Low accuracy using functional API + CNN and CIFAR10; incorrect initialization?

Question

I'm new to using CNNs but I'm trying to make one using the functional API with the CIFAR10 dataset. The only thing is I'm getting very very low accuracy. I've looked over my textbook examples and documentation but can't figure out why it's so low when it should be starting way higher. This is my setup using DenseNet201 and tf version 2.7:

#load in data 
(X_train, y_train), (X_test, y_test) = tf.keras.datasets.cifar10.load_data()

# Normalize pixel values to be between 0 and 1
X_train, X_test = X_train / 255.0, X_test / 255.0

# one hot encode target values/labels
y_train = tf.keras.utils.to_categorical(y_train)
y_test = tf.keras.utils.to_categorical(y_test)

# have to preprocess before using functional API
X_testP = tf.keras.applications.densenet.preprocess_input(X_test)
X_trainP = tf.keras.applications.densenet.preprocess_input(X_train)

# data size we start with
inputs = tf.keras.Input(shape=(32,32,3))
# densenet expects 224x224 so use lambda layer
resized_images = tf.keras.layers.Lambda(lambda image: tf.image.resize(image, (224, 224)))(inputs)

# initialize model
transfer = keras.applications.DenseNet201(include_top=False, weights='imagenet', pooling='max', input_tensor = resized_images,input_shape=(224,224,3), classes=1000)

# add your layers
x = transfer.output
x = tf.keras.layers.Flatten()(x)
x = tf.keras.layers.Dense(256, activation='relu')(x)
x = tf.keras.layers.BatchNormalization() (x)
x = tf.keras.layers.Dense(200, activation='relu')(x)
x = tf.keras.layers.Dense(128, activation='relu')(x)
x = tf.keras.layers.Dense(64, activation='relu')(x)
output = tf.keras.layers.Dense(10, activation='softmax')(x)

transfer_model = keras.Model(inputs=transfer.input, outputs=output)
transfer_model.trainable = False;

# here I try SGD but I also tried Adam to no better results
optimizer = keras.optimizers.SGD(learning_rate=0.2, momentum=0.9, decay=0.01)

transfer_model.compile(optimizer=optimizer,loss='categorical_crossentropy',metrics=['accuracy'])

history_transfer = transfer_model.fit(X_trainP, y_train,epochs=20)

I feel like all the examples I've seen start way higher and that's even without additional layers. Am I misunderstanding something in the initialization?

score 0 · Answer 1 · answered Dec 10 '21 at 05:06

Your problem is that the pixel values are being rescalled twice. The function tf.keras.applications.densenet.preprocess_input rescales the pixel values so you should not rescale them. In your code you have

transfer = keras.applications.DenseNet201(include_top=False, weights='imagenet', pooling='max', input_tensor = resized_images,input_shape=(224,224,3), classes=1000)

first thing is - if you set pooling=max the output is a one dimensional tensor so you do not need a flatten layer in your model. second is dense net only has to have 224 X 224 images if include top=True. So you do not need the to resize the input with the lambda function. Set the input_shape=(32,32,3) set classes=None remove the parameter input_tensor Your model will probably overfit. Add a dropout layer to the model. Use rate=.4 Use the adam optimizer it works well.

Low accuracy using functional API + CNN and CIFAR10; incorrect initialization?

1 Answers1