# get latent vector
with torch.no_grad():
# mu, logvar = sModel.encode(torch.from_numpy(skels[:-1]).unsqueeze(0).float())
mu, logvar = sModel.encode(
torch.from_numpy(skels[:]).unsqueeze(0).float())
z = sModel.reparameterize(mu, logvar)
z = z.cpu().numpy()
with torch.no_grad():
value, action, action_log_prob, states, _, _ = actor_critic.act(
current_obs, current_state, masks, deterministic=True)
# get trajectory of NAO by decoding
with torch.no_grad():
x_hat_rj = rModel.decode(action).reshape(-1, 14)
# send motor values to Real NAO
# torch -> numpy -> list -> bytes -> sendall
filtered_action = alpha * x_hat_rj + (1 - alpha) * prev_actions # TODO
send_data = filtered_action.squeeze(0).cpu().numpy().tolist()
nao_s.write_socket({'header': 'SetMotor', 'data': send_data})
print('send: nao_angle (deg): ', np.rad2deg(send_data))
# time.sleep(1/20) # 20 Hz
next_obs = env.robot._obfilt(z)
update_current_obs(next_obs, current_obs)
prev_actions = filtered_action
rollouts.observations[step], rollouts.states[step],
rollouts.masks[step])
# get next observation
data, last_scene, first_scene = sdm.get_random_scene_frame()
# get latent vector
with torch.no_grad():
mu, logvar = sModel.encode(
torch.from_numpy(data[:-1]).unsqueeze(0).float())
zs = sModel.reparameterize(mu, logvar)
# cpu_actions = action.data.squeeze(1).cpu().numpy()
# get robot motor value by decoding
with torch.no_grad():
x_hat = rModel.decode(action.cpu()).reshape(-1, 14)
if cyclic_policy:
with torch.no_grad():
rSkel = sModel.decode(zs).reshape(75, -1).squeeze()
rSkel = rSkel.numpy()
else:
rSkel = data
next_obs, next_state, reward, done, info, true_rew = env.step(
x_hat,
zs.cpu().numpy(), rSkel)
# To separate the Q-value of the current scene from others
# if last_scene:
# done[0] = 1 # done is np array, [0]
oirgin_imgs.append(test_faceDataset[i][0])
# for img in test_faceDataset[:10]:
for img in oirgin_imgs:
x = Variable(img.unsqueeze_(0).cuda())
out_img = model(x)[0]
recon_imgs.append(out_img)
oirgin_imgs = torch.cat(oirgin_imgs)
recon_imgs = torch.cat(recon_imgs)
compare = torch.cat((oirgin_imgs, recon_imgs.cpu().data))
save_image(compare.cpu(), output_path+'/fig1_3.jpg', nrow=10, normalize=True)
# fig1_4
imgs = []
for i in range(32):
z = Variable(torch.randn(1024).cuda())
out_img = model.decode(z)
imgs.append(out_img)
imgs = torch.cat(imgs)
save_image(imgs.cpu().data, output_path+'/fig1_4.jpg', nrow=8, normalize=True)
# fig1_5
data_for_tsne = torch.cat(data_for_tsne)
label_for_tsne = torch.cat(label_for_tsne).numpy()
mu, logvar = model.encode(Variable(data_for_tsne.cuda()))
latent_code = mu.cpu().data.numpy()
latent_emmbedded = TSNE(random_state=2).fit_transform(latent_code)
plt.figure()
# for i in [0,1]:
# if i:
# gender = 'Male'
# else:
if o.sum() != 0 and o_next.sum() != 0:
with torch.no_grad():
if not is_image:
s = [np.array([1])]
s_next = [np.array([1])]
else:
s = model.encode(
Variable(
torch.cuda.FloatTensor(
np.transpose(o, (2, 0, 1))[None])))[1]
s_next = model.encode(
Variable(
torch.cuda.FloatTensor(
np.transpose(o_next, (2, 0, 1))[None])))[1]
if i == 0 and t == 0:
recon = model.decode(s)
import ipdb
ipdb.set_trace()
save_image(
torch.cat([
torch.cuda.FloatTensor(
np.transpose(o,
(2, 0, 1))[None]).cpu() /
255,
recon.data.cpu()
],
dim=0),
os.path.join(save_path,
'verified_img_preprocessing.png'))
s = s.cpu().numpy()
s_next = s_next.cpu().numpy()