keras指定gpu_keras搭建模型训练教程

import tensorflow as tf
from keras.backend.tensorflow_backend import set_session
config= tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = set_session(tf.Session(config=config))

from keras.utils import multi_gpu_model
multi_train_model = multi_gpu_model(model, gpus=, cpu_relocation=False)#对原来的model外层包一下并返回

from keras import Input, Model#搭建模型输入库，模型库
from keras.applications.vgg16/resnet50 import VGG16/ResNet50#网络模型库
from keras.layers import Concatenate, Conv2D, upSampling2D, BatchNormalization#网络层
from keras.utils import multi_gpu_model#多GPU训练
from keras.optimizers import Adam, SGD#优化器
from keras.callbacks import EarlyStopping, ModelCheckpoint, TensorBoard#训练时后端callbacks
import tensorflow as tf

from keras import Input, Model
from keras.applications.vgg import VGG16
class Network:
    def __init__(self, *argv, **kwargs):
        self._input = Input(name',#定义一个Input的实例
                        shape=(),
                        dtype='float')
        vgg = VGG16(input_tensor=self.input_img#送入输入的input_tensor
                    weights='imagenet',
                    dtype='float32')                
    def build_graph(self):
        Conv2D(**kwargs)(vgg)...#此处定义自己的网络层
        return Model(inputs=self._input, outputs=)#
net_instance = Network(...)
model = net_instance.build_graph()

model.compile(loss=, optimizer=)

model.summary()#打印网络信息，包括网络层数，中间层尺寸变化，以及参数量等信息

model.fit_generator(generator=data_gen_func(),
                    steps_per_epoch=,#每个epoch迭代多少次，注意 step * batchsize = total_nums
                    epochs=,#共训练多少个Epoch
                    validation_data=,
                    verbose=,
                    initial_epoch=,#开始训练的epoch
                    callbacks= [ModelCheckpoint(filepath=,#存储文件路径，str最后预留剩余%s可以在命名时保存训练的一些loss
                                                save_best_only=,#之存储
                                                verbose=),
                                TensorBoard(log_dir=,
                                            histogram_freq=0,
                                            write_graph=True,
                                            writer_images=False)])
model.save(filepath=)#保存网络对象
model.save_weights(filepath=)#只保存网络权重

pred = model.predict(x, batch_size=None, steps=None)