Starting today, I'll be here learning about the record depth study of the footprints, the purpose is very simple, there are nearly at hand has been marked 3w a good worth code correctly, you want to train from scratch a model that can be used,
Although I know the relevant models and demo online a lot, but still very much hope that they can personally engage in a usable out, learning books are: "Getting Tensorflow depth study of the principles and advanced combat" Li Jinhong teacher.
In addition, I will also verify the source code, models, and nearly 3w a fully open source code out of the code recognition model trained. Mutual encouragement.
. 1 # ! / Usr / bin / Python the env 2 # - * - Coding: UTF-. 8 - * - . 3 # @time: 2019/9/23 21:27 . 4 # @author: SongSa . 5 # @Desc: . 6 # @ File: fitting two-dimensional data .py . 7 # @Software: PyCharm . 8 . 9 Import tensorflow TF AS 10 Import numpy AS NP . 11 Import matplotlib.pyplot AS PLT 12 is 13 is "" " 14 depth study is divided into four steps: 15 prepare the data 16 construction model 17 iterative training 18 Use Model 19 "" " 20 is 21 is # ####### preparing data ######## 22 is train_X np.linspace = (-1,. 1, 100 ) 23 is train_Y = 2 * + np.random train_X .randn (100) 0.3 * # Y = 2x but noise added 24 plt.plot (train_X, train_Y, ' RO ' , label = ' Original data ' ) # shows the analog data points 25 plt.legend () # has display Legend label 26 is plt.show () 27 28 29 plotdata = { " BatchSize " : [], " Loss ": []} 30 DEF MOVING_AVERAGE (A, W = 10 ): 31 is IF len (A) < W: 32 return A [:] 33 is return [Val IF IDX <W the else SUM (A [(IDX-W): IDX ]) / W for IDX, Val in the enumerate (A)] 34 is 35 36 # ####### ######## construction model 37 # model is divided into two directions: forward and reverse 38 # Create model 39 X-tf.placeholder = ( " a float " ) # placeholder 40 the Y = tf.placeholder ( "a float " ) 41 is # model parameter 42 is W is = tf.Variable (tf.random_normal ([1]), name = " weight " ) # W is is initialized to [-1, 1] of the random number, the shape of one-dimensional digital 43 is B = tf.Variable (tf.zeros ([. 1]), name = " BIAS " ) # B is initialized to 0. 44 is # forward structure 45 Z = tf.multiply (X-, W is) + B # tf.multiply () function is multiplied * X-W is = + Z B 46 is 47 # reverse construction model 48 # neural network training data flow in two directions, a first value generated by the positive, then the true value observed gap, then the parameters inside adjusted by the reverse process, 49 #Next to generate positive predictive value in the real value comparison, so the cycle, we know the parameters adjusted to get the right value, will introduce back-propagation algorithm to achieve proper adjustment of parameters. 50 cost = tf.reduce_mean (tf.square (the Y - Z)) # cost equal to the square of the difference from the true value generated value 51 is # tf.reduce_mean () for calculating an average value of the tensor along the axis designated 52 is # tf.square ( ) for calculating the square Yz 53 is learning_rate = 0.01 # learning rate (a higher number indicating greater speed of adjustment, accurate but not vice versa) 54 is Optimizer = tf.train.GradientDescentOptimizer (learning_rate) .minimize (cost) # encapsulated gradient descent algorithm 55 56 is 57 is # ####### iterative training ######## 58 # Tensorflow tasks are carried out by the session 59 the init = tf.global_variables_initializer () # initialize all variable 60 # custom parameter 61 is training_epochs = 20 is 62 is display_stop 2 = 63 is # start the session 64 with tf.Session () AS Sess: 65 sess.run (the init) 66 plotdata = { " BatchSize " : [], " Loss " : [] } # store batch value and loss value 67 for Epoch in Range (training_epochs): 68 for (X, Y) in ZIP (train_X, train_Y): 69 sess.run (Optimizer, feed_dict = {X-: X, the Y: Y }) 70 71 is #Training details display 72 IF Epoch% display_stop == : 0 73 is Loss = sess.run (cost, feed_dict = {X-: train_X, the Y: train_Y}) 74 Print ( " Epoch: " , + Epoch. 1, " cost = " , Loss, " W is = " , sess.run (W is), " B = " , sess.run (B)) 75 IF Not (Loss == " NA " ): 76 plotdata [ ' BatchSize ' ]. append(epoch) 77 plotdata["loss"].append(loss) 78 79 print("Finished!") 80 print("cost=", sess.run(cost, feed_dict={X:train_X, Y:train_Y}), "W=", sess.run(W), "b=", sess.run(b)) 81 82 # 训练模型可视化 83 plt.plot(train_X, train_Y, 'ro', label="Original data") 84 plt.plot(train_X, sess.run(W) * train_X + sess.run(b), label="Fittedline") 85 plt.legend() 86 plt.show() 87 88 plotdata['avgloss'] = moving_average(plotdata["loss"]) 89 plt.figure(1) 90 plt.subplot(211) 91 plt.plot(plotdata["batchsize"], plotdata['avgloss'], 'b--') 92 plt.ylabel("Loss") 93 plt.title ( " Minibatch FC vs. RUN Training Loss " ) 94 plt.show () 95 96 97 # ####### usage model ######## 98 Print ( ' using the model: \ n \ n- ' ) 99 Print ( " X = 0.2, Z = " , sess.run (Z, X-feed_dict = {:} 0.2))
Finally, make ad: To learn more about the Python content crawlers, data analysis, data source to obtain a large number of reptiles crawling to welcome everyone concerned about my micro-channel public number: enlightenment Python
