eval_samples = iter(eval_loader)
num_points = 400
file_dir = 'ours_fix_alpha_' + str(num_points)
os.makedirs(file_dir, exist_ok=True)
feature_list = np.random.randn(num_points * opt.batch_size * 2, opt.vector_dim)
label = []
for total_step in tqdm(range(num_points)):
batch = next(train_samples)
to_device(batch, device)
# feature = encoder(batch['img'])
feature = encoder(batch['img'], batch['v_0'])
check_shape(feature, 'feature')
set_mute(True)
feature_list[total_step * opt.batch_size:(total_step + 1) *
opt.batch_size] = feature.data.cpu().numpy()
label = [0] * num_points * opt.batch_size
for total_step in tqdm(range(num_points, 2 * num_points)):
batch = next(eval_samples)
to_device(batch, device)
feature = encoder(batch['img'], batch['v_0'])
check_shape(feature, 'feature')
set_mute(True)
feature_list[total_step * opt.batch_size:(total_step + 1) *
opt.batch_size] = feature.data.cpu().numpy()
label.extend([1] * num_points * opt.batch_size)
np_feature_list = np.array(feature_list)
def eval_decision(logger, total_steps, algorithm='WCM'):
batch = next(eval_samples)
to_device(batch, device)
check_shape(batch)
set_mute(True)
x = model_cluster[0]._decoder._base_dist.mean.clone().detach().view(-1, opt.points_num, 2)
x.requires_grad = True
zs = [model._params(
velocity=batch['v_0'].view(-1,1),
visual_features=batch['img'],
) for model in model_cluster]
optimizer = torch.optim.Adam(params=[x], lr=1e-1)
x_best = x.clone()
loss_best = torch.ones(()).to(x.device) * 1000.0
for _ in range(50):
optimizer.zero_grad()
y, _ = model_cluster[0]._decoder._forward(x=x, z=zs[0])
imitation_posteriors = list()
for model, z in zip(model_cluster, zs):
_, log_prob, logabsdet = model._decoder._inverse(y=y, z=z)
imitation_prior = torch.mean(log_prob - logabsdet)
imitation_posteriors.append(imitation_prior)
imitation_posteriors = torch.stack(imitation_posteriors, dim=0)
if algorithm == "WCM":
loss, _ = torch.min(-imitation_posteriors, dim=0)
elif algorithm == "BCM":
loss, _ = torch.max(-imitation_posteriors, dim=0)
else:
loss = torch.mean(-imitation_posteriors, dim=0)
loss.backward(retain_graph=True)
optimizer.step()
if loss < loss_best:
x_best = x.clone()
loss_best = loss.clone()
plan, _ = model_cluster[0]._decoder._forward(x=x_best, z=zs[0])
xy = plan.detach().cpu().numpy()[0]*opt.max_dist
real_xy = batch['xy'].view(-1, 2).data.cpu().numpy()*opt.max_dist
fake_x = xy[:,0]
fake_y = xy[:,1]
real_x = real_xy[:,0]
real_y = real_xy[:,1]
time = batch['t'].data.cpu().numpy()[0]*opt.max_t
# time_list = [0.0, 0.75, 1.5, 2.25] #oxford training time
time_list = [0.0000, 0.1875, 0.3750, 0.5625, 0.7500, 0.9375, 1.1250, 1.3125, 1.5001,1.6875, 1.8750, 2.0625, 2.2501, 2.4376, 2.6250, 2.8126]
xs = []
ys = []
for t in time:
x, y = interpolation(time_list, fake_x, fake_y, t)
xs.append(x)
ys.append(y)
fake_x = np.array(xs)
fake_y = np.array(ys)
# max_x = 30.
# max_y = 30.
# fig = plt.figure(figsize=(7, 7))
# ax1 = fig.add_subplot(111)
# ax1.plot(real_x, real_y, label='real-trajectory', color = 'b', linewidth=3, linestyle='--')
# ax1.plot(fake_x, fake_y, label='fake-trajectory', color = 'r', linewidth=3)
# ax1.set_xlabel('Forward/(m)')
# ax1.set_ylabel('Sideways/(m)')
# ax1.set_xlim([0., max_x])
# ax1.set_ylim([-max_y/2, max_y/2])
# plt.legend(loc='lower right')
# plt.legend(loc='lower left')
# plt.savefig('result/output/%s/' % opt.dataset_name+'/'+str(total_steps)+'.png')
# plt.close('all')
ex = np.mean(np.abs(fake_x-real_x))
ey = np.mean(np.abs(fake_y-real_y))
fde = np.hypot(fake_x - real_x, fake_y - real_y)[-1]
ade = np.mean(np.hypot(fake_x - real_x, fake_y - real_y))
logger.add_scalar('eval/ex', ex.item(), total_steps)
logger.add_scalar('eval/ey', ey.item(), total_steps)
logger.add_scalar('eval/fde', fde.item(), total_steps)
logger.add_scalar('eval/ade', ade.item(), total_steps)