def evaluate(self, X, y): """ Evaluate, i.e., compute metrics on given inputs and targets. Parameters ---------- X: ndarray Input data. y: 1darray Targets. Returns ------- result: dict Dictionary with result metrics. """ X = _ensure_float32(X) stop_criterion = MaxEpochs(0) train_set = SignalAndTarget(X, y) model_constraint = None valid_set = None test_set = None loss_function = self.loss if self.cropped: loss_function = lambda outputs, targets: self.loss( th.mean(outputs, dim=2), targets ) # reset runtime monitor if exists... for monitor in self.monitors: if hasattr(monitor, "last_call_time"): monitor.last_call_time = time.time() exp = Experiment( self.network, train_set, valid_set, test_set, iterator=self.iterator, loss_function=loss_function, optimizer=self.optimizer, model_constraint=model_constraint, monitors=self.monitors, stop_criterion=stop_criterion, remember_best_column=None, run_after_early_stop=False, cuda=self.is_cuda, log_0_epoch=True, do_early_stop=False, ) exp.monitor_epoch({"train": train_set}) result_dict = dict( [ (key.replace("train_", ""), val) for key, val in dict(exp.epochs_df.iloc[0]).items() ] ) return result_dict
def activate_model_evaluation(self, model, state=None, subject=None, final_evaluation=False, ensemble=False): print( f'free params in network:{NAS_utils.pytorch_count_params(model)}') if subject is None: subject = self.subject_id if self.cuda: torch.cuda.empty_cache() if final_evaluation: self.stop_criterion = Or([ MaxEpochs(global_vars.get('final_max_epochs')), NoIncreaseDecrease( 'valid_accuracy', global_vars.get('final_max_increase_epochs')) ]) if global_vars.get('cropping'): self.set_cropping_for_model(model) dataset = self.get_single_subj_dataset(subject, final_evaluation) nn_trainer = NN_Trainer(self.iterator, self.loss_function, self.stop_criterion, self.monitors) return nn_trainer.train_and_evaluate_model(model, dataset, state=state)
def train(config): cuda = True model = config['model'] if model == 'deep': model = Deep4Net(n_chans, n_classes, input_time_length=input_time_length, final_conv_length=2, config=config).create_network() to_dense_prediction_model(model) if cuda: model.cuda() log.info("Model: \n{:s}".format(str(model))) dummy_input = np_to_var(train_set.X[:1, :, :, None]) if cuda: dummy_input = dummy_input.cuda() out = model(dummy_input) n_preds_per_input = out.cpu().data.numpy().shape[2] optimizer = optim.Adam(model.parameters()) iterator = CropsFromTrialsIterator(batch_size=60, input_time_length=input_time_length, n_preds_per_input=n_preds_per_input) stop_criterion = Or([MaxEpochs(20), NoDecrease('valid_misclass', 80)]) monitors = [ LossMonitor(), MisclassMonitor(col_suffix='sample_misclass'), CroppedTrialMisclassMonitor(input_time_length=input_time_length), RuntimeMonitor() ] model_constraint = MaxNormDefaultConstraint() loss_function = lambda preds, targets: F.nll_loss( th.mean(preds, dim=2, keepdim=False), targets) exp = Experiment(model, train_set, valid_set, test_set, iterator=iterator, loss_function=loss_function, optimizer=optimizer, model_constraint=model_constraint, monitors=monitors, stop_criterion=stop_criterion, remember_best_column='valid_misclass', run_after_early_stop=True, cuda=cuda) exp.run() print(exp.rememberer) return exp.rememberer.lowest_val
def setup_after_stop_training(self, model, final_evaluation): self.rememberer.reset_to_best_model(self.epochs_df, model, self.optimizer) if final_evaluation: loss_to_reach = float(self.epochs_df['train_loss'].iloc[-1]) self.stop_criterion = Or(stop_criteria=[ MaxEpochs(max_epochs=global_vars.get('final_max_epochs')), ColumnBelow(column_name='valid_loss', target_value=loss_to_reach) ])
def get_normal_settings(): if global_vars.get('problem') == 'regression': loss_function = F.mse_loss else: loss_function = F.nll_loss stop_criterion = Or([MaxEpochs(global_vars.get('max_epochs')), NoIncreaseDecrease(f'valid_{global_vars.get("nn_objective")}', global_vars.get('max_increase_epochs'), oper=get_oper_by_loss_function(loss_function))]) iterator = BalancedBatchSizeIterator(batch_size=global_vars.get('batch_size')) monitors = [LossMonitor(), GenericMonitor('accuracy'), RuntimeMonitor()] for metric in global_vars.get('evaluation_metrics'): monitors.append(GenericMonitor(metric)) return stop_criterion, iterator, loss_function, monitors
def setUp(self): args = parse_args( ['-e', 'tests', '-c', '../configurations/config.ini']) init_config(args.config) configs = get_configurations(args.experiment) assert (len(configs) == 1) global_vars.set_config(configs[0]) global_vars.set('eeg_chans', 22) global_vars.set('num_subjects', 9) global_vars.set('input_time_len', 1125) global_vars.set('n_classes', 4) set_params_by_dataset() input_shape = (50, global_vars.get('eeg_chans'), global_vars.get('input_time_len')) class Dummy: def __init__(self, X, y): self.X = X self.y = y dummy_data = Dummy(X=np.ones(input_shape, dtype=np.float32), y=np.ones(50, dtype=np.longlong)) self.iterator = BalancedBatchSizeIterator( batch_size=global_vars.get('batch_size')) self.loss_function = F.nll_loss self.monitors = [ LossMonitor(), MisclassMonitor(), GenericMonitor('accuracy', acc_func), RuntimeMonitor() ] self.stop_criterion = Or([ MaxEpochs(global_vars.get('max_epochs')), NoDecrease('valid_misclass', global_vars.get('max_increase_epochs')) ]) self.naiveNAS = NaiveNAS(iterator=self.iterator, exp_folder='../tests', exp_name='', train_set=dummy_data, val_set=dummy_data, test_set=dummy_data, stop_criterion=self.stop_criterion, monitors=self.monitors, loss_function=self.loss_function, config=global_vars.config, subject_id=1, fieldnames=None, model_from_file=None)
def setup_after_stop_training(self): """ Setup training after first stop. Resets parameters to best parameters and updates stop criterion. """ # also remember old monitor chans, will be put back into # monitor chans after experiment finished self.before_stop_df = deepcopy(self.epochs_df) self.rememberer.reset_to_best_model(self.epochs_df, self.model, self.optimizer) loss_to_reach = float(self.epochs_df['train_loss'].iloc[-1]) self.stop_criterion = Or(stop_criteria=[ MaxEpochs(max_epochs=self.rememberer.best_epoch * 2), ColumnBelow(column_name='valid_loss', target_value=loss_to_reach)]) log.info("Train loss to reach {:.5f}".format(loss_to_reach))
def get_cropped_settings(): stop_criterion = Or([MaxEpochs(global_vars.get('max_epochs')), NoIncreaseDecrease(f'valid_{global_vars.get("nn_objective")}', global_vars.get('max_increase_epochs'))]) iterator = CropsFromTrialsIterator(batch_size=global_vars.get('batch_size'), input_time_length=global_vars.get('input_time_len'), n_preds_per_input=global_vars.get('n_preds_per_input')) loss_function = lambda preds, targets: F.nll_loss(torch.mean(preds, dim=2, keepdim=False), targets) monitors = [LossMonitor(), GenericMonitor('accuracy', acc_func), CroppedTrialGenericMonitor('accuracy', acc_func, input_time_length=global_vars.get('input_time_len')), RuntimeMonitor()] if global_vars.get('dataset') in ['NER15', 'Cho', 'SonarSub']: monitors.append(CroppedTrialGenericMonitor('auc', auc_func, input_time_length=global_vars.get('input_time_len'))) if global_vars.get('dataset') in ['BCI_IV_2b']: monitors.append(CroppedGenericMonitorPerTimeStep('kappa', kappa_func, input_time_length=global_vars.get('input_time_len'))) return stop_criterion, iterator, loss_function, monitors
def network_model(model, train_set, test_set, valid_set, n_chans, input_time_length, cuda): max_epochs = 30 max_increase_epochs = 10 batch_size = 64 init_block_size = 1000 set_random_seeds(seed=20190629, cuda=cuda) n_classes = 2 n_chans = n_chans input_time_length = input_time_length if model == 'deep': model = Deep4Net(n_chans, n_classes, input_time_length=input_time_length, final_conv_length='auto').create_network() elif model == 'shallow': model = ShallowFBCSPNet(n_chans, n_classes, input_time_length=input_time_length, final_conv_length='auto').create_network() if cuda: model.cuda() log.info("%s model: ".format(str(model))) optimizer = AdamW(model.parameters(), lr=0.00625, weight_decay=0) iterator = BalancedBatchSizeIterator(batch_size=batch_size) stop_criterion = Or([MaxEpochs(max_epochs), NoDecrease('valid_misclass', max_increase_epochs)]) monitors = [LossMonitor(), MisclassMonitor(), RuntimeMonitor()] model_constraint = None print(train_set.X.shape[0]) model_test = Experiment(model, train_set, valid_set, test_set, iterator=iterator, loss_function=F.nll_loss, optimizer=optimizer, model_constraint=model_constraint, monitors=monitors, stop_criterion=stop_criterion, remember_best_column='valid_misclass', run_after_early_stop=True, cuda=cuda) model_test.run() return model_test
def setup_after_stop_training(self): """ Setup training after first stop. Resets parameters to best parameters and updates stop criterion. """ # also remember old monitor chans, will be put back into # monitor chans after experiment finished self.before_stop_df = deepcopy(self.epochs_df) self.rememberer.reset_to_best_model(self.epochs_df, self.model, self.optimizer) loss_to_reach = float(self.epochs_df['train_loss'].iloc[-1]) self.stop_criterion = Or(stop_criteria=[ MaxEpochs(max_epochs=self.rememberer.best_epoch * 2), ColumnBelow(column_name='valid_loss', target_value=loss_to_reach) ]) log.info("Train loss to reach {:.5f}".format(loss_to_reach)) # def exp_lr_scheduler(optimizer, global_step, init_lr, decay_steps, decay_rate, lr_clip, staircase=True): # if staircase: # lr = init_lr * decay_rate # else: # lr = init_lr * decay_rate**(global_step / decay_steps) # lr = max(lr, lr_clip) # #print(lr) # if global_step % decay_steps == 0: # print('LR is set to {}'.format(lr)) # lr=0.1 # for param_group in optimizer.param_groups: # param_group['lr'] = lr # return optimizer # def adjust_lr(optimizer, epoch): # lr = init_lr * (0.1 ** (epoch // 20)) # for param_group in optimizer.param_groups: # param_group['lr'] = lr # return optimizer
def train(train_set, test_set, model, iterator, monitors, loss_function, max_epochs, cuda): if cuda: model.cuda() optimizer = AdamW(model.parameters(), lr=1 * 0.01, weight_decay=0.5 * 0.001) # these are good values for the deep model stop_criterion = MaxEpochs(max_epochs) model_constraint = MaxNormDefaultConstraint() n_updates_per_epoch = sum( [1 for _ in iterator.get_batches(train_set, shuffle=True)]) n_updates_per_period = n_updates_per_epoch * max_epochs scheduler = CosineAnnealing(n_updates_per_period) optimizer = ScheduledOptimizer(scheduler, optimizer, schedule_weight_decay=True) exp = Experiment(model, train_set, None, test_set, iterator=iterator, loss_function=loss_function, optimizer=optimizer, model_constraint=model_constraint, monitors=monitors, remember_best_column=None, stop_criterion=stop_criterion, cuda=cuda, run_after_early_stop=False, do_early_stop=False) exp.run() return exp
def evaluate(self, X, y, batch_size=32): # Create a dummy experiment for the evaluation iterator = BalancedBatchSizeIterator(batch_size=batch_size, seed=0) # seed irrelevant stop_criterion = MaxEpochs(0) train_set = SignalAndTarget(X, y) model_constraint = None valid_set = None test_set = None loss_function = self.loss if self.cropped: loss_function = lambda outputs, targets: \ self.loss(th.mean(outputs, dim=2), targets) exp = Experiment(self.network, train_set, valid_set, test_set, iterator=iterator, loss_function=loss_function, optimizer=self.optimizer, model_constraint=model_constraint, monitors=self.monitors, stop_criterion=stop_criterion, remember_best_column=None, run_after_early_stop=False, cuda=self.cuda, print_0_epoch=False, do_early_stop=False) exp.monitor_epoch({'train': train_set}) result_dict = dict([ (key.replace('train_', ''), val) for key, val in dict(exp.epochs_df.iloc[0]).items() ]) return result_dict
def train_model(self, train_set, val_set, test_set, save_model): """ :param train_set: EEG data (n_trials*n_channels*n_samples) :param val_set: EEG data (n_trials*n_channels*n_samples) :param test_set: EEG data (n_trials*n_channels*n_samples) - can be None when training on inner-fold :param save_model: Boolean: True if trained model is to be saved :return: Accuracy and loss scores for the model trained with a given set of hyper-parameters """ predictions = None model = None model = self.call_model() set_random_seeds(seed=20190629, cuda=self.cuda) if self.cuda: model.cuda() torch.backends.cudnn.deterministic = True log.info("%s model: ".format(str(model))) optimizer = optim.Adam(model.parameters(), lr=self.learning_rate, weight_decay=0, eps=1e-8, amsgrad=False) stop_criterion = Or([ MaxEpochs(self.epochs), NoDecrease('valid_misclass', self.max_increase_epochs) ]) model_loss_function = None #####Setup to run the selected model##### model_test = Experiment(model, train_set, val_set, test_set=test_set, iterator=self.iterator, loss_function=self.loss, optimizer=optimizer, model_constraint=self.model_constraint, monitors=self.monitors, stop_criterion=stop_criterion, remember_best_column='valid_misclass', run_after_early_stop=True, model_loss_function=model_loss_function, cuda=self.cuda, data_type=self.data_type, model_type=self.model_type, subject_id=self.subject, model_number=str(self.model_number), save_model=save_model) model_test.run() model_acc = model_test.epochs_df['valid_misclass'].astype('float') model_loss = model_test.epochs_df['valid_loss'].astype('float') current_val_acc = 1 - current_acc(model_acc) current_val_loss = current_loss(model_loss) test_accuracy = None if test_set is not None: test_accuracy = round( (1 - model_test.epochs_df['test_misclass'].min()) * 100, 3) predictions = model_test.predictions probabilities = model_test.probabilites return current_val_acc, current_val_loss, test_accuracy, model_test, predictions, probabilities
def run_exp(max_recording_mins, n_recordings, sec_to_cut, duration_recording_mins, max_abs_val, max_min_threshold, max_min_expected, shrink_val, max_min_remove, batch_set_zero_val, batch_set_zero_test, sampling_freq, low_cut_hz, high_cut_hz, exp_demean, exp_standardize, moving_demean, moving_standardize, channel_demean, channel_standardize, divisor, n_folds, i_test_fold, input_time_length, final_conv_length, pool_stride, n_blocks_to_add, sigmoid, model_constraint, batch_size, max_epochs, only_return_exp): cuda = True preproc_functions = [] preproc_functions.append(lambda data, fs: ( data[:, int(sec_to_cut * fs):-int(sec_to_cut * fs)], fs)) preproc_functions.append(lambda data, fs: (data[:, :int( duration_recording_mins * 60 * fs)], fs)) if max_abs_val is not None: preproc_functions.append( lambda data, fs: (np.clip(data, -max_abs_val, max_abs_val), fs)) if max_min_threshold is not None: preproc_functions.append(lambda data, fs: (clean_jumps( data, 200, max_min_threshold, max_min_expected, cuda), fs)) if max_min_remove is not None: window_len = 200 preproc_functions.append(lambda data, fs: (set_jumps_to_zero( data, window_len=window_len, threshold=max_min_remove, cuda=cuda, clip_min_max_to_zero=True), fs)) if shrink_val is not None: preproc_functions.append(lambda data, fs: (shrink_spikes( data, shrink_val, 1, 9, ), fs)) preproc_functions.append(lambda data, fs: (resampy.resample( data, fs, sampling_freq, axis=1, filter='kaiser_fast'), sampling_freq)) preproc_functions.append(lambda data, fs: (bandpass_cnt( data, low_cut_hz, high_cut_hz, fs, filt_order=4, axis=1), fs)) if exp_demean: preproc_functions.append(lambda data, fs: (exponential_running_demean( data.T, factor_new=0.001, init_block_size=100).T, fs)) if exp_standardize: preproc_functions.append( lambda data, fs: (exponential_running_standardize( data.T, factor_new=0.001, init_block_size=100).T, fs)) if moving_demean: preproc_functions.append(lambda data, fs: (padded_moving_demean( data, axis=1, n_window=201), fs)) if moving_standardize: preproc_functions.append(lambda data, fs: (padded_moving_standardize( data, axis=1, n_window=201), fs)) if channel_demean: preproc_functions.append(lambda data, fs: (demean(data, axis=1), fs)) if channel_standardize: preproc_functions.append(lambda data, fs: (standardize(data, axis=1), fs)) if divisor is not None: preproc_functions.append(lambda data, fs: (data / divisor, fs)) dataset = DiagnosisSet(n_recordings=n_recordings, max_recording_mins=max_recording_mins, preproc_functions=preproc_functions) if not only_return_exp: X, y = dataset.load() splitter = Splitter( n_folds, i_test_fold, ) if not only_return_exp: train_set, valid_set, test_set = splitter.split(X, y) del X, y # shouldn't be necessary, but just to make sure else: train_set = None valid_set = None test_set = None set_random_seeds(seed=20170629, cuda=cuda) if sigmoid: n_classes = 1 else: n_classes = 2 in_chans = 21 net = Deep4Net( in_chans=in_chans, n_classes=n_classes, input_time_length=input_time_length, final_conv_length=final_conv_length, pool_time_length=pool_stride, pool_time_stride=pool_stride, n_filters_2=50, n_filters_3=80, n_filters_4=120, ) model = net_with_more_layers(net, n_blocks_to_add, nn.MaxPool2d) if sigmoid: model = to_linear_plus_minus_net(model) optimizer = optim.Adam(model.parameters()) to_dense_prediction_model(model) log.info("Model:\n{:s}".format(str(model))) if cuda: model.cuda() # determine output size test_input = np_to_var( np.ones((2, in_chans, input_time_length, 1), dtype=np.float32)) if cuda: test_input = test_input.cuda() out = model(test_input) n_preds_per_input = out.cpu().data.numpy().shape[2] log.info("{:d} predictions per input/trial".format(n_preds_per_input)) iterator = CropsFromTrialsIterator(batch_size=batch_size, input_time_length=input_time_length, n_preds_per_input=n_preds_per_input) if sigmoid: loss_function = lambda preds, targets: binary_cross_entropy_with_logits( th.mean(preds, dim=2)[:, 1, 0], targets.type_as(preds)) else: loss_function = lambda preds, targets: F.nll_loss( th.mean(preds, dim=2)[:, :, 0], targets) if model_constraint is not None: model_constraint = MaxNormDefaultConstraint() monitors = [ LossMonitor(), MisclassMonitor(col_suffix='sample_misclass'), CroppedTrialMisclassMonitor(input_time_length), RuntimeMonitor(), ] stop_criterion = MaxEpochs(max_epochs) batch_modifier = None if batch_set_zero_val is not None: batch_modifier = RemoveMinMaxDiff(batch_set_zero_val, clip_max_abs=True, set_zero=True) if (batch_set_zero_val is not None) and (batch_set_zero_test == True): iterator = ModifiedIterator( iterator, batch_modifier, ) batch_modifier = None exp = Experiment(model, train_set, valid_set, test_set, iterator, loss_function, optimizer, model_constraint, monitors, stop_criterion, remember_best_column='valid_misclass', run_after_early_stop=True, batch_modifier=batch_modifier, cuda=cuda) if not only_return_exp: exp.run() else: exp.dataset = dataset exp.splitter = splitter return exp
def build_exp(model_name, cuda, data, batch_size, max_epochs, max_increase_epochs): log.info("==============================") log.info("Loading Data...") log.info("==============================") train_set = data.train_set valid_set = data.validation_set test_set = data.test_set log.info("==============================") log.info("Setting Up Model...") log.info("==============================") set_random_seeds(seed=20190706, cuda=cuda) n_classes = 4 n_chans = int(train_set.X.shape[1]) input_time_length = train_set.X.shape[2] if model_name == "shallow": model = NewShallowNet( n_chans, n_classes, input_time_length, final_conv_length="auto" ) # model = ShallowFBCSPNet( # n_chans, # n_classes, # input_time_length=input_time_length, # final_conv_length="auto", # ).create_network() elif model_name == "deep": model = NewDeep4Net(n_chans, n_classes, input_time_length, "auto") # model = Deep4Net( # n_chans, # n_classes, # input_time_length=input_time_length, # final_conv_length="auto", # ).create_network() elif model_name == "eegnet": # model = EEGNet(n_chans, n_classes, # input_time_length=input_time_length) # model = EEGNetv4(n_chans, n_classes, # input_time_length=input_time_length).create_network() model = NewEEGNet(n_chans, n_classes, input_time_length=input_time_length) if cuda: model.cuda() log.info("==============================") log.info("Logging Model Architecture:") log.info("==============================") log.info("Model: \n{:s}".format(str(model))) log.info("==============================") log.info("Building Experiment:") log.info("==============================") optimizer = optim.Adam(model.parameters()) iterator = BalancedBatchSizeIterator(batch_size=batch_size) stop_criterion = Or( [MaxEpochs(max_epochs), NoDecrease("valid_misclass", max_increase_epochs)] ) monitors = [LossMonitor(), MisclassMonitor(), RuntimeMonitor()] model_constraint = MaxNormDefaultConstraint() exp = Experiment( model, train_set, valid_set, test_set, iterator=iterator, loss_function=F.nll_loss, optimizer=optimizer, model_constraint=model_constraint, monitors=monitors, stop_criterion=stop_criterion, remember_best_column="valid_misclass", run_after_early_stop=True, cuda=cuda, ) return exp
def run_exp(epoches, batch_size, subject_num, model_type, cuda, single_subject, single_subject_num): # ival = [-500, 4000] max_increase_epochs = 160 # Preprocessing X, y = loadSubjects(subject_num, single_subject, single_subject_num) X = X.astype(np.float32) y = y.astype(np.int64) X, y = shuffle(X, y) trial_length = X.shape[2] print("trial_length " + str(trial_length)) print("trying to run with {} sec trials ".format((trial_length - 1) / 256)) print("y") print(y) trainingSampleSize = int(len(X) * 0.6) valudationSampleSize = int(len(X) * 0.2) testSampleSize = int(len(X) * 0.2) print("INFO : Training sample size: {}".format(trainingSampleSize)) print("INFO : Validation sample size: {}".format(valudationSampleSize)) print("INFO : Test sample size: {}".format(testSampleSize)) train_set = SignalAndTarget(X[:trainingSampleSize], y=y[:trainingSampleSize]) valid_set = SignalAndTarget( X[trainingSampleSize:(trainingSampleSize + valudationSampleSize)], y=y[trainingSampleSize:(trainingSampleSize + valudationSampleSize)]) test_set = SignalAndTarget(X[(trainingSampleSize + valudationSampleSize):], y=y[(trainingSampleSize + valudationSampleSize):]) set_random_seeds(seed=20190706, cuda=cuda) n_classes = 3 n_chans = int(train_set.X.shape[1]) input_time_length = train_set.X.shape[2] if model_type == 'shallow': model = ShallowFBCSPNet(n_chans, n_classes, input_time_length=input_time_length, final_conv_length='auto').create_network() elif model_type == 'deep': model = Deep4Net(n_chans, n_classes, input_time_length=input_time_length, final_conv_length='auto').create_network() elif model_type == 'eegnet': model = EEGNetv4(n_chans, n_classes, input_time_length=input_time_length, final_conv_length='auto').create_network() if cuda: model.cuda() log.info("Model: \n{:s}".format(str(model))) optimizer = optim.Adam(model.parameters()) iterator = BalancedBatchSizeIterator(batch_size=batch_size) stop_criterion = Or([ MaxEpochs(max_epochs), NoDecrease('valid_misclass', max_increase_epochs) ]) monitors = [LossMonitor(), MisclassMonitor(), RuntimeMonitor()] model_constraint = MaxNormDefaultConstraint() exp = Experiment(model, train_set, valid_set, test_set, iterator=iterator, loss_function=F.nll_loss, optimizer=optimizer, model_constraint=model_constraint, monitors=monitors, stop_criterion=stop_criterion, remember_best_column='valid_misclass', run_after_early_stop=True, cuda=cuda) exp.run() # th.save(model, "models\{}-cropped-singleSubjectNum{}-{}sec-{}epoches-torch_model".format(model_type, single_subject_num, ((trial_length - 1) / 256), epoches)) return exp
def fit( self, train_X, train_y, epochs, batch_size, input_time_length=None, validation_data=None, model_constraint=None, remember_best_column=None, scheduler=None, log_0_epoch=True, ): """ Fit the model using the given training data. Will set `epochs_df` variable with a pandas dataframe to the history of the training process. Parameters ---------- train_X: ndarray Training input data train_y: 1darray Training labels epochs: int Number of epochs to train batch_size: int input_time_length: int, optional Super crop size, what temporal size is pushed forward through the network, see cropped decoding tuturial. validation_data: (ndarray, 1darray), optional X and y for validation set if wanted model_constraint: object, optional You can supply :class:`.MaxNormDefaultConstraint` if wanted. remember_best_column: string, optional In case you want to do an early stopping/reset parameters to some "best" epoch, define here the monitored value whose minimum determines the best epoch. scheduler: 'cosine' or None, optional Whether to use cosine annealing (:class:`.CosineAnnealing`). log_0_epoch: bool Whether to compute the metrics once before training as well. Returns ------- exp: Underlying braindecode :class:`.Experiment` """ if (not hasattr(self, "compiled")) or (not self.compiled): raise ValueError( "Compile the model first by calling model.compile(loss, optimizer, metrics)" ) if self.cropped and input_time_length is None: raise ValueError( "In cropped mode, need to specify input_time_length," "which is the number of timesteps that will be pushed through" "the network in a single pass.") train_X = _ensure_float32(train_X) if self.cropped: self.network.eval() test_input = np_to_var( np.ones( (1, train_X[0].shape[0], input_time_length) + train_X[0].shape[2:], dtype=np.float32, )) while len(test_input.size()) < 4: test_input = test_input.unsqueeze(-1) if self.is_cuda: test_input = test_input.cuda() out = self.network(test_input) n_preds_per_input = out.cpu().data.numpy().shape[2] self.iterator = CropsFromTrialsIterator( batch_size=batch_size, input_time_length=input_time_length, n_preds_per_input=n_preds_per_input, seed=self.seed_rng.randint(0, np.iinfo(np.int32).max - 1), ) else: self.iterator = BalancedBatchSizeIterator( batch_size=batch_size, seed=self.seed_rng.randint(0, np.iinfo(np.int32).max - 1), ) if log_0_epoch: stop_criterion = MaxEpochs(epochs) else: stop_criterion = MaxEpochs(epochs - 1) train_set = SignalAndTarget(train_X, train_y) optimizer = self.optimizer if scheduler is not None: assert (scheduler == "cosine" ), "Supply either 'cosine' or None as scheduler." n_updates_per_epoch = sum([ 1 for _ in self.iterator.get_batches(train_set, shuffle=True) ]) n_updates_per_period = n_updates_per_epoch * epochs if scheduler == "cosine": scheduler = CosineAnnealing(n_updates_per_period) schedule_weight_decay = False if optimizer.__class__.__name__ == "AdamW": schedule_weight_decay = True optimizer = ScheduledOptimizer( scheduler, self.optimizer, schedule_weight_decay=schedule_weight_decay, ) loss_function = self.loss if self.cropped: loss_function = lambda outputs, targets: self.loss( th.mean(outputs, dim=2), targets) if validation_data is not None: valid_X = _ensure_float32(validation_data[0]) valid_y = validation_data[1] valid_set = SignalAndTarget(valid_X, valid_y) else: valid_set = None test_set = None self.monitors = [LossMonitor()] if self.cropped: self.monitors.append( CroppedTrialMisclassMonitor(input_time_length)) else: self.monitors.append(MisclassMonitor()) if self.extra_monitors is not None: self.monitors.extend(self.extra_monitors) self.monitors.append(RuntimeMonitor()) exp = Experiment( self.network, train_set, valid_set, test_set, iterator=self.iterator, loss_function=loss_function, optimizer=optimizer, model_constraint=model_constraint, monitors=self.monitors, stop_criterion=stop_criterion, remember_best_column=remember_best_column, run_after_early_stop=False, cuda=self.is_cuda, log_0_epoch=log_0_epoch, do_early_stop=(remember_best_column is not None), ) exp.run() self.epochs_df = exp.epochs_df return exp
def test_experiment_class(): import mne from mne.io import concatenate_raws # 5,6,7,10,13,14 are codes for executed and imagined hands/feet subject_id = 1 event_codes = [5, 6, 9, 10, 13, 14] # This will download the files if you don't have them yet, # and then return the paths to the files. physionet_paths = mne.datasets.eegbci.load_data(subject_id, event_codes) # Load each of the files parts = [mne.io.read_raw_edf(path, preload=True, stim_channel='auto', verbose='WARNING') for path in physionet_paths] # Concatenate them raw = concatenate_raws(parts) # Find the events in this dataset events, _ = mne.events_from_annotations(raw) # Use only EEG channels eeg_channel_inds = mne.pick_types(raw.info, meg=False, eeg=True, stim=False, eog=False, exclude='bads') # Extract trials, only using EEG channels epoched = mne.Epochs(raw, events, dict(hands=2, feet=3), tmin=1, tmax=4.1, proj=False, picks=eeg_channel_inds, baseline=None, preload=True) import numpy as np from braindecode.datautil.signal_target import SignalAndTarget from braindecode.datautil.splitters import split_into_two_sets # Convert data from volt to millivolt # Pytorch expects float32 for input and int64 for labels. X = (epoched.get_data() * 1e6).astype(np.float32) y = (epoched.events[:, 2] - 2).astype(np.int64) # 2,3 -> 0,1 train_set = SignalAndTarget(X[:60], y=y[:60]) test_set = SignalAndTarget(X[60:], y=y[60:]) train_set, valid_set = split_into_two_sets(train_set, first_set_fraction=0.8) from braindecode.models.shallow_fbcsp import ShallowFBCSPNet from torch import nn from braindecode.torch_ext.util import set_random_seeds from braindecode.models.util import to_dense_prediction_model # Set if you want to use GPU # You can also use torch.cuda.is_available() to determine if cuda is available on your machine. cuda = False set_random_seeds(seed=20170629, cuda=cuda) # This will determine how many crops are processed in parallel input_time_length = 450 n_classes = 2 in_chans = train_set.X.shape[1] # final_conv_length determines the size of the receptive field of the ConvNet model = ShallowFBCSPNet(in_chans=in_chans, n_classes=n_classes, input_time_length=input_time_length, final_conv_length=12).create_network() to_dense_prediction_model(model) if cuda: model.cuda() from torch import optim optimizer = optim.Adam(model.parameters()) from braindecode.torch_ext.util import np_to_var # determine output size test_input = np_to_var( np.ones((2, in_chans, input_time_length, 1), dtype=np.float32)) if cuda: test_input = test_input.cuda() out = model(test_input) n_preds_per_input = out.cpu().data.numpy().shape[2] print("{:d} predictions per input/trial".format(n_preds_per_input)) from braindecode.experiments.experiment import Experiment from braindecode.datautil.iterators import CropsFromTrialsIterator from braindecode.experiments.monitors import RuntimeMonitor, LossMonitor, \ CroppedTrialMisclassMonitor, MisclassMonitor from braindecode.experiments.stopcriteria import MaxEpochs import torch.nn.functional as F import torch as th from braindecode.torch_ext.modules import Expression # Iterator is used to iterate over datasets both for training # and evaluation iterator = CropsFromTrialsIterator(batch_size=32, input_time_length=input_time_length, n_preds_per_input=n_preds_per_input) # Loss function takes predictions as they come out of the network and the targets # and returns a loss loss_function = lambda preds, targets: F.nll_loss( th.mean(preds, dim=2, keepdim=False), targets) # Could be used to apply some constraint on the models, then should be object # with apply method that accepts a module model_constraint = None # Monitors log the training progress monitors = [LossMonitor(), MisclassMonitor(col_suffix='sample_misclass'), CroppedTrialMisclassMonitor(input_time_length), RuntimeMonitor(), ] # Stop criterion determines when the first stop happens stop_criterion = MaxEpochs(4) exp = Experiment(model, train_set, valid_set, test_set, iterator, loss_function, optimizer, model_constraint, monitors, stop_criterion, remember_best_column='valid_misclass', run_after_early_stop=True, batch_modifier=None, cuda=cuda) # need to setup python logging before to be able to see anything import logging import sys logging.basicConfig(format='%(asctime)s %(levelname)s : %(message)s', level=logging.DEBUG, stream=sys.stdout) exp.run() import pandas as pd from io import StringIO compare_df = pd.read_csv(StringIO( 'train_loss,valid_loss,test_loss,train_sample_misclass,valid_sample_misclass,' 'test_sample_misclass,train_misclass,valid_misclass,test_misclass\n' '14.167170524597168,13.910758018493652,15.945781707763672,0.5,0.5,' '0.5333333333333333,0.5,0.5,0.5333333333333333\n' '1.1735659837722778,1.4342904090881348,1.8664429187774658,0.4629567736185384,' '0.5120320855614973,0.5336007130124778,0.5,0.5,0.5333333333333333\n' '1.3168460130691528,1.60431969165802,1.9181344509124756,0.49298128342245995,' '0.5109180035650625,0.531729055258467,0.5,0.5,0.5333333333333333\n' '0.8465543389320374,1.280307412147522,1.439755916595459,0.4413435828877005,' '0.5461229946524064,0.5283422459893048,0.47916666666666663,0.5,' '0.5333333333333333\n0.6977059841156006,1.1762590408325195,1.2779350280761719,' '0.40290775401069523,0.588903743315508,0.5307486631016043,0.5,0.5,0.5\n' '0.7934166193008423,1.1762590408325195,1.2779350280761719,0.4401069518716577,' '0.588903743315508,0.5307486631016043,0.5,0.5,0.5\n0.5982189178466797,' '0.8581563830375671,0.9598925113677979,0.32032085561497325,0.47660427807486627,' '0.4672905525846702,0.31666666666666665,0.5,0.4666666666666667\n0.5044312477111816,' '0.7133197784423828,0.8164243102073669,0.2591354723707665,0.45699643493761144,' '0.4393048128342246,0.16666666666666663,0.41666666666666663,0.43333333333333335\n' '0.4815250039100647,0.6736412644386292,0.8016976714134216,0.23413547237076648,' '0.39505347593582885,0.42932263814616756,0.15000000000000002,0.41666666666666663,0.5\n')) for col in compare_df: np.testing.assert_allclose(np.array(compare_df[col]), exp.epochs_df[col], rtol=1e-3, atol=1e-4)
# Loss function takes predictions as they come out of the network and the targets # and returns a loss loss_function = lambda preds, targets: log_categorical_crossentropy( preds, targets) # Could be used to apply some constraint on the models, then should be object # with apply method that accepts a module model_constraint = None # Monitors log the training progress monitors = [ LossMonitor(), MisclassMonitor(col_suffix='misclass'), SeizureMonitor(input_time_length), RuntimeMonitor(), ] # Stop criterion determines when the first stop happens stop_criterion = MaxEpochs(5) exp = Experiment(model, train_set, valid_set, test_set, iterator, loss_function, optimizer, model_constraint, monitors, stop_criterion, remember_best_column='valid_misclass', run_after_early_stop=True, batch_modifier=None, cuda=cuda)
def run_exp(data_folder, subject_id, low_cut_hz, model, cuda): ival = [-500, 4000] max_epochs = 1600 max_increase_epochs = 160 batch_size = 60 high_cut_hz = 38 factor_new = 1e-3 init_block_size = 1000 valid_set_fraction = 0.2 train_filename = "A{:02d}T.gdf".format(subject_id) test_filename = "A{:02d}E.gdf".format(subject_id) train_filepath = os.path.join(data_folder, train_filename) test_filepath = os.path.join(data_folder, test_filename) train_label_filepath = train_filepath.replace(".gdf", ".mat") test_label_filepath = test_filepath.replace(".gdf", ".mat") train_loader = BCICompetition4Set2A( train_filepath, labels_filename=train_label_filepath ) test_loader = BCICompetition4Set2A( test_filepath, labels_filename=test_label_filepath ) train_cnt = train_loader.load() test_cnt = test_loader.load() # Preprocessing train_cnt = train_cnt.drop_channels( ["EOG-left", "EOG-central", "EOG-right"] ) assert len(train_cnt.ch_names) == 22 # lets convert to millvolt for numerical stability of next operations train_cnt = mne_apply(lambda a: a * 1e6, train_cnt) train_cnt = mne_apply( lambda a: bandpass_cnt( a, low_cut_hz, high_cut_hz, train_cnt.info["sfreq"], filt_order=3, axis=1, ), train_cnt, ) train_cnt = mne_apply( lambda a: exponential_running_standardize( a.T, factor_new=factor_new, init_block_size=init_block_size, eps=1e-4, ).T, train_cnt, ) test_cnt = test_cnt.drop_channels(["EOG-left", "EOG-central", "EOG-right"]) assert len(test_cnt.ch_names) == 22 test_cnt = mne_apply(lambda a: a * 1e6, test_cnt) test_cnt = mne_apply( lambda a: bandpass_cnt( a, low_cut_hz, high_cut_hz, test_cnt.info["sfreq"], filt_order=3, axis=1, ), test_cnt, ) test_cnt = mne_apply( lambda a: exponential_running_standardize( a.T, factor_new=factor_new, init_block_size=init_block_size, eps=1e-4, ).T, test_cnt, ) marker_def = OrderedDict( [ ("Left Hand", [1]), ("Right Hand", [2]), ("Foot", [3]), ("Tongue", [4]), ] ) train_set = create_signal_target_from_raw_mne(train_cnt, marker_def, ival) test_set = create_signal_target_from_raw_mne(test_cnt, marker_def, ival) train_set, valid_set = split_into_two_sets( train_set, first_set_fraction=1 - valid_set_fraction ) set_random_seeds(seed=20190706, cuda=cuda) n_classes = 4 n_chans = int(train_set.X.shape[1]) input_time_length = train_set.X.shape[2] if model == "shallow": model = ShallowFBCSPNet( n_chans, n_classes, input_time_length=input_time_length, final_conv_length="auto", ).create_network() elif model == "deep": model = Deep4Net( n_chans, n_classes, input_time_length=input_time_length, final_conv_length="auto", ).create_network() if cuda: model.cuda() log.info("Model: \n{:s}".format(str(model))) optimizer = optim.Adam(model.parameters()) iterator = BalancedBatchSizeIterator(batch_size=batch_size) stop_criterion = Or( [ MaxEpochs(max_epochs), NoDecrease("valid_misclass", max_increase_epochs), ] ) monitors = [LossMonitor(), MisclassMonitor(), RuntimeMonitor()] model_constraint = MaxNormDefaultConstraint() exp = Experiment( model, train_set, valid_set, test_set, iterator=iterator, loss_function=F.nll_loss, optimizer=optimizer, model_constraint=model_constraint, monitors=monitors, stop_criterion=stop_criterion, remember_best_column="valid_misclass", run_after_early_stop=True, cuda=cuda, ) exp.run() return exp
def network_model(subject_id, model_type, data_type, cropped, cuda, parameters, hyp_params): best_params = dict() # dictionary to store hyper-parameter values #####Parameter passed to funciton##### max_epochs = parameters['max_epochs'] max_increase_epochs = parameters['max_increase_epochs'] batch_size = parameters['batch_size'] #####Constant Parameters##### best_loss = 100.0 # instatiate starting point for loss iterator = BalancedBatchSizeIterator(batch_size=batch_size) stop_criterion = Or([MaxEpochs(max_epochs), NoDecrease('valid_misclass', max_increase_epochs)]) monitors = [LossMonitor(), MisclassMonitor(), RuntimeMonitor()] model_constraint = MaxNormDefaultConstraint() epoch = 4096 #####Collect and format data##### if data_type == 'words': data, labels = format_data(data_type, subject_id, epoch) data = data[:,:,768:1280] # within-trial window selected for classification elif data_type == 'vowels': data, labels = format_data(data_type, subject_id, epoch) data = data[:,:,512:1024] elif data_type == 'all_classes': data, labels = format_data(data_type, subject_id, epoch) data = data[:,:,768:1280] x = lambda a: a * 1e6 # improves numerical stability data = x(data) data = normalize(data) data, labels = balanced_subsample(data, labels) # downsampling the data to ensure equal classes data, _, labels, _ = train_test_split(data, labels, test_size=0, random_state=42) # redundant shuffle of data/labels #####model inputs##### unique, counts = np.unique(labels, return_counts=True) n_classes = len(unique) n_chans = int(data.shape[1]) input_time_length = data.shape[2] #####k-fold nested corss-validation##### num_folds = 4 skf = StratifiedKFold(n_splits=num_folds, shuffle=True, random_state=10) out_fold_num = 0 # outer-fold number cv_scores = [] #####Outer=Fold##### for inner_ind, outer_index in skf.split(data, labels): inner_fold, outer_fold = data[inner_ind], data[outer_index] inner_labels, outer_labels = labels[inner_ind], labels[outer_index] out_fold_num += 1 # list for storing cross-validated scores loss_with_params = dict()# for storing param values and losses in_fold_num = 0 # inner-fold number #####Inner-Fold##### for train_idx, valid_idx in skf.split(inner_fold, inner_labels): X_Train, X_val = inner_fold[train_idx], inner_fold[valid_idx] y_train, y_val = inner_labels[train_idx], inner_labels[valid_idx] in_fold_num += 1 train_set = SignalAndTarget(X_Train, y_train) valid_set = SignalAndTarget(X_val, y_val) loss_with_params[f"Fold_{in_fold_num}"] = dict() ####Nested cross-validation##### for drop_prob in hyp_params['drop_prob']: for loss_function in hyp_params['loss']: for i in range(len(hyp_params['lr_adam'])): model = None # ensure no duplication of models # model, learning-rate and optimizer setup according to model_type if model_type == 'shallow': model = ShallowFBCSPNet(in_chans=n_chans, n_classes=n_classes, input_time_length=input_time_length, n_filters_time=80, filter_time_length=40, n_filters_spat=80, pool_time_length=75, pool_time_stride=25, final_conv_length='auto', conv_nonlin=square, pool_mode='max', pool_nonlin=safe_log, split_first_layer=True, batch_norm=True, batch_norm_alpha=0.1, drop_prob=drop_prob).create_network() lr = hyp_params['lr_ada'][i] optimizer = optim.Adadelta(model.parameters(), lr=lr, rho=0.9, weight_decay=0.1, eps=1e-8) elif model_type == 'deep': model = Deep4Net(in_chans=n_chans, n_classes=n_classes, input_time_length=input_time_length, final_conv_length='auto', n_filters_time=20, n_filters_spat=20, filter_time_length=10, pool_time_length=3, pool_time_stride=3, n_filters_2=50, filter_length_2=15, n_filters_3=100, filter_length_3=15, n_filters_4=400, filter_length_4=10, first_nonlin=leaky_relu, first_pool_mode='max', first_pool_nonlin=safe_log, later_nonlin=leaky_relu, later_pool_mode='max', later_pool_nonlin=safe_log, drop_prob=drop_prob, double_time_convs=False, split_first_layer=False, batch_norm=True, batch_norm_alpha=0.1, stride_before_pool=False).create_network() #filter_length_4 changed from 15 to 10 lr = hyp_params['lr_ada'][i] optimizer = optim.Adadelta(model.parameters(), lr=lr, weight_decay=0.1, eps=1e-8) elif model_type == 'eegnet': model = EEGNetv4(in_chans=n_chans, n_classes=n_classes, final_conv_length='auto', input_time_length=input_time_length, pool_mode='mean', F1=16, D=2, F2=32, kernel_length=64, third_kernel_size=(8,4), drop_prob=drop_prob).create_network() lr = hyp_params['lr_adam'][i] optimizer = optim.Adam(model.parameters(), lr=lr, weight_decay=0, eps=1e-8, amsgrad=False) set_random_seeds(seed=20190629, cuda=cuda) if cuda: model.cuda() torch.backends.cudnn.deterministic = True model = torch.nn.DataParallel(model) log.info("%s model: ".format(str(model))) loss_function = loss_function model_loss_function = None #####Setup to run the selected model##### model_test = Experiment(model, train_set, valid_set, test_set=None, iterator=iterator, loss_function=loss_function, optimizer=optimizer, model_constraint=model_constraint, monitors=monitors, stop_criterion=stop_criterion, remember_best_column='valid_misclass', run_after_early_stop=True, model_loss_function=model_loss_function, cuda=cuda, data_type=data_type, subject_id=subject_id, model_type=model_type, cropped=cropped, model_number=str(out_fold_num)) model_test.run() model_loss = model_test.epochs_df['valid_loss'].astype('float') current_val_loss = current_loss(model_loss) loss_with_params[f"Fold_{in_fold_num}"][f"{drop_prob}/{loss_function}/{lr}"] = current_val_loss ####Select and train optimized model##### df = pd.DataFrame(loss_with_params) df['mean'] = df.mean(axis=1) # compute mean loss across k-folds writer_df = f"results_folder\\results\\S{subject_id}\\{model_type}_parameters.xlsx" df.to_excel(writer_df) best_dp, best_loss, best_lr = df.loc[df['mean'].idxmin()].__dict__['_name'].split("/") # extract best param values if str(best_loss[10:13]) == 'nll': best_loss = F.nll_loss elif str(best_loss[10:13]) == 'cro': best_loss = F.cross_entropy print(f"Best parameters: dropout: {best_dp}, loss: {str(best_loss)[10:13]}, lr: {best_lr}") #####Train model on entire inner fold set##### torch.backends.cudnn.deterministic = True model = None #####Create outer-fold validation and test sets##### X_valid, X_test, y_valid, y_test = train_test_split(outer_fold, outer_labels, test_size=0.5, random_state=42, stratify=outer_labels) train_set = SignalAndTarget(inner_fold, inner_labels) valid_set = SignalAndTarget(X_valid, y_valid) test_set = SignalAndTarget(X_test, y_test) if model_type == 'shallow': model = ShallowFBCSPNet(in_chans=n_chans, n_classes=n_classes, input_time_length=input_time_length, n_filters_time=60, filter_time_length=5, n_filters_spat=40, pool_time_length=50, pool_time_stride=15, final_conv_length='auto', conv_nonlin=relu6, pool_mode='mean', pool_nonlin=safe_log, split_first_layer=True, batch_norm=True, batch_norm_alpha=0.1, drop_prob=0.1).create_network() #50 works better than 75 optimizer = optim.Adadelta(model.parameters(), lr=2.0, rho=0.9, weight_decay=0.1, eps=1e-8) elif model_type == 'deep': model = Deep4Net(in_chans=n_chans, n_classes=n_classes, input_time_length=input_time_length, final_conv_length='auto', n_filters_time=20, n_filters_spat=20, filter_time_length=5, pool_time_length=3, pool_time_stride=3, n_filters_2=20, filter_length_2=5, n_filters_3=40, filter_length_3=5, n_filters_4=1500, filter_length_4=10, first_nonlin=leaky_relu, first_pool_mode='mean', first_pool_nonlin=safe_log, later_nonlin=leaky_relu, later_pool_mode='mean', later_pool_nonlin=safe_log, drop_prob=0.1, double_time_convs=False, split_first_layer=True, batch_norm=True, batch_norm_alpha=0.1, stride_before_pool=False).create_network() optimizer = AdamW(model.parameters(), lr=0.1, weight_decay=0) elif model_type == 'eegnet': model = EEGNetv4(in_chans=n_chans, n_classes=n_classes, final_conv_length='auto', input_time_length=input_time_length, pool_mode='mean', F1=16, D=2, F2=32, kernel_length=64, third_kernel_size=(8,4), drop_prob=0.1).create_network() optimizer = optim.Adam(model.parameters(), lr=0.1, weight_decay=0, eps=1e-8, amsgrad=False) if cuda: model.cuda() torch.backends.cudnn.deterministic = True #model = torch.nn.DataParallel(model) log.info("Optimized model") model_loss_function=None #####Setup to run the optimized model##### optimized_model = op_exp(model, train_set, valid_set, test_set=test_set, iterator=iterator, loss_function=best_loss, optimizer=optimizer, model_constraint=model_constraint, monitors=monitors, stop_criterion=stop_criterion, remember_best_column='valid_misclass', run_after_early_stop=True, model_loss_function=model_loss_function, cuda=cuda, data_type=data_type, subject_id=subject_id, model_type=model_type, cropped=cropped, model_number=str(out_fold_num)) optimized_model.run() log.info("Last 5 epochs") log.info("\n" + str(optimized_model.epochs_df.iloc[-5:])) writer = f"results_folder\\results\\S{subject_id}\\{data_type}_{model_type}_{str(out_fold_num)}.xlsx" optimized_model.epochs_df.iloc[-30:].to_excel(writer) accuracy = 1 - np.min(np.array(optimized_model.class_acc)) cv_scores.append(accuracy) # k accuracy scores for this param set. #####Print and store fold accuracies and mean accuracy##### print(f"Class Accuracy: {np.mean(np.array(cv_scores))}") results_df = pd.DataFrame(dict(cv_scores=cv_scores, cv_mean=np.mean(np.array(cv_scores)))) writer2 = f"results_folder\\results\\S{subject_id}\\{data_type}_{model_type}_cvscores.xlsx" results_df.to_excel(writer2) return optimized_model, np.mean(np.array(cv_scores))
batch_size = 60 max_epochs = 20000 max_increase_epochs = 360 model = ShallowFBCSPNet(in_chan, db.n_classes, input_time_length=time_steps, final_conv_length="auto").create_network() log.info("Model: \n{:s}".format(str(model))) optimizer = optim.Adam(model.parameters()) iterator = BalancedBatchSizeIterator(batch_size=batch_size) stop_criterion = Or([ MaxEpochs(max_epochs), NoDecrease("valid_misclass", max_increase_epochs), ]) monitors = [LossMonitor(), MisclassMonitor(), RuntimeMonitor()] model_constraint = MaxNormDefaultConstraint() exp = Experiment( model, train_set, valid_set, test_set, iterator=iterator, loss_function=F.nll_loss, optimizer=optimizer,
def run_exp( data_folders, n_recordings, sensor_types, n_chans, max_recording_mins, sec_to_cut, duration_recording_mins, test_recording_mins, max_abs_val, sampling_freq, divisor, test_on_eval, n_folds, i_test_fold, shuffle, model_name, n_start_chans, n_chan_factor, input_time_length, final_conv_length, model_constraint, init_lr, batch_size, max_epochs, cuda, ): import torch.backends.cudnn as cudnn cudnn.benchmark = True preproc_functions = [] preproc_functions.append(lambda data, fs: ( data[:, int(sec_to_cut * fs):-int(sec_to_cut * fs)], fs)) preproc_functions.append(lambda data, fs: (data[:, :int( duration_recording_mins * 60 * fs)], fs)) if max_abs_val is not None: preproc_functions.append( lambda data, fs: (np.clip(data, -max_abs_val, max_abs_val), fs)) preproc_functions.append(lambda data, fs: (resampy.resample( data, fs, sampling_freq, axis=1, filter='kaiser_fast'), sampling_freq)) if divisor is not None: preproc_functions.append(lambda data, fs: (data / divisor, fs)) dataset = DiagnosisSet(n_recordings=n_recordings, max_recording_mins=max_recording_mins, preproc_functions=preproc_functions, data_folders=data_folders, train_or_eval='train', sensor_types=sensor_types) if test_on_eval: if test_recording_mins is None: test_recording_mins = duration_recording_mins test_preproc_functions = copy(preproc_functions) test_preproc_functions[1] = lambda data, fs: (data[:, :int( test_recording_mins * 60 * fs)], fs) test_dataset = DiagnosisSet(n_recordings=n_recordings, max_recording_mins=None, preproc_functions=test_preproc_functions, data_folders=data_folders, train_or_eval='eval', sensor_types=sensor_types) X, y = dataset.load() max_shape = np.max([list(x.shape) for x in X], axis=0) assert max_shape[1] == int(duration_recording_mins * sampling_freq * 60) if test_on_eval: test_X, test_y = test_dataset.load() max_shape = np.max([list(x.shape) for x in test_X], axis=0) assert max_shape[1] == int(test_recording_mins * sampling_freq * 60) if not test_on_eval: splitter = TrainValidTestSplitter(n_folds, i_test_fold, shuffle=shuffle) train_set, valid_set, test_set = splitter.split(X, y) else: splitter = TrainValidSplitter(n_folds, i_valid_fold=i_test_fold, shuffle=shuffle) train_set, valid_set = splitter.split(X, y) test_set = SignalAndTarget(test_X, test_y) del test_X, test_y del X, y # shouldn't be necessary, but just to make sure set_random_seeds(seed=20170629, cuda=cuda) n_classes = 2 if model_name == 'shallow': model = ShallowFBCSPNet( in_chans=n_chans, n_classes=n_classes, n_filters_time=n_start_chans, n_filters_spat=n_start_chans, input_time_length=input_time_length, final_conv_length=final_conv_length).create_network() elif model_name == 'deep': model = Deep4Net(n_chans, n_classes, n_filters_time=n_start_chans, n_filters_spat=n_start_chans, input_time_length=input_time_length, n_filters_2=int(n_start_chans * n_chan_factor), n_filters_3=int(n_start_chans * (n_chan_factor**2.0)), n_filters_4=int(n_start_chans * (n_chan_factor**3.0)), final_conv_length=final_conv_length, stride_before_pool=True).create_network() elif (model_name == 'deep_smac'): if model_name == 'deep_smac': do_batch_norm = False else: assert model_name == 'deep_smac_bnorm' do_batch_norm = True double_time_convs = False drop_prob = 0.244445 filter_length_2 = 12 filter_length_3 = 14 filter_length_4 = 12 filter_time_length = 21 final_conv_length = 1 first_nonlin = elu first_pool_mode = 'mean' first_pool_nonlin = identity later_nonlin = elu later_pool_mode = 'mean' later_pool_nonlin = identity n_filters_factor = 1.679066 n_filters_start = 32 pool_time_length = 1 pool_time_stride = 2 split_first_layer = True n_chan_factor = n_filters_factor n_start_chans = n_filters_start model = Deep4Net(n_chans, n_classes, n_filters_time=n_start_chans, n_filters_spat=n_start_chans, input_time_length=input_time_length, n_filters_2=int(n_start_chans * n_chan_factor), n_filters_3=int(n_start_chans * (n_chan_factor**2.0)), n_filters_4=int(n_start_chans * (n_chan_factor**3.0)), final_conv_length=final_conv_length, batch_norm=do_batch_norm, double_time_convs=double_time_convs, drop_prob=drop_prob, filter_length_2=filter_length_2, filter_length_3=filter_length_3, filter_length_4=filter_length_4, filter_time_length=filter_time_length, first_nonlin=first_nonlin, first_pool_mode=first_pool_mode, first_pool_nonlin=first_pool_nonlin, later_nonlin=later_nonlin, later_pool_mode=later_pool_mode, later_pool_nonlin=later_pool_nonlin, pool_time_length=pool_time_length, pool_time_stride=pool_time_stride, split_first_layer=split_first_layer, stride_before_pool=True).create_network() elif model_name == 'shallow_smac': conv_nonlin = identity do_batch_norm = True drop_prob = 0.328794 filter_time_length = 56 final_conv_length = 22 n_filters_spat = 73 n_filters_time = 24 pool_mode = 'max' pool_nonlin = identity pool_time_length = 84 pool_time_stride = 3 split_first_layer = True model = ShallowFBCSPNet( in_chans=n_chans, n_classes=n_classes, n_filters_time=n_filters_time, n_filters_spat=n_filters_spat, input_time_length=input_time_length, final_conv_length=final_conv_length, conv_nonlin=conv_nonlin, batch_norm=do_batch_norm, drop_prob=drop_prob, filter_time_length=filter_time_length, pool_mode=pool_mode, pool_nonlin=pool_nonlin, pool_time_length=pool_time_length, pool_time_stride=pool_time_stride, split_first_layer=split_first_layer, ).create_network() elif model_name == 'linear': model = nn.Sequential() model.add_module("conv_classifier", nn.Conv2d(n_chans, n_classes, (600, 1))) model.add_module('softmax', nn.LogSoftmax()) model.add_module('squeeze', Expression(lambda x: x.squeeze(3))) else: assert False, "unknown model name {:s}".format(model_name) to_dense_prediction_model(model) log.info("Model:\n{:s}".format(str(model))) if cuda: model.cuda() # determine output size test_input = np_to_var( np.ones((2, n_chans, input_time_length, 1), dtype=np.float32)) if cuda: test_input = test_input.cuda() log.info("In shape: {:s}".format(str(test_input.cpu().data.numpy().shape))) out = model(test_input) log.info("Out shape: {:s}".format(str(out.cpu().data.numpy().shape))) n_preds_per_input = out.cpu().data.numpy().shape[2] log.info("{:d} predictions per input/trial".format(n_preds_per_input)) iterator = CropsFromTrialsIterator(batch_size=batch_size, input_time_length=input_time_length, n_preds_per_input=n_preds_per_input) optimizer = optim.Adam(model.parameters(), lr=init_lr) loss_function = lambda preds, targets: F.nll_loss( th.mean(preds, dim=2, keepdim=False), targets) if model_constraint is not None: assert model_constraint == 'defaultnorm' model_constraint = MaxNormDefaultConstraint() monitors = [ LossMonitor(), MisclassMonitor(col_suffix='sample_misclass'), CroppedDiagnosisMonitor(input_time_length, n_preds_per_input), RuntimeMonitor(), ] stop_criterion = MaxEpochs(max_epochs) batch_modifier = None run_after_early_stop = True exp = Experiment(model, train_set, valid_set, test_set, iterator, loss_function, optimizer, model_constraint, monitors, stop_criterion, remember_best_column='valid_misclass', run_after_early_stop=run_after_early_stop, batch_modifier=batch_modifier, cuda=cuda) exp.run() return exp
def run_exp(data_folder, subject_id, low_cut_hz, model, cuda): ival = [-500, 4000] input_time_length = 1000 max_epochs = 800 max_increase_epochs = 80 batch_size = 60 high_cut_hz = 38 factor_new = 1e-3 init_block_size = 1000 valid_set_fraction = 0.2 train_filename = 'A{:02d}T.gdf'.format(subject_id) test_filename = 'A{:02d}E.gdf'.format(subject_id) train_filepath = os.path.join(data_folder, train_filename) test_filepath = os.path.join(data_folder, test_filename) train_label_filepath = train_filepath.replace('.gdf', '.mat') test_label_filepath = test_filepath.replace('.gdf', '.mat') train_loader = BCICompetition4Set2A(train_filepath, labels_filename=train_label_filepath) test_loader = BCICompetition4Set2A(test_filepath, labels_filename=test_label_filepath) train_cnt = train_loader.load() test_cnt = test_loader.load() # Preprocessing train_cnt = train_cnt.drop_channels( ['STI 014', 'EOG-left', 'EOG-central', 'EOG-right']) assert len(train_cnt.ch_names) == 22 # lets convert to millvolt for numerical stability of next operations train_cnt = mne_apply(lambda a: a * 1e6, train_cnt) train_cnt = mne_apply( lambda a: bandpass_cnt(a, low_cut_hz, high_cut_hz, train_cnt.info['sfreq'], filt_order=3, axis=1), train_cnt) train_cnt = mne_apply( lambda a: exponential_running_standardize(a.T, factor_new=factor_new, init_block_size= init_block_size, eps=1e-4).T, train_cnt) test_cnt = test_cnt.drop_channels( ['STI 014', 'EOG-left', 'EOG-central', 'EOG-right']) assert len(test_cnt.ch_names) == 22 test_cnt = mne_apply(lambda a: a * 1e6, test_cnt) test_cnt = mne_apply( lambda a: bandpass_cnt(a, low_cut_hz, high_cut_hz, test_cnt.info['sfreq'], filt_order=3, axis=1), test_cnt) test_cnt = mne_apply( lambda a: exponential_running_standardize(a.T, factor_new=factor_new, init_block_size= init_block_size, eps=1e-4).T, test_cnt) marker_def = OrderedDict([('Left Hand', [1]), ( 'Right Hand', [2], ), ('Foot', [3]), ('Tongue', [4])]) train_set = create_signal_target_from_raw_mne(train_cnt, marker_def, ival) test_set = create_signal_target_from_raw_mne(test_cnt, marker_def, ival) train_set, valid_set = split_into_two_sets(train_set, first_set_fraction=1 - valid_set_fraction) set_random_seeds(seed=20190706, cuda=cuda) n_classes = 4 n_chans = int(train_set.X.shape[1]) if model == 'shallow': model = ShallowFBCSPNet(n_chans, n_classes, input_time_length=input_time_length, final_conv_length=30).create_network() elif model == 'deep': model = Deep4Net(n_chans, n_classes, input_time_length=input_time_length, final_conv_length=2).create_network() to_dense_prediction_model(model) if cuda: model.cuda() log.info("Model: \n{:s}".format(str(model))) dummy_input = np_to_var(train_set.X[:1, :, :, None]) if cuda: dummy_input = dummy_input.cuda() out = model(dummy_input) n_preds_per_input = out.cpu().data.numpy().shape[2] optimizer = optim.Adam(model.parameters()) iterator = CropsFromTrialsIterator(batch_size=batch_size, input_time_length=input_time_length, n_preds_per_input=n_preds_per_input) stop_criterion = Or([ MaxEpochs(max_epochs), NoDecrease('valid_misclass', max_increase_epochs) ]) monitors = [ LossMonitor(), MisclassMonitor(col_suffix='sample_misclass'), CroppedTrialMisclassMonitor(input_time_length=input_time_length), RuntimeMonitor() ] model_constraint = MaxNormDefaultConstraint() loss_function = lambda preds, targets: F.nll_loss( th.mean(preds, dim=2, keepdim=False), targets) exp = Experiment(model, train_set, valid_set, test_set, iterator=iterator, loss_function=loss_function, optimizer=optimizer, model_constraint=model_constraint, monitors=monitors, stop_criterion=stop_criterion, remember_best_column='valid_misclass', run_after_early_stop=True, cuda=cuda) exp.run() return exp
# %% Loss function takes predictions as they come out of the network and the targets and returns a loss loss_function = F.mse_loss # Could be used to apply some constraint on the models, then should be object with apply method that accepts a module model_constraint = None # %% Monitors log the training progress monitors = [ LossMonitor(), CorrelationMonitor1d(input_time_length), RuntimeMonitor(), ] # %% Stop criterion determines when the first stop happens stop_criterion = MaxEpochs(maxTrainEpochs) # %% re-initialize model model.load_state_dict(deepcopy(start_param_values)) if adam: optimizer = optim.Adam(model.parameters()) else: weight_decay = np.float32(2.0 * 0.001) init_lr = np.float32((1 / 32.0) * 0.01) np_th_seed = 321312 scheduler_name = 'cosine' schedule_weight_decay = True optimizer_name = 'adamw' if optimizer_name == 'adam':
def train_model(self, model, dataset, state=None, final_evaluation=False, ensemble=False): if type(model) == list: model = AveragingEnsemble(model) if self.cuda: for mod in model.models: mod.cuda() if self.cuda: torch.cuda.empty_cache() if final_evaluation: self.stop_criterion = Or([ MaxEpochs(global_vars.get('final_max_epochs')), NoIncreaseDecrease( f'valid_{global_vars.get("nn_objective")}', global_vars.get('final_max_increase_epochs'), oper=get_oper_by_loss_function(self.loss_function)) ]) if global_vars.get('cropping'): self.set_cropping_for_model(model) self.epochs_df = pd.DataFrame() if global_vars.get('do_early_stop') or global_vars.get( 'remember_best'): self.rememberer = RememberBest( f"valid_{global_vars.get('nn_objective')}", oper=get_oper_by_loss_function(self.loss_function, equals=True)) self.optimizer = optim.Adam(model.parameters()) if self.cuda: assert torch.cuda.is_available(), "Cuda not available" if torch.cuda.device_count() > 1 and global_vars.get( 'parallel_gpu'): model.cuda() with torch.cuda.device(0): model = nn.DataParallel( model.cuda(), device_ids=[ int(s) for s in global_vars.get('gpu_select').split(',') ]) else: model.cuda() try: if global_vars.get('inherit_weights_normal') and state is not None: current_state = model.state_dict() for k, v in state.items(): if k in current_state and current_state[k].shape == v.shape: current_state.update({k: v}) model.load_state_dict(current_state) except Exception as e: print(f'failed weight inheritance\n,' f'state dict: {state.keys()}\n' f'current model state: {model.state_dict().keys()}') print('load state dict failed. Exception message: %s' % (str(e))) pdb.set_trace() self.monitor_epoch(dataset, model) if global_vars.get('log_epochs'): self.log_epoch() if global_vars.get('remember_best'): self.rememberer.remember_epoch(self.epochs_df, model, self.optimizer) self.iterator.reset_rng() start = time.time() num_epochs = self.run_until_stop(model, dataset) self.setup_after_stop_training(model, final_evaluation) if final_evaluation: dataset_train_backup = deepcopy(dataset['train']) if ensemble: self.run_one_epoch(dataset, model) self.rememberer.remember_epoch(self.epochs_df, model, self.optimizer, force=ensemble) num_epochs += 1 else: dataset['train'] = concatenate_sets( [dataset['train'], dataset['valid']]) num_epochs += self.run_until_stop(model, dataset) self.rememberer.reset_to_best_model(self.epochs_df, model, self.optimizer) dataset['train'] = dataset_train_backup end = time.time() self.final_time = end - start self.num_epochs = num_epochs return model
def run_exp(data_folder, session_id, subject_id, low_cut_hz, model, cuda): ival = [-500, 4000] max_epochs = 1600 max_increase_epochs = 160 batch_size = 10 high_cut_hz = 38 factor_new = 1e-3 init_block_size = 1000 valid_set_fraction = .2 ''' # BCIcompetition train_filename = 'A{:02d}T.gdf'.format(subject_id) test_filename = 'A{:02d}E.gdf'.format(subject_id) train_filepath = os.path.join(data_folder, train_filename) test_filepath = os.path.join(data_folder, test_filename) train_label_filepath = train_filepath.replace('.gdf', '.mat') test_label_filepath = test_filepath.replace('.gdf', '.mat') train_loader = BCICompetition4Set2A( train_filepath, labels_filename=train_label_filepath) test_loader = BCICompetition4Set2A( test_filepath, labels_filename=test_label_filepath) train_cnt = train_loader.load() test_cnt = test_loader.load() ''' # GIGAscience filename = 'sess{:02d}_subj{:02d}_EEG_MI.mat'.format( session_id, subject_id) filepath = os.path.join(data_folder, filename) train_variable = 'EEG_MI_train' test_variable = 'EEG_MI_test' train_loader = GIGAscience(filepath, train_variable) test_loader = GIGAscience(filepath, test_variable) train_cnt = train_loader.load() test_cnt = test_loader.load() # Preprocessing ''' channel ['Fp1', 'Fp2', 'F7', 'F3', 'Fz', 'F4', 'F8', 'FC5', 'FC1', 'FC2', 'FC6', 'T7', 'C3', 'Cz', 'C4', 'T8', 'TP9', 'CP5', 'CP1', 'CP2', 'CP6', 'TP10', 'P7', 'P3', 'Pz', 'P4', 'P8', 'PO9', 'O1', 'Oz', 'O2', 'PO10', 'FC3', 'FC4', 'C5', 'C1', 'C2', 'C6', 'CP3', 'CPz', 'CP4', 'P1', 'P2', 'POz', 'FT9', 'FTT9h', 'TTP7h', 'TP7', 'TPP9h', 'FT10', 'FTT10h', 'TPP8h', 'TP8', 'TPP10h', 'F9', 'F10', 'AF7', 'AF3', 'AF4', 'AF8', 'PO3', 'PO4'] ''' train_cnt = train_cnt.pick_channels([ 'FC5', 'FC3', 'FC1', 'Fz', 'FC2', 'FC4', 'FC6', 'C5', 'C3', 'C1', 'Cz', 'C2', 'C4', 'C6', 'CP5', 'CP3', 'CP1', 'CPz', 'CP2', 'CP4', 'CP6', 'Pz' ]) train_cnt, train_cnt.info['events'] = train_cnt.copy().resample( 250, npad='auto', events=train_cnt.info['events']) assert len(train_cnt.ch_names) == 22 # lets convert to millvolt for numerical stability of next operations train_cnt = mne_apply(lambda a: a * 1e6, train_cnt) train_cnt = mne_apply( lambda a: bandpass_cnt(a, low_cut_hz, high_cut_hz, train_cnt.info['sfreq'], filt_order=3, axis=1), train_cnt) train_cnt = mne_apply( lambda a: exponential_running_standardize(a.T, factor_new=factor_new, init_block_size= init_block_size, eps=1e-4).T, train_cnt) test_cnt = test_cnt.pick_channels([ 'FC5', 'FC3', 'FC1', 'Fz', 'FC2', 'FC4', 'FC6', 'C5', 'C3', 'C1', 'Cz', 'C2', 'C4', 'C6', 'CP5', 'CP3', 'CP1', 'CPz', 'CP2', 'CP4', 'CP6', 'Pz' ]) test_cnt, test_cnt.info['events'] = test_cnt.copy().resample( 250, npad='auto', events=test_cnt.info['events']) assert len(test_cnt.ch_names) == 22 test_cnt = mne_apply(lambda a: a * 1e6, test_cnt) test_cnt = mne_apply( lambda a: bandpass_cnt(a, low_cut_hz, high_cut_hz, test_cnt.info['sfreq'], filt_order=3, axis=1), test_cnt) test_cnt = mne_apply( lambda a: exponential_running_standardize(a.T, factor_new=factor_new, init_block_size= init_block_size, eps=1e-4).T, test_cnt) marker_def = OrderedDict([('Right Hand', [1]), ('Left Hand', [2])]) train_set = create_signal_target_from_raw_mne(train_cnt, marker_def, ival) test_set = create_signal_target_from_raw_mne(test_cnt, marker_def, ival) train_set, valid_set = split_into_two_sets(train_set, first_set_fraction=1 - valid_set_fraction) set_random_seeds(seed=20190706, cuda=cuda) n_classes = 2 n_chans = int(train_set.X.shape[1]) input_time_length = train_set.X.shape[2] if model == 'shallow': model = ShallowFBCSPNet(n_chans, n_classes, input_time_length=input_time_length, final_conv_length='auto').create_network() elif model == 'deep': model = Deep4Net(n_chans, n_classes, input_time_length=input_time_length, final_conv_length='auto').create_network() if cuda: model.cuda() log.info("Model: \n{:s}".format(str(model))) optimizer = optim.Adam(model.parameters()) iterator = BalancedBatchSizeIterator(batch_size=batch_size) stop_criterion = Or([ MaxEpochs(max_epochs), NoDecrease('valid_misclass', max_increase_epochs) ]) monitors = [LossMonitor(), MisclassMonitor(), RuntimeMonitor()] model_constraint = MaxNormDefaultConstraint() exp = Experiment(model, train_set, valid_set, test_set, iterator=iterator, loss_function=F.nll_loss, optimizer=optimizer, model_constraint=model_constraint, monitors=monitors, stop_criterion=stop_criterion, remember_best_column='valid_misclass', run_after_early_stop=True, cuda=cuda) exp.run() return exp
def run_exp_on_high_gamma_dataset(train_filename, test_filename, low_cut_hz, model_name, max_epochs, max_increase_epochs, np_th_seed, debug): train_set, valid_set, test_set = load_train_valid_test( train_filename=train_filename, test_filename=test_filename, low_cut_hz=low_cut_hz, debug=debug) if debug: max_epochs = 4 set_random_seeds(np_th_seed, cuda=True) #torch.backends.cudnn.benchmark = True# sometimes crashes? n_classes = int(np.max(train_set.y) + 1) n_chans = int(train_set.X.shape[1]) input_time_length = 1000 if model_name == 'deep': model = Deep4Net(n_chans, n_classes, input_time_length=input_time_length, final_conv_length=2).create_network() elif model_name == 'shallow': model = ShallowFBCSPNet(n_chans, n_classes, input_time_length=input_time_length, final_conv_length=30).create_network() to_dense_prediction_model(model) model.cuda() model.eval() out = model(np_to_var(train_set.X[:1, :, :input_time_length, None]).cuda()) n_preds_per_input = out.cpu().data.numpy().shape[2] optimizer = optim.Adam(model.parameters(), weight_decay=0, lr=1e-3) iterator = CropsFromTrialsIterator(batch_size=60, input_time_length=input_time_length, n_preds_per_input=n_preds_per_input, seed=np_th_seed) monitors = [ LossMonitor(), MisclassMonitor(col_suffix='sample_misclass'), CroppedTrialMisclassMonitor(input_time_length=input_time_length), RuntimeMonitor() ] model_constraint = MaxNormDefaultConstraint() loss_function = lambda preds, targets: F.nll_loss(th.mean(preds, dim=2), targets) run_after_early_stop = True do_early_stop = True remember_best_column = 'valid_misclass' stop_criterion = Or([ MaxEpochs(max_epochs), NoDecrease('valid_misclass', max_increase_epochs) ]) exp = Experiment(model, train_set, valid_set, test_set, iterator=iterator, loss_function=loss_function, optimizer=optimizer, model_constraint=model_constraint, monitors=monitors, stop_criterion=stop_criterion, remember_best_column=remember_best_column, run_after_early_stop=run_after_early_stop, cuda=True, do_early_stop=do_early_stop) exp.run() return exp
def run_exp(max_recording_mins, n_recordings, sec_to_cut_at_start, sec_to_cut_at_end, duration_recording_mins, max_abs_val, clip_before_resample, sampling_freq, divisor, n_folds, i_test_fold, shuffle, merge_train_valid, model, input_time_length, optimizer, learning_rate, weight_decay, scheduler, model_constraint, batch_size, max_epochs, only_return_exp, time_cut_off_sec, start_time, test_on_eval, test_recording_mins, sensor_types, log_dir, np_th_seed, cuda=True): import torch.backends.cudnn as cudnn cudnn.benchmark = True if optimizer == 'adam': assert merge_train_valid == False else: assert optimizer == 'adamw' assert merge_train_valid == True preproc_functions = create_preproc_functions( sec_to_cut_at_start=sec_to_cut_at_start, sec_to_cut_at_end=sec_to_cut_at_end, duration_recording_mins=duration_recording_mins, max_abs_val=max_abs_val, clip_before_resample=clip_before_resample, sampling_freq=sampling_freq, divisor=divisor) dataset = DiagnosisSet(n_recordings=n_recordings, max_recording_mins=max_recording_mins, preproc_functions=preproc_functions, train_or_eval='train', sensor_types=sensor_types) if test_on_eval: if test_recording_mins is None: test_recording_mins = duration_recording_mins test_preproc_functions = create_preproc_functions( sec_to_cut_at_start=sec_to_cut_at_start, sec_to_cut_at_end=sec_to_cut_at_end, duration_recording_mins=test_recording_mins, max_abs_val=max_abs_val, clip_before_resample=clip_before_resample, sampling_freq=sampling_freq, divisor=divisor) test_dataset = DiagnosisSet(n_recordings=n_recordings, max_recording_mins=None, preproc_functions=test_preproc_functions, train_or_eval='eval', sensor_types=sensor_types) if not only_return_exp: X, y = dataset.load() max_shape = np.max([list(x.shape) for x in X], axis=0) assert max_shape[1] == int(duration_recording_mins * sampling_freq * 60) if test_on_eval: test_X, test_y = test_dataset.load() max_shape = np.max([list(x.shape) for x in test_X], axis=0) assert max_shape[1] == int(test_recording_mins * sampling_freq * 60) if not test_on_eval: splitter = TrainValidTestSplitter(n_folds, i_test_fold, shuffle=shuffle) else: splitter = TrainValidSplitter(n_folds, i_valid_fold=i_test_fold, shuffle=shuffle) if not only_return_exp: if not test_on_eval: train_set, valid_set, test_set = splitter.split(X, y) else: train_set, valid_set = splitter.split(X, y) test_set = SignalAndTarget(test_X, test_y) del test_X, test_y del X, y # shouldn't be necessary, but just to make sure if merge_train_valid: train_set = concatenate_sets([train_set, valid_set]) # just reduce valid for faster computations valid_set.X = valid_set.X[:8] valid_set.y = valid_set.y[:8] # np.save('/data/schirrmr/schirrmr/auto-diag/lukasrepr/compare/mne-0-16-2/train_X.npy', train_set.X) # np.save('/data/schirrmr/schirrmr/auto-diag/lukasrepr/compare/mne-0-16-2/train_y.npy', train_set.y) # np.save('/data/schirrmr/schirrmr/auto-diag/lukasrepr/compare/mne-0-16-2/valid_X.npy', valid_set.X) # np.save('/data/schirrmr/schirrmr/auto-diag/lukasrepr/compare/mne-0-16-2/valid_y.npy', valid_set.y) # np.save('/data/schirrmr/schirrmr/auto-diag/lukasrepr/compare/mne-0-16-2/test_X.npy', test_set.X) # np.save('/data/schirrmr/schirrmr/auto-diag/lukasrepr/compare/mne-0-16-2/test_y.npy', test_set.y) else: train_set = None valid_set = None test_set = None log.info("Model:\n{:s}".format(str(model))) if cuda: model.cuda() model.eval() in_chans = 21 # determine output size test_input = np_to_var( np.ones((2, in_chans, input_time_length, 1), dtype=np.float32)) if cuda: test_input = test_input.cuda() out = model(test_input) n_preds_per_input = out.cpu().data.numpy().shape[2] log.info("{:d} predictions per input/trial".format(n_preds_per_input)) iterator = CropsFromTrialsIterator(batch_size=batch_size, input_time_length=input_time_length, n_preds_per_input=n_preds_per_input, seed=np_th_seed) assert optimizer in ['adam', 'adamw'], ("Expect optimizer to be either " "adam or adamw") schedule_weight_decay = optimizer == 'adamw' if optimizer == 'adam': optim_class = optim.Adam assert schedule_weight_decay == False assert merge_train_valid == False else: optim_class = AdamW assert schedule_weight_decay == True assert merge_train_valid == True optimizer = optim_class(model.parameters(), lr=learning_rate, weight_decay=weight_decay) if scheduler is not None: assert scheduler == 'cosine' n_updates_per_epoch = sum( [1 for _ in iterator.get_batches(train_set, shuffle=True)]) # Adapt if you have a different number of epochs n_updates_per_period = n_updates_per_epoch * max_epochs scheduler = CosineAnnealing(n_updates_per_period) optimizer = ScheduledOptimizer( scheduler, optimizer, schedule_weight_decay=schedule_weight_decay) loss_function = nll_loss_on_mean if model_constraint is not None: assert model_constraint == 'defaultnorm' model_constraint = MaxNormDefaultConstraint() monitors = [ LossMonitor(), MisclassMonitor(col_suffix='sample_misclass'), CroppedDiagnosisMonitor(input_time_length, n_preds_per_input), RuntimeMonitor(), ] stop_criterion = MaxEpochs(max_epochs) loggers = [Printer(), TensorboardWriter(log_dir)] batch_modifier = None exp = Experiment(model, train_set, valid_set, test_set, iterator, loss_function, optimizer, model_constraint, monitors, stop_criterion, remember_best_column='valid_misclass', run_after_early_stop=True, batch_modifier=batch_modifier, cuda=cuda, loggers=loggers) if not only_return_exp: # Until first stop exp.setup_training() exp.monitor_epoch(exp.datasets) exp.log_epoch() exp.rememberer.remember_epoch(exp.epochs_df, exp.model, exp.optimizer) exp.iterator.reset_rng() while not exp.stop_criterion.should_stop(exp.epochs_df): if (time.time() - start_time) > time_cut_off_sec: log.info( "Ran out of time after {:.2f} sec.".format(time.time() - start_time)) return exp log.info("Still in time after {:.2f} sec.".format(time.time() - start_time)) exp.run_one_epoch(exp.datasets, remember_best=True) if (time.time() - start_time) > time_cut_off_sec: log.info("Ran out of time after {:.2f} sec.".format(time.time() - start_time)) return exp if not merge_train_valid: exp.setup_after_stop_training() # Run until second stop datasets = exp.datasets datasets['train'] = concatenate_sets( [datasets['train'], datasets['valid']]) exp.monitor_epoch(datasets) exp.log_epoch() exp.iterator.reset_rng() while not exp.stop_criterion.should_stop(exp.epochs_df): if (time.time() - start_time) > time_cut_off_sec: log.info("Ran out of time after {:.2f} sec.".format( time.time() - start_time)) return exp log.info("Still in time after {:.2f} sec.".format(time.time() - start_time)) exp.run_one_epoch(datasets, remember_best=False) else: exp.dataset = dataset exp.splitter = splitter if test_on_eval: exp.test_dataset = test_dataset return exp
def run_exp(data_folder, subject_id, low_cut_hz, model, cuda): train_filename = 'A{:02d}T.gdf'.format(subject_id) test_filename = 'A{:02d}E.gdf'.format(subject_id) train_filepath = os.path.join(data_folder, train_filename) test_filepath = os.path.join(data_folder, test_filename) train_label_filepath = train_filepath.replace('.gdf', '.mat') test_label_filepath = test_filepath.replace('.gdf', '.mat') train_loader = BCICompetition4Set2A(train_filepath, labels_filename=train_label_filepath) test_loader = BCICompetition4Set2A(test_filepath, labels_filename=test_label_filepath) train_cnt = train_loader.load() test_cnt = test_loader.load() # Preprocessing train_cnt = train_cnt.drop_channels( ['STI 014', 'EOG-left', 'EOG-central', 'EOG-right']) assert len(train_cnt.ch_names) == 22 # lets convert to millvolt for numerical stability of next operations train_cnt = mne_apply(lambda a: a * 1e6, train_cnt) train_cnt = mne_apply( lambda a: bandpass_cnt( a, low_cut_hz, 38, train_cnt.info['sfreq'], filt_order=3, axis=1), train_cnt) train_cnt = mne_apply( lambda a: exponential_running_standardize( a.T, factor_new=1e-3, init_block_size=1000, eps=1e-4).T, train_cnt) test_cnt = test_cnt.drop_channels( ['STI 014', 'EOG-left', 'EOG-central', 'EOG-right']) assert len(test_cnt.ch_names) == 22 test_cnt = mne_apply(lambda a: a * 1e6, test_cnt) test_cnt = mne_apply( lambda a: bandpass_cnt( a, low_cut_hz, 38, test_cnt.info['sfreq'], filt_order=3, axis=1), test_cnt) test_cnt = mne_apply( lambda a: exponential_running_standardize( a.T, factor_new=1e-3, init_block_size=1000, eps=1e-4).T, test_cnt) marker_def = OrderedDict([('Left Hand', [1]), ( 'Right Hand', [2], ), ('Foot', [3]), ('Tongue', [4])]) ival = [-500, 4000] train_set = create_signal_target_from_raw_mne(train_cnt, marker_def, ival) test_set = create_signal_target_from_raw_mne(test_cnt, marker_def, ival) train_set, valid_set = split_into_two_sets(train_set, first_set_fraction=0.8) set_random_seeds(seed=20190706, cuda=cuda) n_classes = 4 n_chans = int(train_set.X.shape[1]) input_time_length = train_set.X.shape[2] if model == 'shallow': model = ShallowFBCSPNet(n_chans, n_classes, input_time_length=input_time_length, final_conv_length='auto').create_network() elif model == 'deep': model = Deep4Net(n_chans, n_classes, input_time_length=input_time_length, final_conv_length='auto').create_network() if cuda: model.cuda() log.info("Model: \n{:s}".format(str(model))) optimizer = optim.Adam(model.parameters()) iterator = BalancedBatchSizeIterator(batch_size=60) stop_criterion = Or([MaxEpochs(1600), NoDecrease('valid_misclass', 160)]) monitors = [LossMonitor(), MisclassMonitor(), RuntimeMonitor()] model_constraint = MaxNormDefaultConstraint() exp = Experiment(model, train_set, valid_set, test_set, iterator=iterator, loss_function=F.nll_loss, optimizer=optimizer, model_constraint=model_constraint, monitors=monitors, stop_criterion=stop_criterion, remember_best_column='valid_misclass', run_after_early_stop=True, cuda=cuda) exp.run() return exp