def forward(self, data_src, data_tgt=0): loss = 0 data_src_feature = self.sharedNet(data_src) # data_src_feature = self.DSFE(data_src_feature) if self.training == True: data_tgt_feature = self.sharedNet(data_tgt) loss += utils.mmd(data_src_feature, data_tgt_feature) data_src_cls = self.cls_fc(data_src_feature) return data_src_cls, loss, data_src_feature
def loss_function(recon_x1, recon_x2, x1, x2, mu1, mu2, logvar1, logvar2, VTlabel, z1, z2): x1 = x1.float() x2 = x2.float() reconLoss = (recon_loss(recon_x1, x1) + recon_loss(recon_x2, x2)) / mb_size KLLoss = torch.mean(0.5 * torch.sum( torch.exp(logvar1) + mu1**2 - 1. - logvar1, 1)) + torch.mean( 0.5 * torch.sum(torch.exp(logvar2) + mu2**2 - 1. - logvar2, 1)) #mmd x1x1, x1x2, x2x2 = mmd(z1, z2, 1.0) labels = (VTlabel - 1) / (-2) mmdLoss = (torch.mean(x1x1 * labels) - 2 * torch.mean(x1x2 * labels) + torch.mean(x2x2 * labels)) return reconLoss, KLLoss, mmdLoss
def loss_actor(self, o): o2 = o.repeat(self.expand_batch, 1) a2 = self.actor_critic.actor(o2) q1 = self.actor_critic.critic1(o2, a2) q2 = self.actor_critic.critic2(o2, a2) a2 = a2.view(self.expand_batch, -1, a2.shape[-1]).transpose(0, 1) with torch.no_grad(): a3 = (2 * torch.rand_like(a2) - 1) mmd_entropy = mmd(a2, a3, kernel=self.kernel) # Entropy-regularized policy loss loss = -torch.min(q1, q2).mean() + self.alpha * mmd_entropy return loss, mmd_entropy.detach()
def main(unused_argv): # Data generating process parameters exp_parameter_grid = { 'model': ["dlvm", "lrmf"] if FLAGS.model is None else [FLAGS.model], 'citcio': [ False, ], 'n': [1000, 5000, 10000] if FLAGS.n_observations is None else [FLAGS.n_observations], 'p': [10, 50] if FLAGS.p_ambient is None else [FLAGS.p_ambient], 'y_snr': [5.] if FLAGS.y_snr is None else [FLAGS.y_snr], 'x_snr': [2.] if FLAGS.x_snr is None else [FLAGS.x_snr], 'mu_z': [0.] if FLAGS.mu_z is None else [FLAGS.mu_z], 'sig_z': [1.] if FLAGS.sig_z is None else [FLAGS.sig_z], 'sig_xgivenz': [0.001] if FLAGS.sig_xgivenz is None else [FLAGS.sig_xgivenz], 'prop_miss': [ 0.0, ] if FLAGS.prop_miss is None else [FLAGS.prop_miss], 'regularize': [False] if FLAGS.regularize is None else [FLAGS.regularize], 'seed': np.arange(FLAGS.n_seeds), } range_d_over_p = [ 0.002, 0.01, 0.1 ] if FLAGS.d_over_p is None and FLAGS.d_latent is None else [ FLAGS.d_over_p ] range_d = None if range_d_over_p is not None and FLAGS.d_latent is None else FLAGS.d_latent # MDC parameters range_d_offset = [0, 5, 10] if FLAGS.miwae_d_offset is None else [ FLAGS.miwae_d_offset ] mdc_parameter_grid = { 'mu_prior': [0.] if FLAGS.miwae_mu_prior is None else [FLAGS.miwae_mu_prior], 'sig_prior': [1.] if FLAGS.miwae_sig_prior is None else [FLAGS.miwae_sig_prior], 'num_samples_zmul': [500] if FLAGS.miwae_n_samples_zmul is None else [FLAGS.miwae_n_samples_zmul], 'learning_rate': [ 0.0001, ] if FLAGS.miwae_learning_rate is None else [FLAGS.miwae_learning_rate], 'n_epochs': [ 5000, ] if FLAGS.miwae_n_epochs is None else [FLAGS.miwae_n_epochs], } test_seeds = np.arange(FLAGS.n_test_seeds) + 1000 save_test_data = True if FLAGS.save_test_data is None else FLAGS.save_test_data # Experiment and output file name output = f'results/{FLAGS.exp_name}.csv' if FLAGS.output is None else FLAGS.output FLAGS.log_dir = './sessions/logging/' if FLAGS.log_path is None else FLAGS.log_path logging.get_absl_handler().use_absl_log_file() logging.info('*' * 20) logging.info(f'Starting exp: {FLAGS.exp_name}') logging.info('*' * 20) exp_arguments = [ dict(zip(exp_parameter_grid.keys(), vals)) for vals in itertools.product(*exp_parameter_grid.values()) ] previous_runs = set() if tf.io.gfile.exists(output): with tf.io.gfile.GFile(output, mode='r') as f: reader = csv.DictReader(f) for row in reader: # Note: we need to do this conversion because DictReader creates an # OrderedDict, and reads all values as str instead of bool or int. previous_runs.add( str({ 'model': row['model'], 'citcio': row['citcio'] == 'True', 'n': int(row['n']), 'p': int(row['p']), 'y_snr': float(row['y_snr']), 'x_snr': float(row['x_snr']), 'mu_z': float(row['mu_z']), 'sig_z': float(row['sig_z']), 'prop_miss': float(row['prop_miss']), 'regularize': row['regularize'] == 'True', 'seed': int(row['seed']), 'd': int(row['d']), 'sig_xgivenz': float(row['sig_xgivenz']) })) logging.info('Previous runs') logging.info(previous_runs) for args in exp_arguments: # For given p, if range_d is not yet specified, # create range for d such that 1 < d < p # starting with given ratios for d/p if range_d is None: range_d = [ np.maximum(2, int(np.floor(args['p'] * x))) for x in range_d_over_p ] range_d = np.unique( np.array(range_d)[np.array(range_d) < args['p']].tolist()) exp_time = time.time() for args['d'] in range_d: # We only consider cases where latent dimension <= ambient dimension if args['d'] > args['p']: continue res = [] if str(args) in previous_runs: logging.info(f'Skipped {args}') continue else: logging.info(f'running exp with {args}') if args['model'] == "lrmf": Z, X, w, y, ps, mu0, mu1 = gen_lrmf( n=args['n'], d=args['d'], p=args['p'], y_snr=args['y_snr'], x_snr=args['x_snr'], citcio=args['citcio'], prop_miss=args['prop_miss'], seed=args['seed']) elif args['model'] == "dlvm": Z, X, w, y, ps, mu0, mu1 = gen_dlvm( n=args['n'], d=args['d'], p=args['p'], y_snr=args['y_snr'], citcio=args['citcio'], prop_miss=args['prop_miss'], seed=args['seed'], mu_z=args['mu_z'], sig_z=args['sig_z'], x_snr=args['x_snr'], sig_xgivenz=args['sig_xgivenz']) X_miss = ampute(X, prop_miss=args['prop_miss'], seed=args['seed']) # MIWAE mdc_parameter_grid['d_miwae'] = [ args['d'] + x for x in range_d_offset ] mdc_arguments = [ dict(zip(mdc_parameter_grid.keys(), vals)) for vals in itertools.product(*mdc_parameter_grid.values()) ] for mdc_arg in mdc_arguments: t0 = time.time() mdc_arg['mu_prior'] = args['mu_z'] session_file = './sessions/' + \ args['model'] + '_'+ \ '_sigXgivenZ' + str(args['sig_xgivenz']) + \ '_n' + str(args['n']) + \ '_p' + str(args['p']) + \ '_d' + str(args['d']) + \ '_ysnr' + str(args['y_snr']) +\ '_xsnr' + str(args['x_snr']) +\ '_propNA' + str(args['prop_miss']) + \ '_seed' + str(args['seed']) session_file_complete = session_file + \ '_dmiwae' + str(mdc_arg['d_miwae']) + \ '_sigprior' + str(mdc_arg['sig_prior']) epochs = -1 tmp = glob.glob(session_file_complete + '.*') sess = tf.Session(graph=tf.reset_default_graph()) if len(tmp) > 0: new_saver = tf.train.import_meta_graph( session_file_complete + '.meta') new_saver.restore(sess, session_file_complete) #with open(session_file_complete+'.pkl', 'rb') as f: # xhat, zhat, zhat_mul, elbo, epochs = pickle.load(f) else: xhat, zhat, zhat_mul, elbo, epochs = miwae_es( X_miss, d_miwae=mdc_arg['d_miwae'], mu_prior=mdc_arg['mu_prior'], sig_prior=mdc_arg['sig_prior'], num_samples_zmul=mdc_arg['num_samples_zmul'], l_rate=mdc_arg['learning_rate'], n_epochs=mdc_arg['n_epochs'], save_session=True, session_file=session_file) new_saver = tf.train.import_meta_graph( session_file_complete + '.meta') new_saver.restore(sess, session_file_complete ) #tf.train.latest_checkpoint('./')) with open(session_file_complete + '.pkl', 'wb') as file_data: # Python 3: open(..., 'wb') pickle.dump([xhat, zhat, zhat_mul, elbo, epochs], file_data) args['training_time'] = int(time.time() - t0) # Evaluate performance of trained model on new testsets graph = tf.get_default_graph() K = graph.get_tensor_by_name('K:0') x = graph.get_tensor_by_name('x:0') batch_size = tf.shape(x)[0] xms = graph.get_tensor_by_name('xms:0') imp_weights = graph.get_tensor_by_name('imp_weights:0') xm = tf.einsum('ki,kij->ij', imp_weights, xms, name='xm') zgivenx_flat = graph.get_tensor_by_name('zgivenx_flat:0') zgivenx = tf.reshape(zgivenx_flat, [K, batch_size, zgivenx_flat.shape[1]]) z_hat = tf.einsum('ki,kij->ij', imp_weights, zgivenx, name='z_hat') sir_logits = graph.get_tensor_by_name('sir_logits:0') sirz = tfd.Categorical(logits=sir_logits).sample( mdc_arg['num_samples_zmul']) zmul = graph.get_tensor_by_name('zmul:0') for test_seed in test_seeds: if args['model'] == "lrmf": (Z_test, X_test, w_test, y_test, ps_test, mu0_test, mu1_test) = gen_lrmf(n=args['n'], d=args['d'], p=args['p'], y_snr=args['y_snr'], citcio=args['citcio'], prop_miss=args['prop_miss'], seed=test_seed) elif args['model'] == "dlvm": ( Z_test, X_test, w_test, y_test, ps_test, mu0_test, mu1_test ) = gen_dlvm( n=args['n'], d=args['d'], p=args['p'], y_snr=args['y_snr'], citcio=args['citcio'], prop_miss=args[ 'prop_miss'], # this argument is only used if citcio=True seed=test_seed, mu_z=args['mu_z'], sig_z=args['sig_z'], x_snr=args['x_snr'], sig_xgivenz=args['sig_xgivenz']) X_miss_test = ampute(X_test, prop_miss=args['prop_miss'], seed=args['seed']) mask_test = np.isfinite( X_miss_test ) # binary mask that indicates which values are missing t0 = time.time() tmp_elm_pkl = glob.glob(session_file_complete + '_testset_eval' + str(test_seed) + '.pkl') if len(tmp_elm_pkl) > 0: with open( session_file_complete + '_testset_eval' + str(test_seed) + '.pkl', 'rb') as f: xhat_test, zhat_test, zgivenx_test, zhat_mul_test = pickle.load( f) else: x_test_imp0 = np.copy(X_miss_test) x_test_imp0[np.isnan(X_miss_test)] = 0 n_test = X_test.shape[0] xhat_test = np.copy(x_test_imp0) zhat_test = np.zeros([n_test, mdc_arg['d_miwae']]) zgivenx_test = np.tile( zhat_test, [mdc_arg['num_samples_zmul'], 1, 1]) zhat_mul_test = np.tile( zhat_test, [mdc_arg['num_samples_zmul'], 1, 1]) for i in range(n_test): zgivenx_test[:, i, :] = np.squeeze( zgivenx.eval(session=sess, feed_dict={ 'x:0': x_test_imp0[i, :].reshape( [1, args['p']]), 'K:0': mdc_arg['num_samples_zmul'], 'xmask:0': mask_test[i, :].reshape( [1, args['p']]) })).reshape([ mdc_arg['num_samples_zmul'], mdc_arg['d_miwae'] ]) xhat_test[i, :] = xm.eval( session=sess, feed_dict={ 'x:0': x_test_imp0[i, :].reshape([1, args['p']]), 'K:0': 10000, 'xmask:0': mask_test[i, :].reshape([1, args['p']]) }) zhat_test[i, :] = z_hat.eval( session=sess, feed_dict={ 'x:0': x_test_imp0[i, :].reshape([1, args['p']]), 'K:0': 10000, 'xmask:0': mask_test[i, :].reshape([1, args['p']]) }) si, zmu = sess.run( [sirz, zmul], feed_dict={ 'x:0': x_test_imp0[i, :].reshape([1, args['p']]), 'K:0': 10000, 'xmask:0': mask_test[i, :].reshape([1, args['p']]) }) zhat_mul_test[:, i, :] = np.squeeze( zmu[si, :, :]).reshape( (mdc_arg['num_samples_zmul'], mdc_arg['d_miwae'])) if save_test_data: with open( session_file_complete + '_testset_eval' + str(test_seed) + '.pkl', 'wb' ) as file_data: # Python 3: open(..., 'wb') pickle.dump([ xhat_test, zhat_test, zgivenx_test, zhat_mul_test ], file_data) evaluation_time = int(time.time() - t0) if args['d'] == 1 and mdc_arg['d_miwae'] == 1: row = { 'Z_cor': pearsonr(Z_test.reshape([ args['n'], ]), zhat_test.reshape([ args['n'], ]))[0] } else: row = {'Z_cor': np.NaN} if args['d'] == mdc_arg['d_miwae']: row.update({'Z_mmd': mmd(Z_test, zhat_test, beta=1.)}) row.update({'Z_rvcoef': compute_rv(Z_test, zhat_test)}) else: row.update({'Z_mmd': np.NaN}) row.update({'Z_rvcoef': np.NaN}) row.update( {'X_mse': mean_squared_error(X_test, xhat_test)}) row.update({'X_mmd': mmd(X_test, xhat_test, beta=1.)}) row.update({'X_rvcoef': compute_rv(X_test, xhat_test)}) row.update(args) row.update(mdc_arg) row.update({'epochs': epochs}) row.update({'test_seed': test_seed}) row.update({'evaluation_time': evaluation_time}) res.append(row) log_res( output, res, l_metrics + list(args.keys()) + list(mdc_arg.keys()) + ['epochs', 'test_seed', 'evaluation_time']) logging.info('........... DONE') logging.info(f'in {time.time() - exp_time} s \n\n') logging.info('*' * 20) logging.info(f'Exp: {FLAGS.exp_name} succesfully ended.') logging.info('*' * 20)