0a68639f4c33323274e7829b9349d0170dc6c8ea,train.py,,classifier,#,67

Before Change

    best_va = 0.0

    with tf.device(ARGS.train_device):
        global_step = tf.Variable(0, trainable=False, name="global_step")

        // Get images and labels
        with tf.device("/cpu:0"):
            images, labels = DATASET.distorted_inputs(ARGS.batch_size)
        log_io(images)
        // Build a Graph that computes the logits predictions from the
        // inference model.
        is_training_, logits = MODEL.get(
            images,
            DATASET.num_classes(),
            train_phase=True,
            l2_penalty=ARGS.l2_penalty)

        // Calculate loss.
        loss = MODEL.loss(logits, labels)
        tf_log(tf.summary.scalar("loss", loss))

        // Create optimizer and log learning rate
        optimizer = build_optimizer(global_step)
        train_op = optimizer.minimize(
            loss,
            global_step=global_step,
            var_list=variables_to_train(ARGS.trainable_scopes))

        train_accuracy = utils.accuracy_op(logits, labels)
        // General validation summary
        accuracy_value_ = tf.placeholder(tf.float32, shape=())
        accuracy_summary = tf.summary.scalar("accuracy", accuracy_value_)

        // read collection after that every op added its own
        // summaries in the train_summaries collection
        train_summaries = tf.summary.merge(
            tf.get_collection_ref(MODEL_SUMMARIES))

        // Build an initialization operation to run below.
        init = [
            tf.variables_initializer(tf.global_variables() +
                                     tf.local_variables()),
            tf.tables_initializer()
        ]

        // Start running operations on the Graph.
        with tf.Session(config=tf.ConfigProto(
                allow_soft_placement=True)) as sess:
            sess.run(init)

            // Start the queue runners with a coordinator
            coord = tf.train.Coordinator()
            threads = tf.train.start_queue_runners(sess=sess, coord=coord)

            // Create the savers.
            train_saver, best_saver = build_train_savers([global_step])
            restore_or_restart(sess, global_step)
            train_log, validation_log = build_loggers(sess.graph)

            // Extract previous global step value
            old_gs = sess.run(global_step)

            // Restart from where we were
            for step in range(old_gs, MAX_STEPS):
                start_time = time.time()
                _, loss_value = sess.run(
                    [train_op, loss], feed_dict={is_training_: True})

                duration = time.time() - start_time

                if np.isnan(loss_value):
                    print("Model diverged with loss = NaN")
                    break

                // update logs every 10 iterations
                if step % 10 == 0:
                    examples_per_sec = ARGS.batch_size / duration
                    sec_per_batch = float(duration)

                    format_str = ("{}: step {}, loss = {:.4f} "
                                  "({:.1f} examples/sec; {:.3f} sec/batch)")
                    print(
                        format_str.format(datetime.now(), step, loss_value,
                                          examples_per_sec, sec_per_batch))
                    // log train values
                    summary_lines = sess.run(
                        train_summaries, feed_dict={is_training_: True})
                    train_log.add_summary(summary_lines, global_step=step)

                // Save the model checkpoint at the end of every epoch
                // evaluate train and validation performance
                if (step > 0 and
                        step % STEPS_PER_EPOCH == 0) or (step + 1) == MAX_STEPS:
                    checkpoint_path = os.path.join(LOG_DIR, "model.ckpt")
                    train_saver.save(sess, checkpoint_path, global_step=step)

                    // validation accuracy
                    va_value = eval_model(LOG_DIR, InputType.validation)

                    summary_line = sess.run(
                        accuracy_summary, feed_dict={accuracy_value_: va_value})
                    validation_log.add_summary(summary_line, global_step=step)

                    // train accuracy
                    ta_value = sess.run(
                        train_accuracy, feed_dict={is_training_: False})
                    summary_line = sess.run(
                        accuracy_summary, feed_dict={accuracy_value_: ta_value})
                    train_log.add_summary(summary_line, global_step=step)

                    print(
                        "{} ({}): train accuracy = {:.3f} validation accuracy = {:.3f}".
                        format(datetime.now(),
                               int(step / STEPS_PER_EPOCH), ta_value, va_value))
                    // save best model
                    if va_value > best_va:
                        best_va = va_value
                        best_saver.save(
                            sess,
                            os.path.join(BEST_MODEL_DIR, "model.ckpt"),
                            global_step=step)
            // end of for

            validation_log.close()
            train_log.close()

            // When done, ask the threads to stop.
            coord.request_stop()
            // Wait for threads to finish.
            coord.join(threads)


def autoencoder():
    Train the autoencoder and saves the best model:
    that"s the model with the lower validation error.
    

After Change

    best_va = 0.0

    with tf.Graph().as_default(), tf.device(ARGS.train_device):
        global_step = tf.Variable(0, trainable=False, name="global_step")

        // Get images and labels
        with tf.device("/cpu:0"):
            images, labels = DATASET.distorted_inputs(ARGS.batch_size)
        log_io(images)
        // Build a Graph that computes the logits predictions from the
        // inference model.
        is_training_, logits = MODEL.get(
            images,
            DATASET.num_classes,
            train_phase=True,
            l2_penalty=ARGS.l2_penalty)

        // Calculate loss.
        loss = MODEL.loss(logits, labels)
        tf_log(tf.summary.scalar("loss", loss))

        // Create optimizer and log learning rate
        optimizer = build_optimizer(global_step)
        train_op = optimizer.minimize(
            loss,
            global_step=global_step,
            var_list=variables_to_train(ARGS.trainable_scopes))

        train_accuracy = utils.accuracy_op(logits, labels)
        // General validation summary
        accuracy_value_ = tf.placeholder(tf.float32, shape=())
        accuracy_summary = tf.summary.scalar("accuracy", accuracy_value_)

        // read collection after that every op added its own
        // summaries in the train_summaries collection
        train_summaries = tf.summary.merge(
            tf.get_collection_ref(MODEL_SUMMARIES))

        // Build an initialization operation to run below.
        init = [
            tf.variables_initializer(tf.global_variables() +
                                     tf.local_variables()),
            tf.tables_initializer()
        ]

        // Start running operations on the Graph.
        with tf.Session(config=tf.ConfigProto(
                allow_soft_placement=True)) as sess:
            sess.run(init)

            // Start the queue runners with a coordinator
            coord = tf.train.Coordinator()
            threads = tf.train.start_queue_runners(sess=sess, coord=coord)

            // Create the savers.
            train_saver, best_saver = build_train_savers([global_step])
            restore_or_restart(sess, global_step)
            train_log, validation_log = build_loggers(sess.graph)

            // Extract previous global step value
            old_gs = sess.run(global_step)

            // Restart from where we were
            for step in range(old_gs, MAX_STEPS):
                start_time = time.time()
                _, loss_value = sess.run(
                    [train_op, loss], feed_dict={is_training_: True})

                duration = time.time() - start_time

                if np.isnan(loss_value):
                    print("Model diverged with loss = NaN")
                    break

                // update logs every 10 iterations
                if step % 10 == 0:
                    examples_per_sec = ARGS.batch_size / duration
                    sec_per_batch = float(duration)

                    format_str = ("{}: step {}, loss = {:.4f} "
                                  "({:.1f} examples/sec; {:.3f} sec/batch)")
                    print(
                        format_str.format(datetime.now(), step, loss_value,
                                          examples_per_sec, sec_per_batch))
                    // log train values
                    summary_lines = sess.run(
                        train_summaries, feed_dict={is_training_: True})
                    train_log.add_summary(summary_lines, global_step=step)

                // Save the model checkpoint at the end of every epoch
                // evaluate train and validation performance
                if (step > 0 and
                        step % STEPS_PER_EPOCH == 0) or (step + 1) == MAX_STEPS:
                    checkpoint_path = os.path.join(LOG_DIR, "model.ckpt")
                    train_saver.save(sess, checkpoint_path, global_step=step)

                    // validation accuracy
                    va_value = eval_model(LOG_DIR, InputType.validation)

                    summary_line = sess.run(
                        accuracy_summary, feed_dict={accuracy_value_: va_value})
                    validation_log.add_summary(summary_line, global_step=step)

                    // train accuracy
                    ta_value = sess.run(
                        train_accuracy, feed_dict={is_training_: False})
                    summary_line = sess.run(
                        accuracy_summary, feed_dict={accuracy_value_: ta_value})
                    train_log.add_summary(summary_line, global_step=step)

                    print(
                        "{} ({}): train accuracy = {:.3f} validation accuracy = {:.3f}".
                        format(datetime.now(),
                               int(step / STEPS_PER_EPOCH), ta_value, va_value))
                    // save best model
                    if va_value > best_va:
                        best_va = va_value
                        best_saver.save(
                            sess,
                            os.path.join(BEST_MODEL_DIR, "model.ckpt"),
                            global_step=step)
            // end of for

            validation_log.close()
            train_log.close()

            // When done, ask the threads to stop.
            coord.request_stop()
            // Wait for threads to finish.
            coord.join(threads)


def autoencoder():
    Train the autoencoder and saves the best model:
    that"s the model with the lower validation error.
    

Instances