构建一个字母ABC的手写识别网络,
要求给出算法误差收敛曲线，所给程序要有图片导入接口。
import osimport cv2import numpy as npimport pandas as pdimport matplotlib.pyplot as pltimport tensorflow as tfimport tqdmfrom tensorflow import kerasfrom keras import Input, Model, Sequentialfrom tensorflow.keras.regularizers import l2from keras.layers import Dense, Flatten, InputLayer, Reshape, BatchNormalization, Dropout, Conv2D, MaxPooling2Dfrom tensorflow.keras.utils import plot_modeldata_dir = './data' # Dict of labels 'A' : 0 ,'B' : 1 ,'C' : 2 def load_images (images_folder, img_size = (128 ,128 ) , scale=False) :# Store paths to images for dirname, _, filenames in os.walk(images_folder):for filename in filenames:"There are {} images in {}" .format(len(image_path), images_folder))# Load images and associated labels for path in tqdm.tqdm(image_path):# Resize the images '/' )[-2 ]]) # last folder before the image name is the category if scale:255 # scale return image_path, images, np.asarray(labels)img_size = (128 ,128 )# Resize # images = np.array(images).reshape(-1,128,128,1) There are 600 images in ./data# 查看图片 10 ,10 ))36 )for i in range(36 ):6 ,6 ,i+1 )
)labels_df = pd.DataFrame(labels) 2 201dataset=[]0 for name in tqdm(os.listdir(data_dir)):for im in os.listdir(path):50 ,50 ,3 ))1 100%|██████████| 3/3 [00:03<00:00, 1.06s/it]data=np.array(dataset)data[0 ].shape(50, 50, 3)print(pd.Series(dataname).value_counts())1 202len(categories)3from tensorflow.keras.layers import Dense, Conv2D, Flatten, Dropout, MaxPooling2D, BatchNormalizationdef build_cnn_model () :32 ,kernel_size=(3 ,3 ),activation='relu' ,input_shape=images.shape[1 :]),2 ,2 ),0.4 ),64 ,kernel_size=(3 ,3 ),activation='relu' , padding='same' ),64 ,kernel_size=(3 ,3 ),activation='relu' , padding='same' ),2 ,2 )),0.4 ),128 ,kernel_size=(3 ,3 ),activation='relu' , padding='same' ),128 ,kernel_size=(3 ,3 ),activation='relu' , padding='same' ),2 ,2 ),0.4 ),256 ,kernel_size=(3 ,3 ),activation='relu' , padding='same' ),256 ,kernel_size=(3 ,3 ),activation='relu' , padding='same' ),2 ,2 ),0.4 ),128
3 ,3 ),activation='relu' , padding='same' ),128 ,kernel_size=(3 ,3 ),activation='relu' , padding='same' ),2 ,2 ),0.4 ),64 ,kernel_size=(3 ,3 ),activation='relu' , padding='same' ),64 ,kernel_size=(3 ,3 ),activation='relu' , padding='same' ),2 ,2 )),0.4 ),'softmax' )return cnn_model# Initialize the model by passing some data through 0 ]])# Print the summary of the layers in the model. Model: "sequential_3"tf.keras.utils.plot_model(model, show_shapes=True )from tensorflow.keras.utils import to_categoricalfrom sklearn.preprocessing import LabelEncoderfrom
import shuffle10 ]array([[0., 1., 0.],model.compile(optimizer = "adam" , loss = "binary_crossentropy" , metrics = ["accuracy" ])checkpoint_filepath = '/checkpoint.hdf5' True ,'val_accuracy' ,'max' ,True ,500 )from tensorflow.keras.callbacks import EarlyStopping'val_accuracy' , 10 , 0.001 , 'max' ,True datagen = ImageDataGenerator(horizontal_flip=True ,vertical_flip=True ,rotation_range=20 ,zoom_range=0.2 ,0.2 ,height_shift_range=0.2 ,shear_range=0.1 ,fill_mode="nearest" )from tensorflow.keras.callbacks import ReduceLROnPlateau"val_accuracy" ,factor = 0.3 , patience = 3 ,0.001 ,mode = 'auto' ,verbose=1 )from sklearn.model_selection import train_test_split# Train, validation and test split 0.10 , random_state=7 )0.20 , random_state=1 ) print("*-*-*-*-*-*" )"Train" )"*-*-*-*-*-*" )"Validation" )"*-*-*-*-*-*" )"Test" )*-*-*-*-*-*history = model.fit(X_train, y_train, 32 , 100 , 1 , Epoch 1/1000plt.plot(history.history["accuracy" ])"val_accuracy" ])"Model accuracy" )"Accuracy" )"Epoch" )"Train" , "Test" ], loc = "upper left" )
plt.plot(history.history["loss" ])"val_loss" ])"Model loss" )"Loss" )"Epoch" )"Train" , "Test" ], loc = "upper left" )def predict_class (img) :# Resize 1 ,128 ,128 ,3 )# Predict for pred in predictions]# Return label corresponding to predicted index return list(categories.keys())[list(categories.values()).index(true_prediction)]# Predict on test set # From categorical outputs to discrete values for y in y_pred]for y in y_test]from sklearn.metrics import classification_report precision recall f1-score supportplt.figure(figsize=(10 ,10 ))0 ],36 )for i in range(36 ):6 ,6 ,i+1 )False )# Predict and get label model.save("model.h5" )# 识别接口 def predict (path,model_str,img_size = (128 ,128 ) ) :# Resize the images # Resize 1 ,128 ,128 ,3 )# Predict for pred in predictions]# Return label corresponding to predicted index return list(categories.keys())[list(categories.values()).index(true_prediction)]"./data/A/051.jpg" ,"model.h5" )'A'predict("./data/B/048.jpg" ,"model.h5" )'B'predict("./data/C/050.jpg" ,"model.h5" )'C'
