1. 程式人生 > >【TensorFlow】多GPU訓練:示例程式碼解析

【TensorFlow】多GPU訓練:示例程式碼解析

使用多GPU有助於提升訓練速度和調參效率。
本文主要對tensorflow的示例程式碼進行註釋解析:cifar10_multi_gpu_train.py


1080Ti下加速效果如下(batch=128)
單卡:
在這裡插入圖片描述
兩個GPU比單個GPU加速了近一倍 :
在這裡插入圖片描述

在這裡插入圖片描述

1.簡介

多GPU訓練分為:
資料並行和模型並行
單機多卡和多機多卡

2.示例程式碼解讀

官方示例程式碼給出了使用多個GPU計算的流程:

  • CPU 做為引數伺服器
  • 多個GPU計算彙總更新

#--------------------------Multi-GPUs-code------------------------#

1.demo檔案的說明部分
# Copyright 2015 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
# # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License.
# ============================================================================== """A binary to train CIFAR-10 using multiple GPUs with synchronous updates. 在100k大概256epochs後可以達到約86%的精度 Accuracy: cifar10_multi_gpu_train.py achieves ~86% accuracy after 100K steps (256 epochs of data) as judged by cifar10_eval.py. Speed: With batch_size 128. 下面是一些訓練參考時間: System | Step Time (sec/batch) | Accuracy -------------------------------------------------------------------- 1 Tesla K20m | 0.35-0.60 | ~86% at 60K steps (5 hours) 1 Tesla K40m | 0.25-0.35 | ~86% at 100K steps (4 hours) 2 Tesla K20m | 0.13-0.20 | ~84% at 30K steps (2.5 hours) 3 Tesla K20m | 0.13-0.18 | ~84% at 30K steps 4 Tesla K20m | ~0.10 | ~84% at 30K steps """ from __future__ import absolute_import from __future__ import division from __future__ import print_function #匯入版本支援 from datetime import datetime #匯入時間模組 import os.path #路徑模組用於穿件資料夾 import re #正則表示式模組 import time import numpy as np from six.moves import xrange # pylint: disable=redefined-builtin #這句類似python range,py2/py3相容模組,也可將文中的xrange替換為range import tensorflow as tf #匯入tensorflow import cifar10 #匯入自定義的cifar10.py,包含了各種資料初始化、模型構建、損失和訓練函式
2.定義一些flags

這裡包含了對於資料目錄、最大batch步數、gpu數目和日誌檔案定義等

FLAGS = tf.app.flags.FLAGS    #定義引數flags,隨後利用FLAGS讀取引數
#https://blog.csdn.net/m0_37041325/article/details/77448971
#https://blog.csdn.net/weiqi_fan/article/details/72722510

#定義引數對應的預設值
tf.app.flags.DEFINE_string('train_dir', './your/path/to/data/cifar10_train',
                           """Directory where to write event logs """
                           """and checkpoint.""")
tf.app.flags.DEFINE_integer('max_steps', 1000000,
                            """Number of batches to run.""")
tf.app.flags.DEFINE_integer('num_gpus', 1,
                            """How many GPUs to use.""")
tf.app.flags.DEFINE_boolean('log_device_placement', False,
                            """Whether to log device placement.""")

3.定義損失彙總函式和梯度平均函式

主要定義了各個GPU上的損失函式及其合併


def tower_loss(scope, images, labels):
  """Calculate the total loss on a single tower running the CIFAR model.
     計算單個tower上的總損失

  Args:
    scope: 特定tower的名稱空間, e.g. 'tower_0'
    images: Images. 4D tensor of shape [batch_size, height, width, 3].
    labels: Labels. 1D tensor of shape [batch_size].

  Returns:
     Tensor of shape [] containing the 某個批次資料的總損失
  """
	
	# 計算圖構建的輸出
	logits = cifar10.inference(images)
	
	# 呼叫函式計算loss
	_ = cifar10.loss(logits, labels)
	
	# 綜合tower的loss
	losses = tf.get_collection('losses', scope)
	
	# 計算當前tower的總loss
	total_loss = tf.add_n(losses, name='total_loss')
	
	# Attach a scalar summary to all individual losses and the total loss; do the
	# same for the averaged version of the losses.
	for l in losses + [total_loss]:
		# Remove 'tower_[0-9]/' from the name in case this is a multi-GPU training
		# session. 清理tensorboard
		loss_name = re.sub('%s_[0-9]*/' % cifar10.TOWER_NAME, '', l.op.name)
		tf.summary.scalar(loss_name, l)    #tensorboard視覺化
	
	return total_loss

	"""
	#最後得到的total_loss
	#每呼叫一次得到一個GPU的loss
    Tensor("tower_0/total_loss_1:0", shape=(), dtype=float32, device=/device:GPU:0)
	Tensor("tower_1/total_loss_1:0", shape=(), dtype=float32, device=/device:GPU:1)
    """

這部分梯度的綜合比較複雜,把它拆分出來分析,主要過程可以總結為

-首先讀入每個GPU(Tower)中的(梯度,變數),這些變數按照GPU 分為多個字列表儲存,[[GPUi],.......,[GPUn]]
-每個子列表中包含了一整個模型,對應了一整套的[(梯度,變數),........,(梯度,變數)]<-gpui
-將不同GPU中的同一個變數及其梯度((grad0_gpu0, var0_gpu0),.....,(grad0_gpun, var0_gpun))抽取出來,

#定義梯度,這些梯度來自於各個GPU的綜合
def average_gradients(tower_grads):
  """Calculate the average gradient for each shared variable across all towers.

  #這個函式對塔式伺服器中的GPU提供了同步點
  Note that this function provides a synchronization point across all towers.

  Args:
    #輸入引數為list格式,包含了由一系列元組(梯度,變數)組成的子列表
    #外部的list計算獨立梯度,內部計算綜合梯度
    tower_grads: List of lists of (gradient, variable) tuples. The outer list
      is over individual gradients. The inner list is over the gradient
      calculation for each tower.
  Returns:
     #在所有節點上平均後返回
     List of pairs of (gradient, variable) where the gradient has been averaged
     across all towers.
  """

"""例項
對於兩個GPU來說,就是兩個tower,針對這裡例子,tower_gpu中包含了下面這些內容
tower_grads = [[tower0_grad],[tower1_grads]]>>>包含了第一塊gpu的變數梯度和第二塊GPU的變數梯度,他們被放在一個大的列表裡outer-list;
而其中的每一個tower-n_grads 又是一個小的列表inner-list,包含了整個模型的梯度和變數。
[tower-n_grads] = [(grad0,variable0),.......,(gradn,variablen)

#我們將輸入的變數打印出來觀察
>>> tower_grads:
[
    [
        (<tf.Tensor 'tower_0/gradients/tower_0/conv1/Conv2D_grad/tuple/control_dependency_1:0' shape=(5, 5, 3, 64) dtype=float32>, <tf.Variable 'conv1/weights:0' shape=(5, 5, 3, 64) dtype=float32_ref>), 
        (<tf.Tensor 'tower_0/gradients/tower_0/conv1/BiasAdd_grad/tuple/control_dependency_1:0' shape=(64,) dtype=float32>, <tf.Variable 'conv1/biases:0' shape=(64,) dtype=float32_ref>), 
        (<tf.Tensor 'tower_0/gradients/tower_0/conv2/Conv2D_grad/tuple/control_dependency_1:0' shape=(5, 5, 64, 64) dtype=float32>, <tf.Variable 'conv2/weights:0' shape=(5, 5, 64, 64) dtype=float32_ref>),
        (<tf.Tensor 'tower_0/gradients/tower_0/conv2/BiasAdd_grad/tuple/control_dependency_1:0' shape=(64,) dtype=float32>, <tf.Variable 'conv2/biases:0' shape=(64,) dtype=float32_ref>), 
        (<tf.Tensor 'tower_0/gradients/AddN_1:0' shape=(2304, 384) dtype=float32>, <tf.Variable 'local3/weights:0' shape=(2304, 384) dtype=float32_ref>), 
        (<tf.Tensor 'tower_0/gradients/tower_0/local3/add_grad/tuple/control_dependency_1:0' shape=(384,) dtype=float32>, <tf.Variable 'local3/biases:0' shape=(384,) dtype=float32_ref>), 
        (<tf.Tensor 'tower_0/gradients/AddN:0' shape=(384, 192) dtype=float32>, <tf.Variable 'local4/weights:0' shape=(384, 192) dtype=float32_ref>), 
        (<tf.Tensor 'tower_0/gradients/tower_0/local4/add_grad/tuple/control_dependency_1:0' shape=(192,) dtype=float32>, <tf.Variable 'local4/biases:0' shape=(192,) dtype=float32_ref>), 
        (<tf.Tensor 'tower_0/gradients/tower_0/softmax_linear/MatMul_grad/tuple/control_dependency_1:0' shape=(192, 10) dtype=float32>, <tf.Variable 'softmax_linear/weights:0' shape=(192, 10) dtype=float32_ref>), 
        (<tf.Tensor 'tower_0/gradients/tower_0/softmax_linear/softmax_linear_grad/tuple/control_dependency_1:0' shape=(10,) dtype=float32>, <tf.Variable 'softmax_linear/biases:0' shape=(10,) dtype=float32_ref>)],
    [
        (<tf.Tensor 'tower_1/gradients/tower_1/conv1/Conv2D_grad/tuple/control_dependency_1:0' shape=(5, 5, 3, 64) dtype=float32>, <tf.Variable 'conv1/weights:0' shape=(5, 5, 3, 64) dtype=float32_ref>), 
        (<tf.Tensor 'tower_1/gradients/tower_1/conv1/BiasAdd_grad/tuple/control_dependency_1:0' shape=(64,) dtype=float32>, <tf.Variable 'conv1/biases:0' shape=(64,) dtype=float32_ref>), 
        (<tf.Tensor 'tower_1/gradients/tower_1/conv2/Conv2D_grad/tuple/control_dependency_1:0' shape=(5, 5, 64, 64) dtype=float32>, <tf.Variable 'conv2/weights:0' shape=(5, 5, 64, 64) dtype=float32_ref>), 
        (<tf.Tensor 'tower_1/gradients/tower_1/conv2/BiasAdd_grad/tuple/control_dependency_1:0' shape=(64,) dtype=float32>, <tf.Variable 'conv2/biases:0' shape=(64,) dtype=float32_ref>), 
        (<tf.Tensor 'tower_1/gradients/AddN_1:0' shape=(2304, 384) dtype=float32>, <tf.Variable 'local3/weights:0' shape=(2304, 384) dtype=float32_ref>), 
        (<tf.Tensor 'tower_1/gradients/tower_1/local3/add_grad/tuple/control_dependency_1:0' shape=(384,) dtype=float32>, <tf.Variable 'local3/biases:0' shape=(384,) dtype=float32_ref>), 
        (<tf.Tensor 'tower_1/gradients/AddN:0' shape=(384, 192) dtype=float32>, <tf.Variable 'local4/weights:0' shape=(384, 192) dtype=float32_ref>), 
        (<tf.Tensor 'tower_1/gradients/tower_1/local4/add_grad/tuple/control_dependency_1:0' shape=(192,) dtype=float32>, <tf.Variable 'local4/biases:0' shape=(192,) dtype=float32_ref>), 
        (<tf.Tensor 'tower_1/gradients/tower_1/softmax_linear/MatMul_grad/tuple/control_dependency_1:0' shape=(192, 10) dtype=float32>, <tf.Variable 'softmax_linear/weights:0' shape=(192, 10) dtype=float32_ref>), 
        (<tf.Tensor 'tower_1/gradients/tower_1/softmax_linear/softmax_linear_grad/tuple/control_dependency_1:0' shape=(10,) dtype=float32>, <tf.Variable 'softmax_linear/biases:0' shape=(10,) dtype=float32_ref>)
    ]
]

"""

	average_grads = []
	#對輸入元組進行解壓
		for grad_and_vars in zip(*tower_grads):    #在各個變數var上迴圈
		#   grad_and_vars: ((grad0_gpu0, var0_gpu0), ... , (grad0_gpuN, var0_gpuN))
		#   遍歷var0及其梯度在不同GPU上的分佈,此例子中
		#((<tf.Tensor 'tower_0/gradients/tower_0/conv1/Conv2D_grad/tuple/control_dependency_1:0' shape=(5, 5, 3, 64) dtype=float32>, <tf.Variable 'conv1/weights:0' shape=(5, 5, 3, 64) dtype=float32_ref>), 
		#(<tf.Tensor 'tower_1/gradients/tower_1/conv1/Conv2D_grad/tuple/control_dependency_1:0' shape=(5, 5, 3, 64) dtype=float32>, <tf.Variable 'conv1/weights:0' shape=(5, 5, 3, 64) dtype=float32_ref>))

		grads = []
		for g, _ in grad_and_vars:    #對所有GPU上的同一變數的梯度進行組合
			# Add 0 dimension to the gradients to represent the tower.
			expanded_g = tf.expand_dims(g, 0)
			
			# Append on a 'tower' dimension which we will average over below.
			#加上tower維度
			grads.append(expanded_g)
		
		#在tower維度上進行平均
		grad = tf.concat(axis=0, values=grads)    #在tower維度上,對不同的GPU求均值
		grad = tf.reduce_mean(grad, 0)     #得到所有變數及其梯度的均值
		
		# 引數由於共享冗餘,所以只需要返回變數在首個tower的指標
		v = grad_and_vars[0][1]              #指標varxx-gpuxx
		grad_and_var = (grad, v)             #合併為元組  得到某個變數綜合後的平均梯度,及變數名指標。
		average_grads.append(grad_and_var)   #新增新的梯度和v指標,新增各個var
	return average_grads
	
    """最後我們觀察返回的引數
     >>> print(average_grads)
    [(<tf.Tensor 'Mean:0' shape=(5, 5, 3, 64) dtype=float32>, <tf.Variable 'conv1/weights:0' shape=(5, 5, 3, 64) dtype=float32_ref>),
	 (<tf.Tensor 'Mean_1:0' shape=(64,) dtype=float32>, <tf.Variable 'conv1/biases:0' shape=(64,) dtype=float32_ref>),
	 (<tf.Tensor 'Mean_2:0' shape=(5, 5, 64, 64) dtype=float32>, <tf.Variable 'conv2/weights:0' shape=(5, 5, 64, 64) dtype=float32_ref>),
	 (<tf.Tensor 'Mean_3:0' shape=(64,) dtype=float32>, <tf.Variable 'conv2/biases:0' shape=(64,) dtype=float32_ref>),
	 (<tf.Tensor 'Mean_4:0' shape=(2304, 384) dtype=float32>, <tf.Variable 'local3/weights:0' shape=(2304, 384) dtype=float32_ref>), 
	 (<tf.Tensor 'Mean_5:0' shape=(384,) dtype=float32>, <tf.Variable 'local3/biases:0' shape=(384,) dtype=float32_ref>), 
	 (<tf.Tensor 'Mean_6:0' shape=(384, 192) dtype=float32>, <tf.Variable 'local4/weights:0' shape=(384, 192) dtype=float32_ref>), 
	 (<tf.Tensor 'Mean_7:0' shape=(192,) dtype=float32>, <tf.Variable 'local4/biases:0' shape=(192,) dtype=float32_ref>), 
	 (<tf.Tensor 'Mean_8:0' shape=(192, 10) dtype=float32>, <tf.Variable 'softmax_linear/weights:0' shape=(192, 10) dtype=float32_ref>), 
	 (<tf.Tensor 'Mean_9:0' shape=(10,) dtype=float32>, <tf.Variable 'softmax_linear/biases:0' shape=(10,) dtype=float32_ref>)
	]
    可以看到是多gpu平均後的梯度和對應的變數
    """



4.訓練

訓練部分主要包括了構建計算圖、定義計算引數、優化器、


def train():
	"""Train CIFAR-10 for a number of steps."""
	with tf.Graph().as_default(), tf.device('/cpu:0'):
		# Create a variable to count the number of train() calls. This equals the
		# number of batches processed * FLAGS.num_gpus.
		global_step = tf.get_variable(
		    'global_step', [],
		    initializer=tf.constant_initializer(0), trainable=False)
		
		# Calculate the learning rate schedule.
		num_batches_per_epoch = (cifar10.NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN /
		                         FLAGS.batch_size / FLAGS.num_gpus)
		decay_steps = int(num_batches_per_epoch * cifar10.NUM_EPOCHS_PER_DECAY)
		
		# Decay the learning rate exponentially based on the number of steps.
		lr = tf.train.exponential_decay(cifar10.INITIAL_LEARNING_RATE,
		                                global_step,
		                                decay_steps,
		                                cifar10.LEARNING_RATE_DECAY_FACTOR,
		                                staircase=True)
		
		# Create an optimizer that performs gradient descent.
		opt = tf.train.GradientDescentOptimizer(lr)
		#-----------------------------上面定義引數、定義優化器-----------------------#
		
		# 影象和標籤的batch輸入
		images, labels = cifar10.distorted_inputs()
		batch_queue = tf.contrib.slim.prefetch_queue.prefetch_queue(
		      [images, labels], capacity=2 * FLAGS.num_gpus)


		# 計算每一個gpu上的梯度,放入tower_grads中.
		tower_grads = []
		with tf.variable_scope(tf.get_variable_scope()):
			for i in xrange(FLAGS.num_gpus):
				with tf.device('/gpu:%d' % i):
					with tf.name_scope('%s_%d' % (cifar10.TOWER_NAME, i)) as scope:
						# Dequeues one batch for the GPU
						image_batch, label_batch = batch_queue.dequeue()
						# Calculate the loss for one tower of the CIFAR model. This function
						# constructs the entire CIFAR model but shares the variables across
						# all towers.
						loss = tower_loss(scope, image_batch, label_batch)
						
						# Reuse variables for the next tower.
						tf.get_variable_scope().reuse_variables()
						
						# Retain the summaries from the final tower.
						summaries = tf.get_collection(tf.GraphKeys.SUMMARIES, scope)
						
						# Calculate the gradients for the batch of data on this CIFAR tower.
						grads = opt.compute_gradients(loss)
						
						# Keep track of the gradients across all towers.
						tower_grads.append(grads)
	
		# 計算平均梯度
		# 注意同步指標.
		grads = average_gradients(tower_grads)
		
		# tensorboard顯示學習率
		summaries.append(tf.summary.scalar('learning_rate', lr))
		
		# 各種梯度的tensorboard直方圖顯示
		for grad, var in grads:
			if grad is not None:
				summaries.append(tf.summary.histogram(var.op.name + '/gradients', grad))
		
		# 利用計算出的平均梯度來進行優化
		apply_gradient_op = opt.apply_gradients(grads, global_step=global_step)
		
		# 各種變數的直方圖
		for var in tf.trainable_variables():
			summaries.append(tf.summary.histogram(var.op.name, var))
		
		# 跟蹤所有變數的移動平均
		variable_averages = tf.train.ExponentialMovingAverage(
		    cifar10.MOVING_AVERAGE_DECAY, global_step)
		variables_averages_op = variable_averages.apply(tf.trainable_variables())
		
		# 將所有操作組合進單一操作
		train_op = tf.group(apply_gradient_op, variables_averages_op)
		
		# 儲存相關操作
		saver = tf.train.Saver(tf.global_variables())
		
		# 建立綜合操作
		summary_op = tf.summary.merge(summaries)
		
		# 初始化
		init = tf.global_variables_initializer()
		
		# 開始計算
		# Start running operations on the Graph. allow_soft_placement must be set to
		# True to build towers on GPU, as some of the ops do not have GPU
		# implementations.
		sess = tf.Session(config=tf.ConfigProto(
		    allow_soft_placement=True,
		    log_device_placement=FLAGS.log_device_placement))
		sess.run(init)
		
		# Start the queue runners.
		tf.train.start_queue_runners(sess=sess)
		
		#將訓練過程記錄下來,tensorboard視覺化
		summary_writer = tf.summary.FileWriter(FLAGS.train_dir, sess.graph)
		
		#最大步數迭代訓練,顯示時間和loss
		for step in xrange(FLAGS.max_steps):
			start_time = time.time()
			_, loss_value = sess.run([train_op, loss])
			duration = time.time() - start_time
		
			assert not np.isnan(loss_value), 'Model diverged with loss = NaN'
		#---------------------------下面是不同check steps的時候顯示的資訊-----------------#
			if step % 10 == 0:
			    num_examples_per_step = FLAGS.batch_size * FLAGS.num_gpus
			    examples_per_sec = num_examples_per_step / duration
			    sec_per_batch = duration / FLAGS.num_gpus
			
			    format_str = ('%s: step %d, loss = %.2f (%.1f examples/sec; %.3f '
			                  'sec/batch)')
			    print (format_str % (datetime.now(), step, loss_value,
			                         examples_per_sec, sec_per_batch))
		
		  	if step % 100 == 0:
			  	summary_str = sess.run(summary_op)
			  	summary_writer.add_summary(summary_str, step)
		
		  # Save the model checkpoint periodically.
		  	if step % 1000 == 0 or (step + 1) == FLAGS.max_steps:
			  	checkpoint_path = os.path.join(FLAGS.train_dir, 'model.ckpt')
			  	saver.save(sess, checkpoint_path, global_step=step)

#注,此處程式碼較長,執行時需要注意tab鍵/空格鍵是否正確---indent
啟動主函式訓練

def main(argv=None):  # pylint: disable=unused-argument
  cifar10.maybe_download_and_extract()    #沒資料需要下載,這個函式在cifar10.py裡
  if tf.gfile.Exists(FLAGS.train_dir):
    tf.gfile.DeleteRecursively(FLAGS.train_dir)
  tf.gfile.MakeDirs(FLAGS.train_dir)
  train()


if __name__ == '__main__':
  tf.app.run()
  #可以愉快的運行了

在這裡插入圖片描述
pic from pexels.com


ref:
demo:https://github.com/tensorflow/models/blob/master/tutorials/image/cifar10/cifar10_multi_gpu_train.py
https://blog.csdn.net/lqfarmer/article/details/70339330
https://blog.csdn.net/weixin_40546602/article/details/81414321
https://blog.csdn.net/guotong1988/article/details/74355637