Error while making CAM model from MobileNetv3 with tranfer learning

TensorFlow
1

I’m quite newbie to tensorflow so pls help me. I build a tranferlearning model with MobileNetV3Large from Tensorflow with the first layer is InputLayer(224,224,3), then MobileNetV3Large and some layers for classification. Here are the structure:

image
image750×311 16.2 KB

Now I want to make the CAM model to see what my model saw and keep getting errors like this:

cam_model  = Model(inputs=epoch70_model.inputs[0],outputs=(epoch70_model.layers[-6].output,epoch70_model.layers[-1].output))

KeyError: 'Exception encountered when calling Functional.call().\n\n\x1b[1m133568923055728\x1b[0m\n\nArguments received by Functional.call():\n  • inputs=tf.Tensor(shape=(32, 224, 224, 3), dtype=float32)\n  • training=False\n  • mask=None'

I believe the problem is on epoch70_model.layers[-6].output as I remove this output the model working fine. But I cannot figure out why there is a input shape (32, 224, 224, 3) and how to solve this. The model training and predicing fine with dataset has batch size (256, 224, 224, 3).

3

Hi @NLT22, Welcome to the forum!

Can you provide some code snippet to reproduce?

1 Like
5

Thank for your mention, I have find out my problem. To summarize it, I build my model this way:

IMG_SHAPE =(224, 224, 3)
base_model = tf.keras.applications.MobileNetV3Large(input_shape=IMG_SHAPE,
                                                    include_top=False,
                                                    weights='imagenet',
                                                    include_preprocessing=False)
base_model.trainable = False

def feature_extractor(inputs):

  feature_extractor = base_model(inputs)
    
  return feature_extractor

def classifier(inputs):
    x = GlobalAveragePooling2D()(inputs)
    x = Dense(128, activation="silu")(x)
    x = Dropout(0.2)(x)
    x = BatchNormalization()(x)
    x = Dense(1, activation="sigmoid", name="classification")(x)
    return x

def final_model():

    inputs = tf.keras.layers.Input(shape=(224,224,3)) 

    mobilev3_feature_extractor = feature_extractor(inputs)
    
    classification_output = classifier(mobilev3_feature_extractor)

    model = tf.keras.Model(inputs=inputs, outputs = classification_output)

    return model

model = final_model()
model.summary()

And it looks like MobileNetV3 is embedded as a layer within a new model make the error. Instead I find the solution to rather extend the MobileNetV3 model and it working:

IMG_SHAPE =(224, 224, 3)
inp = tf.keras.layers.Input(shape=(IMG_SHAPE))
base_model = tf.keras.applications.MobileNetV3Large(input_shape=IMG_SHAPE,
                                                    include_top=False,
                                                    weights='imagenet',
                                                    include_preprocessing=False,
                                                    input_tensor=inp)
base_model.trainable = False

last_layer = base_model.get_layer('activation_19')
x = GlobalAveragePooling2D()(last_layer.output)
x = Dense(128, activation="silu")(x)
x = Dropout(0.2)(x)
x = BatchNormalization()(x)
x = Dense(1, activation="sigmoid", name="classification")(x)

model = tf.keras.models.Model(inputs = inp, outputs = x)
model.summary()