'num_hid': 20, 'num_hid_dec': 100, 'output_dist': 'Bernoulli', # Bernoulli or Gaussian # 'output_dist': 'Gaussian', 'non_linear': 'tf.nn.relu', # 'non_linear': 'tf.nn.tanh', 'weight_decay': 5e-5 } # Train loop options loop_config = { 'num_steps': args.num_steps, 'steps_per_ckpt': args.steps_per_ckpt } dataset = mnist.read_data_sets("../MNIST_data/", one_hot=True) m = get_train_model(opt) sess = tf.Session() sess.run(tf.initialize_all_variables()) saver = tf.train.Saver(tf.all_variables()) task_name = 'vae_mnist_half' time_obj = datetime.datetime.now() model_id = timestr = '{}-{:04d}{:02d}{:02d}{:02d}{:02d}{:02d}'.format( task_name, time_obj.year, time_obj.month, time_obj.day, time_obj.hour, time_obj.minute, time_obj.second) results_folder = args.results logs_folder = args.logs exp_folder = os.path.join(results_folder, model_id) exp_logs_folder = os.path.join(logs_folder, model_id)
filter_x = (1 / np.sqrt(np.exp(lg_var)) / np.sqrt(2 * np.pi) * np.exp(-0.5 * (span_x - mu_x) * (span_x - mu_x) / np.exp(lg_var))) # [H, F] filter_y = (1 / np.sqrt(np.exp(lg_var)) / np.sqrt(2 * np.pi) * np.exp(-0.5 * (span_y - mu_y) * (span_y - mu_y) / np.exp(lg_var))) read = filter_y.transpose().dot(img).dot(filter_x) return read if not os.path.exists('img.npy'): from data_api import mnist dataset = mnist.read_data_sets("../MNIST_data/", one_hot=True) img = dataset.train.images[0].reshape([im_height, im_width]) np.save('img.npy', img) else: img = np.load('img.npy') # Controller variables. num_img = 8 f, axarr = plt.subplots(2, num_img) for ii in xrange(num_img): gg_x = 0.0 gg_y = 0.0 ddelta = 0.1 * (ii + 1) log.info('ddelta: {}'.format(ddelta)) lg_var = 0.01