Nsryjdtyk

Recently, I have tried to use ENAS code to automatically design a network on my own datasets.

The code first load data as numpy in main.py, then assign data to model.py for example

# main.py

images, labels = read_data(path)

Then in the model.py it inits the self.x_train and self.y_train as follows:

# model.py

class Model(object):

    ...

    with tf.device("/cpu:0"):

    # training data

    self.num_train_examples = np.shape(images["train"])[0]

    self.num_train_batches = (

        self.num_train_examples + self.batch_size - 1) // self.batch_size



    x_train, y_train = tf.train.shuffle_batch(

        [images["train"], labels["train"]], # images['train'] and labels['train'] are both numpy。array

        batch_size=self.batch_size,

        capacity=50000,

        enqueue_many=True,

        num_threads=16,

        allow_smaller_final_batch=True,

    )

Then in the main.py, the part of running graph is as follows:

# main.py

with tf.train.SingularMonitoredSession(

        config=config, hooks=hooks, checkpoint_dir=FLAGS.output_dir) as sess:

    start_time = time.time()

    while True:

        #####################################

        ######  calculate child ops  ########

        #####################################



        run_ops = [

            child_ops["loss"],

            child_ops["lr"],

            child_ops["grad_norm"],

            child_ops["train_acc"],

            child_ops["train_op"],

        ]

        loss, lr, gn, tr_acc, _ = sess.run(run_ops)

        global_step = sess.run(child_ops["global_step"])

        print(sess.run(child_ops['y_train']))

        if FLAGS.child_sync_replicas:

            actual_step = global_step * FLAGS.num_aggregate

        else:

            actual_step = global_step

        epoch = actual_step // ops["num_train_batches"] # ops["num_train_batches"] 

        print('Epoch:{}, step:{}'.format(epoch, actual_step))

        curr_time = time.time()

What confused me is that the code doesn't define operations, such as self.x_train_next=self.x_train.get_next() or tf.train.Coordinator() to load the next iter data in any .py files.

So the follows are my question:

1.Does tf.train.shuffle_batch automatically load the next batch?

2.What's the difference between tf.train.batch() and tf.data.Datasets.from_tensor_slices.batch()?

3.The original code uses CIFAR10, when I try to use my own dataset, the image size can only set less than 160*160, otherwise it will raise ValueError: GraphDef cannot be larger than 2GB. I had tried to use placeholder or TFRecord to load data, but I don't know when the next batch data is loaded, I have no idea how to change the code. So is there any suggestion to load data?

Many thanks!