我正在使用一台有 2 个 GPU 的机器 Titan Black 来训练我的深度学习模型，该模型有 3 层(3x3、3x3 和 5x5(。
训练运行得很好，但是当我观看 nvidia-smi(每 1 秒观看一次(时，我意识到我的程序只使用一个 GPU 进行计算，即使第一个达到 100%，第二个也始终为 0%。
我正在尝试使用 tf.device 为每个任务分配特定任务，但随后它们一个接一个地运行，而不是并行运行，并且总时间甚至增加了，而不是减少(我猜是因为 2 个 GPU 必须相互交换值(
下面是我的程序。这很混乱，也许你只需要注意我使用 tf.device 的图表就足够了......
非常感谢！
import tensorflow as tf
import numpy as np
from six.moves import cPickle as pickle
import matplotlib.pyplot as plt
from os import listdir, sys
from os.path import isfile, join
from time import gmtime, strftime
import time
def validatePath(path):
path = path.replace("\","/")
if (path[len(path)-1] != "/"):
path = path + "/"
return path
hidden_size_default = np.array([16, 32, 64, 32])
cnn1_default = 3
cnn2_default = 3
cnn3_default = 5
SIZE_BATCH_VALID = 200
input_path = 'ARCHIVES-sub-dataset'
output_path = 'ARCHIVES-model'
log_address = "trainlog.txt"
tf.app.flags.DEFINE_integer('h0', hidden_size_default[0], 'Size of hidden layer 0th')
tf.app.flags.DEFINE_integer('h1', hidden_size_default[1], 'Size of hidden layer 1st')
tf.app.flags.DEFINE_integer('h2', hidden_size_default[2], 'Size of hidden layer 2nd')
tf.app.flags.DEFINE_integer('h3', hidden_size_default[3], 'Size of hidden layer 3rd')
tf.app.flags.DEFINE_integer('k1', cnn1_default , 'Size of kernel 1st')
tf.app.flags.DEFINE_integer('k2', cnn2_default , 'Size of kernel 2nd')
tf.app.flags.DEFINE_integer('k3', cnn3_default , 'Size of kernel 3rd')
tf.app.flags.DEFINE_string('input_path', input_path, 'The parent directory which contains 2 directories: dataset and label')
tf.app.flags.DEFINE_string('output_path', output_path, 'The directory which will store models (you have to create)')
tf.app.flags.DEFINE_string('log_address', log_address, 'The file name which will store the log')
FLAGS = tf.app.flags.FLAGS
load_path = FLAGS.input_path
save_model_path = FLAGS.output_path
log_addr = FLAGS.log_address
load_path = validatePath(load_path)
save_model_path = validatePath(save_model_path)
cnn1 = FLAGS.k1
cnn2 = FLAGS.k2
cnn3 = FLAGS.k3
hidden_size = np.array([FLAGS.h0, FLAGS.h1, FLAGS.h2, FLAGS.h3])
# Shuffle the dataset and its label
def randomize(dataset, labels):
permutation = np.random.permutation(labels.shape[0])
shuffled_dataset = dataset[permutation,:]
shuffled_labels = labels[permutation]
return shuffled_dataset, shuffled_labels
def writemyfile(mystring):
with open(log_addr, "a") as myfile:
myfile.write(str(mystring + "n"))
num_labels = 5
def accuracy(predictions, labels):
return (100.0 * np.sum(np.argmax(predictions, 1) == np.argmax(labels, 1))/ predictions.shape[0])
def weight_variable(shape):
initial = tf.truncated_normal(shape, stddev=0.1)
return tf.Variable(initial)
def bias_variable(shape):
initial = tf.constant(0.1, shape=shape)
return tf.Variable(initial)
def conv2d(x, W):
return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')
def max_pool_2x2(x):
return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
def DivideSets(input_set):
length_set = input_set.shape[0]
index_70 = int(length_set*0.7)
index_90 = int(length_set*0.9)
set_train = input_set[0:index_70]
set_valid = input_set[index_70:index_90]
set_test = input_set[index_90:length_set]
return np.float32(set_train), np.float32(set_valid), np.float32(set_test)
# from 1-value labels to 5 values of (0 and 1)
def LabelReconstruct(label_set):
label_set = label_set.astype(int)
new_label_set = np.zeros(shape=(len(label_set),num_labels))
for i in range(len(label_set)):
new_label_set[i][label_set[i]] = 1
return new_label_set.astype(int)
def LoadDataSet(load_path):
list_data = [f for f in listdir(load_path + "dataset/") if isfile(join(load_path + "dataset/", f))]
list_label = [f for f in listdir(load_path + "label/") if isfile(join(load_path + "dataset/", f))]
if list_data.sort() == list_label.sort():
return list_data
else:
print("data and labels are not suitable")
return 0
# load, randomize, normalize images and reconstruct labels
def PrepareData(*arg):
filename = arg[0]
loaded_dataset = pickle.load( open( load_path + "dataset/" + filename, "rb" ))
loaded_labels = pickle.load( open( load_path + "label/" + filename, "rb" ))
if len(arg) == 1:
datasize = len(loaded_labels)
elif len(arg) == 2:
datasize = int(arg[1])
else:
print("not more than 2 arguments please!")
dataset_full,labels_full = randomize(loaded_dataset[0:datasize], loaded_labels[0:datasize])
return NormalizeData(dataset_full), LabelReconstruct(labels_full)
def NormalizeData(dataset):
dataset = dataset - (dataset.mean())
dataset = dataset / (dataset.std())
return dataset
### LOAD DATA
listfiles = LoadDataSet(load_path)
# divide
listfiles_train = listfiles[0:15]
listfiles_valid = listfiles[15:25]
listfiles_test = listfiles[25:len(listfiles)]
graphCNN = tf.Graph()
with graphCNN.as_default():
with tf.device('/gpu:0'):
x = tf.placeholder(tf.float32, shape=(None, 224,224,3)) # X
y_ = tf.placeholder(tf.float32, shape=(None, num_labels)) # Y_
dropout = tf.placeholder(tf.float32)
if dropout == 1.0:
keep_prob = tf.constant([0.2, 0.3, 0.5], dtype=tf.float32)
else:
keep_prob = tf.constant([1.0, 1.0, 1.0], dtype=tf.float32)
weights_1 = weight_variable([cnn1,cnn1,3, hidden_size[0]])
biases_1 = bias_variable([hidden_size[0]])
weights_2 = weight_variable([cnn2,cnn2,hidden_size[0], hidden_size[1]])
biases_2 = bias_variable([hidden_size[1]])
weights_3 = weight_variable([cnn3,cnn3,hidden_size[1], hidden_size[2]])
biases_3 = bias_variable([hidden_size[2]])
weights_4 = weight_variable([56 * 56 * hidden_size[2], hidden_size[3]])
biases_4 = bias_variable([hidden_size[3]])
weights_5 = weight_variable([hidden_size[3], num_labels])
biases_5 = bias_variable([num_labels])
def model(data):
with tf.device('/gpu:1'):
train_hidden_1 = tf.nn.relu(conv2d(data, weights_1) + biases_1)
train_hidden_2 = max_pool_2x2(tf.nn.relu(conv2d(train_hidden_1, weights_2) + biases_2))
train_hidden_2_drop = tf.nn.dropout(train_hidden_2, keep_prob[0])
train_hidden_3 = max_pool_2x2(tf.nn.relu(conv2d(train_hidden_2_drop, weights_3) + biases_3))
train_hidden_3_drop = tf.nn.dropout(train_hidden_3, keep_prob[1])
train_hidden_3_drop = tf.reshape(train_hidden_3_drop,[-1, 56 * 56 * hidden_size[2]])
train_hidden_4 = tf.nn.relu(tf.matmul(train_hidden_3_drop, weights_4) + biases_4)
train_hidden_4_drop = tf.nn.dropout(train_hidden_4, keep_prob[2])
logits = tf.matmul(train_hidden_4_drop, weights_5) + biases_5
return logits
t_train_labels = tf.argmax(y_, 1) # From one-hot (one and zeros) vectors to values
loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=model(x), labels=t_train_labels))
optimizer = tf.train.AdamOptimizer(0.01).minimize(loss)
y = tf.nn.softmax(model(x))
### RUNNING
print("log address: %s" % (log_addr))
#num_steps = 10001
times_repeat = 20 # number of epochs
batch_size = 100
with tf.Session(graph=graphCNN,config=tf.ConfigProto(log_device_placement=True)) as session:
tf.initialize_all_variables().run()
saver = tf.train.Saver(max_to_keep=0)
writemyfile("---ARCHIVES_M1----")
mytime = strftime("%Y-%m-%d %H:%M:%S", time.localtime())
writemyfile(str("nTime: %s nLayers: %d,%d,%d epochs: %d" % (mytime,cnn1,cnn2,cnn3,times_repeat)))
writemyfile("Train files:" + str(listfiles_train))
writemyfile("Valid files:" + str(listfiles_valid))
writemyfile("Test files:" + str(listfiles_test))
print("Model will be saved in file: %s" % save_model_path)
writemyfile(str("Model will be saved in file: %s" % save_model_path))
### TRAINING & VALIDATION
valid_accuracies_epochs = np.array([])
for time_repeat in range(times_repeat):
print("- time_repeat:",time_repeat)
writemyfile("- time_repeat:"+str(time_repeat))
for file_train in listfiles_train:
file_train_id = int(file_train[0:len(file_train)-4])
time_start_this_file = time.time()
#LOAD DATA
print("- - file:",file_train_id, end=' ')
writemyfile("- - file:" + str(file_train_id))
Data_train, Label_train= PrepareData(file_train)
for step in range(0,len(Data_train)-batch_size,batch_size):
batch_data = Data_train[step:step+batch_size]
batch_labels = Label_train[step:step+batch_size]
feed_dict = {x : batch_data, y_ : batch_labels, dropout: 1.0}
opti, l, predictions = session.run([optimizer, loss, y], feed_dict=feed_dict)
train_accuracies = np.array([])
for index_tr_accu in range(0,len(Data_train)-SIZE_BATCH_VALID,SIZE_BATCH_VALID):
current_predictions = y.eval(feed_dict={x: Data_train[index_tr_accu:index_tr_accu+SIZE_BATCH_VALID],dropout: 0.0})
current_accuracy = accuracy(current_predictions, Label_train[index_tr_accu:index_tr_accu+SIZE_BATCH_VALID])
train_accuracies = np.r_[train_accuracies,current_accuracy]
train_accuracy = train_accuracies.mean()
print("batch accu: %.2f%%" %(train_accuracy),end=" | ")
writemyfile("batch accu: %.2f%%" %(train_accuracy))
time_done_this_file = time.time() - time_start_this_file
print("time: %.2fs" % (time_done_this_file))
writemyfile("time: %.2fs" % (time_done_this_file))
# save model
model_addr = save_model_path + "model335" + "-epoch-" + str(time_repeat) + ".ckpt"
save_path = saver.save(session, model_addr,) # max_to_keep default was 5
mytime = strftime("%Y-%m-%d %H:%M:%S", time.localtime())
print("epoch finished at %s n model address: %s" % (mytime,model_addr))
writemyfile("epoch finished at %s n model address: %s" % (mytime,model_addr))
# validation
valid_accuracies = np.array([])
for file_valid in listfiles_valid:
file_valid_id = int(file_valid[0:len(file_valid)-4])
Data_valid, Label_valid = PrepareData(file_valid)
for index_vl_accu in range(0,len(Data_valid)-SIZE_BATCH_VALID,SIZE_BATCH_VALID):
current_predictions = y.eval(feed_dict={x: Data_valid[index_vl_accu:index_vl_accu+SIZE_BATCH_VALID],dropout: 0.0})
current_accuracy = accuracy(current_predictions, Label_valid[index_vl_accu:index_vl_accu+SIZE_BATCH_VALID])
valid_accuracies = np.r_[valid_accuracies,current_accuracy]
valid_accuracy = valid_accuracies.mean()
print("epoch %d - valid accu: %.2f%%" %(time_repeat,valid_accuracy))
writemyfile("epoch %d - valid accu: %.2f%%" %(time_repeat,valid_accuracy))
valid_accuracies_epochs = np.hstack([valid_accuracies_epochs,valid_accuracy])
print('Done!!')
writemyfile(str('Done!!'))
session.close()
更新：我发现cifar10_multi_gpu_train.py似乎是使用多GPU进行训练的一个很好的例子，但老实说，我不知道如何应用于我的情况。
我认为你需要改变
def model(data):
with tf.device('/gpu:1'):
自：
def model(data):
for d in ['/gpu:0', '/gpu:1']:
with tf.device(d):
并抛弃这条线with tf.device('/gpu:0'):
因为在第一with tf.device...你只是在做启蒙变量，然后您将使用下一个with tf.device重置设备。
让我知道这是否有效，因为我无法测试它。