Low GPU usage in machine learning computing - tensorflow

Question

My friend and I challenged a kaggle competition named "Plant Seedlings Classification". He built a keras kernel as follow :
import numpy as np
import pandas as pd
import datetime as dt
import matplotlib.pyplot as plt
from os import listdir, makedirs
from os.path import join, exists, expanduser
from keras import models
from keras.preprocessing import image
from keras.applications import xception
from keras.preprocessing.image import ImageDataGenerator
from keras.callbacks import TensorBoard
from keras.callbacks import ModelCheckpoint
from keras import optimizers
from keras import layers
from keras.applications.xception import Xception
train_dir = './Plant_Seedlings_Classification/train/'
valid_dir = './Plant_Seedlings_Classification/validation/'
im_size = 299
batch_size = 10
train_num = 3803
valid_num = 947
conv_base = Xception(weights='imagenet',\
include_top=False,
input_shape=(im_size, im_size, 3))
conv_base.trainable = False
model = models.Sequential()
model.add(conv_base)
model.add(layers.Flatten()) # Flatten
model.add(layers.Dense(128, activation='relu'))
model.add(layers.Dropout(0.5))
model.add(layers.Dense(12, activation='softmax'))
model.summary()
train_datagen = ImageDataGenerator(rescale=1./255,\
rotation_range=30,\
width_shift_range=0.2,\
height_shift_range=0.2,\
zoom_range=0.2,\
horizontal_flip=True,\
fill_mode='nearest')
valid_datagen = ImageDataGenerator(rescale=1./255)
print("train gen")
train_generator = train_datagen.flow_from_directory(
train_dir,
class_mode='categorical',
target_size=(im_size, im_size),
color_mode='rgb',
batch_size=batch_size,)
print("validation gen")
validation_generator = valid_datagen.flow_from_directory(
valid_dir,
class_mode='categorical',
target_size=(im_size, im_size),
color_mode='rgb',
batch_size=batch_size,)
print("train indices",train_generator.class_indices)
print("validation indices", validation_generator.class_indices)
model.compile(loss='categorical_crossentropy',
optimizer=optimizers.Nadam(lr=1e-4,
beta_1=0.9, \
beta_2=0.999, \
epsilon=1e-08, \
schedule_decay=0.004),
metrics=['acc'])
steps_per_epoch = int(train_num/batch_size)+50
validation_step = int(valid_num/batch_size)+1
print("steps_per_epoch", steps_per_epoch)
print("validation_step", validation_step)
model_save_path = 'xception_299_0304_nomask_epoch{epoch:02d}_vacc{val_acc:.4f}.h5'
history = model.fit_generator(
train_generator,
steps_per_epoch=steps_per_epoch,
epochs=50,
validation_data=validation_generator,
validation_steps=validation_step,
callbacks=[TensorBoard(log_dir='/tmp/tensorflow/log/1'), \
ModelCheckpoint(filepath=model_save_path, \
monitor='val_acc',\
save_best_only=True,\
mode='max')])
The code will run 50 epochs.
When the kernel ran on my friend's PC (I7-7700, GTX 1060, 8G DDR4-2400), it took about 90 seconds per epoch.
And it took about 120 seconds per epoch when it ran on my PC (I5-7400, GTX 1070Ti, 16G DDR4-2400)
We both used tensorflow_gpu to run this kernel. My question is, why the GPU computing speed on my PC is much slower than my friend's?
We checked the GPU usage using nvidia-smi. When the kernel was running, the Volatile GPU-Util is only 30~60% on my GPU, and >95% on my friend's.
The Volatile GPU-Util can reach ~100% on my GPU when I run other ML frameworks like Caffe or Tensorflow without keras API.
Any pointer? Thanks.