def length_accuracy(dataset): """ generates the lengthwise accuracy of a dataset :param dataset: dataset to consider :return: """ model = load_model(constants.text_recognition) for x, y in dataset: # create the accuracy evaluation object accuracy = SparseCategoricalAccuracy() # make a prediction and update the state of the accuracy using it prediction = model.predict(x) accuracy.update_state(y, prediction) print(y[0].numpy()) print(np.argmax(prediction, axis=-1)[0]) print("sequences of length", x[1].shape[1] - 1, "had an accuracy of", accuracy.result().numpy())
def evaluate(self): results = [] for x_test, y_test in self.gen.evaluate(): m = SparseCategoricalAccuracy() predictions = [] predictions.append(self.models[0].predict(np.array(x_test))) predictions.append(self.models[0].predict(np.array(x_test))) # calculate average outcomes = average(predictions).numpy() print(outcomes) print(y_test) if len(results) > 0: curr_avg = sum(results) / len(results) print("Accuracy:", curr_avg) m.update_state( # We have changed y_true = [[2], [1], [3]] to the following y_true=y_test, y_pred=outcomes, sample_weight=[1, 1, 1]) results.append(m.result().numpy()) avg = sum(results) / len(results) return avg
def evaluate(self, dataset): accuracy_metric = SparseCategoricalAccuracy() for x, y in dataset: y_hat = self.predict(x) accuracy_metric.update_state(y, y_hat) return accuracy_metric.result().numpy()
def on_epoch_end(self, epoch, logs={}): y_pred = self.model.predict(self.X_val, verbose=0) scce = SparseCategoricalCrossentropy(from_logits=True) score = scce(self.y_val, y_pred).numpy() acc = SparseCategoricalAccuracy() acc.update_state(self.y_val, y_pred) print("\n") print("The loss is : {}, the accuracy is: {}".format( score, acc.result().numpy())) gc.collect() K.clear_session()
class SBVAT(SemiSupervisedModel): """ Implementation of sample-based Batch Virtual Adversarial Training Graph Convolutional Networks (SBVAT). `Batch Virtual Adversarial Training for Graph Convolutional Networks <https://arxiv.org/abs/1902.09192>` Tensorflow 1.x implementation: <https://github.com/thudzj/BVAT> """ def __init__(self, *graph, n_samples=50, adj_transform="normalize_adj", attr_transform=None, device='cpu:0', seed=None, name=None, **kwargs): """Create a sample-based Batch Virtual Adversarial Training Graph Convolutional Networks (SBVAT) model. This can be instantiated in several ways: model = SBVAT(graph) with a `graphgallery.data.Graph` instance representing A sparse, attributed, labeled graph. model = SBVAT(adj_matrix, attr_matrix, labels) where `adj_matrix` is a 2D Scipy sparse matrix denoting the graph, `attr_matrix` is a 2D Numpy array-like matrix denoting the node attributes, `labels` is a 1D Numpy array denoting the node labels. Parameters: ---------- graph: An instance of `graphgallery.data.Graph` or a tuple (list) of inputs. A sparse, attributed, labeled graph. n_samples (Positive integer, optional): The number of sampled subset nodes in the graph where the length of the shortest path between them is at least `4`. (default :obj: `50`) adj_transform: string, `transform`, or None. optional How to transform the adjacency matrix. See `graphgallery.transforms` (default: :obj:`'normalize_adj'` with normalize rate `-0.5`. i.e., math:: \hat{A} = D^{-\frac{1}{2}} A D^{-\frac{1}{2}}) attr_transform: string, `transform`, or None. optional How to transform the node attribute matrix. See `graphgallery.transforms` (default :obj: `None`) device: string. optional The device where the model is running on. You can specified `CPU` or `GPU` for the model. (default: :str: `CPU:0`, i.e., running on the 0-th `CPU`) seed: interger scalar. optional Used in combination with `tf.random.set_seed` & `np.random.seed` & `random.seed` to create a reproducible sequence of tensors across multiple calls. (default :obj: `None`, i.e., using random seed) name: string. optional Specified name for the model. (default: :str: `class.__name__`) kwargs: other customized keyword Parameters. """ super().__init__(*graph, device=device, seed=seed, name=name, **kwargs) self.adj_transform = T.get(adj_transform) self.attr_transform = T.get(attr_transform) self.n_samples = n_samples self.process() def process_step(self): graph = self.graph adj_matrix = self.adj_transform(graph.adj_matrix) attr_matrix = self.attr_transform(graph.attr_matrix) self.neighbors = find_4o_nbrs(adj_matrix) self.feature_inputs, self.structure_inputs = T.astensors( attr_matrix, adj_matrix, device=self.device) # use decorator to make sure all list arguments have the same length @EqualVarLength() def build(self, hiddens=[16], activations=['relu'], dropout=0.5, lr=0.01, l2_norm=5e-4, use_bias=False, p1=1., p2=1., n_power_iterations=1, epsilon=0.03, xi=1e-6): with tf.device(self.device): x = Input(batch_shape=[None, self.graph.n_attrs], dtype=self.floatx, name='attr_matrix') adj = Input(batch_shape=[None, None], dtype=self.floatx, sparse=True, name='adj_matrix') index = Input(batch_shape=[None], dtype=self.intx, name='node_index') GCN_layers = [] dropout_layers = [] for hidden, activation in zip(hiddens, activations): GCN_layers.append( GraphConvolution( hidden, activation=activation, use_bias=use_bias, kernel_regularizer=regularizers.l2(l2_norm))) dropout_layers.append(Dropout(rate=dropout)) GCN_layers.append( GraphConvolution(self.graph.n_classes, use_bias=use_bias)) self.GCN_layers = GCN_layers self.dropout_layers = dropout_layers logit = self.forward(x, adj) output = Gather()([logit, index]) model = Model(inputs=[x, adj, index], outputs=output) self.model = model self.train_metric = SparseCategoricalAccuracy() self.test_metric = SparseCategoricalAccuracy() self.loss_fn = SparseCategoricalCrossentropy(from_logits=True) self.optimizer = Adam(lr=lr) self.p1 = p1 # Alpha self.p2 = p2 # Beta self.xi = xi # Small constant for finite difference # Norm length for (virtual) adversarial training self.epsilon = epsilon self.n_power_iterations = n_power_iterations # Number of power iterations def forward(self, x, adj, training=True): h = x for dropout_layer, GCN_layer in zip(self.dropout_layers, self.GCN_layers[:-1]): h = GCN_layer([h, adj]) h = dropout_layer(h, training=training) h = self.GCN_layers[-1]([h, adj]) return h @tf.function def train_step(self, sequence): with tf.device(self.device): self.train_metric.reset_states() for inputs, labels in sequence: x, adj, index, adv_mask = inputs with tf.GradientTape() as tape: logit = self.forward(x, adj) output = tf.gather(logit, index) loss = self.loss_fn(labels, output) entropy_loss = entropy_y_x(logit) vat_loss = self.virtual_adversarial_loss(x, adj, logit=logit, adv_mask=adv_mask) loss += self.p1 * vat_loss + self.p2 * entropy_loss self.train_metric.update_state(labels, output) trainable_variables = self.model.trainable_variables gradients = tape.gradient(loss, trainable_variables) self.optimizer.apply_gradients( zip(gradients, trainable_variables)) return loss, self.train_metric.result() @tf.function def test_step(self, sequence): with tf.device(self.device): self.test_metric.reset_states() for inputs, labels in sequence: x, adj, index, _ = inputs logit = self.forward(x, adj, training=False) output = tf.gather(logit, index) loss = self.loss_fn(labels, output) self.test_metric.update_state(labels, output) return loss, self.test_metric.result() def virtual_adversarial_loss(self, x, adj, logit, adv_mask): d = tf.random.normal(shape=tf.shape(x), dtype=self.floatx) for _ in range(self.n_power_iterations): d = get_normalized_vector(d) * self.xi logit_p = logit with tf.GradientTape() as tape: tape.watch(d) logit_m = self.forward(x + d, adj) dist = kl_divergence_with_logit(logit_p, logit_m, adv_mask) grad = tape.gradient(dist, d) d = tf.stop_gradient(grad) r_vadv = get_normalized_vector(d) * self.epsilon logit_p = tf.stop_gradient(logit) logit_m = self.forward(x + r_vadv, adj) loss = kl_divergence_with_logit(logit_p, logit_m, adv_mask) return tf.identity(loss) def train_sequence(self, index): index = T.asintarr(index) labels = self.graph.labels[index] sequence = SBVATSampleSequence( [self.feature_inputs, self.structure_inputs, index], labels, neighbors=self.neighbors, n_samples=self.n_samples, device=self.device) return sequence def test_sequence(self, index): index = T.asintarr(index) labels = self.graph.labels[index] sequence = SBVATSampleSequence( [self.feature_inputs, self.structure_inputs, index], labels, neighbors=self.neighbors, n_samples=self.n_samples, resample=False, device=self.device) return sequence def predict_step(self, sequence): with tf.device(self.device): for inputs, _ in sequence: x, adj, index, adv_mask = inputs output = self.forward(x, adj, training=False) logit = tf.gather(output, index) if tf.is_tensor(logit): logit = logit.numpy() return logit
def train(input_params, train, test, valid, class_cnt): current_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # tensorboard train_log_dir = 'logs/gradient_tape/' + current_time + '/train' valid_log_dir = 'logs/gradient_tape/' + current_time + '/valid' test_log_dir = 'logs/gradient_tape/' + current_time + '/test' train_summary_writer = tf.summary.create_file_writer(train_log_dir) valid_summary_writer = tf.summary.create_file_writer(valid_log_dir) # test_summary_writer = tf.summary.create_file_writer(test_log_dir) # todo: create model with hyperparams with model_dir = '../data/models/params/current_time/' model_dir = '../data/models/model-' + current_time # Instantiate an optimizer. optimizer = Adam(learning_rate=0.001) # Instantiate a loss function. loss_fn = SparseCategoricalCrossentropy(from_logits=True) train_step = test_step = 0 # Prepare the metrics. #todo use same variable for all the acc_metrics. acc_metric = SparseCategoricalAccuracy() if utility.dir_empty(model_dir): # model definition mobilenet = MOBILENET(include_top=False, input_shape=(224, 224, 3), weights='imagenet', pooling='avg', dropout=0.001) mobilenet.summary() # select till which layer use mobilenet. base_model = Model(inputs=mobilenet.input, outputs=mobilenet.output) base_model.summary() model = Sequential([ base_model, Dropout(0.2), Dense(units=class_cnt, activation='softmax'), ]) model.summary() epochs = 200 for epoch in range(epochs): print("\nStart of epoch %d" % (epoch,)) for batch_idx, (x_batch_train, y_batch_train) in enumerate(train): with tf.GradientTape() as tape: # forward pass logits = model(x_batch_train, training=True) # compute loss for mini batch loss_value = loss_fn(y_batch_train, logits) grads = tape.gradient(loss_value, model.trainable_weights) optimizer.apply_gradients(zip(grads, model.trainable_weights)) # Update training metric. acc_metric.update_state(y_batch_train, logits) with train_summary_writer.as_default(): # import code; code.interact(local=dict(globals(), **locals())) #TODO: add the metrics for test too. #TODO: take the mean of the losses in every batch and then show, #TODO loss_value is last loss of the batch(only 1). tf.summary.scalar('loss', loss_value, step=train_step) tf.summary.scalar('accuracy', acc_metric.result(), step=train_step) train_step += 1 if batch_idx % 10 == 0: print("training loss for one batch at step %d: %.4f" % (batch_idx, float(loss_value))) # Display metrics at the end of each epoch. print("Training acc over epoch: %.4f" % (float(acc_metric.result()),)) # Reset training metrics at the end of each epoch acc_metric.reset_states() # iterate on validation for batch_idx, (x_batch_val, y_batch_val) in enumerate(valid): # val_logits: y_pred of the validation. val_logits = model(x_batch_val, training=False) loss = loss_fn(y_batch_val, val_logits) # Update val metrics acc_metric.update_state(y_batch_val, val_logits) with valid_summary_writer.as_default(): tf.summary.scalar('loss', loss, step=test_step) tf.summary.scalar('accuracy', acc_metric.result(), step=test_step) test_step += 1 print("Validation acc: %.4f" % (float(acc_metric.result()),)) # print(classification_report(y_batch_val, val_logits, target_names=labels)) acc_metric.reset_states() acc_metric.reset_states() model.save(model_dir + 'model') else: # if model_dir is not empty print("model already exist. loading model...") model = load_model(model_dir+'model')
class Trainer(): def __init__(self, fbnet, input_shape, initial_temperature=5, temperature_decay_rate=0.956, temperature_decay_steps=1, latency_alpha=0.2, latency_beta=0.6, weight_lr=0.01, weight_momentum=0.9, weight_decay=1e-4, theta_lr=1e-3, theta_beta1=0.9, theta_beta2=0.999, theta_decay=5e-4): self._epoch = 0 self.initial_temperature = initial_temperature self.temperature = initial_temperature self.latency_alpha = latency_alpha self.latency_beta = latency_beta self.exponential_decay = lambda step: exponential_decay( initial_temperature, temperature_decay_rate, temperature_decay_steps, step) fbnet.build(input_shape) self.fbnet = fbnet self.weights = [] self.thetas = [] for trainable_weight in fbnet.trainable_weights: if 'theta' in trainable_weight.name: self.thetas.append(trainable_weight) else: self.weights.append(trainable_weight) self.weight_opt = SGD(learning_rate=weight_lr, momentum=weight_momentum, decay=weight_decay) self.theta_opt = Adam(learning_rate=theta_lr, beta_1=theta_beta1, beta_2=theta_beta2, decay=theta_decay) self.loss_fn = SparseCategoricalCrossentropy(from_logits=True) self.accuracy_metric = SparseCategoricalAccuracy() self.loss_metric = Mean() @property def epoch(self): return self._epoch @epoch.setter def epoch(self, epoch): self._epoch = epoch self.temperature = self.exponential_decay(epoch) def reset_metrics(self): self.accuracy_metric.reset_states() self.loss_metric.reset_states() def _train(self, x, y, weights, opt, training=True): with tf.GradientTape() as tape: y_hat = self.fbnet(x, self.temperature, training=training) loss = self.loss_fn(y, y_hat) latency = sum(self.fbnet.losses) loss += latency_loss(latency, self.latency_alpha, self.latency_beta) grads = tape.gradient(loss, weights) opt.apply_gradients(zip(grads, weights)) self.accuracy_metric.update_state(y, y_hat) self.loss_metric.update_state(loss) @tf.function def train_weights(self, x, y): self._train(x, y, self.weights, self.weight_opt) @tf.function def train_thetas(self, x, y): self._train(x, y, self.thetas, self.theta_opt, training=False) @property def training_accuracy(self): return self.accuracy_metric.result().numpy() @property def training_loss(self): return self.loss_metric.result().numpy() @tf.function def predict(self, x): y_hat = self.fbnet(x, self.temperature, training=False) return y_hat def evaluate(self, dataset): accuracy_metric = SparseCategoricalAccuracy() for x, y in dataset: y_hat = self.predict(x) accuracy_metric.update_state(y, y_hat) return accuracy_metric.result().numpy() def sample_sequential_config(self): ops = [ op.sample(self.temperature) if isinstance(op, MixedOperation) else op for op in self.fbnet.ops ] sequential_config = { 'name': 'sampled_fbnet', 'layers': [{ 'class_name': type(op).__name__, 'config': op.get_config() } for op in ops if not isinstance(op, Identity)] } return sequential_config def save_weights(self, checkpoint): self.fbnet.save_weights(checkpoint, save_format='tf') def load_weights(self, checkpoint): self.fbnet.load_weights(checkpoint)
def sparse_categorical_accuracy(y_true, y_pred): m = SparseCategoricalAccuracy() m.update_state(y_true, y_pred) return m.result().numpy()
class DualStudent(Model): """" Dual Student for Automatic Speech Recognition (ASR). How to train: 1) set the optimizer by means of compile(), 2) use train() How to test: use test() Remarks: - Do not use fit() by Keras, use train() - Do not use evaluate() by Keras, use test() - Compiled metrics and loss (i.e. set by means of compile()) are not used Original proposal for image classification: https://arxiv.org/abs/1909.01804 """ def __init__(self, n_classes, n_hidden_layers=3, n_units=96, consistency_loss='mse', consistency_scale=10, stabilization_scale=100, xi=0.6, padding_value=0., sigma=0.01, schedule='rampup', schedule_length=5, version='mono_directional'): """ Constructs a Dual Student model. :param n_classes: number of classes (i.e. number of units in the last layer of each student) :param n_hidden_layers: number of hidden layers in each student (i.e. LSTM layers) :param n_units: number of units for each hidden layer :param consistency_loss: one of 'mse', 'kl' :param consistency_scale: maximum value of weight for consistency constraint :param stabilization_scale: maximum value of weight for stabilization constraint :param xi: threshold for stable sample :param padding_value: value used to pad input sequences (used as mask_value for Masking layer) :param sigma: standard deviation for noisy augmentation :param schedule: type of schedule for lambdas, one of 'rampup', 'triangular_cycling', 'sinusoidal_cycling' :param schedule_length: :param version: one of: - 'mono_directional': both students have mono-directional LSTM layers - 'bidirectional: both students have bidirectional LSTM layers - 'imbalanced': one student has mono-directional LSTM layers, the other one bidirectional """ super(DualStudent, self).__init__() # store parameters self.n_classes = n_classes self.padding_value = padding_value self.n_units = n_units self.n_hidden_layers = n_hidden_layers self.xi = xi self.consistency_scale = consistency_scale self.stabilization_scale = stabilization_scale self.sigma = sigma self.version = version self.schedule = schedule self.schedule_length = schedule_length self._lambda1 = None self._lambda2 = None # schedule for lambdas if schedule == 'rampup': self.schedule_fn = sigmoid_rampup elif schedule == 'triangular_cycling': self.schedule_fn = triangular_cycling elif schedule == 'sinusoidal_cycling': self.schedule_fn = sinusoidal_cycling else: raise ValueError('Invalid schedule') # loss self._loss_cls = SparseCategoricalCrossentropy() # classification loss self._loss_sta = MeanSquaredError() # stabilization loss if consistency_loss == 'mse': self._loss_con = MeanSquaredError() # consistency loss elif consistency_loss == 'kl': self._loss_con = KLDivergence() else: raise ValueError('Invalid consistency metric') # metrics for training self._loss1 = Mean( name='loss1') # we want to average the loss for each batch self._loss2 = Mean(name='loss2') self._loss1_cls = Mean(name='loss1_cls') self._loss2_cls = Mean(name='loss2_cls') self._loss1_con = Mean(name='loss1_con') self._loss2_con = Mean(name='loss2_con') self._loss1_sta = Mean(name='loss1_sta') self._loss2_sta = Mean(name='loss2_sta') self._acc1 = SparseCategoricalAccuracy(name='acc1') self._acc2 = SparseCategoricalAccuracy(name='acc2') # metrics for testing self._test_loss1 = Mean(name='test_loss1') self._test_loss2 = Mean(name='test_loss2') self._test_acc1_train_phones = SparseCategoricalAccuracy( name='test_acc1_train_phones') self._test_acc2_train_phones = SparseCategoricalAccuracy( name='test_acc2_train_phones') self._test_acc1 = Accuracy(name='test_acc1') self._test_acc2 = Accuracy(name='test_acc2') self._test_per1 = PhoneErrorRate(name='test_per1') self._test_per2 = PhoneErrorRate(name='test_per2') # compose students if version == 'mono_directional': lstm_types = ['mono_directional', 'mono_directional'] elif version == 'bidirectional': lstm_types = ['bidirectional', 'bidirectional'] elif version == 'imbalanced': lstm_types = ['mono_directional', 'bidirectional'] else: raise ValueError('Invalid student version') self.student1 = self._get_student('student1', lstm_types[0]) self.student2 = self._get_student('student2', lstm_types[1]) # masking layer (just to use compute_mask and remove padding) self.mask = Masking(mask_value=self.padding_value) def _get_student(self, name, lstm_type): student = Sequential(name=name) student.add(Masking(mask_value=self.padding_value)) if lstm_type == 'mono_directional': for i in range(self.n_hidden_layers): student.add(LSTM(units=self.n_units, return_sequences=True)) elif lstm_type == 'bidirectional': for i in range(self.n_hidden_layers): student.add( Bidirectional( LSTM(units=self.n_units, return_sequences=True))) else: raise ValueError('Invalid LSTM version') student.add(Dense(units=self.n_classes, activation="softmax")) return student def _noisy_augment(self, x): return x + tf.random.normal(shape=x.shape, stddev=self.sigma) def call(self, inputs, training=False, student='student1', **kwargs): """ Feed-forwards inputs to one of the students. This function is called internally by __call__(). Do not use it directly, use the model as callable. You may prefer to use pad_and_predict() instead of this, because it pads the sequences and splits in batches. For a big dataset, it is strongly suggested that you use pad_and_predict(). :param inputs: tensor of shape (batch_size, n_frames, n_features) :param training: boolean, whether the call is in inference mode or training mode :param student: one of 'student1', 'student2' :return: tensor of shape (batch_size, n_frames, n_classes), softmax activations (probabilities) """ if student == 'student1': return self.student1(inputs, training=training) elif student != 'student1': return self.student2(inputs, training=training) else: raise ValueError('Invalid student') def build(self, input_shape): super(DualStudent, self).build(input_shape) self.student1.build(input_shape) self.student2.build(input_shape) def train(self, x_labeled, x_unlabeled, y_labeled, x_val=None, y_val=None, n_epochs=10, batch_size=32, shuffle=True, evaluation_mapping=None, logs_path=None, checkpoints_path=None, initial_epoch=0, seed=None): """ Trains the students with both labeled and unlabeled data (semi-supervised learning). :param x_labeled: numpy array of numpy arrays (n_frames, n_features), features corresponding to y_labeled. 'n_frames' can vary, padding is added to make x_labeled a tensor. :param x_unlabeled: numpy array of numpy arrays of shape (n_frames, n_features), features without labels. 'n_frames' can vary, padding is added to make x_unlabeled a tensor. :param y_labeled: numpy array of numpy arrays of shape (n_frames,), labels corresponding to x_labeled. 'n_frames' can vary, padding is added to make y_labeled a tensor. :param x_val: like x_labeled, but for validation set :param y_val: like y_labeled, but for validation set :param n_epochs: integer, number of training epochs :param batch_size: integer, batch size :param shuffle: boolean, whether to shuffle at each epoch or not :param evaluation_mapping: dictionary {training label -> test label}, the test phones should be a subset of the training phones :param logs_path: path where to save logs for TensorBoard :param checkpoints_path: path to a directory. If the directory contains checkpoints, the latest checkpoint is restored. :param initial_epoch: int, initial epoch from which to start the training. It can be used together with checkpoints_path to resume the training from a previous run. :param seed: seed for the random number generator """ # set seed if seed is not None: np.random.seed(seed) tf.random.set_seed(seed) # show summary self.build(input_shape=(None, ) + x_labeled[0].shape) self.student1.summary() self.student2.summary() # setup for logs train_summary_writer = None if logs_path is not None: train_summary_writer = tf.summary.create_file_writer(logs_path) # setup for checkpoints checkpoint = None if checkpoints_path is not None: checkpoint = tf.train.Checkpoint(optimizer=self.optimizer, model=self) checkpoint_path = tf.train.latest_checkpoint(checkpoints_path) if checkpoint_path is not None: checkpoint.restore(checkpoint_path) checkpoint_path = Path(checkpoints_path) / 'ckpt' checkpoint_path = str(checkpoint_path) # compute batch sizes labeled_batch_size = ceil( len(x_labeled) / (len(x_unlabeled) + len(x_labeled)) * batch_size) unlabeled_batch_size = batch_size - labeled_batch_size n_batches = min(ceil(len(x_unlabeled) / unlabeled_batch_size), ceil(len(x_labeled) / labeled_batch_size)) # training loop for epoch in trange(initial_epoch, n_epochs, desc='epochs'): # ramp up lambda1 and lambda2 self._lambda1 = self.consistency_scale * self.schedule_fn( epoch, self.schedule_length) self._lambda2 = self.stabilization_scale * self.schedule_fn( epoch, self.schedule_length) # shuffle training set if shuffle: indices = np.arange( len(x_labeled) ) # get indices to shuffle coherently features and labels np.random.shuffle(indices) x_labeled = x_labeled[indices] y_labeled = y_labeled[indices] np.random.shuffle(x_unlabeled) for i in trange(n_batches, desc='batches'): # select batch x_labeled_batch = select_batch(x_labeled, i, labeled_batch_size) x_unlabeled_batch = select_batch(x_unlabeled, i, unlabeled_batch_size) y_labeled_batch = select_batch(y_labeled, i, labeled_batch_size) # pad batch x_labeled_batch = pad_sequences(x_labeled_batch, padding='post', value=self.padding_value, dtype='float32') x_unlabeled_batch = pad_sequences(x_unlabeled_batch, padding='post', value=self.padding_value, dtype='float32') y_labeled_batch = pad_sequences(y_labeled_batch, padding='post', value=-1) # convert to tensors x_labeled_batch = tf.convert_to_tensor(x_labeled_batch) x_unlabeled_batch = tf.convert_to_tensor(x_unlabeled_batch) y_labeled_batch = tf.convert_to_tensor(y_labeled_batch) # train step self._train_step(x_labeled_batch, x_unlabeled_batch, y_labeled_batch) # put metrics in dictionary (easy management) train_metrics = { self._loss1.name: self._loss1.result(), self._loss2.name: self._loss2.result(), self._loss1_cls.name: self._loss1_cls.result(), self._loss2_cls.name: self._loss2_cls.result(), self._loss1_con.name: self._loss1_con.result(), self._loss2_con.name: self._loss2_con.result(), self._loss1_sta.name: self._loss1_sta.result(), self._loss2_sta.name: self._loss2_sta.result(), self._acc1.name: self._acc1.result(), self._acc2.name: self._acc2.result(), } metrics = {'train': train_metrics} # test on validation set if x_val is not None and y_val is not None: val_metrics = self.test(x_val, y_val, evaluation_mapping=evaluation_mapping) metrics['val'] = val_metrics # print metrics for dataset, metrics_ in metrics.items(): print(f'Epoch {epoch + 1} - ', dataset, ' - ', sep='', end='') for k, v in metrics_.items(): print(f'{k}: {v}, ', end='') print() # save logs if train_summary_writer is not None: with train_summary_writer.as_default(): for dataset, metrics_ in metrics.items(): for k, v in metrics_.items(): tf.summary.scalar(k, v, step=epoch) # save checkpoint if checkpoint is not None: checkpoint.save(file_prefix=checkpoint_path) # reset metrics self._loss1.reset_states() self._loss2.reset_states() self._loss1_cls.reset_states() self._loss2_cls.reset_states() self._loss1_con.reset_states() self._loss2_con.reset_states() self._loss1_sta.reset_states() self._loss2_sta.reset_states() self._acc1.reset_states() self._acc2.reset_states() """ If you want to use graph execution, pad the whole dataset externally and uncomment the decorator below. If you uncomment the decorator without padding the dataset, the graph will be compiled for each batch, because train() pads at batch level and so the batches have different shapes. This would result in worse performance compared to eager execution. """ # @tf.function def _train_step(self, x_labeled, x_unlabeled, y_labeled): # noisy augmented batches (TODO: improvement with data augmentation instead of noise) B1_labeled = self._noisy_augment(x_labeled) B2_labeled = self._noisy_augment(x_labeled) B1_unlabeled = self._noisy_augment(x_unlabeled) B2_unlabeled = self._noisy_augment(x_unlabeled) # compute masks (to remove padding) mask_labeled = self.mask.compute_mask(x_labeled) mask_unlabeled = self.mask.compute_mask(x_unlabeled) y_labeled = y_labeled[mask_labeled] # remove padding from labels # forward pass with tf.GradientTape(persistent=True) as tape: # predict augmented labeled samples (for classification and consistency constraint) prob1_labeled_B1 = self.student1(B1_labeled, training=True) prob1_labeled_B2 = self.student1(B2_labeled, training=True) prob2_labeled_B1 = self.student2(B1_labeled, training=True) prob2_labeled_B2 = self.student2(B2_labeled, training=True) # predict augmented unlabeled samples (for consistency and stabilization constraints) prob1_unlabeled_B1 = self.student1(B1_unlabeled, training=True) prob1_unlabeled_B2 = self.student1(B2_unlabeled, training=True) prob2_unlabeled_B1 = self.student2(B1_unlabeled, training=True) prob2_unlabeled_B2 = self.student2(B2_unlabeled, training=True) # remove padding prob1_labeled_B1 = prob1_labeled_B1[mask_labeled] prob1_labeled_B2 = prob1_labeled_B2[mask_labeled] prob2_labeled_B1 = prob2_labeled_B1[mask_labeled] prob2_labeled_B2 = prob2_labeled_B2[mask_labeled] prob1_unlabeled_B1 = prob1_unlabeled_B1[mask_unlabeled] prob1_unlabeled_B2 = prob1_unlabeled_B2[mask_unlabeled] prob2_unlabeled_B1 = prob2_unlabeled_B1[mask_unlabeled] prob2_unlabeled_B2 = prob2_unlabeled_B2[mask_unlabeled] # compute classification losses L1_cls = self._loss_cls(y_labeled, prob1_labeled_B1) L2_cls = self._loss_cls(y_labeled, prob2_labeled_B2) # concatenate labeled and unlabeled probability predictions (for consistency loss) prob1_labeled_unlabeled_B1 = tf.concat( [prob1_labeled_B1, prob1_unlabeled_B1], axis=0) prob1_labeled_unlabeled_B2 = tf.concat( [prob1_labeled_B2, prob1_unlabeled_B2], axis=0) prob2_labeled_unlabeled_B1 = tf.concat( [prob2_labeled_B1, prob2_unlabeled_B1], axis=0) prob2_labeled_unlabeled_B2 = tf.concat( [prob2_labeled_B2, prob2_unlabeled_B2], axis=0) # compute consistency losses L1_con = self._loss_con(prob1_labeled_unlabeled_B1, prob1_labeled_unlabeled_B2) L2_con = self._loss_con(prob2_labeled_unlabeled_B1, prob2_labeled_unlabeled_B2) # prediction P1_unlabeled_B1 = tf.argmax(prob1_unlabeled_B1, axis=-1) P1_unlabeled_B2 = tf.argmax(prob1_unlabeled_B2, axis=-1) P2_unlabeled_B1 = tf.argmax(prob2_unlabeled_B1, axis=-1) P2_unlabeled_B2 = tf.argmax(prob2_unlabeled_B2, axis=-1) # confidence (probability of predicted class) M1_unlabeled_B1 = tf.reduce_max(prob1_unlabeled_B1, axis=-1) M1_unlabeled_B2 = tf.reduce_max(prob1_unlabeled_B2, axis=-1) M2_unlabeled_B1 = tf.reduce_max(prob2_unlabeled_B1, axis=-1) M2_unlabeled_B2 = tf.reduce_max(prob2_unlabeled_B2, axis=-1) # stable samples (masks to index probabilities) R1 = tf.logical_and( P1_unlabeled_B1 == P1_unlabeled_B2, tf.logical_or(M1_unlabeled_B1 > self.xi, M1_unlabeled_B2 > self.xi)) R2 = tf.logical_and( P2_unlabeled_B1 == P2_unlabeled_B2, tf.logical_or(M2_unlabeled_B1 > self.xi, M2_unlabeled_B2 > self.xi)) R12 = tf.logical_and(R1, R2) # stabilities epsilon1 = MSE(prob1_unlabeled_B1[R12], prob1_unlabeled_B2[R12]) epsilon2 = MSE(prob2_unlabeled_B1[R12], prob2_unlabeled_B2[R12]) # compute stabilization losses L1_sta = self._loss_sta( prob1_unlabeled_B1[R12][epsilon1 > epsilon2], prob2_unlabeled_B1[R12][epsilon1 > epsilon2]) L2_sta = self._loss_sta( prob1_unlabeled_B2[R12][epsilon1 < epsilon2], prob2_unlabeled_B2[R12][epsilon1 < epsilon2]) L1_sta += self._loss_sta( prob1_unlabeled_B1[tf.logical_and(tf.logical_not(R1), R2)], prob2_unlabeled_B1[tf.logical_and(tf.logical_not(R1), R2)]) L2_sta += self._loss_sta( prob1_unlabeled_B2[tf.logical_and(R1, tf.logical_not(R2))], prob2_unlabeled_B2[tf.logical_and(R1, tf.logical_not(R2))]) # compute complete losses L1 = L1_cls + self._lambda1 * L1_con + self._lambda2 * L1_sta L2 = L2_cls + self._lambda1 * L2_con + self._lambda2 * L2_sta # backward pass gradients1 = tape.gradient(L1, self.student1.trainable_variables) gradients2 = tape.gradient(L2, self.student2.trainable_variables) self.optimizer.apply_gradients( zip(gradients1, self.student1.trainable_variables)) self.optimizer.apply_gradients( zip(gradients2, self.student2.trainable_variables)) del tape # to release memory (persistent tape) # update metrics self._loss1.update_state(L1) self._loss2.update_state(L2) self._loss1_cls.update_state(L1_cls) self._loss2_cls.update_state(L2_cls) self._loss1_con.update_state(L1_con) self._loss2_con.update_state(L2_con) self._loss1_sta.update_state(L1_sta) self._loss2_sta.update_state(L2_sta) self._acc1.update_state(y_labeled, prob1_labeled_B1) self._acc2.update_state(y_labeled, prob2_labeled_B2) def test(self, x, y, batch_size=32, evaluation_mapping=None): """ Tests the model (both students). :param x: numpy array of numpy arrays (n_frames, n_features), features corresponding to y_labeled. 'n_frames' can vary, padding is added to make x a tensor. :param y: numpy array of numpy arrays of shape (n_frames,), labels corresponding to x_labeled. 'n_frames' can vary, padding is added to make y a tensor. :param batch_size: integer, batch size :param evaluation_mapping: dictionary {training label -> test label}, the test phones should be a subset of the training phones :return: dictionary {metric_name -> value} """ # test batch by batch n_batches = ceil(len(x) / batch_size) for i in trange(n_batches, desc='test batches'): # select batch x_batch = select_batch(x, i, batch_size) y_batch = select_batch(y, i, batch_size) # pad batch x_batch = pad_sequences(x_batch, padding='post', value=self.padding_value, dtype='float32') y_batch = pad_sequences(y_batch, padding='post', value=-1) # convert to tensors x_batch = tf.convert_to_tensor(x_batch) y_batch = tf.convert_to_tensor(y_batch) # test step self._test_step(x_batch, y_batch, evaluation_mapping) # put metrics in dictionary (easy management) test_metrics = { self._test_loss1.name: self._test_loss1.result(), self._test_loss2.name: self._test_loss2.result(), self._test_acc1_train_phones.name: self._test_acc1_train_phones.result(), self._test_acc2_train_phones.name: self._test_acc2_train_phones.result(), self._test_acc1.name: self._test_acc1.result(), self._test_acc2.name: self._test_acc2.result(), self._test_per1.name: self._test_per1.result(), self._test_per2.name: self._test_per2.result(), } # reset metrics self._test_loss1.reset_states() self._test_loss2.reset_states() self._test_acc1_train_phones.reset_states() self._test_acc2_train_phones.reset_states() self._test_acc1.reset_states() self._test_acc2.reset_states() self._test_per1.reset_states() self._test_per2.reset_states() return test_metrics # @tf.function # see note in _train_step() def _test_step(self, x, y, evaluation_mapping): # compute mask (to remove padding) mask = self.mask.compute_mask(x) # forward pass y_prob1_train_phones = self.student1(x, training=False) y_prob2_train_phones = self.student2(x, training=False) y_pred1_train_phones = tf.argmax(y_prob1_train_phones, axis=-1) y_pred2_train_phones = tf.argmax(y_prob2_train_phones, axis=-1) y_train_phones = tf.identity(y) # map labels to set of test phones if evaluation_mapping is not None: y = tf.numpy_function(map_labels, [y_train_phones, evaluation_mapping], [tf.float32]) y_pred1 = tf.numpy_function( map_labels, [y_pred1_train_phones, evaluation_mapping], [tf.float32]) y_pred2 = tf.numpy_function( map_labels, [y_pred2_train_phones, evaluation_mapping], [tf.float32]) else: y = y_train_phones y_pred1 = y_pred1_train_phones y_pred2 = y_pred2_train_phones # update phone error rate self._test_per1.update_state(y, y_pred1, mask) self._test_per2.update_state(y, y_pred2, mask) # remove padding y_pred1 = y_pred1[mask] y_pred2 = y_pred2[mask] y_prob1_train_phones = y_prob1_train_phones[mask] y_prob2_train_phones = y_prob2_train_phones[mask] y_train_phones = y_train_phones[mask] y = y[mask] # compute loss loss1 = self._loss_cls(y_train_phones, y_prob1_train_phones) loss2 = self._loss_cls(y_train_phones, y_prob2_train_phones) # update loss self._test_loss1.update_state(loss1) self._test_loss2.update_state(loss2) # update accuracy using training phones self._test_acc1_train_phones.update_state(y_train_phones, y_prob1_train_phones) self._test_acc2_train_phones.update_state(y_train_phones, y_prob2_train_phones) # update accuracy using test phones self._test_acc1.update_state(y, y_pred1) self._test_acc2.update_state(y, y_pred2)
class SBVAT(SupervisedModel): """ Implementation of sample-based Batch Virtual Adversarial Training Graph Convolutional Networks (SBVAT). [Batch Virtual Adversarial Training for Graph Convolutional Networks](https://arxiv.org/pdf/1902.09192) Tensorflow 1.x implementation: https://github.com/thudzj/BVAT Arguments: ---------- adj: `scipy.sparse.csr_matrix` (or `csc_matrix`) with shape (N, N) The input `symmetric` adjacency matrix, where `N` is the number of nodes in graph. features: `np.array` with shape (N, F) The input node feature matrix, where `F` is the dimension of node features. labels: `np.array` with shape (N,) The ground-truth labels for all nodes in graph. n_samples (Positive integer, optional): The number of sampled subset nodes in the graph where the shortest path length between them is at least 4. (default :obj: `50`) normalize_rate (Float scalar, optional): The normalize rate for adjacency matrix `adj`. (default: :obj:`-0.5`, i.e., math:: \hat{A} = D^{-\frac{1}{2}} A D^{-\frac{1}{2}}) normalize_features (Boolean, optional): Whether to use row-normalize for node feature matrix. (default :obj: `True`) device (String, optional): The device where the model is running on. You can specified `CPU` or `GPU` for the model. (default: :obj: `CPU:0`, i.e., the model is running on the 0-th device `CPU`) seed (Positive integer, optional): Used in combination with `tf.random.set_seed & np.random.seed & random.seed` to create a reproducible sequence of tensors across multiple calls. (default :obj: `None`, i.e., using random seed) name (String, optional): Name for the model. (default: name of class) """ def __init__(self, adj, features, labels, n_samples=100, normalize_rate=-0.5, normalize_features=True, device='CPU:0', seed=None, **kwargs): super().__init__(adj, features, labels, device=device, seed=seed, **kwargs) self.normalize_rate = normalize_rate self.normalize_features = normalize_features self.preprocess(adj, features) self.n_samples = n_samples def preprocess(self, adj, features): if self.normalize_rate is not None: adj = self._normalize_adj(adj, self.normalize_rate) if self.normalize_features: features = self._normalize_features(features) self.neighbors = list(find_4o_nbrs(adj.indices, adj.indptr, np.arange(self.n_nodes))) with self.device: self.features, self.adj = self._to_tensor([features, adj]) def build(self, hidden_layers=[16], activations=['relu'], dropout=0.5, learning_rate=0.01, l2_norm=5e-4, p1=1., p2=1., n_power_iterations=1, epsilon=0.03, xi=1e-6): with self.device: x = Input(batch_shape=[self.n_nodes, self.n_features], dtype=tf.float32, name='features') adj = Input(batch_shape=[self.n_nodes, self.n_nodes], dtype=tf.float32, sparse=True, name='adj_matrix') index = Input(batch_shape=[None], dtype=tf.int32, name='index') self.GCN_layers = [GraphConvolution(hidden_layers[0], activation=activations[0], kernel_regularizer=regularizers.l2(l2_norm)), GraphConvolution(self.n_classes)] self.dropout_layer = Dropout(dropout) logit = self.propagation(x, adj) output = tf.gather(logit, index) output = Softmax()(output) model = Model(inputs=[x, adj, index], outputs=output) self.model = model self.train_metric = SparseCategoricalAccuracy() self.test_metric = SparseCategoricalAccuracy() self.optimizer = Adam(lr=learning_rate) self.built = True self.p1 = p1 # Alpha self.p2 = p2 # Beta self.xi = xi # Small constant for finite difference self.epsilon = epsilon # Norm length for (virtual) adversarial training self.n_power_iterations = n_power_iterations # Number of power iterations def propagation(self, x, adj, training=True): h = x for layer in self.GCN_layers: h = self.dropout_layer(h, training=training) h = layer([h, adj]) return h @tf.function def do_train_forward(self, sequence): with self.device: self.train_metric.reset_states() for inputs, labels in sequence: x, adj, index, adv_mask = inputs with tf.GradientTape() as tape: logit = self.propagation(x, adj) output = tf.gather(logit, index) output = softmax(output) loss = tf.reduce_mean(sparse_categorical_crossentropy(labels, output)) entropy_loss = entropy_y_x(logit) vat_loss = self.virtual_adversarial_loss(x, adj, logit=logit, adv_mask=adv_mask) loss += self.p1 * vat_loss + self.p2 * entropy_loss self.train_metric.update_state(labels, output) trainable_variables = self.model.trainable_variables gradients = tape.gradient(loss, trainable_variables) self.optimizer.apply_gradients(zip(gradients, trainable_variables)) return loss, self.train_metric.result() @tf.function def do_test_forward(self, sequence): with self.device: self.test_metric.reset_states() for inputs, labels in sequence: x, adj, index, _ = inputs logit = self.propagation(x, adj, training=False) output = tf.gather(logit, index) output = softmax(output) loss = tf.reduce_mean(sparse_categorical_crossentropy(labels, output)) self.test_metric.update_state(labels, output) return loss, self.test_metric.result() def do_forward(self, sequence, training=True): if training: loss, accuracy = self.do_train_forward(sequence) else: loss, accuracy = self.do_test_forward(sequence) return loss.numpy(), accuracy.numpy() def virtual_adversarial_loss(self, x, adj, logit, adv_mask): d = tf.random.normal(shape=tf.shape(x)) for _ in range(self.n_power_iterations): d = get_normalized_vector(d) * self.xi logit_p = logit with tf.GradientTape() as tape: tape.watch(d) logit_m = self.propagation(x + d, adj) dist = kl_divergence_with_logit(logit_p, logit_m, adv_mask) grad = tape.gradient(dist, d) d = tf.stop_gradient(grad) r_vadv = get_normalized_vector(d) * self.epsilon logit_p = tf.stop_gradient(logit) logit_m = self.propagation(x + r_vadv, adj) loss = kl_divergence_with_logit(logit_p, logit_m, adv_mask) return tf.identity(loss) def train_sequence(self, index): index = self._check_and_convert(index) labels = self.labels[index] with self.device: sequence = NodeSampleSequence([self.features, self.adj, index], labels, neighbors=self.neighbors, n_samples=self.n_samples) return sequence def test_sequence(self, index): index = self._check_and_convert(index) labels = self.labels[index] with self.device: sequence = NodeSampleSequence([self.features, self.adj, index], labels, neighbors=self.neighbors, n_samples=self.n_samples, resample=False) return sequence def predict(self, index): super().predict(index) index = self._check_and_convert(index) with self.device: sequence = NodeSampleSequence([self.features, self.adj, index], None, neighbors=self.neighbors, n_samples=self.n_samples, resample=False) for inputs, _ in sequence: x, adj, index, adv_mask = inputs output = self.propagation(x, adj, training=False) logit = softmax(tf.gather(output, index)) return logit.numpy()
# forward pass of student model student_pred = stud_model(x, training=True) assert stud_model.trainable == True, 'Student model should be trainable' # hard labels loss loss_hard = cross_entr(y, student_pred) # soft labels loss with temperature temp loss_soft = kl_div(tf.nn.softmax(teacher_pred / temp), tf.nn.softmax(student_pred / temp)) # final loss value loss = alpha * loss_hard + (1 - alpha) * loss_soft # update train accuracy metric train_acc_metric.update_state(y, student_pred) # calculate gradients grads = tape.gradient(loss, stud_model.weights) # gradient descent optimizer.apply_gradients(zip(grads, stud_model.trainable_weights)) # print some info print("Epoch {}, step {}, loss {:5f}".format(ep, step, loss)) # get result of train accuracy metric print("Train accuracy is {:4f}".format(train_acc_metric.result())) # reset metric train_acc_metric.reset_states()