var_acc = 0
    accuraces = np.array(accuraces)
    for i in range(accuraces.shape[0]):
        var_acc += (accuraces[i] - mean_acc) * (accuraces[i] - mean_acc)

    var_acc /= float(n_tests)

    print('\nModel: {}'.format(MODEL))
    print('Mean testing accuracy for {} runs: {}'.format(n_tests, mean_acc))
    print('Variance of testing accuracy for {} runs: {}'.format(
        n_tests, var_acc))


# Prepare data
_, (x_test_all, y_test_all) = utils.load_mnist_excluded()

if not is_only_3_and_4:
    test_model(n_tests, x_test_all, y_test_all, ang_min, ang_max)

else:
    nImg = x_test_all.shape[0]
    x_test = np.empty([1, sy, sx, 1], dtype=np.float32)
    y_test = np.empty([1, NCLASSES])
    k = 0
    for i in range(nImg):
        y_i = y_test_all[i, :]
        y_i = np.argmax(y_i)
        if (y_i == 3) or (y_i == 4):
            if k == 0:
                x_test[0, :, :, :] = x_test_all[i, :, :, :]
def train():
    with tf.variable_scope(tf.get_variable_scope()):
        basenet_output = model(x_input, phase_train=phase_train,
                                 keep_prob=keep_prob, reuse = False)
        basenet_output_test = model(x_input, phase_train=phase_train,
                                      keep_prob=keep_prob, reuse=True)

    loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits = basenet_output, labels=y_target), name="error_cost")
    correct_predictions_train = tf.equal(tf.argmax(basenet_output, 1), tf.argmax(y_target, 1))
    correct_predictions_test = tf.equal(tf.argmax(basenet_output_test, 1), tf.argmax(y_target, 1))
    accuracy_mean_train = tf.reduce_mean(tf.cast(correct_predictions_train, "float"))
    accuracy_mean_test = tf.reduce_mean(tf.cast(correct_predictions_test, "float"))
    accuracy_sum_train = tf.reduce_sum(tf.cast(correct_predictions_train, "float"))
    accuracy_sum_test = tf.reduce_sum(tf.cast(correct_predictions_test, "float"))

    update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
    # var = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='Basenet')
    with tf.control_dependencies(update_ops):
        optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate, beta1=beta1).minimize(loss, global_step=global_step)

    init = tf.global_variables_initializer()
    loss_summary = tf.summary.scalar(name='Loss', tensor=loss)
    accuracy_train_summary = tf.summary.scalar(name='Train Accuracy', tensor=accuracy_mean_train)
    accuracy_test_summary = tf.summary.scalar(name='Test Accuracy', tensor=accuracy_mean_test)

    summary_op = tf.summary.merge([loss_summary, accuracy_train_summary, accuracy_test_summary])

    saver = tf.train.Saver()


    # Prepare Training Data
    (x_train, y_train), (x_test, y_test) = utils.load_mnist_excluded()

    nTrain = x_train.shape[0]
    nTrainBatches = int(nTrain / BATCH_SIZE)

    nTest = x_test.shape[0]
    nTestBatches = int(nTest / BATCH_SIZE)
    step = 0
    with tf.Session() as sess:
        sess.run(init)
        writer = tf.summary.FileWriter(logdir=LOGDIR, graph=sess.graph)
        maxAcc = -1
        for i in range(n_epochs):
            print("------------------Epoch {}/{}------------------".format(i, n_epochs))
            mean_train_acc = 0
            for b in range(nTrainBatches):
                batch_x, batch_y = utils.get_random_mnist_batch(x_train, y_train, BATCH_SIZE)
                # Check what images are fed to the network
                if i==0 and b == 0:
                    utils.save_img_batch(batch_x, './DEBUG_TRAIN_BATCH.png')
                sess.run(optimizer, feed_dict={x_input: batch_x,
                                               y_target: batch_y,
                                               phase_train:True,
                                               keep_prob: 0.5})
                loss_val, step, lr, acc_mean_batch, acc_sum_batch, summary = sess.run(
                    [loss, global_step, learning_rate, accuracy_mean_train, accuracy_sum_train, summary_op],
                          feed_dict={x_input: batch_x,
                                     y_target: batch_y,
                                     phase_train: False,
                                     keep_prob: 1.0 })
                mean_train_acc +=  acc_sum_batch
                writer.add_summary(summary, global_step=step)
                sys.stdout.write(ERASE_LINE)
                sys.stdout.write("\r\rEpoch: {}, Global step: {}, Learning rate: {}, Iteration: {}/{}\tmean batch accuracy: {}".format(i, step, lr, b, nTrainBatches, acc_mean_batch))
                sys.stdout.flush()
                time.sleep(0.005)
                step += 1

            print("\n------------------Evaluation after epoch {}/{}------------------".format(i, n_epochs))
            print("TRAIN ACCURACY: {}".format(mean_train_acc/float(nTrainBatches*BATCH_SIZE)))
            mean_test_acc = 0
            for b in range(nTestBatches):
                batch_x = x_test[b*BATCH_SIZE:(b+1)*BATCH_SIZE, :, :, :]
                batch_y = y_test[b*BATCH_SIZE:(b+1)*BATCH_SIZE, :]
                batch_x = utils.rotate_batch_for_test(batch_x, batch_y, [ang_min, ang_max])
                if i==0 and b == 0:
                    utils.save_img_batch(batch_x, './DEBUG_TEST_BATCH.png')
                acc_sum_test_batch = sess.run( accuracy_sum_test, feed_dict={
                    x_input: batch_x,
                    y_target: batch_y,
                    phase_train: False,
                    keep_prob: 1.0 })
                mean_test_acc += acc_sum_test_batch

            nleft = nTest - int(nTestBatches * BATCH_SIZE)
            if nleft > 0:
                batch_x = x_test[(nTest - nleft):, :, :, :]
                batch_x = utils.rotate_batch_for_test(batch_x, batch_y, [ang_min, ang_max])
                batch_y = y_test[(nTest - nleft):, :]
                acc_sum_test_batch = sess.run(accuracy_sum_test, feed_dict={
                    x_input: batch_x,
                    y_target: batch_y,
                    phase_train: False,
                    keep_prob: 1.0})
                mean_test_acc += acc_sum_test_batch

            mean_test_acc = mean_test_acc / float(nTest)
            print("TEST ACCURACY: {}".format(mean_test_acc))

            if( mean_test_acc > maxAcc ):
                maxAcc = mean_test_acc
                saver.save(sess, SAVEDIR + '/intermediate_model')

        saver.save(sess, SAVEDIR + '/final-model')
示例#3
0
def main(args):
    """Get dataset hyperparameters."""
    assert len(args) == 2 and isinstance(args[1], str)
    dataset_name = args[1]
    logger.info('Using dataset: {}'.format(dataset_name))
    """Set reproduciable random seed"""
    tf.set_random_seed(1234)

    coord_add = get_coord_add(dataset_name)
    dataset_size = get_dataset_size_train(dataset_name)
    num_classes = get_num_classes(dataset_name)

    # Prepare Training Data
    (x_train, y_train), (x_test, y_test) = utils.load_mnist_excluded()

    with tf.Graph().as_default():  #, tf.device('/cpu:0'):

        # Placeholders for input data and the targets
        x_input = tf.placeholder(tf.float32, (None, *IMG_DIM), name='Input')
        y_target = tf.placeholder(tf.int32, [
            None,
        ], name='Target')
        """Get global_step."""
        global_step = tf.get_variable('global_step', [],
                                      initializer=tf.constant_initializer(0),
                                      trainable=False)
        """Get batches per epoch."""
        num_batches_per_epoch = int(dataset_size / cfg.batch_size)
        """Use exponential decay leanring rate?"""
        lrn_rate = tf.maximum(
            tf.train.exponential_decay(1e-3, global_step,
                                       num_batches_per_epoch, 0.8), 1e-5)
        tf.summary.scalar('learning_rate', lrn_rate)
        opt = tf.train.AdamOptimizer()  # lrn_rate
        """Define the dataflow graph."""
        m_op = tf.placeholder(dtype=tf.float32, shape=())
        with tf.device('/gpu:0'):
            with slim.arg_scope([slim.variable]):  #, device='/cpu:0'):
                sample_batch = tf.identity(x_input)
                batch_labels = tf.identity(y_target)
                batch_squash = tf.divide(sample_batch, 255.)
                batch_x = slim.batch_norm(sample_batch,
                                          center=False,
                                          is_training=True,
                                          trainable=True)
                output, pose_out = net.build_arch(batch_x,
                                                  coord_add,
                                                  is_train=True,
                                                  num_classes=num_classes)

                tf.logging.debug(pose_out.get_shape())
                loss, spread_loss, mse, reconstruction = net.spread_loss(
                    output, pose_out, batch_squash, batch_labels, m_op)
                sample_batch = tf.squeeze(sample_batch)
                decode_res_op = tf.concat([
                    sample_batch,
                    255 * tf.reshape(reconstruction,
                                     [cfg.batch_size, IMAGE_SIZE, IMAGE_SIZE])
                ],
                                          axis=0)
                acc = net.test_accuracy(output, batch_labels)
                tf.summary.scalar('spread_loss', spread_loss)
                tf.summary.scalar('reconstruction_loss', mse)
                tf.summary.scalar('all_loss', loss)
                tf.summary.scalar('train__batch_acc', acc)
            """Compute gradient."""
            grad = opt.compute_gradients(loss)
            # See: https://stackoverflow.com/questions/40701712/how-to-check-nan-in-gradients-in-tensorflow-when-updating
            grad_check = [
                tf.check_numerics(g, message='Gradient NaN Found!')
                for g, _ in grad if g is not None
            ] + [tf.check_numerics(loss, message='Loss NaN Found')]
        """Apply graident."""
        with tf.control_dependencies(grad_check):
            update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
            with tf.control_dependencies(update_ops):
                train_op = opt.apply_gradients(grad, global_step=global_step)
        """Set Session settings."""
        sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
                                                log_device_placement=False))
        sess.run(tf.local_variables_initializer())
        sess.run(tf.global_variables_initializer())
        """Set Saver."""
        var_to_save = [
            v for v in tf.global_variables() if 'Adam' not in v.name
        ]  # Don't save redundant Adam beta/gamma
        saver = tf.train.Saver(var_list=var_to_save, max_to_keep=cfg.epoch)
        """Display parameters"""
        total_p = np.sum([
            np.prod(v.get_shape().as_list()) for v in var_to_save
        ]).astype(np.int32)
        train_p = np.sum([
            np.prod(v.get_shape().as_list()) for v in tf.trainable_variables()
        ]).astype(np.int32)
        logger.info('Total Parameters: {}'.format(total_p))
        logger.info('Trainable Parameters: {}'.format(train_p))

        # read snapshot
        # latest = os.path.join(cfg.logdir, 'model.ckpt-4680')
        # saver.restore(sess, latest)
        """Set summary op."""
        summary_op = tf.summary.merge_all()
        """Start coord & queue."""
        coord = tf.train.Coordinator()
        threads = tf.train.start_queue_runners(sess=sess, coord=coord)
        """Set summary writer"""
        if not os.path.exists(cfg.logdir +
                              '/caps/{}/train_log/'.format(dataset_name)):
            os.makedirs(cfg.logdir +
                        '/caps/{}/train_log/'.format(dataset_name))
        summary_writer = tf.summary.FileWriter(
            cfg.logdir + '/caps/{}/train_log/'.format(dataset_name),
            graph=sess.graph)  # graph = sess.graph, huge!

        if not os.path.exists(cfg.logdir +
                              '/caps/{}/images/'.format(dataset_name)):
            os.makedirs(cfg.logdir + '/caps/{}/images/'.format(dataset_name))
        """Main loop."""
        m_min = 0.2
        m_max = 0.9
        m = m_min
        max_iter = cfg.epoch * num_batches_per_epoch + 1

        for step in range(max_iter):
            tic = time.time()
            """"TF queue would pop batch until no file"""

            batch_x, batch_y = utils.get_random_mnist_batch(
                x_train, y_train, cfg.batch_size)

            try:
                _, loss_value, train_acc_val, summary_str, mse_value = sess.run(
                    [train_op, loss, acc, summary_op, mse],
                    feed_dict={
                        m_op: m,
                        x_input: batch_x,
                        y_target: batch_y
                    })

                sys.stdout.write(ERASE_LINE)
                sys.stdout.write('\r\r%d/%d iteration finishes in ' %
                                 (step, max_iter) + '%f second' %
                                 (time.time() - tic) +
                                 ' training accuracy = %f' % train_acc_val +
                                 ' loss=%f' % loss_value +
                                 '\treconstruction_loss=%f' % mse_value)
                sys.stdout.flush()
                time.sleep(0.001)

            except KeyboardInterrupt:
                sess.close()
                sys.exit()
            except tf.errors.InvalidArgumentError:
                logger.warning(
                    '%d iteration contains NaN gradients. Discard.' % step)
                continue
            else:
                """Write to summary."""
                if step % 10 == 0:
                    summary_writer.add_summary(summary_str, step)

                if step % 200 == 0:
                    images = sess.run(decode_res_op,
                                      feed_dict={
                                          m_op: m,
                                          x_input: batch_x,
                                          y_target: batch_y
                                      })
                    image = combine_images(images)
                    img_name = cfg.logdir + '/caps/{}/images/'.format(
                        dataset_name) + "/step_{}.png".format(str(step))
                    Image.fromarray(image.astype(np.uint8)).save(img_name)
                """Epoch wise linear annealling."""
                if (step % num_batches_per_epoch) == 0:
                    if step > 0:
                        m += (m_max - m_min) / (cfg.epoch * cfg.m_schedule)
                        if m > m_max:
                            m = m_max
                    """Save model periodically """
                    ckpt_path = os.path.join(
                        cfg.logdir + '/caps/{}/'.format(dataset_name),
                        'model-{:.4f}.ckpt'.format(loss_value))
                    saver.save(sess, ckpt_path, global_step=step)

        ckpt_path = os.path.join(cfg.logdir + '/caps/{}/'.format(dataset_name),
                                 'finall-model-{:.4f}.ckpt'.format(loss_value))
        saver.save(sess, ckpt_path, global_step=step)

        print('Training is finished!')