Accuracy loss after transfer keras (.h5) model to tf frozen graph (.pd) format

I'm using tensorflow 1.12.0 and keras 2.1.4 (I used 2.2.4 at beginning, the problem is same.)
I am working on some model transferring task and one of the step is to convert keras model (.h5) to tensorflow frozen graph format (.pb). The model accuracy was 98% while loading .h5 and test in keras, but only remain 10% while loading frozen graph (.pb) and test in tensorflow.
I transfer the h5 to frozen graph through following code.

import tensorflow.keras as keras
import tensorflow as tf
import os
def freeze_session(session, keep_var_names=None, output_names=None, clear_devices=True):
    """
    Freezes the state of a session into a pruned computation graph.
    Creates a new computation graph where variable nodes are replaced by
    constants taking their current value in the session. The new graph will be
    pruned so subgraphs that are not necessary to compute the requested
    outputs are removed.
    @param session The TensorFlow session to be frozen.
    @param keep_var_names A list of variable names that should not be frozen,
                          or None to freeze all the variables in the graph.
    @param output_names Names of the relevant graph outputs.
    @param clear_devices Remove the device directives from the graph for better portability.
    @return The frozen graph definition.
    """
    session.run(tf.initialize_all_variables())
    graph = session.graph
    with graph.as_default():
        freeze_var_names = list(set(v.op.name for v in tf.global_variables()).difference(keep_var_names or []))
        output_names = output_names or []
        output_names += [v.op.name for v in tf.global_variables()]
        input_graph_def = graph.as_graph_def()
        if clear_devices:
            for node in input_graph_def.node:
                node.device = ''
        frozen_graph = tf.graph_util.convert_variables_to_constants(
            session, input_graph_def, output_names, freeze_var_names)
        return frozen_graph
 
 
#keras h5 file
input_path='./'
input_file = './CNN_Mnist.h5'
weight_file_path = os.path.join(input_path, input_file)
output_graph_name = 'frozen_graph.pb'

#Load Module
keras.backend.set_learning_phase(0)
h5_model = keras.models.load_model(weight_file_path)

frozen_graph = freeze_session(keras.backend.get_session(), output_names=[out.op.name for out in h5_model.outputs])
tf.train.write_graph(frozen_graph, input_path, output_graph_name, as_text=False)
print('finish!')

Then I'm trying to load the frozen graph in tensorflow and evaluate it again, but I got very worse result. Here is my evaluation code.

mport os
import argparse
import shutil
import tensorflow as tf
import numpy as np

from tensorflow.python.platform import gfile
import tensorflow.contrib.decent_q
def graph_eval(input_graph_def, input_node, output_node):

    # MNIST dataset    
    (x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data()
    x_test = (x_test/255.0).astype(np.float32)
    x_test = np.reshape(x_test, [-1, 28, 28, 1])
    y_test = tf.keras.utils.to_categorical(y_test, num_classes=10)

    tf.import_graph_def(input_graph_def,name = '')

    # Get input placeholders & tensors
    images_in = tf.get_default_graph().get_tensor_by_name(input_node+':0')
    labels = tf.placeholder(tf.int32,shape = [None,10])

    # get output tensors
    logits = tf.get_default_graph().get_tensor_by_name(output_node+':0')

    # top 5 and top 1 accuracy
    in_top5 = tf.nn.in_top_k(predictions=logits, targets=tf.argmax(labels, 1), k=5)
    in_top1 = tf.nn.in_top_k(predictions=logits, targets=tf.argmax(labels, 1), k=1)
    top5_acc = tf.reduce_mean(tf.cast(in_top5, tf.float32))
    top1_acc = tf.reduce_mean(tf.cast(in_top1, tf.float32))
    
    # Create the Computational graph
    with tf.Session() as sess:
        
        sess.run(tf.initializers.global_variables())
      
        feed_dict={images_in: x_test, labels: y_test}
        t5_acc,t1_acc = sess.run([top5_acc,top1_acc], feed_dict)
    
        print (' Top 1 accuracy with validation set: {:1.4f}'.format(t1_acc))
        print (' Top 5 accuracy with validation set: {:1.4f}'.format(t5_acc))

    print ('FINISHED!')

#define arguments
graph='./frozen_graph_opt.pb'#output_graph #'./frozen_graph.pb'
input_node='conv2d_1_input_5'
output_node='dense_2_3/Softmax'
####
input_graph_def = tf.Graph().as_graph_def()
input_graph_def.ParseFromString(tf.gfile.FastGFile(graph, "rb").read())
graph_eval(input_graph_def,input_node,output_node)

The data pre-processing in training and evaluation are both use the same function. Could someone please help to point me out what part should I check or something should be noticed while transferring .h5 to .pd file?
Or is there any problem in my evaluation code?

Here is my model read in Netron.
keras .h5 model keras .h5 model

Tensorflow .pb model Tensorflow .pb model