如何利用微信監管你的TF訓練

雷鋒網 AI科技評論按：本文作者Coldwings，雷鋒網 AI科技評論獲其授權發布。

之前回答問題【在機器學習模型的訓練期間，大概幾十分鐘到幾小時不等，大家都會在等實驗的時候做什么？】的時候，說到可以用微信來管著訓練，完全不用守著。沒想到這么受歡迎……

原問題下的回答如下

不知道有哪些朋友是在TF/keras/chainer/mxnet等框架下用python擼的….…

這可是python啊……上itchat，弄個微信號加自己為好友（或者自己發自己），訓練進展跟著一路發消息給自己就好了，做了可視化的話順便把圖也一并發過來。

然后就能安心睡覺/逛街/泡妞/寫答案了。

講道理，甚至簡單的參數調整都可以照著用手機來……

大體效果如下

當然可以做得更全面一些。最可靠的辦法自然是干脆地做一個http服務或者一個rpc，然而這樣往往太麻煩。本著簡單高效的原則，幾行代碼能起到效果方便自己當然是最好的，接入微信或者web真就是不錯的選擇了。只是查看的話，TensorBoard就很好，但是如果想加入一些自定義操作，還是自行定制的。echat.js做成web，或者itchat做個微信服務，都是挺不賴的選擇。         

正文如下

這里折騰一個例子。以TensorFlow的example中，利用CNN處理MNIST的程序為例，我們做一點點小小的修改。

首先這里放上寫完的代碼：

#!/usr/bin/env python

# coding: utf-8

'''

A Convolutional Network implementation example using TensorFlow library.

This example is using the MNIST database of handwritten digits

(http://yann.lecun.com/exdb/mnist/)

Author: Aymeric Damien

Project: https://github.com/aymericdamien/TensorFlow-Examples/

Add a itchat controller with multi thread

'''

from __future__ import print_function

import tensorflow as tf

# Import MNIST data

from tensorflow.examples.tutorials.mnist import input_data

# Import itchat & threading

import itchat

import threading

# Create a running status flag

lock = threading.Lock()

running = False

# Parameters

learning_rate = 0.001

training_iters = 200000

batch_size = 128

display_step = 10

def nn_train(wechat_name, param):

   global lock, running
   # Lock
   with lock:
       running = True

   # mnist data reading
   mnist = input_data.read_data_sets("data/", one_hot=True)

   # Parameters
   # learning_rate = 0.001
   # training_iters = 200000
   # batch_size = 128
   # display_step = 10
   learning_rate, training_iters, batch_size, display_step = param

   # Network Parameters
   n_input = 784 # MNIST data input (img shape: 28*28)
   n_classes = 10 # MNIST total classes (0-9 digits)
   dropout = 0.75 # Dropout, probability to keep units

   # tf Graph input
   x = tf.placeholder(tf.float32, [None, n_input])
   y = tf.placeholder(tf.float32, [None, n_classes])
   keep_prob = tf.placeholder(tf.float32) #dropout (keep probability)


   # Create some wrappers for simplicity
   def conv2d(x, W, b, strides=1):
       # Conv2D wrapper, with bias and relu activation
       x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')
       x = tf.nn.bias_add(x, b)
       return tf.nn.relu(x)


   def maxpool2d(x, k=2):
       # MaxPool2D wrapper
       return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, k, k, 1],
                           padding='SAME')


   # Create model
   def conv_net(x, weights, biases, dropout):
       # Reshape input picture
       x = tf.reshape(x, shape=[-1, 28, 28, 1])

       # Convolution Layer
       conv1 = conv2d(x, weights['wc1'], biases['bc1'])
       # Max Pooling (down-sampling)
       conv1 = maxpool2d(conv1, k=2)

       # Convolution Layer
       conv2 = conv2d(conv1, weights['wc2'], biases['bc2'])
       # Max Pooling (down-sampling)
       conv2 = maxpool2d(conv2, k=2)

       # Fully connected layer
       # Reshape conv2 output to fit fully connected layer input
       fc1 = tf.reshape(conv2, [-1, weights['wd1'].get_shape().as_list()[0]])
       fc1 = tf.add(tf.matmul(fc1, weights['wd1']), biases['bd1'])
       fc1 = tf.nn.relu(fc1)
       # Apply Dropout
       fc1 = tf.nn.dropout(fc1, dropout)

       # Output, class prediction
       out = tf.add(tf.matmul(fc1, weights['out']), biases['out'])
       return out

   # Store layers weight & bias
   weights = {
       # 5x5 conv, 1 input, 32 outputs
       'wc1': tf.Variable(tf.random_normal([5, 5, 1, 32])),
       # 5x5 conv, 32 inputs, 64 outputs
       'wc2': tf.Variable(tf.random_normal([5, 5, 32, 64])),
       # fully connected, 7*7*64 inputs, 1024 outputs
       'wd1': tf.Variable(tf.random_normal([7*7*64, 1024])),
       # 1024 inputs, 10 outputs (class prediction)
       'out': tf.Variable(tf.random_normal([1024, n_classes]))
   }

   biases = {
       'bc1': tf.Variable(tf.random_normal([32])),
       'bc2': tf.Variable(tf.random_normal([64])),
       'bd1': tf.Variable(tf.random_normal([1024])),
       'out': tf.Variable(tf.random_normal([n_classes]))
   }

   # Construct model
   pred = conv_net(x, weights, biases, keep_prob)

   # Define loss and optimizer
   cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=pred, labels=y))
   optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)

   # Evaluate model
   correct_pred = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1))
   accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))


   # Initializing the variables
   init = tf.global_variables_initializer()

   # Launch the graph
   with tf.Session() as sess:
       sess.run(init)
       step = 1
       # Keep training until reach max iterations
       print('Wait for lock')
       with lock:
           run_state = running
       print('Start')
       while step * batch_size < training_iters and run_state:
           batch_x, batch_y = mnist.train.next_batch(batch_size)
           # Run optimization op (backprop)
           sess.run(optimizer, feed_dict={x: batch_x, y: batch_y,
                                       keep_prob: dropout})
           if step % display_step == 0:
               # Calculate batch loss and accuracy
               loss, acc = sess.run([cost, accuracy], feed_dict={x: batch_x,
                                                               y: batch_y,
                                                               keep_prob: 1.})
               print("Iter " + str(step*batch_size) + ", Minibatch Loss= " + \ 

                    "{:.6f}".format(loss) + ", Training Accuracy= " + \

                    "{:.5f}".format(acc))
               itchat.send("Iter " + str(step*batch_size) + ", Minibatch Loss= " + \

                    "{:.6f}".format(loss) + ", Training Accuracy= " + \

                            "{:.5f}".format(acc), wechat_name)
           step += 1
           with lock:
               run_state = running
       print("Optimization Finished!")
       itchat.send("Optimization Finished!", wechat_name)

       # Calculate accuracy for 256 mnist test images
       print("Testing Accuracy:", \

            sess.run(accuracy, feed_dict={x: mnist.test.images[:256],
                                       y: mnist.test.labels[:256],
                                       keep_prob: 1.}))
       itchat.send("Testing Accuracy: %s" %
           sess.run(accuracy, feed_dict={x: mnist.test.images[:256],
                                       y: mnist.test.labels[:256],
                                         keep_prob: 1.}), wechat_name)

   with lock:
       running = False

@itchat.msg_register([itchat.content.TEXT])

def chat_trigger(msg):
   global lock, running, learning_rate, training_iters, batch_size, display_step
   if msg['Text'] == u'開始':
       print('Starting')
       with lock:
           run_state = running
       if not run_state:
           try:
               threading.Thread(target=nn_train, args=(msg['FromUserName'], (learning_rate, training_iters, batch_size, display_step))).start()
           except:
               msg.reply('Running')
   elif msg['Text'] == u'停止':
       print('Stopping')
       with lock:
           running = False
   elif msg['Text'] == u'參數':
       itchat.send('lr=%f, ti=%d, bs=%d, ds=%d'%(learning_rate, training_iters, batch_size, display_step),msg['FromUserName'])
   else:
       try:
           param = msg['Text'].split()
           key, value = param
           print(key, value)
           if key == 'lr':
               learning_rate = float(value)
           elif key == 'ti':
               training_iters = int(value)
           elif key == 'bs':
               batch_size = int(value)
           elif key == 'ds':
               display_step = int(value)
       except:
           pass

if __name__ == '__main__':
   itchat.auto_login(hotReload=True)
   itchat.run()

這段代碼里面，我所做的修改主要是：

0.導入了itchat和threading

1. 把原本的腳本里網絡構成和訓練的部分甩到了一個函數nn_train里

def nn_train(wechat_name, param):
   global lock, running
   # Lock
   with lock:
       running = True

   # mnist data reading
   mnist = input_data.read_data_sets("data/", one_hot=True)

   # Parameters
   # learning_rate = 0.001
   # training_iters = 200000
   # batch_size = 128
   # display_step = 10
   learning_rate, training_iters, batch_size, display_step = param

   # Network Parameters
   n_input = 784 # MNIST data input (img shape: 28*28)
   n_classes = 10 # MNIST total classes (0-9 digits)
   dropout = 0.75 # Dropout, probability to keep units

   # tf Graph input
   x = tf.placeholder(tf.float32, [None, n_input])
   y = tf.placeholder(tf.float32, [None, n_classes])
   keep_prob = tf.placeholder(tf.float32) #dropout (keep probability)


   # Create some wrappers for simplicity
   def conv2d(x, W, b, strides=1):
       # Conv2D wrapper, with bias and relu activation
       x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')
       x = tf.nn.bias_add(x, b)
       return tf.nn.relu(x)


   def maxpool2d(x, k=2):
       # MaxPool2D wrapper
       return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, k, k, 1],
                           padding='SAME')


   # Create model
   def conv_net(x, weights, biases, dropout):
       # Reshape input picture
       x = tf.reshape(x, shape=[-1, 28, 28, 1])

       # Convolution Layer
       conv1 = conv2d(x, weights['wc1'], biases['bc1'])
       # Max Pooling (down-sampling)
       conv1 = maxpool2d(conv1, k=2)

       # Convolution Layer
       conv2 = conv2d(conv1, weights['wc2'], biases['bc2'])
       # Max Pooling (down-sampling)
       conv2 = maxpool2d(conv2, k=2)

       # Fully connected layer
       # Reshape conv2 output to fit fully connected layer input
       fc1 = tf.reshape(conv2, [-1, weights['wd1'].get_shape().as_list()[0]])
       fc1 = tf.add(tf.matmul(fc1, weights['wd1']), biases['bd1'])
       fc1 = tf.nn.relu(fc1)
       # Apply Dropout
       fc1 = tf.nn.dropout(fc1, dropout)

       # Output, class prediction
       out = tf.add(tf.matmul(fc1, weights['out']), biases['out'])
       return out

   # Store layers weight & bias
   weights = {
       # 5x5 conv, 1 input, 32 outputs
       'wc1': tf.Variable(tf.random_normal([5, 5, 1, 32])),
       # 5x5 conv, 32 inputs, 64 outputs
       'wc2': tf.Variable(tf.random_normal([5, 5, 32, 64])),
       # fully connected, 7*7*64 inputs, 1024 outputs
       'wd1': tf.Variable(tf.random_normal([7*7*64, 1024])),
       # 1024 inputs, 10 outputs (class prediction)
       'out': tf.Variable(tf.random_normal([1024, n_classes]))
   }

   biases = {
       'bc1': tf.Variable(tf.random_normal([32])),
       'bc2': tf.Variable(tf.random_normal([64])),
       'bd1': tf.Variable(tf.random_normal([1024])),
       'out': tf.Variable(tf.random_normal([n_classes]))
   }

   # Construct model
   pred = conv_net(x, weights, biases, keep_prob)

   # Define loss and optimizer
   cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=pred, labels=y))
   optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)

   # Evaluate model
   correct_pred = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1))
   accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))


   # Initializing the variables
   init = tf.global_variables_initializer()

   # Launch the graph
   with tf.Session() as sess:
       sess.run(init)
       step = 1
       # Keep training until reach max iterations
       print('Wait for lock')
       with lock:
           run_state = running
       print('Start')
       while step * batch_size < training_iters and run_state:
           batch_x, batch_y = mnist.train.next_batch(batch_size)
           # Run optimization op (backprop)
           sess.run(optimizer, feed_dict={x: batch_x, y: batch_y,
                                       keep_prob: dropout})
           if step % display_step == 0:
               # Calculate batch loss and accuracy
               loss, acc = sess.run([cost, accuracy], feed_dict={x: batch_x,
                                                               y: batch_y,
                                                               keep_prob: 1.})
               print("Iter " + str(step*batch_size) + ", Minibatch Loss= " + \

                    "{:.6f}".format(loss) + ", Training Accuracy= " + \

                    "{:.5f}".format(acc))
               itchat.send("Iter " + str(step*batch_size) + ", Minibatch Loss= " + \

                    "{:.6f}".format(loss) + ", Training Accuracy= " + \

                            "{:.5f}".format(acc), wechat_name)
           step += 1
           with lock:
               run_state = running
       print("Optimization Finished!")
       itchat.send("Optimization Finished!", wechat_name)

       # Calculate accuracy for 256 mnist test images
       print("Testing Accuracy:", \

            sess.run(accuracy, feed_dict={x: mnist.test.images[:256],
                                       y: mnist.test.labels[:256],
                                       keep_prob: 1.}))
       itchat.send("Testing Accuracy: %s" %
           sess.run(accuracy, feed_dict={x: mnist.test.images[:256],
                                       y: mnist.test.labels[:256],
                                         keep_prob: 1.}), wechat_name)

   with lock:
       running = False

這里大部分是跟原本的代碼一樣的，不過首先所有print的地方都加了個itchat.send來輸出日志，此外加了個帶鎖的狀態量running用來做運行開關。此外，部分參數是通過函數參數傳入的。

然后呢，寫了個itchat的handler

@itchat.msg_register([itchat.content.TEXT])

def chat_trigger(msg):
   global lock, running, learning_rate, training_iters, batch_size, display_step
   if msg['Text'] == u'開始':
       print('Starting')
       with lock:
           run_state = running
       if not run_state:
           try:
               threading.Thread(target=nn_train, args=(msg['FromUserName'], (learning_rate, training_iters, batch_size, display_step))).start()
           except:
               msg.reply('Running')

作用是，如果收到微信消息，內容為『開始』，那就跑訓練的函數（當然，為了防止阻塞，放在了另一個線程里）

最后再在腳本主流程里使用itchat登錄微信并且啟動itchat的服務，這樣就實現了基本的控制。

if __name__ == '__main__':
   itchat.auto_login(hotReload=True)
   itchat.run()

但是我們不滿足于此，我還希望可以對流程進行一些控制，對參數進行一些修改，于是乎：

@itchat.msg_register([itchat.content.TEXT])

def chat_trigger(msg):
   global lock, running, learning_rate, training_iters, batch_size, display_step
   if msg['Text'] == u'開始':
       print('Starting')
       with lock:
           run_state = running
       if not run_state:
           try:
               threading.Thread(target=nn_train, args=(msg['FromUserName'], (learning_rate, training_iters, batch_size, display_step))).start()
           except:
               msg.reply('Running')
   elif msg['Text'] == u'停止':
       print('Stopping')
       with lock:
           running = False
   elif msg['Text'] == u'參數':
       itchat.send('lr=%f, ti=%d, bs=%d, ds=%d'%(learning_rate, training_iters, batch_size, display_step),msg['FromUserName'])
   else:
       try:
           param = msg['Text'].split()
           key, value = param
           print(key, value)
           if key == 'lr':
               learning_rate = float(value)
           elif key == 'ti':
               training_iters = int(value)
           elif key == 'bs':
               batch_size = int(value)
           elif key == 'ds':
               display_step = int(value)
       except:
           pass

通過這個，我們可以在epoch中途停止（因為nn_train里通過檢查running標志來確定是否需要停下來），也可以在訓練開始前調整learning_rate等幾個參數。

實在是很簡單……

雷峰網版權文章，未經授權禁止轉載。詳情見轉載須知。