {}".format(
epoch, batch_idx, epoch_loss.avg, model.fc1.alphas.data))
def apply_step(self, *args):
loss_g, loss_h = args[:2]
for x in self.params:
g = jacobian(loss_g, x)
h = hessian(loss_h, x)
with torch.no_grad():
g = g.reshape((-1, 1))
h = h.reshape((g.shape[0], g.shape[0]))
dx = conjugate_gradient(h,
g,
n_iterations=self.n_cg,
tol=self.tol).reshape(x.shape)
x.add_(dx, alpha=-self.lr)
def run_irl(world, car, reward, theta, data):
def gen():
for point in data:
for c, x0, u in zip(world.cars, point['x0'], point['u']):
c.traj.x0.set_value(x0)
for cu, uu in zip(c.traj.u, u):
cu.set_value(uu)
yield
r = car.traj.reward(reward)
g = utils.grad(r, car.traj.u)
H = utils.hessian(r, car.traj.u)
I = tt.eye(utils.shape(H)[0])
reg = utils.vector(1)
reg.set_value([1e-1])
H = H-reg[0]*I
L = tt.dot(g, tt.dot(tn.MatrixInverse()(H), g))+tt.log(tn.Det()(-H))
for _ in gen():
pass
optimizer = utils.Maximizer(L, [theta], gen=gen, method='gd', eps=0.1, debug=True, iters=1000, inf_ignore=10)
optimizer.maximize()
print theta.get_value()
