TensorFlow/Keras model 在尝试预测时出错，引用形状不匹配，尽管我提供的形状似乎正确

Question

I am not sure if I don't understand the input tensor ideas well or if I trained the model with wrong input shapes, or if I need to specify images in a certain way.

The model is built as follows:

... import all the usual libraries - TF, Keras, Numpy, OpenCV etc. ...
_MODEL_DIMENSION = 128

def create_model_for_dimension(dimension):
    model = Sequential()
    if dimension >= 256:
        model.add(Conv2D(256, kernel_size=(5, 5), input_shape=(dimension, dimension, 3)))
        model.add(Activation('relu'))
        model.add(MaxPooling2D(pool_size=(2, 2)))
    if dimension >= 128:
        model.add(Conv2D(128, kernel_size=(3, 3)))
        model.add(Activation('relu'))
        model.add(MaxPooling2D(pool_size=(2, 2)))
…… more convolution layers ……..
…………………………………………..
……………………………………………..
    model.add(Flatten())
……… More layers …………………
……………………………………………..
……………………………………………..

    model.add(Dense(2))
    model.add(Activation('sigmoid'))

    model.compile(loss='categorical_crossentropy',
                  optimizer=tf.keras.optimizers.Adam(),
                  metrics=['accuracy'])
    return model

train_datagen = ImageDataGenerator(
        rotation_range=10,
        width_shift_range=0.2,
        height_shift_range=0.2,
        rescale=1./255,
        shear_range=0.2,
        zoom_range=1.2,
        horizontal_flip=True,
        vertical_flip=True,
        fill_mode='nearest')

train_generator = train_datagen.flow_from_directory(
        train_data_directory, 
        target_size=(_MODEL_DIMENSION, _MODEL_DIMENSION),
        batch_size=_BATCH_SIZE,
        class_mode='categorical')

……… validate_generator …………………
……………………………………………………
………………………………………………………
……… test_generator ………………………
………………………………………………………
………………………………………………………

……… train and save the model ……

When I trained and used it to predict on images from a directory, everything worked (or seemed to, at least). But I then tried to feed in an image without using the ImageDataGenerator in the following way:

First, I read the image using OpenCV (I expect to do other processing using OpenCV, so am not using PIL)

ml_model = … read the model back from the saved model ……
snapshot = cv.imread(file_name)
snapshot = cv.cvtColor(snapshot, cv.COLOR_BGR2RGB)

Then I crop the image using ranges

cropped_area = snapshot[y[0]:y[1], x[0]:x[1]]

Then I score using the model as follows. I also tried variants other than np.array conversion.

score_val = model.predict(np.array(cropped_area/255.0))

This is when it errors out as follows:

ValueError: Input 0 of layer sequential is incompatible with the layer: expected ndim=4, found ndim=3. Full shape received: [32, 128, 3]

My image's shape is (128, 128, 3), and all the checks I made seem to confirm that that is what I have been feeding it.

Any pointers would be greatly appreciated, since I have spent the better part of the day trying to troubleshoot this! Thanks!

Answer 1

You set _MODEL_DIMENSION = 128

which leads to

if dimension >= 128:
        model.add(Conv2D(128, kernel_size=(3, 3)))

You do not set an input shape, so according to the documentation the expected input will be:

Input shape

4D tensor with shape: (batch, channels, rows, cols) if data_format is "channels_first" or 4D tensor with shape: (batch, rows, cols, channels) if data_format is "channels_last".

Your input doesn't fit, hence the error

expected ndim=4, found ndim=3

The solution is to define an input shape, which already exists in your code:

if dimension >= 256:
        model.add(Conv2D(256, kernel_size=(5, 5), input_shape=(dimension, dimension, 3)))

Answer 2

Your data is missing the batch dimension. Your model needs 4D data, not 3D.

Your image shape should be (1, 128,128, 3)