def denormalize(self, v):
mean = ptu.from_numpy(self.mean)
std = ptu.from_numpy(self.std)
if v.dim() == 2:
mean = mean.unsqueeze(0)
std = std.unsqueeze(0)
return mean + v * std
def normalize(self, v, clip_range=None):
if clip_range is None:
clip_range = self.default_clip_range
mean = ptu.from_numpy(self.mean)
std = ptu.from_numpy(self.std)
if v.dim() == 2:
# Unsqueeze along the batch use automatic broadcasting
mean = mean.unsqueeze(0)
std = std.unsqueeze(0)
return torch.clamp((v - mean) / std, -clip_range, clip_range)
def _reconstruct_img(self, flat_img):
latent_distribution_params = self.vae.encode(
ptu.from_numpy(flat_img.reshape(1, -1)))
reconstructions, _ = self.vae.decode(latent_distribution_params[0])
imgs = ptu.get_numpy(reconstructions)
imgs = imgs.reshape(1, self.input_channels, self.imsize, self.imsize)
return imgs[0]
def denormalize_scale(self, v):
"""
Only denormalize the scale. Do not add the mean.
"""
std = ptu.from_numpy(self.std)
if v.dim() == 2:
std = std.unsqueeze(0)
return v * std
def normalize_scale(self, v):
"""
Only normalize the scale. Do not subtract the mean.
"""
std = ptu.from_numpy(self.std)
if v.dim() == 2:
std = std.unsqueeze(0)
return v / std
def reconstruct_img(flat_img):
latent_distribution_params = vae.encode(
ptu.from_numpy(flat_img.reshape(1, -1)).cuda())
reconstructions, _ = vae.decode(latent_distribution_params[0])
imgs = ptu.get_numpy(reconstructions)
imgs = imgs.reshape(1, vae.input_channels, vae.imsize,
vae.imsize).transpose(0, 3, 2, 1) # BCWH -> BHWC
img = cv2.cvtColor(imgs[0], cv2.COLOR_RGB2BGR)
return img
def get_latent(raw_image):
"""Get latent variables (mean vector)"""
image = cv2.resize(raw_image, (vae.imsize, vae.imsize))
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = normalize_image(image)
# swap order and reshape
flat_img = torch.from_numpy(image).permute(2, 1,
0).flatten(start_dim=1).numpy()
latent_distribution_params = vae.encode(
ptu.from_numpy(flat_img.reshape(1, -1)).cuda())
latents = ptu.get_numpy(latent_distribution_params[0])
return latents
def _update_info(self, info, obs):
latent_distribution_params = self.vae.encode(
ptu.from_numpy(obs[self.vae_input_observation_key].reshape(1, -1)))
latent_obs, logvar = ptu.get_numpy(latent_distribution_params[0])[0], \
ptu.get_numpy(latent_distribution_params[1])[0]
# assert (latent_obs == obs['latent_observation']).all()
latent_goal = self.desired_goal['latent_desired_goal']
dist = latent_goal - latent_obs
var = np.exp(logvar.flatten())
var = np.maximum(var, self.reward_min_variance)
err = dist * dist / 2 / var
mdist = np.sum(err) # mahalanobis distance
info["vae_mdist"] = mdist
info["vae_success"] = 1 if mdist < self.epsilon else 0
info["vae_dist"] = np.linalg.norm(dist, ord=self.norm_order)
info["vae_dist_l1"] = np.linalg.norm(dist, ord=1)
info["vae_dist_l2"] = np.linalg.norm(dist, ord=2)
def _elem_or_tuple_to_variable(elem_or_tuple):
if isinstance(elem_or_tuple, tuple):
return tuple(_elem_or_tuple_to_variable(e) for e in elem_or_tuple)
return ptu.from_numpy(elem_or_tuple).float()
def torch_ify(np_array_or_other):
if isinstance(np_array_or_other, np.ndarray):
return ptu.from_numpy(np_array_or_other)
else:
return np_array_or_other
def _encode(self, imgs):
latent_distribution_params = self.vae.encode(ptu.from_numpy(imgs))
return ptu.get_numpy(latent_distribution_params[0])
def _decode(self, latents):
reconstructions, _ = self.vae.decode(ptu.from_numpy(latents))
decoded = ptu.get_numpy(reconstructions)
return decoded