基础LENET5模型
import tensorflow as tf from keras.preprocessing.image import ImageDataGenerator from tensorflow.keras.optimizers import RMSprop train_dir="E://PycharmProjects//LearingImageData//DataSets//train" validation_dir="E://PycharmProjects//LearingImageData//DataSets//var" train_datagen= ImageDataGenerator( rescale=1./255, rotation_range=40,#随机旋转，-40°到40° width_shift_range=0.2,#平移宽度 height_shift_range=0.2,#平移高度 shear_range=0.2,#错切0-1 zoom_range=0.2,#放大倍数 horizontal_flip=True,#翻转 fill_mode='nearest') #统一化处理 train_generator=train_datagen.flow_from_directory( train_dir, target_size=(300,300), batch_size=128,#每次抽出128张图片进行训练 class_mode='categorical'#如果只有两类，用binary，多个类别用categorical ) test_datagen= ImageDataGenerator(rescale=1./255) #统一化处理 validation_generator=test_datagen.flow_from_directory( validation_dir, target_size=(300,300), batch_size=128,#每次抽出128张图片进行训练 class_mode='binary'#如果只有两类，用binary，多个类别用categorical ) #搭建一个简单的CNN模型 model= tf.keras.models.Sequential([ tf.keras.layers.Conv2D(16,(3,3),activation='relu', input_shape=(300,300,3)), tf.keras.layers.MaxPool2D(2,2), tf.keras.layers.Conv2D(32,(3,3),activation='relu'),#每个卷积层后面都有一个最大池化层 tf.keras.layers.MaxPool2D(2,2), tf.keras.layers.Conv2D(64,(3,3),activation='relu'), tf.keras.layers.MaxPool2D(2,2), tf.keras.layers.Flatten(), tf.keras.layers.Dense(512,activation='relu'), tf.keras.layers.Dense(1,activation='sigmoid')#3类，0,1，2 ]) #model.compile(loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True), # optimizer=RMSprop(lr=0.001), # metrics=['accuracy']) model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy']) model.summary() history=model.fit( train_generator, epochs=20, validation_data=validation_generator, steps_per_epoch=2, validation_steps=1, verbose=2) model.save("E://PycharmProjects//LearingImageData//DataSets//finalModel")
基于MobileNet的改进模型（自动保存准确度大于0.8的模型并转化成tflite）
import tensorflow as tf import keras import keras_applications.mobilenet_v2 as mobilenet import numpy as np def get_model(): # 获取base-model base_model = mobilenet.MobileNetV2(include_top=False, weights='imagenet', input_shape=(224, 224, 3), backend=keras.backend, layers=keras.layers, models=keras.models, utils=keras.utils) # 冻结base-model所有参数 for layer in base_model.layers: layer.trainable = False # global average pooling & classification layer gm = keras.layers.GlobalAveragePooling2D()(base_model.output) fc = keras.layers.Dense(units=3, activation='softmax', kernel_regularizer=tf.keras.regularizers.l1(0.005))(gm) # define model model = keras.models.Model(base_model.input, fc, name='zc_classifier') return model def get_train_input(): train_image = keras.preprocessing.image.ImageDataGenerator(featurewise_center=False, samplewise_center=False, featurewise_std_normalization=False, samplewise_std_normalization=False, zca_whitening=False, zca_epsilon=1e-06, rotation_range=0.1, width_shift_range=0.1, height_shift_range=0.1, shear_range=0.0, zoom_range=0.1, channel_shift_range=0.0, fill_mode='nearest', cval=0.0, horizontal_flip=True, vertical_flip=False, rescale=None, preprocessing_function=lambda x: mobilenet.preprocess_input(x, data_format="channels_last"), data_format="channels_last") train_gen = train_image.flow_from_directory(r'E:PycharmProjectsLearingImageDataDataSets rain', target_size=(224, 224), batch_size=32, interpolation="bilinear") for img, lab in train_gen: classes = np.argmax(lab, axis=-1) lab[classes == 0] = [0.85, 0.075, 0.075] lab[classes == 1] = [0.05, 0.8, 0.15] lab[classes == 2] = [0.05, 0.15, 0.8] yield img, lab def get_test_input(): test_image = keras.preprocessing.image.ImageDataGenerator(preprocessing_function= lambda x: mobilenet.preprocess_input(x, data_format="channels_last"), data_format="channels_last", ) test_gen = test_image.flow_from_directory(r'E:PycharmProjectsLearingImageDataDataSetsvar', target_size=(224, 224), batch_size=1, shuffle=False, interpolation="bilinear", ) return test_gen def custom_loss(y_true, y_pred): class_weight = [0.33, 0.33, 0.34] #配权重 数量越多，权重越低 # classes = tf.argmax(y_true, axis=1) loss = 0.0 for i in range(3): loss += (y_true[:, i] * tf.math.log(y_pred[:, i]) + (1 - y_true[:, i]) * tf.math.log(1 - y_pred[:, i])) * class_weight[i] return -tf.reduce_mean(loss) def train(model: tf.keras.Model, train_gen, test_gen, save_function): adam = keras.optimizers.Adam(0.01)# model.compile(loss=custom_loss, optimizer=adam, metrics=['accuracy']) model.fit_generator(train_gen, epochs=8, steps_per_epoch=6, validation_data=test_gen, callbacks=[save_function]) model.save('./zc.h5') def test(model: tf.keras.Model, test_gen): # import numpy as np # adam = keras.optimizers.Adam(0.001) # model.compile(loss="categorical_crossentropy", optimizer=adam, metrics=['accuracy']) # model.load_weights('./zc.h5') print(model.evaluate_generator(test_gen, verbose=1)) # for img, lab in test_gen: # print(np.argmax(lab[0]), np.argmax(model.predict(img))) class Save(keras.callbacks.Callback): def __init__(self): super(Save, self).__init__() self.max_acc = 0.0 def on_epoch_begin(self, epoch, logs=None): pass def on_epoch_end(self, epoch, logs=None): self.val_acc = logs["accuracy"] if epoch != 0: if self.val_acc > self.max_acc and self.val_acc > 0.8: model.save("kears_model_" + str(epoch) + "_acc=" + str(self.val_acc) + ".h5") self.max_acc = self.val_acc converter = tf.lite.TFLiteConverter.from_keras_model(model) converter.experimental_new_converter = True tflite_model = converter.convert() open("converted_model_"+str(self.val_acc)+".tflite", "wb").write(tflite_model) if __name__ == "__main__": save_function = Save() model = get_model() train_gen = get_train_input() test_gen = get_test_input() train(model, train_gen, test_gen, save_function) # test(model, test_gen)
使用TensorFlow Lite Model Maker训练模型
pip install -q tflite-model-maker
import os import numpy as np import tensorflow as tf assert tf.__version__.startswith('2') from tflite_model_maker import configs from tflite_model_maker import ExportFormat from tflite_model_maker import image_classifier from tflite_model_maker import ImageClassifierDataLoader from tflite_model_maker import model_spec import matplotlib.pyplot as plt image_path='E://PycharmProjects//LearingImageData//DataSets//train' model_dir='E://PycharmProjects//MLTest//tflite_model' data = ImageClassifierDataLoader.from_folder(image_path) train_data, rest_data = data.split(0.8) validation_data, test_data = rest_data.split(0.5) model = image_classifier.create(train_data, validation_data=validation_data) model.summary() loss,accuracy = model.evaluate(test_data) model.export(export_dir=model_dir)
该卷积神经网络基于EfficientNet-Lite0，效果比前面两个好得多