TensorFlow分布式训练：单机多卡训练MirroredStrategy、多机训练MultiWorkerMirroredStrategy
日萌社
人工智能AI：Keras PyTorch MXNet TensorFlow PaddlePaddle 深度学习实战（不定时更新）
ResNet模型在GPU上的并行实践
TensorFlow分布式训练：单机多卡训练MirroredStrategy、多机训练MultiWorkerMirroredStrategy
4.8 分布式训练
当我们拥有大量计算资源时，通过使用合适的分布式策略，我们可以充分利用这些计算资源，从而大幅压缩模型训练的时间。针对不同的使用场景，TensorFlow 在 tf.distribute.Strategy`中为我们提供了若干种分布式策略，使得我们能够更高效地训练模型。
4.8.1 TensorFlow 分布式的分类
-
图间并行（又称数据并行）
- 每个机器上都会有一个完整的模型，将数据分散到各个机器，分别计算梯度。
-
图内并行（又称模型并行）
- 每个机器分别负责整个模型的一部分计算任务。
1、图间并行用的非常多，会包含两种方式进行更新
同步：收集到足够数量的梯度，一同更新，下图上方
异步：即同步方式下，更新需求的梯度数量为1，下图下方
注：PS(parameter server):维护全局共享的模型参数的服务器
2、实现方式
- 单机多卡
- 多机单卡
TensorFlow多机多卡实现思路。多级多卡的分布式有很多实现方式，比如：
1. 将每个GPU当做一个worker
2. 同一个机器的各个GPU进行图内并行
3. 同一个机器的各个GPU进行图间并行
比如说第三种：模型实现封装成函数、将数据分成GPU数量的份数、在每个GPU下，进行一次模型forward计算，并使用优化器算出梯度、reduce每个GPU下的梯度，并将梯度传入到分布式中的优化器中
4.8.1 单机多卡训练： MirroredStrategy
tf.distribute.MirroredStrategy 是一种简单且高性能的，数据并行的同步式分布式策略，主要支持多个 GPU 在同一台主机上训练。
-
1、MirroredStrategy运行原理：
- 1、训练开始前，该策略在所有 N 个计算设备（GPU）上均各复制一份完整的模型
- 2、每次训练传入一个批次的数据时，将数据分成 N 份，分别传入 N 个计算设备（即数据并行）
- 3、使用分布式计算的 All-reduce 操作，在计算设备间高效交换梯度数据并进行求和，使得最终每个设备都有了所有设备的梯度之和，使用梯度求和的结果更新本地变量
- 当所有设备均更新本地变量后，进行下一轮训练（即该并行策略是同步的）。默认情况下，TensorFlow 中的 MirroredStrategy 策略使用 NVIDIA NCCL 进行 All-reduce 操作。
-
2、构建代码步骤：
1、使用这种策略时，我们只需实例化一个 MirroredStrategy 策略:
strategy = tf.distribute.MirroredStrategy()
2、并将模型构建的代码放入 strategy.scope() 的上下文环境中:
with strategy.scope():
# 模型构建代码
或者可以在参数中指定设备，如:
strategy = tf.distribute.MirroredStrategy(devices=["/gpu:0", "/gpu:1"])即指定只使用第 0、1 号 GPU 参与分布式策略。
4.8.1.1 MirroredStrategy进行分类模型训练
以下代码展示了使用 MirroredStrategy 策略，在 TensorFlow Datasets 中的部分图像数据集上使用 Keras 训练 MobileNetV2 的过程：
import tensorflow as tf
import tensorflow_datasets as tfds
num_epochs = 5
batch_size_per_replica = 64
learning_rate = 0.001
strategy = tf.distribute.MirroredStrategy()
print('Number of devices: %d' % strategy.num_replicas_in_sync) # 输出设备数量
batch_size = batch_size_per_replica * strategy.num_replicas_in_sync
# 载入数据集并预处理
def resize(image, label):
image = tf.image.resize(image, [224, 224]) / 255.0
return image, label
# 当as_supervised为True时，返回image和label两个键值
dataset = tfds.load("cats_vs_dogs", split=tfds.Split.TRAIN, as_supervised=True)
dataset = dataset.map(resize).shuffle(1024).batch(batch_size)
with strategy.scope():
model = tf.keras.applications.MobileNetV2()
model.compile(
optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate),
loss=tf.keras.losses.sparse_categorical_crossentropy,
metrics=[tf.keras.metrics.sparse_categorical_accuracy]
)
model.fit(dataset, epochs=num_epochs)
这个是在线下测试的结果
-
epoch=5,batch_size=64
-
2块 NVIDIA GeForce GTX 1080 Ti 显卡进行单机多卡的模型训练。
| 数据集 | 单机无分布式（Batch Size 为 64） | 单机多卡（总 Batch Size 为 64） | 单机多卡（总 Batch Size 为 256） |
|---|---|---|---|
| cats_vs_dogs | 150s/epoch | 36s/epoch | 30s/epoch |
4.8.2 多机训练： MultiWorkerMirroredStrategy
多机训练的方法和单机多卡类似，将 MirroredStrategy 更换为适合多机训练的 MultiWorkerMirroredStrategy 即可。不过，由于涉及到多台计算机之间的通讯，还需要进行一些额外的设置。
1、需要设置环境变量 TF_CONFIG ，示例如下:
os.environ['TF_CONFIG'] = json.dumps({
'cluster': {
'worker': ["localhost:8888", "localhost:9999"]
},
'task': {'type': 'worker', 'index': 0}
})
2、TF_CONFIG 由 cluster 和 task 两部分组成：
-
cluster 说明了整个多机集群的结构和每台机器的网络地址（IP + 端口号）。对于每一台机器，cluster 的值都是相同的。
-
task 说明了当前机器的角色。例如， {'type': 'worker', 'index': 0} 说明当前机器是 cluster 中的第 0 个 worker（即 localhost:20000 ）。每一台机器的 task 值都需要针对当前主机进行分别的设置。
3、运行过程
-
1、在所有的机器上逐个运行训练代码即可。先运行的代码在尚未与其他主机连接时会进入监听状态，待整个集群的连接建立完毕后，所有的机器即会同时开始训练。
-
2、假设有两台机器，即首先在两台机器上均部署这个程序，唯一的区别是 task 部分，第一台机器设置为 {'type': 'worker', 'index': 0} ，第二台机器设置为 {'type': 'worker', 'index': 1} 。接下来，在两台机器上依次运行程序，待通讯成功后，即会自动开始训练流程。
多机训练代码
同样对于同一个数据及进行训练的时候，指定多机代码：
import tensorflow_datasets as tfds
import os
import json
num_epochs = 5
batch_size_per_replica = 64
learning_rate = 0.001
num_workers = 2
# 1、指定集群环境
os.environ['TF_CONFIG'] = json.dumps({
'cluster': {
'worker': ["localhost:20000", "localhost:20001"]
},
'task': {'type': 'worker', 'index': 0}
})
batch_size = batch_size_per_replica * num_workers
def resize(image, label):
image = tf.image.resize(image, [224, 224]) / 255.0
return image, label
dataset = tfds.load("cats_vs_dogs", split=tfds.Split.TRAIN, as_supervised=True)
dataset = dataset.map(resize).shuffle(1024).batch(batch_size)
# 2、初始化集群
strategy = tf.distribute.experimental.MultiWorkerMirroredStrategy()
# 3、上下文环境定义模型
with strategy.scope():
model = tf.keras.applications.MobileNetV2()
model.compile(
optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate),
loss=tf.keras.losses.sparse_categorical_crossentropy,
metrics=[tf.keras.metrics.sparse_categorical_accuracy]
)
model.fit(dataset, epochs=num_epochs)
测试结果比较：单张 NVIDIA Tesla K80 显卡，使用单机单卡时，Batch Size 设置为 64；使用双机单卡时，测试总 Batch Size 为 64（分发到单台机器的 Batch Size 为 32）和总 Batch Size 为 128（分发到单台机器的 Batch Size 为 64）两种情况。
| 数据集 | 单机单卡（Batch Size 为 64） | 双机单卡（总 Batch Size 为 64） | 双机单卡（总 Batch Size 为 128） |
|---|---|---|---|
| cats_vs_dogs | 1622s | 858s | 755s |
4.8.3 总结
- TensorFlow的分布式训练接口使用
6-3,使用单GPU训练模型
深度学习的训练过程常常非常耗时，一个模型训练几个小时是家常便饭，训练几天也是常有的事情，有时候甚至要训练几十天。
训练过程的耗时主要来自于两个部分，一部分来自数据准备，另一部分来自参数迭代。
当数据准备过程还是模型训练时间的主要瓶颈时，我们可以使用更多进程来准备数据。
当参数迭代过程成为训练时间的主要瓶颈时，我们通常的方法是应用GPU或者Google的TPU来进行加速。
详见《用GPU加速Keras模型——Colab免费GPU使用攻略》
https://zhuanlan.zhihu.com/p/68509398
无论是内置fit方法，还是自定义训练循环，从CPU切换成单GPU训练模型都是非常方便的，无需更改任何代码。当存在可用的GPU时，如果不特意指定device，tensorflow会自动优先选择使用GPU来创建张量和执行张量计算。
但如果是在公司或者学校实验室的服务器环境，存在多个GPU和多个使用者时，为了不让单个同学的任务占用全部GPU资源导致其他同学无法使用（tensorflow默认获取全部GPU的全部内存资源权限，但实际上只使用一个GPU的部分资源），我们通常会在开头增加以下几行代码以控制每个任务使用的GPU编号和显存大小，以便其他同学也能够同时训练模型。
在Colab笔记本中：修改->笔记本设置->硬件加速器 中选择 GPU
注：以下代码只能在Colab 上才能正确执行。
可通过以下colab链接测试效果《tf_单GPU》：
https://colab.research.google.com/drive/1r5dLoeJq5z01sU72BX2M5UiNSkuxsEFe
%tensorflow_version 2.x
import tensorflow as tf
print(tf.__version__)from tensorflow.keras import *
#打印时间分割线
@tf.function
def printbar():
ts = tf.timestamp()
today_ts = ts%(24*60*60)
hour = tf.cast(today_ts//3600+8,tf.int32)%tf.constant(24)
minite = tf.cast((today_ts%3600)//60,tf.int32)
second = tf.cast(tf.floor(today_ts%60),tf.int32)
def timeformat(m):
if tf.strings.length(tf.strings.format("{}",m))==1:
return(tf.strings.format("0{}",m))
else:
return(tf.strings.format("{}",m))
timestring = tf.strings.join([timeformat(hour),timeformat(minite),
timeformat(second)],separator = ":")
tf.print("=========="*8,end = "")
tf.print(timestring)
一，GPU设置
gpus = tf.config.list_physical_devices("GPU")
if gpus:
gpu0 = gpus[0] #如果有多个GPU，仅使用第0个GPU
tf.config.experimental.set_memory_growth(gpu0, True) #设置GPU显存用量按需使用
# 或者也可以设置GPU显存为固定使用量(例如：4G)
#tf.config.experimental.set_virtual_device_configuration(gpu0,
# [tf.config.experimental.VirtualDeviceConfiguration(memory_limit=4096)])
tf.config.set_visible_devices([gpu0],"GPU") 比较GPU和CPU的计算速度
printbar()
with tf.device("/gpu:0"):
tf.random.set_seed(0)
a = tf.random.uniform((10000,100),minval = 0,maxval = 3.0)
b = tf.random.uniform((100,100000),minval = 0,maxval = 3.0)
c = a@b
tf.print(tf.reduce_sum(tf.reduce_sum(c,axis = 0),axis=0))
printbar()================================================================================17:37:01
2.24953778e+11
================================================================================17:37:01printbar()
with tf.device("/cpu:0"):
tf.random.set_seed(0)
a = tf.random.uniform((10000,100),minval = 0,maxval = 3.0)
b = tf.random.uniform((100,100000),minval = 0,maxval = 3.0)
c = a@b
tf.print(tf.reduce_sum(tf.reduce_sum(c,axis = 0),axis=0))
printbar()================================================================================17:37:34
2.24953795e+11
================================================================================17:37:40二，准备数据
MAX_LEN = 300
BATCH_SIZE = 32
(x_train,y_train),(x_test,y_test) = datasets.reuters.load_data()
x_train = preprocessing.sequence.pad_sequences(x_train,maxlen=MAX_LEN)
x_test = preprocessing.sequence.pad_sequences(x_test,maxlen=MAX_LEN)
MAX_WORDS = x_train.max()+1
CAT_NUM = y_train.max()+1
ds_train = tf.data.Dataset.from_tensor_slices((x_train,y_train)) \
.shuffle(buffer_size = 1000).batch(BATCH_SIZE) \
.prefetch(tf.data.experimental.AUTOTUNE).cache()
ds_test = tf.data.Dataset.from_tensor_slices((x_test,y_test)) \
.shuffle(buffer_size = 1000).batch(BATCH_SIZE) \
.prefetch(tf.data.experimental.AUTOTUNE).cache()
三，定义模型
tf.keras.backend.clear_session()
def create_model():
model = models.Sequential()
model.add(layers.Embedding(MAX_WORDS,7,input_length=MAX_LEN))
model.add(layers.Conv1D(filters = 64,kernel_size = 5,activation = "relu"))
model.add(layers.MaxPool1D(2))
model.add(layers.Conv1D(filters = 32,kernel_size = 3,activation = "relu"))
model.add(layers.MaxPool1D(2))
model.add(layers.Flatten())
model.add(layers.Dense(CAT_NUM,activation = "softmax"))
return(model)
model = create_model()
model.summary()
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
embedding (Embedding) (None, 300, 7) 216874
_________________________________________________________________
conv1d (Conv1D) (None, 296, 64) 2304
_________________________________________________________________
max_pooling1d (MaxPooling1D) (None, 148, 64) 0
_________________________________________________________________
conv1d_1 (Conv1D) (None, 146, 32) 6176
_________________________________________________________________
max_pooling1d_1 (MaxPooling1 (None, 73, 32) 0
_________________________________________________________________
flatten (Flatten) (None, 2336) 0
_________________________________________________________________
dense (Dense) (None, 46) 107502
=================================================================
Total params: 332,856
Trainable params: 332,856
Non-trainable params: 0
_________________________________________________________________四，训练模型
optimizer = optimizers.Nadam()
loss_func = losses.SparseCategoricalCrossentropy()
train_loss = metrics.Mean(name='train_loss')
train_metric = metrics.SparseCategoricalAccuracy(name='train_accuracy')
valid_loss = metrics.Mean(name='valid_loss')
valid_metric = metrics.SparseCategoricalAccuracy(name='valid_accuracy')
@tf.function
def train_step(model, features, labels):
with tf.GradientTape() as tape:
predictions = model(features,training = True)
loss = loss_func(labels, predictions)
gradients = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(gradients, model.trainable_variables))
train_loss.update_state(loss)
train_metric.update_state(labels, predictions)
@tf.function
def valid_step(model, features, labels):
predictions = model(features)
batch_loss = loss_func(labels, predictions)
valid_loss.update_state(batch_loss)
valid_metric.update_state(labels, predictions)
def train_model(model,ds_train,ds_valid,epochs):
for epoch in tf.range(1,epochs+1):
for features, labels in ds_train:
train_step(model,features,labels)
for features, labels in ds_valid:
valid_step(model,features,labels)
logs = 'Epoch={},Loss:{},Accuracy:{},Valid Loss:{},Valid Accuracy:{}'
if epoch%1 ==0:
printbar()
tf.print(tf.strings.format(logs,
(epoch,train_loss.result(),train_metric.result(),valid_loss.result(),valid_metric.result())))
tf.print("")
train_loss.reset_states()
valid_loss.reset_states()
train_metric.reset_states()
valid_metric.reset_states()
train_model(model,ds_train,ds_test,10)================================================================================17:13:26
Epoch=1,Loss:1.96735072,Accuracy:0.489200622,Valid Loss:1.64124215,Valid Accuracy:0.582813919
================================================================================17:13:28
Epoch=2,Loss:1.4640888,Accuracy:0.624805152,Valid Loss:1.5559175,Valid Accuracy:0.607747078
================================================================================17:13:30
Epoch=3,Loss:1.20681274,Accuracy:0.68581605,Valid Loss:1.58494771,Valid Accuracy:0.622439921
================================================================================17:13:31
Epoch=4,Loss:0.937500894,Accuracy:0.75361836,Valid Loss:1.77466083,Valid Accuracy:0.621994674
================================================================================17:13:33
Epoch=5,Loss:0.693960547,Accuracy:0.822199941,Valid Loss:2.00267363,Valid Accuracy:0.6197685
================================================================================17:13:35
Epoch=6,Loss:0.519614,Accuracy:0.870296121,Valid Loss:2.23463202,Valid Accuracy:0.613980412
================================================================================17:13:37
Epoch=7,Loss:0.408562034,Accuracy:0.901246965,Valid Loss:2.46969271,Valid Accuracy:0.612199485
================================================================================17:13:39
Epoch=8,Loss:0.339028627,Accuracy:0.920062363,Valid Loss:2.68585229,Valid Accuracy:0.615316093
================================================================================17:13:41
Epoch=9,Loss:0.293798745,Accuracy:0.92930305,Valid Loss:2.88995624,Valid Accuracy:0.613535166
================================================================================17:13:43
Epoch=10,Loss:0.263130337,Accuracy:0.936651051,Valid Loss:3.09705234,Valid Accuracy:0.6126446726-4,使用多GPU训练模型
如果使用多GPU训练模型，推荐使用内置fit方法，较为方便，仅需添加2行代码。
在Colab笔记本中：修改->笔记本设置->硬件加速器 中选择 GPU
注：以下代码只能在Colab 上才能正确执行。
可通过以下colab链接测试效果《tf_多GPU》：
https://colab.research.google.com/drive/1j2kp_t0S_cofExSN7IyJ4QtMscbVlXU-
MirroredStrategy过程简介：
-
训练开始前，该策略在所有 N 个计算设备上均各复制一份完整的模型；
-
每次训练传入一个批次的数据时，将数据分成 N 份，分别传入 N 个计算设备（即数据并行）；
-
N 个计算设备使用本地变量（镜像变量）分别计算自己所获得的部分数据的梯度；
-
使用分布式计算的 All-reduce 操作，在计算设备间高效交换梯度数据并进行求和，使得最终每个设备都有了所有设备的梯度之和；
-
使用梯度求和的结果更新本地变量（镜像变量）；
-
当所有设备均更新本地变量后，进行下一轮训练（即该并行策略是同步的）。
%tensorflow_version 2.x
import tensorflow as tf
print(tf.__version__)
from tensorflow.keras import * #此处在colab上使用1个GPU模拟出两个逻辑GPU进行多GPU训练
gpus = tf.config.experimental.list_physical_devices('GPU')
if gpus:
# 设置两个逻辑GPU模拟多GPU训练
try:
tf.config.experimental.set_virtual_device_configuration(gpus[0],
[tf.config.experimental.VirtualDeviceConfiguration(memory_limit=1024),
tf.config.experimental.VirtualDeviceConfiguration(memory_limit=1024)])
logical_gpus = tf.config.experimental.list_logical_devices('GPU')
print(len(gpus), "Physical GPU,", len(logical_gpus), "Logical GPUs")
except RuntimeError as e:
print(e)一，准备数据
MAX_LEN = 300
BATCH_SIZE = 32
(x_train,y_train),(x_test,y_test) = datasets.reuters.load_data()
x_train = preprocessing.sequence.pad_sequences(x_train,maxlen=MAX_LEN)
x_test = preprocessing.sequence.pad_sequences(x_test,maxlen=MAX_LEN)
MAX_WORDS = x_train.max()+1
CAT_NUM = y_train.max()+1
ds_train = tf.data.Dataset.from_tensor_slices((x_train,y_train)) \
.shuffle(buffer_size = 1000).batch(BATCH_SIZE) \
.prefetch(tf.data.experimental.AUTOTUNE).cache()
ds_test = tf.data.Dataset.from_tensor_slices((x_test,y_test)) \
.shuffle(buffer_size = 1000).batch(BATCH_SIZE) \
.prefetch(tf.data.experimental.AUTOTUNE).cache()
二，定义模型
tf.keras.backend.clear_session()
def create_model():
model = models.Sequential()
model.add(layers.Embedding(MAX_WORDS,7,input_length=MAX_LEN))
model.add(layers.Conv1D(filters = 64,kernel_size = 5,activation = "relu"))
model.add(layers.MaxPool1D(2))
model.add(layers.Conv1D(filters = 32,kernel_size = 3,activation = "relu"))
model.add(layers.MaxPool1D(2))
model.add(layers.Flatten())
model.add(layers.Dense(CAT_NUM,activation = "softmax"))
return(model)
def compile_model(model):
model.compile(optimizer=optimizers.Nadam(),
loss=losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=[metrics.SparseCategoricalAccuracy(),metrics.SparseTopKCategoricalAccuracy(5)])
return(model)三，训练模型
#增加以下两行代码
strategy = tf.distribute.MirroredStrategy()
with strategy.scope():
model = create_model()
model.summary()
model = compile_model(model)
history = model.fit(ds_train,validation_data = ds_test,epochs = 10) WARNING:tensorflow:NCCL is not supported when using virtual GPUs, fallingback to reduction to one device
INFO:tensorflow:Using MirroredStrategy with devices ('/job:localhost/replica:0/task:0/device:GPU:0', '/job:localhost/replica:0/task:0/device:GPU:1')
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
embedding (Embedding) (None, 300, 7) 216874
_________________________________________________________________
conv1d (Conv1D) (None, 296, 64) 2304
_________________________________________________________________
max_pooling1d (MaxPooling1D) (None, 148, 64) 0
_________________________________________________________________
conv1d_1 (Conv1D) (None, 146, 32) 6176
_________________________________________________________________
max_pooling1d_1 (MaxPooling1 (None, 73, 32) 0
_________________________________________________________________
flatten (Flatten) (None, 2336) 0
_________________________________________________________________
dense (Dense) (None, 46) 107502
=================================================================
Total params: 332,856
Trainable params: 332,856
Non-trainable params: 0
_________________________________________________________________
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
Train for 281 steps, validate for 71 steps
Epoch 1/10
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:GPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:GPU:0', '/job:localhost/replica:0/task:0/device:GPU:1').
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:GPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:GPU:0', '/job:localhost/replica:0/task:0/device:GPU:1').
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:GPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:GPU:0', '/job:localhost/replica:0/task:0/device:GPU:1').
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:GPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:GPU:0', '/job:localhost/replica:0/task:0/device:GPU:1').
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:GPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:GPU:0', '/job:localhost/replica:0/task:0/device:GPU:1').
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:GPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:GPU:0', '/job:localhost/replica:0/task:0/device:GPU:1').
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:GPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:GPU:0', '/job:localhost/replica:0/task:0/device:GPU:1').
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:GPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:GPU:0', '/job:localhost/replica:0/task:0/device:GPU:1').
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:GPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:GPU:0', '/job:localhost/replica:0/task:0/device:GPU:1').
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:GPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:GPU:0', '/job:localhost/replica:0/task:0/device:GPU:1').
281/281 [==============================] - 15s 53ms/step - loss: 2.0270 - sparse_categorical_accuracy: 0.4653 - sparse_top_k_categorical_accuracy: 0.7481 - val_loss: 1.7517 - val_sparse_categorical_accuracy: 0.5481 - val_sparse_top_k_categorical_accuracy: 0.7578
Epoch 2/10
281/281 [==============================] - 4s 14ms/step - loss: 1.5206 - sparse_categorical_accuracy: 0.6045 - sparse_top_k_categorical_accuracy: 0.7938 - val_loss: 1.5715 - val_sparse_categorical_accuracy: 0.5993 - val_sparse_top_k_categorical_accuracy: 0.7983
Epoch 3/10
281/281 [==============================] - 4s 14ms/step - loss: 1.2178 - sparse_categorical_accuracy: 0.6843 - sparse_top_k_categorical_accuracy: 0.8547 - val_loss: 1.5232 - val_sparse_categorical_accuracy: 0.6327 - val_sparse_top_k_categorical_accuracy: 0.8112
Epoch 4/10
281/281 [==============================] - 4s 13ms/step - loss: 0.9127 - sparse_categorical_accuracy: 0.7648 - sparse_top_k_categorical_accuracy: 0.9113 - val_loss: 1.6527 - val_sparse_categorical_accuracy: 0.6296 - val_sparse_top_k_categorical_accuracy: 0.8201
Epoch 5/10
281/281 [==============================] - 4s 14ms/step - loss: 0.6606 - sparse_categorical_accuracy: 0.8321 - sparse_top_k_categorical_accuracy: 0.9525 - val_loss: 1.8791 - val_sparse_categorical_accuracy: 0.6158 - val_sparse_top_k_categorical_accuracy: 0.8219
Epoch 6/10
281/281 [==============================] - 4s 14ms/step - loss: 0.4919 - sparse_categorical_accuracy: 0.8799 - sparse_top_k_categorical_accuracy: 0.9725 - val_loss: 2.1282 - val_sparse_categorical_accuracy: 0.6037 - val_sparse_top_k_categorical_accuracy: 0.8112
Epoch 7/10
281/281 [==============================] - 4s 14ms/step - loss: 0.3947 - sparse_categorical_accuracy: 0.9051 - sparse_top_k_categorical_accuracy: 0.9814 - val_loss: 2.3033 - val_sparse_categorical_accuracy: 0.6046 - val_sparse_top_k_categorical_accuracy: 0.8094
Epoch 8/10
281/281 [==============================] - 4s 14ms/step - loss: 0.3335 - sparse_categorical_accuracy: 0.9207 - sparse_top_k_categorical_accuracy: 0.9863 - val_loss: 2.4255 - val_sparse_categorical_accuracy: 0.5993 - val_sparse_top_k_categorical_accuracy: 0.8099
Epoch 9/10
281/281 [==============================] - 4s 14ms/step - loss: 0.2919 - sparse_categorical_accuracy: 0.9304 - sparse_top_k_categorical_accuracy: 0.9911 - val_loss: 2.5571 - val_sparse_categorical_accuracy: 0.6020 - val_sparse_top_k_categorical_accuracy: 0.8126
Epoch 10/10
281/281 [==============================] - 4s 14ms/step - loss: 0.2617 - sparse_categorical_accuracy: 0.9342 - sparse_top_k_categorical_accuracy: 0.9937 - val_loss: 2.6700 - val_sparse_categorical_accuracy: 0.6077 - val_sparse_top_k_categorical_accuracy: 0.8148
CPU times: user 1min 2s, sys: 8.59 s, total: 1min 10s
Wall time: 58.5 s6-5,使用TPU训练模型
如果想尝试使用Google Colab上的TPU来训练模型，也是非常方便，仅需添加6行代码。
在Colab笔记本中：修改->笔记本设置->硬件加速器 中选择 TPU
注：以下代码只能在Colab 上才能正确执行。
可通过以下colab链接测试效果《tf_TPU》：
https://colab.research.google.com/drive/1XCIhATyE1R7lq6uwFlYlRsUr5d9_-r1s
%tensorflow_version 2.x
import tensorflow as tf
print(tf.__version__)
from tensorflow.keras import * 一，准备数据
MAX_LEN = 300
BATCH_SIZE = 32
(x_train,y_train),(x_test,y_test) = datasets.reuters.load_data()
x_train = preprocessing.sequence.pad_sequences(x_train,maxlen=MAX_LEN)
x_test = preprocessing.sequence.pad_sequences(x_test,maxlen=MAX_LEN)
MAX_WORDS = x_train.max()+1
CAT_NUM = y_train.max()+1
ds_train = tf.data.Dataset.from_tensor_slices((x_train,y_train)) \
.shuffle(buffer_size = 1000).batch(BATCH_SIZE) \
.prefetch(tf.data.experimental.AUTOTUNE).cache()
ds_test = tf.data.Dataset.from_tensor_slices((x_test,y_test)) \
.shuffle(buffer_size = 1000).batch(BATCH_SIZE) \
.prefetch(tf.data.experimental.AUTOTUNE).cache()二，定义模型
tf.keras.backend.clear_session()
def create_model():
model = models.Sequential()
model.add(layers.Embedding(MAX_WORDS,7,input_length=MAX_LEN))
model.add(layers.Conv1D(filters = 64,kernel_size = 5,activation = "relu"))
model.add(layers.MaxPool1D(2))
model.add(layers.Conv1D(filters = 32,kernel_size = 3,activation = "relu"))
model.add(layers.MaxPool1D(2))
model.add(layers.Flatten())
model.add(layers.Dense(CAT_NUM,activation = "softmax"))
return(model)
def compile_model(model):
model.compile(optimizer=optimizers.Nadam(),
loss=losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=[metrics.SparseCategoricalAccuracy(),metrics.SparseTopKCategoricalAccuracy(5)])
return(model)三，训练模型
#增加以下6行代码
import os
resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu='grpc://' + os.environ['COLAB_TPU_ADDR'])
tf.config.experimental_connect_to_cluster(resolver)
tf.tpu.experimental.initialize_tpu_system(resolver)
strategy = tf.distribute.experimental.TPUStrategy(resolver)
with strategy.scope():
model = create_model()
model.summary()
model = compile_model(model)
WARNING:tensorflow:TPU system 10.26.134.242:8470 has already been initialized. Reinitializing the TPU can cause previously created variables on TPU to be lost.
WARNING:tensorflow:TPU system 10.26.134.242:8470 has already been initialized. Reinitializing the TPU can cause previously created variables on TPU to be lost.
INFO:tensorflow:Initializing the TPU system: 10.26.134.242:8470
INFO:tensorflow:Initializing the TPU system: 10.26.134.242:8470
INFO:tensorflow:Clearing out eager caches
INFO:tensorflow:Clearing out eager caches
INFO:tensorflow:Finished initializing TPU system.
INFO:tensorflow:Finished initializing TPU system.
INFO:tensorflow:Found TPU system:
INFO:tensorflow:Found TPU system:
INFO:tensorflow:*** Num TPU Cores: 8
INFO:tensorflow:*** Num TPU Cores: 8
INFO:tensorflow:*** Num TPU Workers: 1
INFO:tensorflow:*** Num TPU Workers: 1
INFO:tensorflow:*** Num TPU Cores Per Worker: 8
INFO:tensorflow:*** Num TPU Cores Per Worker: 8
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:localhost/replica:0/task:0/device:CPU:0, CPU, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:localhost/replica:0/task:0/device:CPU:0, CPU, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:localhost/replica:0/task:0/device:XLA_CPU:0, XLA_CPU, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:localhost/replica:0/task:0/device:XLA_CPU:0, XLA_CPU, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:CPU:0, CPU, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:CPU:0, CPU, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:0, TPU, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:0, TPU, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:1, TPU, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:1, TPU, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:2, TPU, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:2, TPU, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:3, TPU, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:3, TPU, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:4, TPU, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:4, TPU, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:5, TPU, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:5, TPU, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:6, TPU, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:6, TPU, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:7, TPU, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:7, TPU, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU_SYSTEM:0, TPU_SYSTEM, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU_SYSTEM:0, TPU_SYSTEM, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:XLA_CPU:0, XLA_CPU, 0, 0)
INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:XLA_CPU:0, XLA_CPU, 0, 0)
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
embedding (Embedding) (None, 300, 7) 216874
_________________________________________________________________
conv1d (Conv1D) (None, 296, 64) 2304
_________________________________________________________________
max_pooling1d (MaxPooling1D) (None, 148, 64) 0
_________________________________________________________________
conv1d_1 (Conv1D) (None, 146, 32) 6176
_________________________________________________________________
max_pooling1d_1 (MaxPooling1 (None, 73, 32) 0
_________________________________________________________________
flatten (Flatten) (None, 2336) 0
_________________________________________________________________
dense (Dense) (None, 46) 107502
=================================================================
Total params: 332,856
Trainable params: 332,856
Non-trainable params: 0
_________________________________________________________________history = model.fit(ds_train,validation_data = ds_test,epochs = 10)Train for 281 steps, validate for 71 steps
Epoch 1/10
281/281 [==============================] - 12s 43ms/step - loss: 3.4466 - sparse_categorical_accuracy: 0.4332 - sparse_top_k_categorical_accuracy: 0.7180 - val_loss: 3.3179 - val_sparse_categorical_accuracy: 0.5352 - val_sparse_top_k_categorical_accuracy: 0.7195
Epoch 2/10
281/281 [==============================] - 6s 20ms/step - loss: 3.3251 - sparse_categorical_accuracy: 0.5405 - sparse_top_k_categorical_accuracy: 0.7302 - val_loss: 3.3082 - val_sparse_categorical_accuracy: 0.5463 - val_sparse_top_k_categorical_accuracy: 0.7235
Epoch 3/10
281/281 [==============================] - 6s 20ms/step - loss: 3.2961 - sparse_categorical_accuracy: 0.5729 - sparse_top_k_categorical_accuracy: 0.7280 - val_loss: 3.3026 - val_sparse_categorical_accuracy: 0.5499 - val_sparse_top_k_categorical_accuracy: 0.7217
Epoch 4/10
281/281 [==============================] - 5s 19ms/step - loss: 3.2751 - sparse_categorical_accuracy: 0.5924 - sparse_top_k_categorical_accuracy: 0.7276 - val_loss: 3.2957 - val_sparse_categorical_accuracy: 0.5543 - val_sparse_top_k_categorical_accuracy: 0.7217
Epoch 5/10
281/281 [==============================] - 5s 19ms/step - loss: 3.2655 - sparse_categorical_accuracy: 0.6008 - sparse_top_k_categorical_accuracy: 0.7290 - val_loss: 3.3022 - val_sparse_categorical_accuracy: 0.5490 - val_sparse_top_k_categorical_accuracy: 0.7231
Epoch 6/10
281/281 [==============================] - 5s 19ms/step - loss: 3.2616 - sparse_categorical_accuracy: 0.6041 - sparse_top_k_categorical_accuracy: 0.7295 - val_loss: 3.3015 - val_sparse_categorical_accuracy: 0.5503 - val_sparse_top_k_categorical_accuracy: 0.7235
Epoch 7/10
281/281 [==============================] - 6s 21ms/step - loss: 3.2595 - sparse_categorical_accuracy: 0.6059 - sparse_top_k_categorical_accuracy: 0.7322 - val_loss: 3.3064 - val_sparse_categorical_accuracy: 0.5454 - val_sparse_top_k_categorical_accuracy: 0.7266
Epoch 8/10
281/281 [==============================] - 6s 21ms/step - loss: 3.2591 - sparse_categorical_accuracy: 0.6063 - sparse_top_k_categorical_accuracy: 0.7327 - val_loss: 3.3025 - val_sparse_categorical_accuracy: 0.5481 - val_sparse_top_k_categorical_accuracy: 0.7231
Epoch 9/10
281/281 [==============================] - 5s 19ms/step - loss: 3.2588 - sparse_categorical_accuracy: 0.6062 - sparse_top_k_categorical_accuracy: 0.7332 - val_loss: 3.2992 - val_sparse_categorical_accuracy: 0.5521 - val_sparse_top_k_categorical_accuracy: 0.7257
Epoch 10/10
281/281 [==============================] - 5s 18ms/step - loss: 3.2577 - sparse_categorical_accuracy: 0.6073 - sparse_top_k_categorical_accuracy: 0.7363 - val_loss: 3.2981 - val_sparse_categorical_accuracy: 0.5516 - val_sparse_top_k_categorical_accuracy: 0.7306
CPU times: user 18.9 s, sys: 3.86 s, total: 22.7 s
Wall time: 1min 1s