def labelfinetune()

in ebm_sandbox.py [0:0]

• 32 lines of code
• 7 McCabe index (conditional complexity)

def labelfinetune(dataloader, test_dataloader, target_vars, sess, savedir, saver, l1val=8, l2val=40):
    X = target_vars['X']
    Y = target_vars['Y']
    Y_GT = target_vars['Y_GT']
    accuracy = target_vars['accuracy']
    train_op = target_vars['train_op']
    l1_norm = target_vars['l1_norm']
    l2_norm = target_vars['l2_norm']

    label_init = np.random.uniform(0, 1, (FLAGS.batch_size, 10))
    label_init = label_init / label_init.sum(axis=1, keepdims=True)

    label_init = np.tile(np.eye(10)[None :, :], (FLAGS.batch_size, 1, 1))
    label_init = np.reshape(label_init, (-1, 10))

    itr = 0

    if FLAGS.train:
        for i in range(1):
            for data_corrupt, data, label_gt in tqdm(dataloader):
                feed_dict = {X: data, Y_GT: label_gt, Y: label_init}
                acc, _ = sess.run([accuracy, train_op], feed_dict)

                itr += 1

                if itr % 10 == 0:
                    print(acc)

        saver.save(sess, osp.join(savedir, "model_supervised"))

    saver.restore(sess, osp.join(savedir, "model_supervised"))


    for i in range(1):
        emp_accuracies = []

        for data_corrupt, data, label_gt in tqdm(test_dataloader):
            feed_dict = {X: data, Y_GT: label_gt, Y: label_init, l1_norm: l1val, l2_norm: l2val}
            emp_accuracy = sess.run([accuracy], feed_dict)
            emp_accuracies.append(emp_accuracy)
            print(np.array(emp_accuracies).mean())


        print("Received total accuracy of {} for li of {} and l2 of {}".format(np.array(emp_accuracies).mean(), l1val, l2val))

    return np.array(emp_accuracies).mean()