siamese networkのペア付けをミニバッチごとに行いたい

Question

行いたいこと：siamese networkのペア付けをミニバッチごとに行えるように変更したいと考えています。
現状：そこで、ペア付けの関数（make_pairs）をfit_generatorで呼び出すようにしたのですがエラーにより実装できていません。
教えて頂きたいこと：ペア付けをミニバッチごとに行って学習できるようにする部分まで教えて頂きたいです。

import random
import numpy as np
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
import matplotlib.pyplot as plt
from time import time
import os
from sklearn.manifold import TSNE


t_start = time() # 開始時間

"""
## Hyperparameters
"""
save_dir ='/home/practice/siamese_model/'
epochs = 100
batch_size = 12
margin = 1  # Margin for constrastive loss.

def make_pairs(batch_size):
    (x_train_val, y_train_val), (x_test, y_test) = keras.datasets.mnist.load_data()
    x_train_val = x_train_val.astype("float32")
    x_test = x_test.astype("float32")

    # Keep 50% of train_val  in validation set
    x_train, x_val = x_train_val[:30000], x_train_val[30000:]
    y_train, y_val = y_train_val[:30000], y_train_val[30000:]
    x = x_train
    y = y_train

    num_classes = max(y) + 1
    digit_indices = [np.where(y == i)[0] for i in range(num_classes)]

    pairs = []
    labels = []
    count = 0
    for idx1 in range(len(x)):
        count += 1
        # add a matching example
        x1 = x[idx1]
        label1 = y[idx1]
        idx2 = random.choice(digit_indices[label1])
        x2 = x[idx2]

        pairs += [[x1, x2]]
        labels += [1]

        # add a non-matching example
        label2 = random.randint(0, num_classes - 1)
        while label2 == label1:
            label2 = random.randint(0, num_classes - 1)

        idx2 = random.choice(digit_indices[label2])
        x2 = x[idx2]

        pairs += [[x1, x2]]
        labels += [0]
    print('All pair is ',count)
    pairs = np.array(pairs)
    x_train_1 = pairs[:, 0]  # x_train_1.shape is (60000, 28, 28)
    x_train_1=x_train_1.reshape(x_train_1.shape[0],x_train_1.shape[1],x_train_1.shape[2],1)
    x_train_2 = pairs[:, 1]
    x_train_2=x_train_2.reshape(x_train_2.shape[0],x_train_2.shape[1],x_train_2.shape[2],1)
    yield [x_train_1,x_train_2], np.array(labels).astype("float32")

# Provided two tensors t1 and t2
# Euclidean distance = sqrt(sum(square(t1-t2)))
def euclidean_distance(vects):
    x, y = vects
    sum_square = tf.math.reduce_sum(tf.math.square(x - y), axis=1, keepdims=True)
    a = tf.math.sqrt(tf.math.maximum(sum_square, tf.keras.backend.epsilon()))
    print(a)
    return a

input = layers.Input((28,28,1))
x = tf.keras.layers.BatchNormalization()(input)
x = layers.Conv2D(4, (5, 5), activation="tanh")(x)
x = layers.AveragePooling2D(pool_size=(2, 2))(x)
x = layers.Conv2D(16, (5, 5), activation="tanh")(x)
x = layers.AveragePooling2D(pool_size=(2, 2))(x)
x = layers.Flatten()(x)

x = tf.keras.layers.BatchNormalization()(x)
x = layers.Dense(10, activation="tanh")(x)
embedding_network = keras.Model(input, x)


input_1 = layers.Input((28,28,1))
input_2 = layers.Input((28,28,1))

# As mentioned above, Siamese Network share weights between
# tower networks (sister networks). To allow this, we will use
# same embedding network for both tower networks.
tower_1 = embedding_network(input_1)
tower_2 = embedding_network(input_2)

merge_layer = layers.Lambda(euclidean_distance)([tower_1, tower_2])

normal_layer = tf.keras.layers.BatchNormalization()(merge_layer)
output_layer = layers.Dense(1, activation="sigmoid")(normal_layer)
siamese = keras.Model(inputs=[input_1, input_2], outputs=output_layer)


"""
## Define the constrastive Loss
"""


def loss(margin=1):

    # Contrastive loss = mean( (1-true_value) * square(prediction) +
    #                         true_value * square( max(margin-prediction, 0) ))
    def contrastive_loss(y_true, y_pred):

        print('y_pred',y_pred)
        print('y_true',y_true)

        square_pred = tf.math.square(y_pred)
        margin_square = tf.math.square(tf.math.maximum(margin - (y_pred), 0))
        return tf.math.reduce_mean(
            (1 - y_true) * square_pred + (y_true) * margin_square
        )

    return contrastive_loss

siamese.compile(loss=loss(margin=margin), optimizer="RMSprop", metrics=["accuracy"])
tf.keras.utils.plot_model(siamese, to_file = 'model2.png', show_shapes = True, show_layer_names = True)

siamese.summary()

history = siamese.fit_generator(
    make_pairs(batch_size=batch_size),
    steps_per_epoch = 2000,
    epochs=epochs
)

siamese.save('/home/practice/siamese_model/model.hdf5')

t_end = time() #終了時間
t_elapsed = t_end - t_start

print("処理時間は{0}".format(t_elapsed))

エラー

All pair is  30000
Epoch 1/100
WARNING:tensorflow:From /home/.local/lib/python3.6/site-packages/tensorflow/python/ops/math_grad.py:1250: add_dispatch_support.<locals>.wrapper (from tensorflow.python.ops.array_ops) is deprecated and will be removed in a future version.
Instructions for updating:
Use tf.where in 2.0, which has the same broadcast rule as np.where
2022-01-04 22:29:48.600988: I tensorflow/stream_executor/platform/default/dso_loader.cc:42] Successfully opened dynamic library libcublas.so.10.0
2022-01-04 22:29:48.902468: I tensorflow/stream_executor/platform/default/dso_loader.cc:42] Successfully opened dynamic library libcudnn.so.7
   1/2000 [..............................] - ETA: 2:32:27 - loss: 0.2255 - acc: 0.6391WARNING:tensorflow:Your dataset iterator ran out of data; interrupting training. Make sure that your iterator can generate at least `steps_per_epoch * epochs` batches (in this case, 200000 batches). You may need touse the repeat() function when building your dataset.