def run_model_duration(model_name, participator, timesteps, stride, nb_epoch, load_weight_from = None): logger_name = model_name + str(participator) + str(timesteps) + str(stride) + str(nb_epoch) + str(load_weight_from) logger = logging.getLogger(logger_name) # so that no multiple loggers input the same data f_time = datetime.datetime.today() output_dir = os.path.join('output', 'dur_'+str(f_time)) if not os.path.exists(output_dir): os.makedirs(output_dir) hdlr = logging.FileHandler(os.path.join(output_dir, 'rnn.log')) logger.addHandler(hdlr) console_handler = logging.StreamHandler() logger.addHandler(console_handler) logger.setLevel(logging.INFO) gal = GAL_data() gal.set_logger(logger) gal.load_data(load_list=['eeg', 'info']) data_description = gal.get_data_description() participator = participator logger.info('participator : {0}'.format(participator)) #event_list=['Dur_Reach', 'Dur_Preload', 'Dur_LoadPhase', 'Dur_Release', 'Dur_Retract'] event_list=['Dur_Reach', 'Dur_LoadReach', 'Dur_LoadMaintain', 'Dur_LoadRetract', 'Dur_Retract'] rnn = EEG_model(event_list) rnn.set_logger(logger) rnn.select_model(model_name) if load_weight_from != None: rnn.load_model_weight(model_name, load_weight_from) logger.info( 'running model data as a whole') data_split_ratio = [0.8, 0.1, 0.1] data = gal.data_event(part=participator, timesteps=timesteps, stride=stride, event_list=event_list, partition_ratio=data_split_ratio, input_dim=32) loss_train, loss_val, loss_test = rnn.run_model_event(data=data, nb_epoch = nb_epoch) with open(os.path.join(output_dir, 'train_loss.json'), 'w') as f: json.dump(loss_train, f) with open(os.path.join(output_dir, 'validate_loss.json'), 'w') as f: json.dump(loss_val, f) with open(os.path.join(output_dir, 'test_loss.json'), 'w') as f: json.dump(loss_test, f) rnn.set_data_description(data_description) rnn.set_model_config('epoch', nb_epoch) generator = gal.data_generator_event(part=participator, timesteps=timesteps, stride=stride, event_list=event_list) rnn.save_event(data=data, event_list=event_list, output_dir=output_dir)
def run_model_duration_classify(model_name, participator, timesteps, stride, max_freq, min_freq, nb_epoch, batch_size, save_by, multi_filter = False, load_weight_from = None, load_data_from= None): logger_name = model_name + str(participator) + str(timesteps) + str(stride) + str(nb_epoch) + str(load_weight_from) logger = logging.getLogger(logger_name) # so that no multiple loggers input the same data f_time = datetime.datetime.today() if load_weight_from != None: load_weight_from_str = load_weight_from.split('/') output_dir = os.path.join('output', '{0}_P{1}_ts{2}_stride{3}_ep{4}_bs_{5}_weight_{6}_maxmin_{7}_saveby_{8}'.format(model_name, participator, timesteps, stride, nb_epoch, batch_size, load_weight_from_str, '{}_{}'.format(max_freq, min_freq), save_by)) else: output_dir = os.path.join('output', '{0}_P{1}_ts{2}_stride{3}_ep{4}_bs_{5}_maxmin_{6}_saveby_{7}'.format(model_name, participator, timesteps, stride, nb_epoch, batch_size, '{}_{}'.format(max_freq, min_freq), save_by)) if not os.path.exists(output_dir): os.makedirs(output_dir) else: print('same configuation already exists!') return if load_data_from != None: assert timesteps == load_data_from[0], 'timesteps does not match' assert stride == load_data_from[1], 'stride does not match' hdlr = logging.FileHandler(os.path.join(output_dir, 'rnn.log')) logger.addHandler(hdlr) console_handler = logging.StreamHandler() logger.addHandler(console_handler) logger.setLevel(logging.INFO) event_list=['Idle', 'Reach_Phase', 'LoadReach_Phase', 'LoadMaintain_Phase', 'LoadRetract_Phase', 'Retract_Phase'] rnn = EEG_model(event_list) rnn.set_logger(logger) rnn.select_model(model_name) if load_weight_from != None: rnn.load_model_weight(model_name, load_weight_from) logger.info('loaded weight') logger.info( 'running model data') data_split_ratio = [0.7, 0.2, 0.1] if load_data_from == None: gal = GAL_data() gal.set_logger(logger) gal.load_data(load_list=['eeg', 'info', 'kin']) if multi_filter ==True: gal.preprocess_filter_multiple() else: gal.preprocess_filter(max_freq=max_freq, min_freq=min_freq, low_pass=False) data_description = gal.get_data_description() participator = participator logger.info('participator : {0}'.format(participator)) data = gal.data_event_classify(part=participator, timesteps=timesteps, stride=stride, event_list=event_list, partition_ratio=data_split_ratio, input_dim=32) else: data_description = dict() data_description['participator'] = participator data_description['timesteps'] = timesteps data_description['stride'] = stride data_description['event_list'] = event_list data_description['preprocess_filter'] = '{}_{}'.format(max_freq, min_freq) ts = load_data_from[0] st = load_data_from[1] data = list() data.append(np.load(os.path.join('data', 'numpy_binary', 'maxfreq_{}_minfreq_{}'.format(max_freq, min_freq), 'train_X_ts_{}_st_{}.npy'.format(ts, st)))) data.append(np.load(os.path.join('data', 'numpy_binary', 'maxfreq_{}_minfreq_{}'.format(max_freq, min_freq), 'train_y_ts_{}_st_{}.npy'.format(ts, st)))) data.append(np.load(os.path.join('data', 'numpy_binary', 'maxfreq_{}_minfreq_{}'.format(max_freq, min_freq), 'validation_X_ts_{}_st_{}.npy'.format(ts, st)))) data.append(np.load(os.path.join('data', 'numpy_binary', 'maxfreq_{}_minfreq_{}'.format(max_freq, min_freq), 'validation_y_ts_{}_st_{}.npy'.format(ts, st)))) data.append(np.load(os.path.join('data', 'numpy_binary', 'maxfreq_{}_minfreq_{}'.format(max_freq, min_freq), 'test_X_ts_{}_st_{}.npy'.format(ts, st)))) data.append(np.load(os.path.join('data', 'numpy_binary', 'maxfreq_{}_minfreq_{}'.format(max_freq, min_freq), 'test_y_ts_{}_st_{}.npy'.format(ts, st)))) for i in range(nb_epoch/save_by): loss_train, loss_test = rnn.run_model_event(data=data, nb_epoch = save_by, batch_size=batch_size) output_dir_temp = os.path.join(output_dir, str(i)) os.makedirs(output_dir_temp) df_loss = pd.DataFrame() df_loss['acc'] = loss_train.history['acc'] df_loss['loss'] = loss_train.history['loss'] df_loss.to_csv(os.path.join(output_dir_temp, 'train_loss_acc.csv')) rnn.set_data_description(data_description) rnn.set_model_config('epoch', nb_epoch/save_by * i) rnn.save_event_classify(data=data, event_list=event_list, output_dir=output_dir_temp)