Keras 模型不是训练层，验证准确率始终为 0.5

Question

My Keras CNN model (based on an implementation of AlexNet ) always has training accuracy close to 0.5 (within +- 0.02) and the validation accuracy is always 0.5 exactly, no matter which epoch. It is a binary classification model where the train/val split is roughly 85/15 and within both those sets the images are split 50/50 for each class.

It doesn't seem to matter which model architecture I use, or whether I initalise with random or ImageNet weights, the validation accuracy is always 0.5. In fact, when I didn't have the images split 50/50 into their binary classifications, the validation accuracy would reflect this (so when I had significantly more images belonging to one class, the validation accuracy would always be 0.85).

Because of this last point, I have a suspicion the problem doesn't lie with the model or weight optimisation, but rather with my instantiation of the ImageDataGenerator class - although this is just an educated hunch at this stage.

I've included my code below, can anyone locate any blindingly obvious errors?

sz=224 # image width = height = 224
batch_size=64
train_data_dir = r"./crack_dataset/train"
validation_data_dir = r"./crack_dataset/validate"
nb_train_samples = 3416
nb_val_samples = 612

train_datagen = ImageDataGenerator(rescale=1./255)

validation_datagen = ImageDataGenerator(rescale=1./255)

train_generator = train_datagen.flow_from_directory(train_data_dir,
                                                    target_size = (sz, sz),
                                                    batch_size=batch_size,
                                                    class_mode='binary')

validation_generator = validation_datagen.flow_from_directory(validation_data_dir,
                                                              target_size = (sz, sz),
                                                              batch_size=batch_size,
                                                              class_mode='binary')

# Create Model 
model = Sequential()

model.add(Conv2D(filters=96, input_shape=input_shape, kernel_size=(11,11), strides=(4,4), padding='valid', activation='relu'))
model.add(MaxPooling2D(pool_size=(2,2), strides=(2,2), padding='valid'))

model.add(Conv2D(filters=256, kernel_size=(11,11), strides=(1,1), padding='valid', activation='relu'))
model.add(MaxPooling2D(pool_size=(2,2), strides=(2,2), padding='valid'))

model.add(Conv2D(filters=384, kernel_size=(3,3), strides=(1,1), padding='valid', activation='relu'))

model.add(Conv2D(filters=384, kernel_size=(3,3), strides=(1,1), padding='valid', activation='relu'))

model.add(Conv2D(filters=256, kernel_size=(3,3), strides=(1,1), padding='valid', activation='relu'))
model.add(MaxPooling2D(pool_size=(2,2), strides=(2,2), padding='valid'))

model.add(Flatten())
model.add(Dense(4096, input_shape=(256,), activation='relu'))
model.add(Dropout(0.4))

model.add(Dense(4096, activation='relu'))
model.add(Dropout(0.4))

model.add(Dense(1, activation='sigmoid'))

model.compile(optimizer=Adam(0.1), loss='binary_crossentropy', metrics=['accuracy'])

model.fit_generator(train_generator,
                    steps_per_epoch = nb_train_samples // batch_size, 
                    epochs=30,
                    validation_data=validation_generator,
                    validation_steps=nb_val_samples // batch_size)

Answer 1

The problem is the learning rate of the optimizer, the value is way too big. As suggested in the comment, it should be set to a low value, close to 0.

You can see how the learning rate might influence the classification accuracy in the image below:

Answer 2

It was a comment, but too long so I switched to a proper answer.

I haven't tried your CNN, but I would revise your layers. I'm not an expert in designing CNN architectures but I don't see a lot of meaning in having 2 layers in a row with the same parameters (5th and 6th layers), it's pretty useless for what I've learned so far. Try with some simple architecture and "scale up" (in terms of complexity of the architecture) from there.

Furthermore, going up and down in terms of "number of units" ( 1st, 3rd, 5th, 6th, 7th layers, Conv2D) might not be the best strategy, everytime you go up you generalize in terms of features to get because you start from less features, you are asking to your network to create features from a smaller set. There are some scenarios where it might be useful, but I don't think this is one of them.

And for the rescaling you don't need to call "ImageDataGenerator", you can reach the same result with:

train_datagen = train_datagen/255.0

Simplify where you can.