Handwritten numeral recognition in deep learning environment

In their own windows The environment of deep learning has been configured under the environment , This paper mainly records the training and use of a simple handwritten numeral recognition model in the environment of deep learning .

1、 stay pycharm Middle configuration conda Environmental Science ：

After the environment is configured , You can start the code of handwritten numeral recognition

2、 load tensorflow and keras The library of

import tensorflow as tf
from   tensorflow.keras import datasets, layers, models

3、 If the local environment is GPU Environment , Need configuration GPU.

If you don't know if the local branch doesn't support GPU, You can print out the current... Through the following code GPU list

gpuList = tf.config.list_physical_devices("GPU")
print(gpuList)

The printed information is as follows ：

[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]

To configure GPU For model training

if gpuList:
    usedGPU = gpuList[0]
    tf.config.experimental.set_memory_growth(usedGPU, True)
    tf.config.set_visible_devices([usedGPU], "GPU")

4、 Import from the network MNIST Handwritten digit recognition data set

This dataset contains 70000 A picture with numbers written by hand , The size of each picture is 28281( Grayscale picture ),MNIST Dataset Links , Local choice 60000 Zhang as a training set ,10000 As validation set , The dataset has been deleted Keras Integrate , therefore , You can go directly through Keras Of datasets.mnist.load_data() Function import data , The code is as follows ：

# trainImageSet  Training picture collection 
# trainLabelSet  Training data tag set 
# testImageSet  Test image collection 
# testLabelSet  Test data tag set 
(trainImageSet, trainLabelSet), (testImageSet, testLabelSet) = datasets.mnist.load_data()

5、 Preprocess the picture

trainImageSet = trainImageSet.reshape((60000, 28, 28, 1))  #  Change the data into a tensor 
testImageSet  = testImageSet.reshape( (10000, 28, 28, 1))  #  Change the data into a tensor 
trainImageSet = trainImageSet / 255.0  #  Change the pixel value of the picture from [0,255] Mapping to  [0,1] within 
testImageSet  = testImageSet  / 255.0  #  Change the pixel value of the picture from [0,255] Mapping to  [0,1] within 

6、 structure Keras Model

Next, build the following depth model ：

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              (None, 26, 26, 32)        320       
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 13, 13, 32)        0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 11, 11, 64)        18496     
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 5, 5, 64)          0         
_________________________________________________________________
flatten (Flatten)            (None, 1600)              0         
_________________________________________________________________
dense (Dense)                (None, 64)                102464    
_________________________________________________________________
dense_1 (Dense)              (None, 10)                650       
=================================================================
Total params: 121,930
Trainable params: 121,930
Non-trainable params: 0
_________________________________________________________________

The implementation code is as follows ：

model = models.Sequential([
    layers.Conv2D(32, (3, 3), activation='relu', input_shape=(28, 28, 1)),  #  Convolution layer 1, Convolution kernel 3*3,  common 32 Convolution kernels 
    layers.MaxPooling2D((2, 2)),                                            #  Pooling layer 1,2*2 sampling , Maximum pooling 
    layers.Conv2D(64, (3, 3), activation='relu'),                           #  Convolution layer 2, Convolution kernel 3*3, common 64 Convolution kernels 
    layers.MaxPooling2D((2, 2)),                                            #  Pooling layer 2,2*2 sampling , Maximum pooling 
    layers.Flatten(),                                                       # Flatten layer , Connect the convolution layer to the full connection layer , Looking up the data, it is found that all the data are converted into one-dimensional vector data , Prepare for the full connection later .
    layers.Dense(64, activation='relu'),                                    #  Fully connected layer , The output dimension is 64
    layers.Dense(10)                                                        #  Output layer , The output dimension is 10, representative 0-9 The value of the corresponding tag 
])

You can print the entire network through the following code ：

model.summary()

7、 Start training model

Set up the optimizer 、 Loss function 、 Evaluation function

model.compile(optimizer='adam',     #  Optimizer 
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True), #  Loss function 
              metrics=['accuracy']) #  Evaluation function 

Loss function reference link

Model training , Save the trained model

history = model.fit(trainImageSet, trainLabelSet, epochs=10,    #  iteration 10 Time 
                    validation_data=(testImageSet, testLabelSet))
model.save('recogNumber.h5')  #  Save the model to  recogNumber.h5 In file 

Only this and nothing more , The whole process of model training has been completed
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Training / Identify split lines ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Next, use the trained model

8、 Read local pictures

adopt matplotlib.image Read local pictures , A value selected locally is 3 Test with pictures of .

import tensorflow as tf
from tensorflow.keras.models import load_model
import scipy.misc as misc
from tensorflow.keras.preprocessing import image
import matplotlib.pyplot as plt
import matplotlib.image as mpimg
import numpy as np

pil_im = mpimg.imread('333.JPG')

9、 Image preprocessing

Preprocess the picture , Be careful ： Modify the local picture to 28 * 28 * 1 A picture of Victoria

im_data2 = pil_im / 255.0                                                #  Data normalization 
print(im_data2)
plt.imshow(im_data2)
plt.show()                                                               #  display picture , After closing the picture , The program will continue to move on 
im_data2 = image.img_to_array(im_data2)                                  
im_data2 = np.expand_dims(im_data2, axis=0)                              #  Convert data to np Array 

10、 Load the model and predict

model = load_model('C://Users//Administrator//Desktop//recogNumber.h5')  #  Load model 
pre = model.predict(im_data2)                                            #  The model predicts 
print(pre)                                                               #  Print the confidence of all current labels 
print(np.argmax(pre))                                                    #  Select the subscript corresponding to the maximum confidence , The label 

After closing the picture , The output result vector and prediction result are as follows ：

#  Output the confidence vector of different labels , Subscript corresponds to label  0-9
[[-11.468813   -1.7106687  -6.7820425  15.381593   -2.7402487  0.973804   -8.244136   -2.8609006  -1.9001967  -0.34003535]] 
#  Predicted results ：
3

From the output results , The recognition rate of model classifier is still very good , except 3 Outside this category , The values of other categories are basically less than 0.

11、 summary

The example process of deep learning is summarized as follows ：
1、 Get and load the dataset （ Training data sets and test data sets ）
2、 Set the model manually
3、 The training model and the saved model
4、 Use the trained model to use
I thought it was hard , It's much easier to take the first step to success , If you see here , Explain that your environment has been set up , When you're happy, just like it ^ _ ^, Please leave a message if you have any questions .