tensorflow example to demonstrate how to do lazy summary addition.

lazy_summary_update_example.py

import numpy as np
import tensorflow as tf
import tensorflow.contrib.slim.nets as nets

# tf version: 1.10.0


def get_random_input_and_label(batch_size, class_size):

    # for demonstration purpose, I'm going to reuse a random input and label
    random_input = np.random.rand(batch_size, 299, 299, 3)
    random_index = np.random.randint(0, 5, batch_size)
    random_output = np.zeros((batch_size, class_size))

    for index, val in enumerate(random_index):
        random_output[index, val] = 1.0

    return random_input, random_output


class_size = 5

# build some model


input_ph = tf.placeholder(tf.float32, [None, 299, 299, 3])
onehot_labels_ph = tf.placeholder(tf.float32, [None, class_size])

# for deatil on the model, check out https://github.com/tensorflow/tensorflow/blob/r1.10/tensorflow/contrib/slim/python/slim/nets/inception_v3.py
logits_ts, end_points = nets.inception.inception_v3(input_ph, num_classes=5)
prediction_ts = end_points['Predictions']

loss_ts = tf.losses.softmax_cross_entropy(
    onehot_labels=onehot_labels_ph, logits=logits_ts)
optimizer_op = tf.train.AdamOptimizer(0.001).minimize(loss_ts)

# calculating accuracy inside the graph
pred_argmax_ts = tf.argmax(prediction_ts, axis=1)
label_argmax_ts = tf.argmax(onehot_labels_ph, axis=1)

# convert bool to float value
compare_ts = tf.to_float(tf.equal(pred_argmax_ts, label_argmax_ts))
same_count = tf.reduce_sum(compare_ts)
number_of_batch = tf.to_float(tf.shape(pred_argmax_ts)[0])

accuracy_ts = same_count / number_of_batch

# in training, we want to log loss value, accuracy value
loss_summary = tf.summary.scalar("loss/loss", loss_ts)
train_accuracy_summary = tf.summary.scalar("metric/acc", accuracy_ts)

# will detect loss_summary and train_accuracy_summary
train_summary_op = tf.summary.merge_all()

test_accuracy_summary = tf.summary.scalar("test/acc", accuracy_ts)
test_summary_op = tf.summary.merge([test_accuracy_summary])

initop = tf.global_variables_initializer()


with tf.Session() as sess:

    writer = tf.summary.FileWriter("tfsummary", session= sess)

    sess.run(initop)

    steps = 20

    train_input, train_label = get_random_input_and_label(4, class_size)
    test_input, test_label = get_random_input_and_label(4, class_size)

    for step in range(steps):

        train_summary, loss_val, prediction, _ = sess.run([train_summary_op, loss_ts, prediction_ts, optimizer_op],
                                                          feed_dict={input_ph: train_input, onehot_labels_ph: train_label})
        writer.add_summary(train_summary, global_step = step)
        print("train done for step={}".format(step))

        if step!=0 and step%5==0:
            test_summary = sess.run(test_summary_op, feed_dict={input_ph: test_input, onehot_labels_ph: test_label})
            writer.add_summary(test_summary, global_step = step)
            print("test done at step={}".format(step))
            


print("end of code")