research/gam/gam/experiments/run_train_gam.py [350:433]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - model_name += '-seed_' + str(FLAGS.seed) model_name += FLAGS.experiment_suffix logging.info('Model name: %s', model_name) # Create directories for model checkpoints, summaries, and # self-labeled data backup. summary_dir = os.path.join(FLAGS.output_dir, 'summaries', FLAGS.dataset_name, model_name) checkpoints_dir = os.path.join(FLAGS.output_dir, 'checkpoints', FLAGS.dataset_name, model_name) data_dir = os.path.join(FLAGS.data_output_dir, 'data_checkpoints', FLAGS.dataset_name, model_name) if not os.path.exists(checkpoints_dir): os.makedirs(checkpoints_dir) if not os.path.exists(data_dir): os.makedirs(data_dir) ############################################################################ # MODEL SETUP # ############################################################################ # Select the model based on the provided FLAGS. model_cls = get_model_cls( model_name=FLAGS.model_cls, data=data, dataset_name=FLAGS.dataset_name, hidden=FLAGS.hidden_cls) # Create agreement model. model_agr = get_model_agr( model_name=FLAGS.model_agr, dataset_name=FLAGS.dataset_name, hidden_aggreg=FLAGS.hidden_aggreg, aggregation_agr_inputs=FLAGS.aggregation_agr_inputs, hidden=FLAGS.hidden_agr) # Train. trainer = TrainerCotraining( model_cls=model_cls, model_agr=model_agr, max_num_iter_cotrain=FLAGS.max_num_iter_cotrain, min_num_iter_cls=FLAGS.min_num_iter_cls, max_num_iter_cls=FLAGS.max_num_iter_cls, num_iter_after_best_val_cls=FLAGS.num_iter_after_best_val_cls, min_num_iter_agr=FLAGS.min_num_iter_agr, max_num_iter_agr=FLAGS.max_num_iter_agr, num_iter_after_best_val_agr=FLAGS.num_iter_after_best_val_agr, num_samples_to_label=FLAGS.num_samples_to_label, min_confidence_new_label=FLAGS.min_confidence_new_label, keep_label_proportions=FLAGS.keep_label_proportions, num_warm_up_iter_agr=FLAGS.num_warm_up_iter_agr, optimizer=tf.train.AdamOptimizer, gradient_clip=FLAGS.gradient_clip, batch_size_agr=FLAGS.batch_size_agr, batch_size_cls=FLAGS.batch_size_cls, learning_rate_cls=FLAGS.learning_rate_cls, learning_rate_agr=FLAGS.learning_rate_agr, enable_summaries=True, enable_summaries_per_model=True, summary_dir=summary_dir, summary_step_cls=FLAGS.summary_step_cls, summary_step_agr=FLAGS.summary_step_agr, logging_step_cls=FLAGS.logging_step_cls, logging_step_agr=FLAGS.logging_step_agr, eval_step_cls=FLAGS.eval_step_cls, eval_step_agr=FLAGS.eval_step_agr, checkpoints_dir=checkpoints_dir, checkpoints_step=1, data_dir=data_dir, abs_loss_chg_tol=1e-10, rel_loss_chg_tol=1e-7, loss_chg_iter_below_tol=30, use_perfect_agr=FLAGS.use_perfect_agreement, use_perfect_cls=FLAGS.use_perfect_classifier, warm_start_cls=FLAGS.warm_start_cls, warm_start_agr=FLAGS.warm_start_agr, ratio_valid_agr=FLAGS.ratio_valid_agr, max_samples_valid_agr=FLAGS.max_samples_valid_agr, weight_decay_cls=FLAGS.weight_decay_cls, weight_decay_schedule_cls=FLAGS.weight_decay_schedule_cls, weight_decay_schedule_agr=FLAGS.weight_decay_schedule_agr, weight_decay_agr=FLAGS.weight_decay_agr, reg_weight_ll=FLAGS.reg_weight_ll, reg_weight_lu=FLAGS.reg_weight_lu, reg_weight_uu=FLAGS.reg_weight_uu, - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - research/gam/gam/experiments/run_train_gam_graph.py [321:404]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - model_name += '-seed_' + str(FLAGS.seed) model_name += FLAGS.experiment_suffix logging.info('Model name: %s', model_name) # Create directories for model checkpoints, summaries, and # self-labeled data backup. summary_dir = os.path.join(FLAGS.output_dir, 'summaries', FLAGS.dataset_name, model_name) checkpoints_dir = os.path.join(FLAGS.output_dir, 'checkpoints', FLAGS.dataset_name, model_name) data_dir = os.path.join(FLAGS.data_output_dir, 'data_checkpoints', FLAGS.dataset_name, model_name) if not os.path.exists(checkpoints_dir): os.makedirs(checkpoints_dir) if not os.path.exists(data_dir): os.makedirs(data_dir) ############################################################################ # MODEL SETUP # ############################################################################ # Create classification model. model_cls = get_model_cls( model_name=FLAGS.model_cls, data=data, dataset_name=FLAGS.dataset_name, hidden=FLAGS.hidden_cls) # Create agreement model. model_agr = get_model_agr( model_name=FLAGS.model_agr, dataset_name=FLAGS.dataset_name, hidden_aggreg=FLAGS.hidden_aggreg, aggregation_agr_inputs=FLAGS.aggregation_agr_inputs, hidden=FLAGS.hidden_agr) # Train. trainer = TrainerCotraining( model_cls=model_cls, model_agr=model_agr, max_num_iter_cotrain=FLAGS.max_num_iter_cotrain, min_num_iter_cls=FLAGS.min_num_iter_cls, max_num_iter_cls=FLAGS.max_num_iter_cls, num_iter_after_best_val_cls=FLAGS.num_iter_after_best_val_cls, min_num_iter_agr=FLAGS.min_num_iter_agr, max_num_iter_agr=FLAGS.max_num_iter_agr, num_iter_after_best_val_agr=FLAGS.num_iter_after_best_val_agr, num_samples_to_label=FLAGS.num_samples_to_label, min_confidence_new_label=FLAGS.min_confidence_new_label, keep_label_proportions=FLAGS.keep_label_proportions, num_warm_up_iter_agr=FLAGS.num_warm_up_iter_agr, optimizer=tf.train.AdamOptimizer, gradient_clip=FLAGS.gradient_clip, batch_size_agr=FLAGS.batch_size_agr, batch_size_cls=FLAGS.batch_size_cls, learning_rate_cls=FLAGS.learning_rate_cls, learning_rate_agr=FLAGS.learning_rate_agr, enable_summaries=True, enable_summaries_per_model=True, summary_dir=summary_dir, summary_step_cls=FLAGS.summary_step_cls, summary_step_agr=FLAGS.summary_step_agr, logging_step_cls=FLAGS.logging_step_cls, logging_step_agr=FLAGS.logging_step_agr, eval_step_cls=FLAGS.eval_step_cls, eval_step_agr=FLAGS.eval_step_agr, checkpoints_dir=checkpoints_dir, checkpoints_step=1, data_dir=data_dir, abs_loss_chg_tol=1e-10, rel_loss_chg_tol=1e-7, loss_chg_iter_below_tol=30, use_perfect_agr=FLAGS.use_perfect_agreement, use_perfect_cls=FLAGS.use_perfect_classifier, warm_start_cls=FLAGS.warm_start_cls, warm_start_agr=FLAGS.warm_start_agr, ratio_valid_agr=FLAGS.ratio_valid_agr, max_samples_valid_agr=FLAGS.max_samples_valid_agr, weight_decay_cls=FLAGS.weight_decay_cls, weight_decay_schedule_cls=FLAGS.weight_decay_schedule_cls, weight_decay_schedule_agr=FLAGS.weight_decay_schedule_agr, weight_decay_agr=FLAGS.weight_decay_agr, reg_weight_ll=FLAGS.reg_weight_ll, reg_weight_lu=FLAGS.reg_weight_lu, reg_weight_uu=FLAGS.reg_weight_uu, - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -