# print([u.subject_id for u in train_subject_instances]) # print([u.subject_id for u in val_subject_instances]) #utility.diagnose_training_subjects(all_subject_instances) #data_generators.diagnose_generator_multiple_signal(train_gen , sig_type_source) #data_generators.diagnose_generator_multiple_signal(val_gen , sig_type_source) #create model if model_type == 'Unetxl': sig_model = network_models.Unet_xl(input_size, kernel_size, filter_number, len(sig_type_source), no_layers) #make model parallel sig_model = nn.DataParallel(sig_model.cuda(), device_ids=[0, 1]) #loss function is negative pearson loss loss = network_models.PearsonRLoss() #training configs cudnn.benchmark = True config['n_epochs'] = 400 config['scheduler_milestones'] = [30, 60, 120] #[50,100,200] config['train_steps'] = 10 config['val_steps'] = 10 config['initial_lr'] = 0.001 config['model_path'] = directory + '/Code Output/best_so_far.pt' config[ 'model_path_for_video'] = directory + '/Models for Video/' + file_name_pre args = { 'lr': config['initial_lr'], 'n_epochs': config['n_epochs'], 'model_path': config['model_path'],
def test_model(gen, sig_model): ''' function tests a model by running it over the test data and calculating error metrics :param gen: test generator :param sig_model: model :return: list_pearson_r_loss: list of pearson correlations between target and estimated segments :return: list_subject_ids: subject ID for each segment :return: list_i_errors: list of R-I interval errors between each target and estimated signal segment :return: list_j_errors: list of R-J interval errors between each target and estimated signal segment :return: list_k_errors: list of R-K interval errors between each target and estimated signal segment ''' criterion = network_models.PearsonRLoss() with torch.no_grad(): finished = False list_pearson_r_loss = [] list_i_errors = [] list_j_errors = [] list_k_errors = [] list_subject_ids = [] list_noise_var_target = [] list_noise_var_estimate = [] list_target_i_points = [] list_target_j_points = [] list_target_k_points = [] list_sdr = [] while not finished: print('Running Testing') torch.cuda.empty_cache() finished, X_batch, Y_batch, subject_id_list= next(gen) if finished: print('Generator Finished') else: #get ecg ecg = X_batch[:, -1, :] #calculate pearson correlation X_batch = torch.from_numpy(X_batch[:, :-1, :]).contiguous() X_batch = network_models.cuda(X_batch) X_batch = X_batch.type(torch.cuda.FloatTensor) Y_batch_predicted= sig_model.forward(X_batch).squeeze() Y_batch = torch.from_numpy(Y_batch) Y_batch = network_models.cuda(Y_batch) Y_batch = Y_batch.type(torch.cuda.FloatTensor) Y_batch = Y_batch.squeeze() if len(Y_batch.size())==1: Y_batch=Y_batch.view(1,-1) if len(Y_batch_predicted.size())==1: Y_batch_predicted=Y_batch_predicted.view(1,-1) loss = criterion.get_induvidual_losses(Y_batch_predicted, Y_batch) list_pearson_r_loss+=loss.cpu().numpy().reshape(-1).tolist() list_subject_ids+=subject_id_list #ijk points Y_batch_predicted = Y_batch_predicted.detach().cpu().numpy() Y_batch = Y_batch.detach().cpu().numpy() for v in range(Y_batch.shape[0]): r_peaks = signal_processing_modules.get_R_peaks(ecg[v, :]) ensemble_avg_target, ensemble_beats_target = signal_processing_modules.get_ensemble_avg(r_peaks, (Y_batch[v, :] - np.mean(Y_batch[v, :]) )/( np.sqrt(np.sum(np.power( Y_batch[v, :] - np.mean(Y_batch[v, :]) ,2)))) , n_samples=500,upsample_factor=1) i_point_target, j_point_target, k_point_target = signal_processing_modules.get_IJK_peaks(ensemble_avg_target, upsample_factor=1) ensemble_avg_estimate, ensemble_beats_estimate = signal_processing_modules.get_ensemble_avg(r_peaks, ( Y_batch_predicted[v,:]-np.mean(Y_batch_predicted[v,:]) )/(np.sqrt(np.sum(np.power( Y_batch_predicted[v,:]-np.mean(Y_batch_predicted[v,:]) , 2 )))) , n_samples=500, upsample_factor=1) i_point_estimate, j_point_estimate, k_point_estimate = signal_processing_modules.get_IJK_peaks(ensemble_avg_estimate, upsample_factor=1) i_error = 1000 * np.abs(i_point_target - i_point_estimate) / (500) if i_point_target != -1 and i_point_estimate != -1 else -1 # 500 is the sampling rate of the signal segments, 1000* for miliseconds j_error = 1000 * np.abs(j_point_target - j_point_estimate) / (500) if j_point_target != -1 and j_point_estimate != -1 else -1 k_error = 1000 * np.abs(k_point_target - k_point_estimate) / (500) if k_point_target != -1 and k_point_estimate != -1 else -1 list_i_errors.append(i_error) list_j_errors.append(j_error) list_k_errors.append(k_error) list_noise_var_target.append( signal_processing_modules.get_noise_variance(ensemble_avg_target, ensemble_beats_target) ) list_noise_var_estimate.append( signal_processing_modules.get_noise_variance(ensemble_avg_estimate, ensemble_beats_estimate ) ) list_target_i_points.append(i_point_target) list_target_j_points.append(j_point_target) list_target_k_points.append(k_point_target) list_sdr.append(signal_processing_modules.get_sdr(( Y_batch_predicted[v,:]-np.mean(Y_batch_predicted[v,:]) )/(np.sqrt(np.sum(np.power( Y_batch_predicted[v,:]-np.mean(Y_batch_predicted[v,:]) , 2 )))) , (Y_batch[v, :] - np.mean(Y_batch[v, :]) )/( np.sqrt(np.sum(np.power( Y_batch[v, :] - np.mean(Y_batch[v, :]) ,2)))) )) del X_batch del Y_batch_predicted del Y_batch return list_pearson_r_loss, list_subject_ids, list_i_errors, list_j_errors, list_k_errors, list_noise_var_target, list_noise_var_estimate, list_target_i_points, list_target_j_points, list_target_k_points, list_sdr
def main(): #sampling rate of training dataset F_SAMPLING=2000 #config parser = argparse.ArgumentParser() arg = parser.add_argument arg('--n_f', type=int, help='n_f') arg('--L', type=int , help='L') args = vars(parser.parse_args()) print(args) #configs config = { 'file_name_pre': 'mdl_00000_a000_dd_mmm_nn_'+str(args['L'])+'_'+str(args['n_f']), 'no_layers': args['L'] , 'filter_number': args['n_f'], 'sig_type_source': ['aX', 'aY', 'aZ'], 'mode':'both', 'eps': 1e-3, 'model_type': 'Unetxl', 'down_sample_factor':4, 'frame_length' : int(4.096*F_SAMPLING), 'kernel_size': 7 , #(3,5) , #5,#(3,5), #5, #(3, 5),#5 'directory':'/media/sinan/9E82D1BB82D197DB/RESEARCH VLAB work on/Gyroscope SCG project/Deep Learning Paper Code and Materials', 'cycle_per_batch':2, 'sig_type_target': 'bcg', 'loss_func':'pearson_r', 'produce_video' : False, 'store_in_ram':True, 'augment_accel':True, 'augment_theta_lim': 10, 'augment_prob':0.5 } # cycle_per_batch = config['cycle_per_batch'] mode= config['mode'] eps = config['eps'] kernel_size = config['kernel_size'] directory= config['directory'] model_type = config['model_type'] no_layers=config['no_layers'] filter_number=config['filter_number'] sig_type_source=config['sig_type_source'] sig_type_target=config['sig_type_target'] down_sample_factor = config['down_sample_factor'] frame_length=config['frame_length'] input_size = frame_length//down_sample_factor normalized = True if model_type!='Unet_multiple_signal_in_not_normalized' else False loss_func = config['loss_func'] produce_video= config['produce_video'] store_in_ram=config['store_in_ram'] augment_accel= config['augment_accel'] augment_theta_lim = config['augment_theta_lim'] augment_prob=config['augment_prob'] file_name_pre = config['file_name_pre'] file_name_pre = file_name_pre[0:4] + model_type[:-1] + str(no_layers) +file_name_pre[9::] axis_string = ''.join(['x' if 'aX' in sig_type_source else '0' , 'y' if 'aY' in sig_type_source else '0' , 'z' if 'aZ' in sig_type_source else '0' ]) file_name_pre = file_name_pre[:12] + axis_string + file_name_pre[15::] print('Model Name: ' + file_name_pre) #get all subject data all_subject_instances = utility.load_subjects(directory + '/Training Data Analog Acc', store_in_ram ) #train test split train_subject_instances, val_subject_instances = train_test_split( all_subject_instances , test_size=0.2, random_state=49 ) #make a train and val generator train_gen = data_generators.make_generator_multiple_signal(list_of_subjects=train_subject_instances, cycle_per_batch=cycle_per_batch, eps=eps,frame_length=frame_length, mode=mode, list_sig_type_source= sig_type_source, sig_type_target= sig_type_target , down_sample_factor =down_sample_factor, normalized=normalized , store_in_ram=store_in_ram , augment_accel = augment_accel , augment_theta_lim = augment_theta_lim , augment_prob=augment_prob) val_gen = data_generators.make_generator_multiple_signal(list_of_subjects=val_subject_instances, cycle_per_batch=cycle_per_batch, eps=eps, frame_length=frame_length, mode=mode, list_sig_type_source= sig_type_source, sig_type_target= sig_type_target, down_sample_factor=down_sample_factor, normalized=normalized, store_in_ram=store_in_ram) #check ! #utility.diagnose_training_subjects(all_subject_instances) #data_generators.diagnose_generator_multiple_signal(train_gen , sig_type_source) #data_generators.diagnose_generator_multiple_signal(val_gen , sig_type_source) #create model if model_type=='Unetxl': sig_model = network_models.Unet_xl(input_size, kernel_size, filter_number, len(sig_type_source) , no_layers ) #make model parallel sig_model = nn.DataParallel(sig_model.cuda(), device_ids=[0,1]) #loss function is negative pearson loss loss = network_models.PearsonRLoss() #training configs cudnn.benchmark = True config['n_epochs'] = 150 config['scheduler_milestones'] = [30,60,120] #[50,100,200] config['train_steps'] = 10 config['val_steps'] = 10 config['initial_lr'] = 0.001 config['model_path'] = directory + '/Code Output/best_so_far.pt' config['model_path_for_video'] = directory + '/Models for Video/' + file_name_pre args = {'lr': config['initial_lr'], 'n_epochs':config['n_epochs'], 'model_path': config['model_path'], 'step_count':config['train_steps'], 'val_steps': config['val_steps'], 'scheduler_milestones': config['scheduler_milestones'], 'model_path_for_video': config['model_path_for_video'] } #train the model start_model_train = time.time() train_history, valid_history , best_val= network_models.train_torch_generator_with_video(args=args, sig_model=sig_model, criterion=loss, train_gen=train_gen, val_gen=val_gen, init_optimizer = lambda lr: Adam(sig_model.parameters(), lr=lr), init_schedule = lambda optimizer,milestones: torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=milestones, gamma=0.5) , produce_video=produce_video) end_model_train = time.time() print('Model Training Duration In Seconds: ' + str(end_model_train - start_model_train)) print('Best Validation Loss: ' + str(best_val)) #save model sig_model = network_models.load_saved_model(model_path=config['model_path'], model_type= model_type, input_size=input_size , kernel_size=kernel_size , filter_number=filter_number, signal_number=len(sig_type_source), no_layers = no_layers) torch.save({ 'model': sig_model.state_dict(), }, directory+'/Code Output/' + file_name_pre + '.pt') #save workspace pickle_list = [config, train_history , valid_history, train_subject_instances, val_subject_instances , best_val] fileObject = open(directory+'/Code Output/' +file_name_pre+'_pickle','wb') pickle.dump(pickle_list,fileObject) fileObject.close() #print configs to a test file with open(directory + '/Code Output/' + file_name_pre + '_config.txt', "w") as text_file: print(config, file=text_file) print('Model Training Duration In Seconds: ' + str(end_model_train - start_model_train), file=text_file) #plot results network_models.show_loss_torch_model(train_history, valid_history, file_name_pre , directory+'/Code Output')