research/gam/gam/models/gcn.py [204:278]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - if self.is_binary_classification: predictions = predictions[:, 0] # Store model variables for easy access. variables = tf.get_collection( tf.GraphKeys.GLOBAL_VARIABLES, scope=tf.get_default_graph().get_name_scope()) all_vars = {var.name: var for var in variables} return predictions, all_vars, reg_params def get_loss(self, predictions, targets, name_scope='loss', reg_params=None, **kwargs): """Returns a loss between the provided targets and predictions. For binary classification, this loss is sigmoid cross entropy. For multi-class classification, it is softmax cross entropy. A weight decay loss is also added to the parameters passed in reg_params. Args: predictions: A tensor of predictions. For multiclass classification its shape is (num_samples, num_classes), where the second dimension contains a logit per class. For binary classification, its shape is (num_samples,), where each element is the probability of class 1 for that sample. targets: A tensor of targets of shape (num_samples,), where each row contains the label index of the corresponding sample. name_scope: A string containing the name scope used in TensorFlow. reg_params: A dictonary of parameters, mapping from name to parameter, for the variables to be included in the weight decay loss. If None, no weight decay is applied. **kwargs: Keyword arguments, potentially containing the weight of the regularization term, passed under the name `weight_decay`. If this is not provided, it defaults to 0.0. Returns: loss: The cummulated loss value. """ reg_params = reg_params if reg_params is not None else {} weight_decay = kwargs['weight_decay'] if 'weight_decay' in kwargs else None with tf.name_scope(name_scope): # Cross entropy error. if self.is_binary_classification: loss = tf.reduce_sum( tf.nn.sigmoid_cross_entropy_with_logits( labels=targets, logits=predictions)) else: loss = tf.losses.softmax_cross_entropy(targets, predictions) # Weight decay loss. if weight_decay is not None: for var in reg_params.values(): loss = loss + weight_decay * tf.nn.l2_loss(var) return loss def normalize_predictions(self, predictions): """Converts predictions to probabilities. Args: predictions: A tensor of logits. For multiclass classification its shape is (num_samples, num_classes), where the second dimension contains a logit per class. For binary classification, its shape is (num_samples,), where each element is the probability of class 1 for that sample. Returns: A tensor of the same shape as predictions, with values between [0, 1] representing probabilities. """ if self.is_binary_classification: return tf.nn.sigmoid(predictions) return tf.nn.softmax(predictions, axis=-1) - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - research/gam/gam/models/mlp.py [175:249]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - if self.is_binary_classification: predictions = predictions[:, 0] # Store model variables for easy access. variables = tf.get_collection( tf.GraphKeys.GLOBAL_VARIABLES, scope=tf.get_default_graph().get_name_scope()) all_vars = {var.name: var for var in variables} return predictions, all_vars, reg_params def get_loss(self, predictions, targets, name_scope='loss', reg_params=None, **kwargs): """Returns a loss between the provided targets and predictions. For binary classification, this loss is sigmoid cross entropy. For multi-class classification, it is softmax cross entropy. A weight decay loss is also added to the parameters passed in reg_params. Args: predictions: A tensor of predictions. For multiclass classification its shape is (num_samples, num_classes), where the second dimension contains a logit per class. For binary classification, its shape is (num_samples,), where each element is the probability of class 1 for that sample. targets: A tensor of targets of shape (num_samples,), where each row contains the label index of the corresponding sample. name_scope: A string containing the name scope used in TensorFlow. reg_params: A dictonary of parameters, mapping from name to parameter, for the variables to be included in the weight decay loss. If None, no weight decay is applied. **kwargs: Keyword arguments, potentially containing the weight of the regularization term, passed under the name `weight_decay`. If this is not provided, it defaults to 0.0. Returns: loss: The cummulated loss value. """ reg_params = reg_params if reg_params is not None else {} weight_decay = kwargs['weight_decay'] if 'weight_decay' in kwargs else None with tf.name_scope(name_scope): # Cross entropy error. if self.is_binary_classification: loss = tf.reduce_sum( tf.nn.sigmoid_cross_entropy_with_logits( labels=targets, logits=predictions)) else: loss = tf.losses.softmax_cross_entropy(targets, predictions) # Weight decay loss. if weight_decay is not None: for var in reg_params.values(): loss = loss + weight_decay * tf.nn.l2_loss(var) return loss def normalize_predictions(self, predictions): """Converts predictions to probabilities. Args: predictions: A tensor of logits. For multiclass classification its shape is (num_samples, num_classes), where the second dimension contains a logit per class. For binary classification, its shape is (num_samples,), where each element is the probability of class 1 for that sample. Returns: A tensor of the same shape as predictions, with values between [0, 1] representing probabilities. """ if self.is_binary_classification: return tf.nn.sigmoid(predictions) return tf.nn.softmax(predictions, axis=-1) - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -