`fit()` missing in your VGG-like model in Facial_Keypoint_Kaggle.ipynb script

[Engeloid]

`fit()` missing in your VGG-like model in Facial_Keypoint_Kaggle.ipynb script

[suvoooo]

Hello, thanks for checking this out.

First, regarding earlystop, I defined it as a function so if you want to include earlystop in callbacks, you need to specify the 'mode' e.g earlystop('acc'). Below I shared a working version of the notebook that I just tried but I would still like to highlight some points regarding training.

1. If we are using only the clean dataset, the sample size is too low for a reasonably deep network to capture features. That's why augmentation is so necessary for this project.
2. You can reduce the number of layers (check the shared notebook) and increase the validation size a bit more so that the network can understand what to penalize. This is because without augmentation the sample size is....
3. Not only mean squared error but you can also try MSE + Log MSE.

      def custom_loss(y_true, y_pred):
         msleObject = tf.keras.losses.MeanSquaredLogarithmicError()
         y_true = tf.cast(y_true, tf.float64)
         y_true = tf.add(y_true, 1e-9)
         y_pred = tf.add(y_pred, 1e-9)
         msleTensor = msleObject(y_true, y_pred)
         mseObject = tf.keras.losses.MeanSquaredError()
         mseTensor = mseObject(y_true, y_pred)
         mseTensor_m = tf.math.scalar_mul(5, mseTensor, name=None) #alpha is hyperparameter
         tot_loss = tf.add(msleTensor, mseTensor_m)
         return tot_loss 
   model.compile(..., loss=custom_loss)

Here, I just went through the VGG16 part to check how bad it gets over epochs.

happy learning and best!