def sample(self): dir_normal = tf.random.normal(shape=tf.shape(self.centers)) dir_normal_norm = helper.safe_tf_sqrt(tf.reduce_sum(dir_normal**2, axis=1, keepdims=True)) sample_dir = dir_normal/dir_normal_norm sample_norm = tf.random_uniform(tf.shape(self.inner_radius), 0, 1, dtype=tf.float32)*(self.outer_radius-self.inner_radius)+self.inner_radius sample = sample_norm*sample_dir+self.centers return sample
def inference(self, batch, additional_inputs_tf):
self.epoch = additional_inputs_tf[0]
self.b_identity = additional_inputs_tf[1]
if len(batch['observed']['properties']['flat']) > 0:
for e in batch['observed']['properties']['flat']:
e['dist'] = 'dirac'
else:
for e in batch['observed']['properties']['image']:
e['dist'] = 'dirac'
self.input_sample = batch['observed']['data']
self.input_dist = distributions.ProductDistribution(
sample_properties=batch['observed']['properties'],
params=self.input_sample)
if not self.bModules: self.generate_modules(batch)
try:
self.n_time = batch['observed']['properties']['flat'][0]['size'][1]
except:
self.n_time = batch['observed']['properties']['image'][0]['size'][
1]
try:
batch_size_tf = tf.shape(self.input_sample['flat'])[0]
except:
batch_size_tf = tf.shape(self.input_sample['image'])[0]
self.prior_param = self.PriorMap.forward(
(tf.zeros(shape=(batch_size_tf, 1)), ))
self.prior_dist = distributions.DiagonalGaussianDistribution(
params=self.prior_param)
self.prior_latent_code = self.prior_dist.sample()
self.neg_ent_prior = self.prior_dist.log_pdf(self.prior_latent_code)
self.mean_neg_ent_prior = tf.reduce_mean(self.neg_ent_prior)
self.uniform_dist = distributions.UniformDistribution(
params=tf.concat([
tf.zeros(shape=(batch_size_tf, 1)),
tf.ones(shape=(batch_size_tf, 1))
],
axis=1))
self.convex_mix = self.uniform_dist.sample()
self.prior_latent_code_expanded = tf.reshape(
self.prior_latent_code,
[-1, 1, *self.prior_latent_code.get_shape().as_list()[1:]])
self.obs_sample_param = self.Decoder.forward(
self.prior_latent_code_expanded)
self.obs_sample_dist = distributions.ProductDistribution(
sample_properties=batch['observed']['properties'],
params=self.obs_sample_param)
self.obs_sample = self.obs_sample_dist.sample(b_mode=True)
#############################################################################
b_use_reconst_as_reg = True
self.posterior_param_expanded = self.Encoder.forward(self.input_sample)
self.posterior_param = self.posterior_param_expanded[:, 0, :]
self.posterior_dist = distributions.DiagonalGaussianDistribution(
params=self.posterior_param)
self.posterior_latent_code = self.posterior_dist.sample()
self.posterior_latent_code_expanded = self.posterior_latent_code[:, np.
newaxis, :]
self.reconst_param = self.Decoder.forward(
self.posterior_latent_code_expanded)
self.reconst_dist = distributions.ProductDistribution(
sample_properties=batch['observed']['properties'],
params=self.reconst_param)
self.reconst_sample = self.reconst_dist.sample(b_mode=True)
if b_use_reconst_as_reg:
self.reg_target_dist = self.reconst_dist
self.reg_target_sample = self.reconst_sample
else:
self.reg_target_dist = self.obs_sample_dist
self.reg_target_sample = self.obs_sample
self.neg_cross_ent_posterior = self.prior_dist.log_pdf(
self.posterior_latent_code)
self.mean_neg_cross_ent_posterior = tf.reduce_mean(
self.neg_cross_ent_posterior)
self.neg_ent_posterior = self.posterior_dist.log_pdf(
self.posterior_latent_code)
self.mean_neg_ent_posterior = tf.reduce_mean(self.neg_ent_posterior)
# self.kl_posterior_prior = distributions.KLDivDiagGaussianVsDiagGaussian().forward(self.posterior_dist, self.prior_dist)
# self.mean_kl_posterior_prior = tf.reduce_mean(self.kl_posterior_prior)
self.mean_kl_posterior_prior = -self.mean_neg_cross_ent_posterior + self.mean_neg_ent_posterior
#############################################################################
self.reg_sample_param = {'image': None, 'flat': None}
try:
self.reg_sample_param['image'] = self.convex_mix[:, np.newaxis, :, np.newaxis, np.newaxis]*self.reg_target_sample['image']+\
(1-self.convex_mix[:, np.newaxis, :, np.newaxis, np.newaxis])*self.input_sample['image']
except:
self.reg_sample_param['flat'] = self.convex_mix[:, np.newaxis, :]*self.reg_target_sample['flat']+\
(1-self.convex_mix[:, np.newaxis, :])*self.input_sample['flat']
self.reg_dist = distributions.ProductDistribution(
sample_properties=batch['observed']['properties'],
params=self.reg_sample_param)
self.reg_sample = self.reg_dist.sample(b_mode=True)
self.critic_real = self.Discriminator.forward(self.input_sample)
self.critic_gen = self.Discriminator.forward(self.obs_sample)
self.critic_reg = self.Discriminator.forward(self.reg_sample)
lambda_t = 1
self.real_reconst_distances_sq = self.metric_distance_sq(
self.input_sample, self.reconst_sample)
self.autoencode_costs = self.real_reconst_distances_sq / 2 + lambda_t * self.mean_kl_posterior_prior
try:
self.convex_grad = tf.gradients(self.critic_reg,
[self.reg_sample['image']])[0]
self.convex_grad_norm = helper.safe_tf_sqrt(
tf.reduce_sum(self.convex_grad**2,
axis=[-1, -2, -3],
keep_dims=False)[:, :, np.newaxis])
except:
self.convex_grad = tf.gradients(self.critic_reg,
[self.reg_sample['flat']])[0]
self.convex_grad_norm = helper.safe_tf_sqrt(
tf.reduce_sum(self.convex_grad**2, axis=[-1], keep_dims=True))
self.gradient_penalties = ((self.convex_grad_norm - 1)**2)
self.mean_critic_real = tf.reduce_mean(self.critic_real)
self.mean_critic_gen = tf.reduce_mean(self.critic_gen)
self.mean_critic_reg = tf.reduce_mean(self.critic_reg)
self.mean_autoencode_cost = tf.reduce_mean(self.autoencode_costs)
self.mean_gradient_penalty = tf.reduce_mean(self.gradient_penalties)
self.regularizer_cost = 10 * self.mean_gradient_penalty
self.discriminator_cost = -self.mean_critic_real + self.mean_critic_gen + self.regularizer_cost
self.generator_cost = -self.mean_critic_gen
self.transporter_cost = self.mean_autoencode_cost
def euclidean_distance(self, a, b):
return helper.safe_tf_sqrt(self.metric_distance_sq(a, b))
def inference(self, batch, additional_inputs_tf):
self.epoch = additional_inputs_tf[0]
self.b_identity = additional_inputs_tf[1]
if len(batch['observed']['properties']['flat'])>0:
for e in batch['observed']['properties']['flat']: e['dist']='dirac'
else:
for e in batch['observed']['properties']['image']: e['dist']='dirac'
self.input_sample = batch['observed']['data']
self.input_dist = distributions.ProductDistribution(sample_properties = batch['observed']['properties'], params = self.input_sample)
if not self.bModules: self.generate_modules(batch)
try: self.n_time = batch['observed']['properties']['flat'][0]['size'][1]
except: self.n_time = batch['observed']['properties']['image'][0]['size'][1]
try: self.batch_size_tf = tf.shape(self.input_sample['flat'])[0]
except: self.batch_size_tf = tf.shape(self.input_sample['image'])[0]
#############################################################################
self.prior_param = self.PriorMap.forward((tf.zeros(shape=(self.batch_size_tf, 1)),))
self.prior_dist = distributions.DiagonalGaussianDistribution(params = self.prior_param)
self.prior_latent_code = self.prior_dist.sample()
self.prior_latent_code_expanded = self.prior_latent_code[:,np.newaxis,:]
# GENERATOR
self.prior_feature_expanded = self.Generator.forward(self.prior_latent_code_expanded)
self.prior_feature = self.prior_feature_expanded[:,0,:]
self.obs_sample_param = self.PostGen.forward(self.prior_feature_expanded)
self.obs_sample_dist = distributions.ProductDistribution(sample_properties = batch['observed']['properties'], params = self.obs_sample_param)
self.obs_sample = self.obs_sample_dist.sample(b_mode=True)
if not os.path.exists(str(Path.home())+'/ExperimentalResults/FixedSamples/'): os.makedirs(str(Path.home())+'/ExperimentalResults/FixedSamples/')
if os.path.exists(str(Path.home())+'/ExperimentalResults/FixedSamples/np_constant_prior_sample_'+str(self.prior_latent_code.get_shape().as_list()[-1])+'.npz'):
np_constant_prior_sample = np.load(str(Path.home())+'/ExperimentalResults/FixedSamples/np_constant_prior_sample_'+str(self.prior_latent_code.get_shape().as_list()[-1])+'.npz')
else:
np_constant_prior_sample = np.random.normal(loc=0., scale=1., size=[400, self.prior_latent_code.get_shape().as_list()[-1]])
np.save(str(Path.home())+'/ExperimentalResults/FixedSamples/np_constant_prior_sample_'+str(self.prior_latent_code.get_shape().as_list()[-1])+'.npz', np_constant_prior_sample)
self.constant_prior_latent_code = tf.constant(np.asarray(np_constant_prior_sample), dtype=np.float32)
self.constant_prior_latent_code_expanded = self.constant_prior_latent_code[:, np.newaxis, :]
self.constant_prior_feature_expanded = self.Generator.forward(self.constant_prior_latent_code_expanded)
self.constant_prior_feature = self.constant_prior_feature_expanded[:,0,:]
self.constant_obs_sample_param = self.PostGen.forward(self.constant_prior_feature_expanded)
self.constant_obs_sample_dist = distributions.ProductDistribution(sample_properties = batch['observed']['properties'], params = self.constant_obs_sample_param)
self.constant_obs_sample = self.constant_obs_sample_dist.sample(b_mode=True)
if self.config['n_latent'] == 2:
grid_scale = 3
x = np.linspace(-grid_scale, grid_scale, 20)
y = np.linspace(grid_scale, -grid_scale, 20)
xv, yv = np.meshgrid(x, y)
np_constant_prior_grid_sample = np.concatenate((xv.flatten()[:, np.newaxis], yv.flatten()[:, np.newaxis][:]), axis=1)
self.constant_prior_grid_latent_code = tf.constant(np.asarray(np_constant_prior_grid_sample), dtype=np.float32)
self.constant_prior_grid_latent_code_expanded = self.constant_prior_grid_latent_code[:, np.newaxis, :]
self.constant_prior_grid_feature_expanded = self.Generator.forward(self.constant_prior_grid_latent_code_expanded)
self.constant_prior_grid_feature = self.constant_prior_grid_feature_expanded[:,0,:]
self.constant_obs_grid_sample_param = self.PostGen.forward(self.constant_prior_grid_feature_expanded)
self.constant_obs_grid_sample_dist = distributions.ProductDistribution(sample_properties = batch['observed']['properties'], params = self.constant_obs_grid_sample_param)
self.constant_obs_grid_sample = self.constant_obs_grid_sample_dist.sample(b_mode=True)
#############################################################################
# ENCODER
self.epsilon_param = self.PriorMap.forward((tf.zeros(shape=(self.batch_size_tf, 1)),))
self.epsilon_dist = distributions.DiagonalGaussianDistribution(params = self.epsilon_param)
if self.config['encoder_mode'] == 'Deterministic':
self.epsilon = None
if self.config['encoder_mode'] == 'Gaussian' or self.config['encoder_mode'] == 'UnivApprox' or self.config['encoder_mode'] == 'UnivApproxNoSpatial' or self.config['encoder_mode'] == 'UnivApproxSine':
self.epsilon = self.epsilon_dist.sample()
self.pre_posterior_feature_expanded = self.PreEnc.forward(self.input_sample, noise=self.epsilon)
self.pre_posterior_feature = self.pre_posterior_feature_expanded[:,0,:]
# self.tiny_perturb_dist = distributions.DiagonalGaussianDistribution(params = tf.zeros(shape=(self.batch_size_tf, 2*512)))
# self.tiny_perturb_sample_expanded = 0.1*self.tiny_perturb_dist.sample()[:,np.newaxis,:]
self.flow_param = self.FlowMap.forward()
self.flow_object = transforms.HouseholdRotationFlow(self.flow_param, 512)
self.transformed_pre_posterior_feature, _ = self.flow_object.transform(self.pre_posterior_feature, tf.zeros(shape=(self.batch_size_tf, 1)))
self.transformed_pre_posterior_feature_abs = helper.relu_abs(self.transformed_pre_posterior_feature)
self.transformed_pre_posterior_feature_abs_means = tf.reduce_mean(self.transformed_pre_posterior_feature_abs, axis=0)
# self.reconst_param = self.PostGen.forward(self.pre_posterior_feature_expanded+self.tiny_perturb_sample_expanded)
self.reconst_param = self.PostGen.forward(self.pre_posterior_feature_expanded)
self.reconst_dist = distributions.ProductDistribution(sample_properties = batch['observed']['properties'], params = self.reconst_param)
self.reconst_sample = self.reconst_dist.sample(b_mode=True)
self.posterior_latent_code_expanded = self.Encoder.forward(self.pre_posterior_feature_expanded)
self.posterior_latent_code = self.posterior_latent_code_expanded[:,0,:]
self.posterior_feature_expanded = self.Generator.forward(self.posterior_latent_code_expanded)
self.posterior_feature = self.posterior_feature_expanded[:,0,:]
self.full_reconst_param = self.PostGen.forward(self.posterior_feature_expanded)
self.full_reconst_dist = distributions.ProductDistribution(sample_properties = batch['observed']['properties'], params = self.full_reconst_param)
self.full_reconst_sample = self.full_reconst_dist.sample(b_mode=True)
self.interpolated_posterior_latent_code = self.interpolate_latent_codes(self.posterior_latent_code, size=self.batch_size_tf//2)
self.interpolated_posterior_latent_code_expanded = self.interpolated_posterior_latent_code[:,np.newaxis,:]
self.interpolated_posterior_feature_expanded = self.Generator.forward(self.interpolated_posterior_latent_code_expanded)
self.interpolated_posterior_feature = self.interpolated_posterior_feature_expanded[:,0,:]
self.interpolated_reconst_param = self.PostGen.forward(self.interpolated_posterior_feature_expanded)
self.interpolated_reconst_dist = distributions.ProductDistribution(sample_properties = batch['observed']['properties'], params = self.interpolated_reconst_param)
self.interpolated_obs = self.interpolated_reconst_dist.sample(b_mode=True)
### Primal Penalty
# self.enc_reg_cost = self.compute_MMD(self.prior_latent_code, self.posterior_latent_code)
# self.div_posterior = self.Diverger.forward(self.posterior_latent_code_expanded)
# self.div_prior = self.Diverger.forward(self.prior_latent_code_expanded)
#############################################################################
# REGULARIZER
self.reg_target_dist = self.reconst_dist
self.reg_target_sample = self.reconst_sample
self.reg_dist = self.reconst_dist
self.reg_sample = self.reconst_sample
#############################################################################
# # OBJECTIVES
# # Divergence
# if self.config['divergence_mode'] == 'GAN' or self.config['divergence_mode'] == 'NS-GAN':
# self.div_cost = -(tf.reduce_mean(tf.log(tf.nn.sigmoid(self.div_posterior)+10e-7))+tf.reduce_mean(tf.log(1-tf.nn.sigmoid(self.div_prior)+10e-7)))
# if self.config['divergence_mode'] == 'WGAN-GP':
# uniform_dist = distributions.UniformDistribution(params = tf.concat([tf.zeros(shape=(self.batch_size_tf, 1)), tf.ones(shape=(self.batch_size_tf, 1))], axis=1))
# uniform_w = uniform_dist.sample()
# self.trivial_line = uniform_w[:,np.newaxis,:]*self.posterior_latent_code_expanded+(1-uniform_w[:,np.newaxis,:])*self.prior_latent_code_expanded
# self.div_trivial_line = self.Diverger.forward(self.trivial_line)
# self.trivial_line_grad = tf.gradients(tf.reduce_sum(self.div_trivial_line), [self.trivial_line])[0]
# self.trivial_line_grad_norm = helper.safe_tf_sqrt(tf.reduce_sum(self.trivial_line_grad**2, axis=-1, keep_dims=False)[:,:,np.newaxis])
# self.trivial_line_grad_norm_1_penalties = ((self.trivial_line_grad_norm-1)**2)
# self.div_reg_cost = tf.reduce_mean(self.trivial_line_grad_norm_1_penalties)
# self.div_cost = -(tf.reduce_mean(self.div_posterior)-tf.reduce_mean(self.div_prior))+10*self.div_reg_cost
# ### Encoder
# b_use_timer, timescale, starttime = False, 10, 5
# self.OT_primal = self.sample_distance_function(self.input_sample, self.reconst_sample)
# self.mean_OT_primal = tf.reduce_mean(self.OT_primal)
# if b_use_timer:
# self.mean_POT_primal = self.mean_OT_primal+helper.hardstep((self.epoch-float(starttime))/float(timescale))*self.config['enc_reg_strength']*self.enc_reg_cost
# else:
# self.mean_POT_primal = self.mean_OT_primal+self.config['enc_reg_strength']*self.enc_reg_cost
# self.enc_cost = self.mean_POT_primal
# self.enc_cost = self.mean_OT_primal
# ### Critic
# # self.cri_cost = self.compute_MMD(self.prior_latent_code, self.prior_latent_code)
# if self.config['divergence_mode'] == 'NS-GAN':
# self.cri_cost = -tf.reduce_mean(tf.log(tf.nn.sigmoid(self.div_prior)+10e-7))+self.config['enc_reg_strength']*self.compute_MMD(self.prior_latent_code, self.prior_latent_code)
# elif self.config['divergence_mode'] == 'GAN':
# self.cri_cost = tf.reduce_mean(tf.log(1-tf.nn.sigmoid(self.div_prior)+10e-7))+self.config['enc_reg_strength']*self.compute_MMD(self.prior_latent_code, self.prior_latent_code)
# elif self.config['divergence_mode'] == 'WGAN-GP':
# self.cri_cost = -tf.reduce_mean(self.div_prior)+self.config['enc_reg_strength']*self.compute_MMD(self.prior_latent_code, self.prior_latent_code)
# self.cri_reg_cost = self.shouldbezero_cost
# self.cri_cost = self.shouldbenormal_cost+self.cri_reg_cost
### Generator
# self.gen_cost = self.mean_OT_primal
# self.div_cost = tf.reduce_mean(tf.reduce_sum(self.transformed_pre_posterior_feature[:, self.config['n_latent']:]**2, axis=1))
self.div_cost = tf.reduce_mean(tf.reduce_sum(helper.relu_abs(self.transformed_pre_posterior_feature[:, self.config['n_latent']:]), axis=1))
self.enc_cost = tf.reduce_mean(helper.safe_tf_sqrt(tf.reduce_sum((self.pre_posterior_feature-self.posterior_feature)**2, axis=1)))
self.gen_cost = tf.reduce_mean(helper.safe_tf_sqrt(tf.reduce_sum((self.pre_posterior_feature-self.posterior_feature)**2, axis=1)))
self.cri_cost = tf.reduce_mean(self.sample_distance_function(self.input_sample, self.reconst_sample))
def inference(self, batch, additional_inputs_tf):
self.epoch = additional_inputs_tf[0]
self.b_identity = additional_inputs_tf[1]
if len(batch['observed']['properties']['flat'])>0:
for e in batch['observed']['properties']['flat']: e['dist']='dirac'
else:
for e in batch['observed']['properties']['image']: e['dist']='dirac'
self.input_sample = batch['observed']['data']
self.input_dist = distributions.ProductDistribution(sample_properties = batch['observed']['properties'], params = self.input_sample)
if not self.bModules: self.generate_modules(batch)
try: self.n_time = batch['observed']['properties']['flat'][0]['size'][1]
except: self.n_time = batch['observed']['properties']['image'][0]['size'][1]
try: self.batch_size_tf = tf.shape(self.input_sample['flat'])[0]
except: self.batch_size_tf = tf.shape(self.input_sample['image'])[0]
#############################################################################
# GENERATOR
self.prior_param = self.PriorMap.forward((tf.zeros(shape=(self.batch_size_tf, 1)),))
self.prior_dist = distributions.DiagonalGaussianDistribution(params = self.prior_param)
self.prior_latent_code = self.prior_dist.sample()
self.prior_latent_code_expanded = tf.reshape(self.prior_latent_code, [-1, 1, *self.prior_latent_code.get_shape().as_list()[1:]])
self.neg_ent_prior = self.prior_dist.log_pdf(self.prior_latent_code)
self.mean_neg_ent_prior = tf.reduce_mean(self.neg_ent_prior)
self.obs_sample_param = self.Decoder.forward(self.prior_latent_code_expanded)
self.obs_sample_dist = distributions.ProductDistribution(sample_properties = batch['observed']['properties'], params = self.obs_sample_param)
self.obs_sample = self.obs_sample_dist.sample(b_mode=True)
if os.path.exists('./np_constant_prior_sample_'+str(self.prior_latent_code.get_shape().as_list()[-1])+'.npz'):
np_constant_prior_sample = np.load('./np_constant_prior_sample_'+str(self.prior_latent_code.get_shape().as_list()[-1])+'.npz')
else:
np_constant_prior_sample = np.random.normal(loc=0., scale=1., size=[400, self.prior_latent_code.get_shape().as_list()[-1]])
np.save('./np_constant_prior_sample_'+str(self.prior_latent_code.get_shape().as_list()[-1])+'.npz', np_constant_prior_sample)
self.constant_prior_latent_code = tf.constant(np.asarray(np_constant_prior_sample), dtype=np.float32)
self.constant_prior_latent_code_expanded = tf.reshape(self.constant_prior_latent_code, [-1, 1, *self.constant_prior_latent_code.get_shape().as_list()[1:]])
self.constant_obs_sample_param = self.Decoder.forward(self.constant_prior_latent_code_expanded)
self.constant_obs_sample_dist = distributions.ProductDistribution(sample_properties = batch['observed']['properties'], params = self.constant_obs_sample_param)
self.constant_obs_sample = self.constant_obs_sample_dist.sample(b_mode=True)
#############################################################################
# ENCODER
b_deterministic = False
if b_deterministic:
self.epsilon = None
else:
self.epsilon_param = self.PriorMap.forward((tf.zeros(shape=(self.batch_size_tf, 1)),))
self.epsilon_dist = distributions.DiagonalGaussianDistribution(params = self.prior_param)
self.epsilon = self.epsilon_dist.sample()
self.posterior_latent_code_expanded = self.EncoderSampler.forward(self.input_sample, noise=self.epsilon)
self.posterior_latent_code = self.posterior_latent_code_expanded[:,0,:]
self.interpolated_posterior_latent_code = self.interpolate_latent_codes(self.posterior_latent_code, size=self.batch_size_tf//2)
self.interpolated_obs = self.Decoder.forward(self.interpolated_posterior_latent_code)
self.reconst_param = self.Decoder.forward(self.posterior_latent_code_expanded)
self.reconst_dist = distributions.ProductDistribution(sample_properties = batch['observed']['properties'], params = self.reconst_param)
self.reconst_sample = self.reconst_dist.sample(b_mode=True)
self.kernel_function = self.inv_multiquadratics_kernel
self.k_post_prior = tf.reduce_mean(self.kernel_function(self.posterior_latent_code, self.prior_dist.sample()))
self.k_post_post = tf.reduce_mean(self.kernel_function(self.posterior_latent_code))
self.k_prior_prior = tf.reduce_mean(self.kernel_function(self.prior_dist.sample()))
self.MMD = self.k_prior_prior+self.k_post_post-2*self.k_post_prior
self.disc_posterior = self.DivergenceLatent.forward(self.posterior_latent_code_expanded)
self.disc_prior = self.DivergenceLatent.forward(self.prior_latent_code_expanded)
self.mean_z_divergence = tf.reduce_mean(tf.log(10e-7+self.disc_prior))+tf.reduce_mean(tf.log(10e-7+1-self.disc_posterior))
self.mean_z_divergence_minimizer = tf.reduce_mean(tf.log(10e-7+1-self.disc_posterior))
#############################################################################
# REGULARIZER
self.uniform_dist = distributions.UniformDistribution(params = tf.concat([tf.zeros(shape=(self.batch_size_tf, 1)), tf.ones(shape=(self.batch_size_tf, 1))], axis=1))
self.uniform_sample = self.uniform_dist.sample()
self.reg_trivial_sample_param = {'image': None, 'flat': None}
try:
self.reg_trivial_sample_param['image'] = self.uniform_sample[:, np.newaxis, :, np.newaxis, np.newaxis]*self.obs_sample['image']+\
(1-self.uniform_sample[:, np.newaxis, :, np.newaxis, np.newaxis])*self.input_sample['image']
except:
self.reg_trivial_sample_param['flat'] = self.uniform_sample[:, np.newaxis, :]*self.obs_sample['flat']+\
(1-self.uniform_sample[:, np.newaxis, :])*self.input_sample['flat']
self.reg_trivial_dist = distributions.ProductDistribution(sample_properties = batch['observed']['properties'], params = self.reg_trivial_sample_param)
self.reg_trivial_sample = self.reg_trivial_dist.sample(b_mode=True)
self.reg_target_dist = self.obs_sample_dist
self.reg_target_sample = self.obs_sample
self.reg_dist = self.reg_trivial_dist
self.reg_sample = self.reg_trivial_sample
#############################################################################
# CRITIC
self.critic_real = self.Discriminator.forward(self.input_sample)
self.critic_gen = self.Discriminator.forward(self.obs_sample)
self.critic_reg_trivial = self.Discriminator.forward(self.reg_trivial_sample)
self.critic_reg = self.critic_reg_trivial
try:
self.trivial_grad = tf.gradients(self.critic_reg_trivial, [self.reg_trivial_sample['image']])[0]
self.trivial_grad_norm = helper.safe_tf_sqrt(tf.reduce_sum(self.trivial_grad**2, axis=[-1,-2,-3], keep_dims=False)[:,:,np.newaxis])
except:
self.trivial_grad = tf.gradients(self.critic_reg_trivial, [self.reg_trivial_sample['flat']])[0]
self.trivial_grad_norm = helper.safe_tf_sqrt(tf.reduce_sum(self.trivial_grad**2, axis=[-1,], keep_dims=True))
self.trivial_grad_norm_1_penalties = ((self.trivial_grad_norm-1)**2)
self.mean_critic_real = tf.reduce_mean(self.critic_real)
self.mean_critic_gen = tf.reduce_mean(self.critic_gen)
self.mean_critic_reg = tf.reduce_mean(self.critic_reg)
self.mean_OT_dual = self.mean_critic_real-self.mean_critic_gen
#############################################################################
# OBJECTIVES
self.enc_reg_strength, self.disc_reg_strength = 100, 10
self.real_reconst_distances_sq = self.metric_distance_sq(self.input_sample, self.reconst_sample)
self.OT_primal = tf.reduce_mean(helper.safe_tf_sqrt(self.real_reconst_distances_sq))
self.mean_OT_primal = tf.reduce_mean(self.OT_primal)
self.enc_reg_cost = self.MMD # self.mean_z_divergence_minimizer
self.mean_POT_primal = self.mean_OT_primal+self.enc_reg_strength*self.enc_reg_cost
self.disc_reg_cost = tf.reduce_mean(self.trivial_grad_norm_1_penalties)
# WGAN-GP
# self.div_cost = -self.mean_z_divergence
self.enc_cost = self.mean_POT_primal
self.disc_cost = -self.mean_OT_dual+self.disc_reg_strength*self.disc_reg_cost
self.gen_cost = -self.mean_critic_gen
def inference(self, batch, additional_inputs_tf):
self.epoch = additional_inputs_tf[0]
self.b_identity = additional_inputs_tf[1]
if len(batch['observed']['properties']['flat']) > 0:
for e in batch['observed']['properties']['flat']:
e['dist'] = 'dirac'
else:
for e in batch['observed']['properties']['image']:
e['dist'] = 'dirac'
self.input_sample = batch['observed']['data']
self.input_dist = distributions.ProductDistribution(
sample_properties=batch['observed']['properties'],
params=self.input_sample)
if not self.bModules: self.generate_modules(batch)
try:
self.n_time = batch['observed']['properties']['flat'][0]['size'][1]
except:
self.n_time = batch['observed']['properties']['image'][0]['size'][
1]
try:
self.batch_size_tf = tf.shape(self.input_sample['flat'])[0]
except:
self.batch_size_tf = tf.shape(self.input_sample['image'])[0]
#############################################################################
# GENERATOR
self.prior_param = self.PriorMap.forward(
(tf.zeros(shape=(self.batch_size_tf, 1)), ))
self.prior_dist = distributions.DiagonalGaussianDistribution(
params=self.prior_param)
self.prior_latent_code = self.prior_dist.sample()
self.prior_latent_code_expanded = tf.reshape(
self.prior_latent_code,
[-1, 1, *self.prior_latent_code.get_shape().as_list()[1:]])
# self.neg_ent_prior = self.prior_dist.log_pdf(self.prior_latent_code)
# self.mean_neg_ent_prior = tf.reduce_mean(self.neg_ent_prior)
self.obs_sample_param = self.Generator.forward(
self.prior_latent_code_expanded)
self.obs_sample_dist = distributions.ProductDistribution(
sample_properties=batch['observed']['properties'],
params=self.obs_sample_param)
self.obs_sample = self.obs_sample_dist.sample(b_mode=True)
if not os.path.exists('./fixed_samples/'):
os.makedirs('./fixed_samples/')
if os.path.exists(
'./fixed_samples/np_constant_prior_sample_' +
str(self.prior_latent_code.get_shape().as_list()[-1]) +
'.npz'):
np_constant_prior_sample = np.load(
'./np_constant_prior_sample_' +
str(self.prior_latent_code.get_shape().as_list()[-1]) + '.npz')
else:
np_constant_prior_sample = np.random.normal(
loc=0.,
scale=1.,
size=[400,
self.prior_latent_code.get_shape().as_list()[-1]])
np.save(
'./fixed_samples/np_constant_prior_sample_' +
str(self.prior_latent_code.get_shape().as_list()[-1]) + '.npz',
np_constant_prior_sample)
self.constant_prior_latent_code = tf.constant(
np.asarray(np_constant_prior_sample), dtype=np.float32)
self.constant_prior_latent_code_expanded = tf.reshape(
self.constant_prior_latent_code, [
-1, 1,
*self.constant_prior_latent_code.get_shape().as_list()[1:]
])
self.constant_obs_sample_param = self.Generator.forward(
self.constant_prior_latent_code_expanded)
self.constant_obs_sample_dist = distributions.ProductDistribution(
sample_properties=batch['observed']['properties'],
params=self.constant_obs_sample_param)
self.constant_obs_sample = self.constant_obs_sample_dist.sample(
b_mode=True)
#############################################################################
# ENCODER
if self.config['encoder_mode'] == 'Deterministic':
self.epsilon = None
if self.config['encoder_mode'] == 'Gaussian' or self.config[
'encoder_mode'] == 'UnivApprox' or self.config[
'encoder_mode'] == 'UnivApproxNoSpatial' or self.config[
'encoder_mode'] == 'UnivApproxSine':
self.epsilon_param = self.PriorMap.forward(
(tf.zeros(shape=(self.batch_size_tf, 1)), ))
self.epsilon_dist = distributions.DiagonalGaussianDistribution(
params=self.epsilon_param)
self.epsilon = self.epsilon_dist.sample()
self.posterior_latent_code_expanded = self.Encoder.forward(
self.input_sample, noise=self.epsilon)
self.posterior_latent_code = self.posterior_latent_code_expanded[:,
0, :]
self.interpolated_posterior_latent_code = self.interpolate_latent_codes(
self.posterior_latent_code, size=self.batch_size_tf // 2)
self.interpolated_obs = self.Generator.forward(
self.interpolated_posterior_latent_code)
self.reconst_param = self.Generator.forward(
self.posterior_latent_code_expanded)
self.reconst_dist = distributions.ProductDistribution(
sample_properties=batch['observed']['properties'],
params=self.reconst_param)
self.reconst_sample = self.reconst_dist.sample(b_mode=True)
### Primal Penalty
if self.config['divergence_mode'] == 'MMD':
self.MMD = self.compute_MMD(self.posterior_latent_code,
self.prior_dist.sample())
self.enc_reg_cost = self.MMD
if self.config['divergence_mode'] == 'INV-MMD':
batch_rand_vectors = tf.random_normal(shape=[
self.config['enc_inv_MMD_n_trans'], self.config['n_latent']
])
batch_rand_dirs = batch_rand_vectors / helper.safe_tf_sqrt(
tf.reduce_sum((batch_rand_vectors**2), axis=1, keep_dims=True))
self.Inv_MMD = self.stable_div(self.compute_MMD,
self.posterior_latent_code,
batch_rand_dirs)
self.MMD = self.compute_MMD(self.posterior_latent_code,
self.prior_dist.sample())
self.enc_reg_cost = self.MMD + self.config[
'enc_inv_MMD_strength'] * self.Inv_MMD
elif self.config['divergence_mode'] == 'GAN' or self.config[
'divergence_mode'] == 'NS-GAN':
self.div_posterior = self.Diverger.forward(
self.posterior_latent_code_expanded)
self.div_prior = self.Diverger.forward(
self.prior_latent_code_expanded)
self.mean_z_divergence = tf.reduce_mean(
tf.log(10e-7 + self.div_prior)) + tf.reduce_mean(
tf.log(10e-7 + 1 - self.div_posterior))
if self.config['divergence_mode'] == 'NS-GAN':
self.enc_reg_cost = -tf.reduce_mean(
tf.log(10e-7 + self.div_posterior))
elif self.config['divergence_mode'] == 'GAN':
self.enc_reg_cost = tf.reduce_mean(
tf.log(10e-7 + 1 - self.div_posterior))
#############################################################################
# REGULARIZER
self.uniform_dist = distributions.UniformDistribution(
params=tf.concat([
tf.zeros(shape=(self.batch_size_tf, 1)),
tf.ones(shape=(self.batch_size_tf, 1))
],
axis=1))
### Visualized regularization sample
self.reg_target_dist = self.reconst_dist
self.reg_dist = None
### Uniform Sample From Geodesic of Trivial Coupling Pairs
if self.compute_all_regularizers or \
'Trivial Gradient Norm' in self.config['critic_reg_mode'] or \
'Trivial Lipschitz' in self.config['critic_reg_mode']:
self.uniform_sample_b = self.uniform_dist.sample()
self.trivial_line_sample_param = {'image': None, 'flat': None}
try:
self.uniform_sample_b_expanded = self.uniform_sample_b[:, np.
newaxis, :,
np.
newaxis,
np.
newaxis]
self.trivial_line_sample_param['image'] = self.uniform_sample_b_expanded*self.obs_sample['image']+\
(1-self.uniform_sample_b_expanded)*self.input_sample['image']
except:
self.uniform_sample_b_expanded = self.uniform_sample_b[:, np.
newaxis, :]
self.trivial_line_sample_param['flat'] = self.uniform_sample_b_expanded*self.obs_sample['flat']+\
(1-self.uniform_sample_b_expanded)*self.input_sample['flat']
self.trivial_line_dist = distributions.ProductDistribution(
sample_properties=batch['observed']['properties'],
params=self.trivial_line_sample_param)
self.trivial_line_sample = self.trivial_line_dist.sample(
b_mode=True)
if self.reg_dist is None: self.reg_dist = self.trivial_line_dist
#############################################################################
# CRITIC
self.critic_real = self.Critic.forward(self.input_sample)
self.critic_gen = self.Critic.forward(self.obs_sample)
self.critic_rec = self.Critic.forward(self.reconst_sample)
self.mean_critic_real = tf.reduce_mean(self.critic_real)
self.mean_critic_gen = tf.reduce_mean(self.critic_gen)
self.mean_critic_rec = tf.reduce_mean(self.critic_rec)
self.mean_critic_reg = None
if self.compute_all_regularizers or \
'Trivial Gradient Norm' in self.config['critic_reg_mode'] or \
'Trivial Lipschitz' in self.config['critic_reg_mode']:
self.critic_trivial_line = self.Critic.forward(
self.trivial_line_sample)
if self.mean_critic_reg is None:
self.mean_critic_reg = tf.reduce_mean(self.critic_trivial_line)
try:
self.trivial_line_grad = tf.gradients(
self.critic_trivial_line,
[self.trivial_line_sample['image']])[0]
self.trivial_line_grad_norm = tf.sqrt(
tf.reduce_sum(tf.square(self.trivial_line_grad),
axis=[-1, -2, -3],
keep_dims=False))[:, np.newaxis, :]
except:
self.trivial_line_grad = tf.gradients(
self.critic_trivial_line,
[self.trivial_line_sample['flat']])[0]
self.trivial_line_grad_norm = helper.safe_tf_sqrt(
tf.reduce_sum(self.trivial_line_grad**2,
axis=[
-1,
],
keep_dims=True))
self.cri_reg_cost = 0
if self.compute_all_regularizers or 'Trivial Gradient Norm' in self.config[
'critic_reg_mode']:
self.trivial_line_grad_norm_1_penalties = ((
self.trivial_line_grad_norm - 1.)**2)
self.mean_trivial_line_grad_norm_1_penalties = tf.reduce_mean(
self.trivial_line_grad_norm_1_penalties)
if 'Trivial Gradient Norm' in self.config['critic_reg_mode']:
print('Adding Trivial Gradient Norm Penalty')
self.cri_reg_cost += self.mean_trivial_line_grad_norm_1_penalties
if len(self.config['critic_reg_mode']) > 0:
self.cri_reg_cost /= len(self.config['critic_reg_mode'])
#############################################################################
# OBJECTIVES
### Divergence
if self.config['divergence_mode'] == 'GAN' or self.config[
'divergence_mode'] == 'NS-GAN':
self.div_cost = -self.config[
'enc_reg_strength'] * self.mean_z_divergence
### Encoder
self.OT_primal = self.sample_distance_function(self.input_sample,
self.reconst_sample)
self.mean_OT_primal = tf.reduce_mean(self.OT_primal)
# self.mean_OT_primal = helper.tf_print(self.mean_OT_primal, [self.mean_OT_primal])
self.mean_POT_primal = self.mean_OT_primal #+self.config['enc_reg_strength']*self.enc_reg_cost
self.enc_cost = self.mean_POT_primal
### Generator
if self.config['dual_dist_mode'] == 'Coupling':
self.gen_cost = -self.mean_critic_rec
elif self.config['dual_dist_mode'] == 'Prior':
self.gen_cost = -self.mean_critic_gen
elif self.config['dual_dist_mode'] == 'CouplingAndPrior':
self.gen_cost = -0.5 * (self.mean_critic_rec +
self.mean_critic_gen)
### Critic
if self.config['dual_dist_mode'] == 'Coupling':
self.mean_OT_dual = self.mean_critic_real - self.mean_critic_rec
elif self.config['dual_dist_mode'] == 'Prior':
self.mean_OT_dual = self.mean_critic_real - self.mean_critic_gen
elif self.config['dual_dist_mode'] == 'CouplingAndPrior':
self.mean_OT_dual = self.mean_critic_real - 0.5 * (
self.mean_critic_rec + self.mean_critic_gen)
self.cri_cost = -self.mean_OT_dual + self.config[
'cri_reg_strength'] * self.cri_reg_cost
def sample(self): dir_normal = tf.random.normal(shape=tf.shape(self.centers)) dir_normal_norm = helper.safe_tf_sqrt(tf.reduce_sum(dir_normal**2, axis=1, keepdims=True)) sample_dir = dir_normal/dir_normal_norm sample = self.radius*sample_dir+self.centers return sample
layer_2_rec = tf.layers.dense(inputs=layer_1_rec,
units=500,
use_bias=True,
activation=tf.nn.relu)
layer_3_rec = tf.layers.dense(inputs=layer_2_rec,
units=500,
use_bias=True,
activation=tf.nn.relu)
# layer_4_rec = tf.layers.dense(inputs = layer_3_rec, units = 500, use_bias = True, activation = tf.nn.relu)
x_rec = tf.layers.dense(inputs=layer_3_rec,
units=input_dim * tile_rate,
use_bias=True,
activation=None)
rec_cost = tf.reduce_mean(
helper.safe_tf_sqrt(tf.reduce_sum((x_rec - x_input)**2, axis=1)))
MMD = helper.compute_MMD(z, z_prior, positive_only=True)
# start, timescale = 200, 1500
start, timescale = 0, 1
lambda_z_comp = 100 * helper.hardstep(
(iter_tf - float(start)) / float(timescale))
cost = MMD + lambda_z_comp * rec_cost
optimizer = tf.train.AdamOptimizer(learning_rate=0.0001,
beta1=0.9,
beta2=0.999,
epsilon=1e-08)
cost_step = optimizer.minimize(cost)
init = tf.initialize_all_variables()