Iterate over tensor as an array Tensorflow

Question

I am trying to save the predicted images on my CNN network which I wrote with Tensorflow. In my code y_pred_cls contain my predicted labels and the y_pred_cls is a tensor of dimensions 1 x batch size. Now, I want to iterate over y_pred_cls as an array and make a file name including pred class, true class, and some index number, then find out images relate to predicted labels and use imsave to save as image.

with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
train_writer.add_graph(sess.graph)



print("{} Start training...".format(datetime.now().strftime('%Y-%m-%d %H:%M:%S')))
print("{} Open Tensorboard at --logdir {}".format(datetime.now().strftime('%Y-%m-%d %H:%M:%S'), tensorboard_dir))

for epoch in range(FLAGS.num_epochs):
    print("{} Epoch number: {}".format(datetime.now().strftime('%Y-%m-%d %H:%M:%S'), epoch + 1))
    step = 1

    # Start training
    while step < train_batches_per_epoch:
        batch_xs, batch_ys = train_preprocessor.next_batch(FLAGS.batch_size)
        opt, train_acc = sess.run([optimizer, accuracy], feed_dict={x: batch_xs, y_true: batch_ys})

        # Logging
        if step % FLAGS.log_step == 0:
            s = sess.run(sum, feed_dict={x: batch_xs, y_true: batch_ys})
            train_writer.add_summary(s, epoch * train_batches_per_epoch + step)

        step += 1

    # Epoch completed, start validation
    print("{} Start validation".format(datetime.now().strftime('%Y-%m-%d %H:%M:%S')))
    val_acc = 0.
    val_count = 0
    cm_running_total = None

    for _ in range(val_batches_per_epoch):
        batch_tx, batch_ty = val_preprocessor.next_batch(FLAGS.batch_size)
        acc, loss , conf_m= sess.run([accuracy, cost, tf.confusion_matrix(y_true_cls, y_pred_cls, FLAGS.num_classes)],
                                      feed_dict={x: batch_tx, y_true: batch_ty})



        if cm_running_total is None:
            cm_running_total = conf_m
        else:
            cm_running_total += conf_m


        val_acc += acc
        val_count += 1

    val_acc /= val_count

    s = tf.Summary(value=[
        tf.Summary.Value(tag="validation_accuracy", simple_value=val_acc),
        tf.Summary.Value(tag="validation_loss", simple_value=loss)
    ])

    val_writer.add_summary(s, epoch + 1)
    print("{} -- Training Accuracy = {:.4%} -- Validation Accuracy = {:.4%} -- Validation Loss = {:.4f}".format(
        datetime.now().strftime('%Y-%m-%d %H:%M:%S'), train_acc, val_acc, loss))

    # Reset the dataset pointers
    val_preprocessor.reset_pointer()
    train_preprocessor.reset_pointer()

    print("{} Saving checkpoint of model...".format(datetime.now().strftime('%Y-%m-%d %H:%M:%S')))

    # save checkpoint of the model
    checkpoint_path = os.path.join(checkpoint_dir, 'model_epoch.ckpt' + str(epoch+1))
    save_path = saver.save(sess, checkpoint_path)
    print("{} Model checkpoint saved at {}".format(datetime.now().strftime('%Y-%m-%d %H:%M:%S'), checkpoint_path))

batch_tx, batch_ty is my RGB data and labels respectively.

Thanks in advance.

Answer 1

To extract data from a tensor into a python-variable use

label = sess.run(y_pred_cls)

This will give you an array for a one-hot-vector label or an int variable for a scalar label.

To save arrays to images you can use the PIL-library

from PIL import Image
img = Image.fromarray(data, 'RGB')
img.save('name.png')

The rest should be straight forward,

1. extract data from your batch_tx, batch_ty and y_pred_cls tensors
2. iterate over each triplet
3. create an RGB image out of current x
4. create a string of the form name = str(y)+'_'+str(y_hat)
5. save your image

If you have trouble applying these steps I can help you out further