Question

作为图像分类项目（猫和狗）的一部分，我尝试删除预训练网络的最后一层，并用xgboost层替换它们。更准确地说，从没有最后层的预训练模型中提取特征，然后将这些特征发送到xgboost层。问题在于，该新模型的准确度不超过50％，而初始预训练模型的准确度为88％。我想知道我在做什么错吗？

def load_pretrained_model(image_size=(IMG_HEIGHT,IMG_WIDTH,3)):
    # generate VGG16  with all layers, without top (no classifier, meaning include_top equals to False), and also set the weights equal to 'imagenet'
    pre_trained_model = VGG16(weights='imagenet', include_top=False, input_shape=(IMG_HEIGHT,IMG_WIDTH,3))
    # freeze the layers except the last 2 layers (re-train the last two), meaning that you want to keep all the layers of the pretrained model except the last two
    for layer in pre_trained_model.layers[:-3]:
        layer.trainable = False
    # check the trainable status of the individual layers
    for layer in pre_trained_model.layers:
        print(layer,layer.trainable)
    # create a sequential model
    pre_trained_model.summary()
    model = Sequential()
    # add the model you just transferred
    model.add(pre_trained_model)
    # add new layers (flatten, dense, dropout and then again, dense)
    model.add(Flatten())
    model.add(Dense(256, activation='relu'))
    model.add(Dropout(0.5))
    model.add(Dense(1, activation='sigmoid'))
    # show a summary of the model. Check the number of trainable parameters
    model.summary()
    # use the params of the former model
    model.compile(# set the loss to binary_crossentropy
    loss='binary_crossentropy',
    # set the optimizer to the one you import
    optimizer=RMSprop(lr=1e-4), 
    # set the metrics to accuracy ('acc')
    metrics=['acc'])
    return model


new_vggmodel = load_pretrained_model(image_size=(IMG_HEIGHT,IMG_WIDTH,3))


history = new_vggmodel.fit_generator(
    # set the generator to the train_data_generator
    train_data_generator,
    # as a rule of thumb, you can use the total number of pictures in the train test and doing a floor division with the batch size
    steps_per_epoch= total_train // batch_size,
    epochs=epochs,
    # set the generator to the test_data_generator
    validation_data=test_data_generator,
    # same as above, but replace the train set by the test set
    validation_steps= total_test // batch_size)


layer_name = 'dense'
FC_layer_model = Model(inputs=new_vggmodel.input,
                                 outputs=new_vggmodel.get_layer(layer_name).output)
FC_layer_model.summary()
FC_layer_model.compile(# set the loss to binary_crossentropy
loss='binary_crossentropy',
# set the optimizer to the one you import
optimizer=RMSprop(lr=1e-4), 
# set the metrics to accuracy ('acc')
metrics=['acc'])


train_features=np.zeros(shape=(total_train,256))
train_data_generator.reset() # reset the generator before predict
FC_output = FC_layer_model.predict_generator(train_data_generator,verbose=1)
train_features=FC_output


test_features=np.zeros(shape=(total_test,256))
test_data_generator.reset() # reset the generator before predict
FC_output = FC_layer_model.predict_generator(test_data_generator,verbose=1)
test_features=FC_output




params = {"objective":"binary:logistic", "max_depth":5,"learning_rate":0.1, "min_child_weight":1,"gamma":0,"subsample":0.8,"colsample_bytree":0.8,
        "nthread":4,"scale_pos_weight":1}
dtrain = xgb.DMatrix(data=train_features, label=train_data_generator.labels)
cv_results = xgb.cv(dtrain=dtrain, params=params, nfold=3, num_boost_round=20, metrics="error", as_pandas=True, seed=123)   #verbose ne marche pas
print(cv_results)


xg_cl=xgb.XGBClassifier(learning_rate =0.1,n_estimators=20,max_depth=5,min_child_weight=1,gamma=0,subsample=0.8,colsample_bytree=0.8,objective= 'binary:logistic',
         nthread=4,scale_pos_weight=1,seed=27)
xg_cl.fit(train_features, train_data_generator.labels)


preds=xg_cl.predict(test_features)
accuracy = float(np.sum(preds==test_data_generator.labels))/test_data_generator.labels.shape[0]
print("accuracy: %f" % (accuracy))

CNN + XGBoost模型的准确性提高不超过50％，准确性不如CNN模型

0 个答案: