Tebsorflow開源實現多GPU訓練cifar10資料集：cifar10_multi_gpu_train.py

Tensorflow開源實現cifar10神經網路：cifar10.py

Tensorflow中的並行分為模型並行和資料並行。模型並行需要根據不同模型設計不同的並行方式，其主要原理是將模型中不同計算節點放在不同硬體資源上運算。比較通用且能簡便地實現大規模並行的方式是資料並行，同時使用多個硬體資源來計算不同batch的資料梯度，然後彙總梯度進行全域性更新。

資料並行幾乎適用於所有深度學習模型，總是可以利用多塊GPU同時訓練多個batch資料，執行在每塊GPU上的模型都基於同一個神經網路，網路結構一樣，並且共享模型引數。

import os
import re
import time
import numpy as np
import tensorflow as tf
import cifar10_input
import cifar10

batch_size = 128
max_steps = 1000
num_gpus = 1 # gpu數量


# 在scope下生成神經網路並返回scope下的loss
def tower_loss(scope):
 # 資料集的路徑可以在cifar10.py中的tf.app.flags.DEFINE_string中定義
 images,labels = cifar10.distorted_inputs()
 logits = cifar10.inference(images) # 生成神經網路
 _ = cifar10.loss(logits,labels) # 不直接返回loss而是放到collection
 losses = tf.get_collection('losses',scope) # 獲取當前GPU上的loss(通過scope限定範圍)
 total_loss = tf.add_n(losses,name='total_loss')
 return total_loss


'''
外層是不同GPU計算的梯度，內層是某個GPU對應的不同var的值
tower_grads = 
[[(grad0_gpu0,var0_gpu0),(grad1_gpu0,var1_gpu0),...],[(grad0_gpu1,var0_gpu1),(grad1_gpu1,var1_gpu1),...]]
zip(*tower_grads)= 相當於轉置了
[[(grad0_gpu0,(grad0_gpu1,var0,gpu1),[(grad1_gpu0,...]]
'''


def average_gradients(tower_grads):
 average_grads = []
 for grad_and_vars in zip(*tower_grads):
  grads = [tf.expand_dims(g,0) for g,_ in grad_and_vars]
  grads = tf.concat(grads,0)
  grad = tf.reduce_mean(grads,0)
  grad_and_var = (grad,grad_and_vars[0][1])
  # [(grad0,var0),(grad1,var1),...]
  average_grads.append(grad_and_var)
 return average_grads


def train():
 # 預設的計算裝置為CPU
 with tf.Graph().as_default(),tf.device('/cpu:0'):
  # []表示沒有維度，為一個數
  # trainable=False,不會加入GraphKeys.TRAINABLE_VARIABLES參與訓練
  global_step = tf.get_variable('global_step',[],initializer=tf.constant_initializer(0),trainable=False)
  num_batches_per_epoch = cifar10.NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN / batch_size
  decay_steps = int(num_batches_per_epoch * cifar10.NUM_EPOCHS_PER_DECAY)
  # https://tensorflow.google.cn/api_docs/python/tf/train/exponential_decay
  # decayed_learning_rate = learning_rate * decay_rate ^ (global_step / decay_steps)
  # staircase is True,then global_step / decay_steps is an integer division
  lr = tf.train.exponential_decay(cifar10.INITIAL_LEARNING_RATE,global_step,decay_steps,cifar10.LEARNING_RATE_DECAY_FACTOR,staircase=True)
  opt = tf.train.GradientDescentOptimizer(lr)

  tower_grads = []
  for i in range(num_gpus):
   with tf.device('/gpu:%d' % i):
    with tf.name_scope('%s_%d' % (cifar10.TOWER_NAME,i)) as scope:
     loss = tower_loss(scope)
     # 讓神經網路的變數可以重用，所有GPU使用完全相同的引數
     # 讓下一個tower重用引數
     tf.get_variable_scope().reuse_variables()
     grads = opt.compute_gradients(loss)
     tower_grads.append(grads)
  grads = average_gradients(tower_grads)
  apply_gradient_op = opt.apply_gradients(grads,global_step=global_step)

  init = tf.global_variables_initializer()
  # True會自動選擇一個存在並且支援的裝置來執行
  sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
  sess.run(init)
  tf.train.start_queue_runners(sess=sess)

  for step in range(max_steps):
   start_time = time.time()
   _,loss_value = sess.run([apply_gradient_op,loss])
   duration = time.time() - start_time

   if step % 10 == 0:
    num_examples_per_step = batch_size * num_gpus
    examples_per_sec = num_examples_per_step / duration
    sec_per_batch = duration / num_gpus

    print('step %d,loss=%.2f(%.1f examples/sec;%.3f sec/batch)'
      % (step,loss_value,examples_per_sec,sec_per_batch))



if __name__ == '__main__':
 train()

以上這篇Tensorflow實現多GPU並行方式就是小編分享給大家的全部內容了，希望能給大家一個參考，也希望大家多多支援我們。