# os.symlink(os.path.abspath(temp_dir), os.path.abspath(global_args.exp_dir)) # global_args.exp_dir = temp_dir _, _, batch = next(data_loader) try: fixed_batch_data = batch['observed']['data']['image'].copy() except: pass with tf.Graph().as_default(): tf.set_random_seed(global_args.seed) model = Model(vars(global_args)) global_step = tf.Variable(0.0, name='global_step', trainable=False) with tf.variable_scope("training"): tf.set_random_seed(global_args.seed) additional_inputs_tf = tf.placeholder(tf.float32, [2]) batch_tf, input_dict_func = helper.tf_batch_and_input_dict(batch, additional_inputs_tf) train_outs_dict, test_outs_dict = model.inference(batch_tf, additional_inputs_tf) generative_dict = model.generative_model(batch_tf) inference_obs_dist = model.obs_dist transport_dist = model.transport_dist rec_dist = model.rec_dist discriminator_vars = [v for v in tf.trainable_variables() if 'Discriminator' in v.name] generator_vars = [v for v in tf.trainable_variables() if 'Decoder' in v.name] transport_vars = [v for v in tf.trainable_variables() if 'TransportPlan' in v.name or 'EncodingPlan' in v.name or 'MixingPlan' in v.name] # Weight clipping discriminator_vars_flat_concat = tf.concat([tf.reshape(e, [-1]) for e in discriminator_vars], axis=0) max_abs_discriminator_vars = tf.reduce_max(tf.abs(discriminator_vars_flat_concat)) clip_op_list = [] for e in discriminator_vars:
print("TENSORBOARD: Mac:\nhttp://0.0.0.0:" + str(20000 + int(global_args.exp_dir[-4:-1], 16))) print("\n\n\n") shutil.copyfile('./models/SLVM.py', global_args.exp_dir + 'SLVM.py') shutil.copyfile('./models/ModelGTM.py', global_args.exp_dir + 'ModelGTM.py') _, _, batch = next(data_loader) with tf.Graph().as_default(): tf.set_random_seed(global_args.seed) model = SLVM(vars(global_args)) global_step = tf.Variable(0.0, name='global_step', trainable=False) with tf.variable_scope("training"): tf.set_random_seed(global_args.seed) batch_tf, input_dict_func = helper.tf_batch_and_input_dict(batch) train_out_list, test_out_list = model.inference(batch_tf, global_step) batch_loss_tf = train_out_list[0] obs_dist = model.obs_dist sample_obs_dist, obs_sample_out_tf, latent_sample_out_tf = model.generative_model( batch_tf) if global_args.optimizer_class == 'RmsProp': train_step_tf = tf.train.RMSPropOptimizer( learning_rate=global_args.learning_rate, momentum=0.9).minimize(batch_loss_tf, global_step=global_step) elif global_args.optimizer_class == 'Adam': train_step_tf = tf.train.AdamOptimizer( learning_rate=global_args.learning_rate, beta1=0.9,