Loss and MAE relation and possible optimization

Course Discussions AI Deep Learning by IIT Roorkee
#1

Hello Team,

I am experimenting with a deep model using Keras where my loss over training and validation set looks like below picture. Does it mean that this is a perfectly fit model ?

image
image1047×494 28.1 KB

MAE on training and validation looks as below:

image
image1096×555 41.2 KB

Why there is a bigger noise in MAE validation set ?

Thanks,
Rakesh

#2

Hello,
can someone please suggest me about above learning curve ?
Does you think this loss curve looks good ? if yes then why MAE is noisy on validation dataset ? How can i reduce the noise ?
I have around 60K samples for training, 5K for validation and 5K for testing.

My sample code looks as below.

def createModel():
model = models.Sequential()
model.add(layers.Dense(11, activation=‘relu’, input_shape=(X_train.shape[1],)))
model.add(layers.Dense(11, activation=‘relu’))
model.add(layers.Dense(1))
return model

model = createModel()
model.compile(optimizer=‘rmsprop’, loss=‘mse’, metrics=[‘mae’])
history = model.fit(X_train,
y_train,
epochs=25,
batch_size=25,
validation_data=(X_val, y_val))

#3

Your model looks like an non overfitted model! as the validations loss is always minimum and converging very fast! as the learning rate may be very high!

Suggestions :–>

  1. As compared to the 60K testing you must provide at least 18K of training data!.
  2. Decrease your learning rate.
  3. are you sure this code is correct?
    The layer name should be different right ? but id always 11 only?
    model.add(layers.Dense(11, activation=‘relu’))
    model.add(layers.Dense(1))

For reference kindly refer the below ?