Step By Step
1、创建GPU实例
2、查看使用GPU卡情况
3、LetNet Code运行示例
4、显存释放问题
一、创建GPU实例
- 登录DSW控制台
二、查看使用GPU卡情况
!nvidia-smi
- 查看tf版本及devices信息
import tensorflow as tf
print(tf.__version__)
tf.config.list_physical_devices()
三、LetNet Code运行示例
- 3.1 Code Sample
import tensorflow as tf
import matplotlib.pyplot as plt
from tensorflow import keras
# 数据预处理函数
def preprocess(x, y):
x = tf.cast(x, dtype=tf.float32) / 255.
x = tf.reshape(x, [-1, 32, 32, 1])
y = tf.one_hot(y, depth=10) # one_hot 编码
return x, y
# 加载数据集
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.fashion_mnist.load_data()
# 样本图像周围补0（上下左右均补2个0），将28*28的图像转成32*32的图像
paddings = tf.constant([[0, 0], [2, 2], [2, 2]])
x_train = tf.pad(x_train, paddings)
x_test = tf.pad(x_test, paddings)
train_db = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_db = train_db.shuffle(10000) # 打乱训练集样本
train_db = train_db.batch(128)
train_db = train_db.map(preprocess)
test_db = tf.data.Dataset.from_tensor_slices((x_test, y_test))
test_db = test_db.shuffle(10000) # 打乱测试集样本
test_db = test_db.batch(128)
test_db = test_db.map(preprocess)
batch = 32
# 创建模型
model = keras.Sequential([
# 卷积层1
keras.layers.Conv2D(6, 5), # 使用6个5*5的卷积核对单通道32*32的图片进行卷积，结果得到6个28*28的特征图
keras.layers.MaxPooling2D(pool_size=2, strides=2), # 对28*28的特征图进行2*2最大池化，得到14*14的特征图
keras.layers.ReLU(), # ReLU激活函数
# 卷积层2
keras.layers.Conv2D(16, 5), # 使用16个5*5的卷积核对6通道14*14的图片进行卷积，结果得到16个10*10的特征图
keras.layers.MaxPooling2D(pool_size=2, strides=2), # 对10*10的特征图进行2*2最大池化，得到5*5的特征图
keras.layers.ReLU(), # ReLU激活函数
# 卷积层3
keras.layers.Conv2D(120, 5), # 使用120个5*5的卷积核对16通道5*5的图片进行卷积，结果得到120个1*1的特征图
keras.layers.ReLU(), # ReLU激活函数
# 将 (None, 1, 1, 120) 的下采样图片拉伸成 (None, 120) 的形状
keras.layers.Flatten(),
# 全连接层1
keras.layers.Dense(84, activation='relu'), # 120*84
# 全连接层2
keras.layers.Dense(10, activation='softmax') # 84*10
])
model.build(input_shape=(batch, 32, 32, 1))
model.summary()
model.compile(optimizer=keras.optimizers.Adam(), loss=keras.losses.CategoricalCrossentropy(), metrics=['accuracy'])
# 训练
history = model.fit(train_db, epochs=50)
# 损失下降曲线
plt.plot(history.history['loss'])
plt.title('model loss')
plt.ylabel('loss')
plt.xlabel('epoch')
plt.show()
# 测试
model.evaluate(test_db)- 运行结果
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/train-labels-idx1-ubyte.gz
32768/29515 [=================================] - 0s 1us/step
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/train-images-idx3-ubyte.gz
26427392/26421880 [==============================] - 1s 0us/step
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/t10k-labels-idx1-ubyte.gz
8192/5148 [===============================================] - 0s 0us/step
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/t10k-images-idx3-ubyte.gz
4423680/4422102 [==============================] - 0s 0us/step
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d (Conv2D) (32, 28, 28, 6) 156
_________________________________________________________________
max_pooling2d (MaxPooling2D) (32, 14, 14, 6) 0
_________________________________________________________________
re_lu (ReLU) (32, 14, 14, 6) 0
_________________________________________________________________
conv2d_1 (Conv2D) (32, 10, 10, 16) 2416
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (32, 5, 5, 16) 0
_________________________________________________________________
re_lu_1 (ReLU) (32, 5, 5, 16) 0
_________________________________________________________________
conv2d_2 (Conv2D) (32, 1, 1, 120) 48120
_________________________________________________________________
re_lu_2 (ReLU) (32, 1, 1, 120) 0
_________________________________________________________________
flatten (Flatten) (32, 120) 0
_________________________________________________________________
dense (Dense) (32, 84) 10164
_________________________________________________________________
dense_1 (Dense) (32, 10) 850
=================================================================
Total params: 61,706
Trainable params: 61,706
Non-trainable params: 0
_________________________________________________________________
Epoch 1/50
469/469 [==============================] - 2s 4ms/step - loss: 0.6662 - accuracy: 0.7553
Epoch 2/50
469/469 [==============================] - 2s 4ms/step - loss: 0.3988 - accuracy: 0.8569
Epoch 3/50
469/469 [==============================] - 2s 4ms/step - loss: 0.3414 - accuracy: 0.8751
Epoch 4/50
469/469 [==============================] - 2s 4ms/step - loss: 0.3081 - accuracy: 0.8861
Epoch 5/50
469/469 [==============================] - 2s 4ms/step - loss: 0.2888 - accuracy: 0.8938
Epoch 6/50
469/469 [==============================] - 2s 4ms/step - loss: 0.2733 - accuracy: 0.8997
Epoch 7/50
469/469 [==============================] - 2s 4ms/step - loss: 0.2590 - accuracy: 0.9039
Epoch 8/50
469/469 [==============================] - 2s 4ms/step - loss: 0.2464 - accuracy: 0.9078
Epoch 9/50
469/469 [==============================] - 2s 4ms/step - loss: 0.2358 - accuracy: 0.9126
Epoch 10/50
469/469 [==============================] - 2s 4ms/step - loss: 0.2292 - accuracy: 0.9155
Epoch 11/50
469/469 [==============================] - 2s 4ms/step - loss: 0.2175 - accuracy: 0.9190
Epoch 12/50
469/469 [==============================] - 2s 4ms/step - loss: 0.2096 - accuracy: 0.9227
Epoch 13/50
469/469 [==============================] - 2s 4ms/step - loss: 0.2022 - accuracy: 0.9239
Epoch 14/50
469/469 [==============================] - 2s 4ms/step - loss: 0.1929 - accuracy: 0.9284
Epoch 15/50
469/469 [==============================] - 2s 4ms/step - loss: 0.1875 - accuracy: 0.9301
Epoch 16/50
469/469 [==============================] - 2s 4ms/step - loss: 0.1776 - accuracy: 0.9342
Epoch 17/50
469/469 [==============================] - 2s 4ms/step - loss: 0.1737 - accuracy: 0.9349
Epoch 18/50
469/469 [==============================] - 2s 4ms/step - loss: 0.1658 - accuracy: 0.9379
Epoch 19/50
469/469 [==============================] - 2s 4ms/step - loss: 0.1592 - accuracy: 0.9407
Epoch 20/50
469/469 [==============================] - 2s 4ms/step - loss: 0.1515 - accuracy: 0.9438
Epoch 21/50
469/469 [==============================] - 2s 4ms/step - loss: 0.1461 - accuracy: 0.9445
Epoch 22/50
469/469 [==============================] - 2s 4ms/step - loss: 0.1403 - accuracy: 0.9467
Epoch 23/50
469/469 [==============================] - 2s 4ms/step - loss: 0.1340 - accuracy: 0.9487
Epoch 24/50
469/469 [==============================] - 2s 4ms/step - loss: 0.1295 - accuracy: 0.9506
Epoch 25/50
469/469 [==============================] - 2s 4ms/step - loss: 0.1235 - accuracy: 0.9536
Epoch 26/50
469/469 [==============================] - 2s 4ms/step - loss: 0.1189 - accuracy: 0.9545
Epoch 27/50
469/469 [==============================] - 2s 4ms/step - loss: 0.1134 - accuracy: 0.9571
Epoch 28/50
469/469 [==============================] - 2s 4ms/step - loss: 0.1059 - accuracy: 0.9602
Epoch 29/50
469/469 [==============================] - 2s 4ms/step - loss: 0.1028 - accuracy: 0.9615
Epoch 30/50
469/469 [==============================] - 2s 4ms/step - loss: 0.1006 - accuracy: 0.9616
Epoch 31/50
469/469 [==============================] - 2s 4ms/step - loss: 0.0936 - accuracy: 0.9644
Epoch 32/50
469/469 [==============================] - 2s 4ms/step - loss: 0.0892 - accuracy: 0.9663
Epoch 33/50
469/469 [==============================] - 2s 4ms/step - loss: 0.0846 - accuracy: 0.9683
Epoch 34/50
469/469 [==============================] - 2s 4ms/step - loss: 0.0807 - accuracy: 0.9692
Epoch 35/50
469/469 [==============================] - 2s 4ms/step - loss: 0.0756 - accuracy: 0.9711
Epoch 36/50
469/469 [==============================] - 2s 4ms/step - loss: 0.0763 - accuracy: 0.9718
Epoch 37/50
469/469 [==============================] - 2s 4ms/step - loss: 0.0700 - accuracy: 0.9732
Epoch 38/50
469/469 [==============================] - 2s 4ms/step - loss: 0.0687 - accuracy: 0.9748
Epoch 39/50
469/469 [==============================] - 2s 4ms/step - loss: 0.0648 - accuracy: 0.9754
Epoch 40/50
469/469 [==============================] - 2s 4ms/step - loss: 0.0611 - accuracy: 0.9771
Epoch 41/50
469/469 [==============================] - 2s 4ms/step - loss: 0.0602 - accuracy: 0.9774
Epoch 42/50
469/469 [==============================] - 2s 4ms/step - loss: 0.0564 - accuracy: 0.9793
Epoch 43/50
469/469 [==============================] - 2s 4ms/step - loss: 0.0533 - accuracy: 0.9807
Epoch 44/50
469/469 [==============================] - 2s 4ms/step - loss: 0.0531 - accuracy: 0.9803
Epoch 45/50
469/469 [==============================] - 2s 4ms/step - loss: 0.0499 - accuracy: 0.9809
Epoch 46/50
469/469 [==============================] - 2s 4ms/step - loss: 0.0452 - accuracy: 0.9827
Epoch 47/50
469/469 [==============================] - 2s 4ms/step - loss: 0.0433 - accuracy: 0.9847
Epoch 48/50
469/469 [==============================] - 2s 4ms/step - loss: 0.0439 - accuracy: 0.9838
Epoch 49/50
469/469 [==============================] - 2s 4ms/step - loss: 0.0367 - accuracy: 0.9864
Epoch 50/50
469/469 [==============================] - 2s 4ms/step - loss: 0.0415 - accuracy: 0.9847
79/79 [==============================] - 0s 3ms/step - loss: 0.5499 - accuracy: 0.9044
[0.5498712658882141, 0.9043999910354614]
四、显存释放问题
- 4.1 问题现象
- 4.2 Terminal fuser安装
sudo apt-get update
sudo apt-get install psmisc
- 4.3 查看僵尸进程
fuser -v /dev/nvidia*
- 4.4 查看具体这个进程调用GPU的情况
pmap -d PID- 4.5 强行关掉所有当前并未执行的僵尸进程
sudo kill -9 PID
更多参考
解决GPU显存未释放问题
Ubuntu：fuser command not found
TF Implementing LeNet-5 in TensorFlow 2.0
