Use Tfrecord to feed multi-input neural network - tensorflow

Question

I have a dataset on Tfrecord that includes images and tabular data, I want to feed those data into a mixed neural network data. I saw examples of how to use multi-input neural network by using numpy arrays
model.fit([X_images,X_tabular],[Y_images,Y_tabular])
but transforming Tfrecord dataset into a numpy array by iterating slows down the process.
Is there a way to train it directly on the Tfrecord data. So far I have a model that train only an images and this is the code that I am using
tfrecords_schema = {
'id': tf.io.FixedLenFeature([], tf.string),
'y_label': tf.io.FixedLenFeature([], tf.int64),
'image': tf.io.FixedLenFeature([], tf.string),
'x_1': tf.io.FixedLenFeature([], tf.float32),
'x_2': tf.io.FixedLenFeature([], tf.float32),
'x_3': tf.io.FixedLenFeature([], tf.int64),
'x_4': tf.io.FixedLenFeature([], tf.int64),
}
def _parse_function(example_proto):
parsed_example = tf.io.parse_single_example(example_proto, tfrecords_schema)
return parsed_example['image'], parsed_example['y_label']
def _parse_error(image, label):
return image != b''
def _preprocess_function(image, label):
image = tf.image.decode_jpeg(image)
image = tf.image.resize(image, (299, 299))
label = label_decoder(label)
return image, label
train_data = (
train_data
.map(_parse_function)
.filter(_parse_error)
.map(_preprocess_function)
.apply(tf.data.experimental.ignore_errors())
.batch(BATCH_SIZE)
)
validation_data = (
validation_data
.map(_parse_function)
.filter(_parse_error)
.map(_preprocess_function)
.apply(tf.data.experimental.ignore_errors())
.batch(BATCH_SIZE)
)
AUTOTUNE = tf.data.experimental.AUTOTUNE
train_data = train_data.prefetch(buffer_size=AUTOTUNE)
validation_data = validation_data.prefetch(buffer_size=AUTOTUNE)
def make_model(image_shape, num_classes):
inputs = tf.keras.Input(shape=image_shape)
x = tf.keras.applications.inception_v3.preprocess_input(inputs)
base_model = tf.keras.applications.InceptionV3(input_shape=IMAGE_SHAPE, include_top=False, weights='imagenet')
base_model.trainable = False
x = base_model(x, training=False)
global_average_layer = tf.keras.layers.GlobalAveragePooling2D()
x = global_average_layer(x)
x = tf.keras.layers.Dropout(0.3)(x)
prediction_layer = tf.keras.layers.Dense(num_classes)
outputs = prediction_layer(x)
return tf.keras.Model(inputs, outputs)
model = make_model(IMAGE_SHAPE, NUM_CLASSES)
learning_rate = 0.0001
model.compile(
optimizer=tf.keras.optimizers.Adam(lr=learning_rate),
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=["accuracy"]
)
callbacks = [
tf.keras.callbacks.ModelCheckpoint(filepath="tf/pvc_data/tmp/weights.hdf5", verbose=1, save_best_only=True)
]
history = model.fit(
train_data, epochs=1, callbacks=callbacks, validation_data=validation_data
)