CODEWITHSUNDEEP
pythontensorflowmachine-learningdeep-learningkeras
franchzilla
franchzilla

Reputation: 55

Tensorflow: Loss and Accuracy curves showing similar behavior

I am using Keras with the Tensorflow backend to train a modified Resnet-50 that classifies objects into 15 categories. I am using the Adam optimizer and I tried learning rates of 0.001 and 0.01 but got similar results.

The problem I am facing is that the loss and accuracy both show similar behavior (in the training and validation datasets). They both go up or down at similar times and I expected to get higher accuracies as loss went down. What can be causing this behavior?

Here are some Tensorboard curves from my last run: Training Set Accuracy and Loss

Validation Set Accuracy and Loss

Edit: The code for the model is the following:

#Model creation:
def create_model(possible_labels):
    rn50 = ResNet50(include_top=True, weights=None)
    layer_name = rn50.layers[-2].name
    model = Model(rn50.input, 
            Dense(len(possible_labels))(rn50.get_layer(layer_name).output))
    adam = Adam(lr=0.0001)
    model.compile(loss='categorical_crossentropy', 
            optimizer=adam, metrics=['accuracy'])
    checkpointer = ModelCheckpoint(filepath='the_best_you_ever_had', 
                             verbose=1, save_best_only=True)
    tensorboard = TensorBoard()
    return model, [checkpointer, tensorboard]

model, checkpointers = create_model(labels)

#Dataset generation:
train_datagen = ImageDataGenerator(
featurewise_center=True,
featurewise_std_normalization=True,
rotation_range=20,
width_shift_range=0.2,
height_shift_range=0.2,
horizontal_flip=True,
vertical_flip=True,
channel_shift_range=0.2,
shear_range=0.2,
zoom_range=0.2
)

val_datagen = ImageDataGenerator()

train_generator = train_datagen.flow_from_directory(
'data\\train',
target_size=(224, 224),
batch_size=32,
class_mode='categorical'
)

val_generator = val_datagen.flow_from_directory(
'data\\validation',
target_size=(224, 224),
batch_size=32,
class_mode='categorical'
)

#Model training:
model.fit_generator(
train_generator,
steps_per_epoch=5000,
epochs=50,
validation_data=val_generator,
callbacks=checkpointers
)

Upvotes: 2

Views: 426

Answers (1)

franchzilla
franchzilla

Reputation: 55

I found the error in my code and it was the fact that I was just using the default (linear) activation in my added last layer. I switched it to a softmax activation (as it is a classification and not a regression problem) by doing this in the code:

model = Model(rn50.input, 
        Dense(len(possible_labels), activation='softmax')
        (rn50.get_layer(layer_name).output))

and then the curves started behaving as expected and I achieved 96% accuracy.

Upvotes: 1

Related Questions