Last active
January 3, 2021 21:59
-
-
Save isspek/0fbeda0de8c0cc6b1448e2f216ff862f to your computer and use it in GitHub Desktop.
Can not get_weights() in custom train step
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| def train_step(self, input_data): | |
| labeled_data, labels = input_data | |
| # Train the both | |
| with tf.GradientTape() as tape: | |
| augmented_data, augmented_labels = augment_data(labeled_data, labels, self.unlabeled_train) | |
| logits_student = self.student(augmented_data) | |
| clf_loss = self.classification_costs(logits_student, augmented_labels) | |
| # we augment the data again | |
| augmented_data, augmented_labels = augment_data(labeled_data, labels, self.unlabeled_train) | |
| logits_student = self.student(augmented_data) | |
| logits_teacher = self.teacher(augmented_data) | |
| consistency_cost = self.compute_consistency_cost(logits_teacher, logits_student) | |
| loss = self.compute_overall_cost(clf_loss, consistency_cost) | |
| grads = tape.gradient(loss, self.student.trainable_weights) | |
| self.optimizer.apply_gradients( | |
| zip(grads, self.student.trainable_weights) | |
| ) | |
| self.ema() | |
| logits_teacher = self.teacher(input_data) | |
| self.compiled_metrics.update_state(labels, logits_teacher) | |
| metrics = {m.name: m.result() for m in self.metrics} | |
| metrics["loss"] = loss | |
| return metrics | |
| def ema(self): | |
| print(self.teacher.get_weights()) | |
| teacher_weights = self.teacher.get_weights() | |
| student_weights = self.student.get_weights() | |
| # length must be equal otherwise it will not work | |
| assert len(student_weights) == len( | |
| teacher_weights), f'length of student and teachers weights are not equal Please check. \n Student: {len(student_weights)}, \n Teacher:{len(teacher_weights)}' | |
| new_layers = [] | |
| for i, layers in enumerate(student_weights): | |
| new_layer = self.alpha * (teacher_weights[i]) + (1 - self.alpha) * layers | |
| new_layers.append(new_layer) | |
| self.teacher.set_weights(new_layers) |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment