conv_split_awa.py [252:318]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
        if do_sampling:
            # List to store important samples from the previous tasks
            last_task_x = None
            last_task_y_ = None

        # Mask for softmax 
        logit_mask = np.zeros(model.total_classes)

        max_batch_dimension = 500

        # Dict to store the number of times a class has already been seen in the training
        class_seen_already = dict()
        for cls in range(TOTAL_CLASSES):
            class_seen_already[cls] = 0

        # Training loop for all the tasks
        for task in range(len(task_labels)):
            print('\t\tTask %d:'%(task))

            # If not the first task then restore weights from previous task
            if(task > 0):
                model.restore(sess)

            # Increment the class seen count
            for cls in task_labels[task]:
                class_seen_already[cls] += 1

            # Load the task specific dataset
            task_train_images, task_train_labels = load_task_specific_data_in_proportion(datasets[0]['train'], task_labels[task], classes_appearing_in_tasks, class_seen_already)
            print('Received {} images, {} labels at task {}'.format(task_train_images.shape[0], task_train_labels.shape[0], task))
            print('Unique labels in the task: {}'.format(np.unique(np.nonzero(task_train_labels)[1])))

            # Assign equal weights to all the examples
            task_sample_weights = np.ones([task_train_labels.shape[0]], dtype=np.float32)

            num_train_examples = task_train_images.shape[0]

            logit_mask[:] = 0
            # Train a task observing sequence of data
            if train_single_epoch:
                # Ceiling operation
                num_iters = (num_train_examples + batch_size - 1) // batch_size
            else:
                num_iters = train_iters

            logit_mask_offset = task * TOTAL_CLASSES
            classes_adjusted_for_head = [cls + logit_mask_offset for cls in task_labels[task]]
            logit_mask[classes_adjusted_for_head] = 1.0

            # Randomly suffle the training examples
            perm = np.arange(num_train_examples)
            np.random.shuffle(perm)
            train_x = task_train_images[perm]
            train_y = task_train_labels[perm]
            task_sample_weights = task_sample_weights[perm]

            # Array to store accuracies when training for task T
            if cross_validate_mode:
                # Because we will evalaute at the end
                ftask = 0
            elif train_single_epoch:
                # Because we will evaluate after every mini-batch of every task
                ftask = np.zeros([max_batch_dimension+1, model.num_tasks])
                batch_dim_count = 0
            else:
                # Because we will evaluate after every task
                ftask = []
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



conv_split_awa_hybrid.py [257:322]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
        if do_sampling:
            # List to store important samples from the previous tasks
            last_task_x = None
            last_task_y_ = None

        # Mask for softmax
        logit_mask = np.zeros(model.total_classes)

        max_batch_dimension = 500

        # Dict to store the number of times a class has already been seen in the training
        class_seen_already = dict()
        for cls in range(TOTAL_CLASSES):
            class_seen_already[cls] = 0

        # Training loop for all the tasks
        for task in range(len(task_labels)):
            print('\t\tTask %d:'%(task))

            # If not the first task then restore weights from previous task
            if(task > 0):
                model.restore(sess)

            # Increment the class seen count
            for cls in task_labels[task]:
                class_seen_already[cls] += 1

            task_train_images, task_train_labels = load_task_specific_data_in_proportion(datasets[0]['train'], task_labels[task], classes_appearing_in_tasks, class_seen_already)
            print('Received {} images, {} labels at task {}'.format(task_train_images.shape[0], task_train_labels.shape[0], task))
            print('Unique labels in the task: {}'.format(np.unique(np.nonzero(task_train_labels)[1])))

            # Assign equal weights to all the examples
            task_sample_weights = np.ones([task_train_labels.shape[0]], dtype=np.float32)

            num_train_examples = task_train_images.shape[0]

            # Train a task observing sequence of data
            logit_mask[:] = 0
            if train_single_epoch:
                # Ceiling operation
                num_iters = (num_train_examples + batch_size - 1) // batch_size
            else:
                num_iters = train_iters

            logit_mask_offset = task * TOTAL_CLASSES
            classes_adjusted_for_head = [cls + logit_mask_offset for cls in task_labels[task]]
            logit_mask[classes_adjusted_for_head] = 1.0

            # Randomly suffle the training examples
            perm = np.arange(num_train_examples)
            np.random.shuffle(perm)
            train_x = task_train_images[perm]
            train_y = task_train_labels[perm]
            task_sample_weights = task_sample_weights[perm]

            # Array to store accuracies when training for task T
            if cross_validate_mode:
                # Because we will evaluate at the end
                ftask = 0
            elif train_single_epoch:
                # Because we will evaluate after every mini-batch of every task
                ftask = np.zeros([max_batch_dimension+1, model.num_tasks])
                batch_dim_count = 0
            else:
                # Multi-epoch because we will evaluate after every task
                ftask = []
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -