Update Keras, use PyTorch instead of Tensorflow

[carschno]

Using Keras/Tensorflow seems to be outdated because most people use PyTorch instead of Tensorflow.
We can keep using Keras, which can use Tensorflow and Pytorch (and Jax) as backends.
Also, Keras 3 should be used (and will be used by default with pip install keras), but the lesson should be updated accordingly.

To check:

• Update setup instructions to refer to Keras version 3
• Update setup instructions to use PyTorch instead of Tensorflow
• Investigate which parts of the code need to be adapted (Keras 3 is highly backwards compatible with Keras 2)
• Update outputs, and potentially explanations, where necessary

As a starting point, I recommend running a few commands and compare the outputs to estimate which episodes need to be updated and how much effort that will be.
As a result of the investigation, splitting this task up into multiple might be reasonable.

[ashwinvis]

I tried to add feature scaling using a StandardScaler, and that indeed works as @cantordust pointed out and model converges. It could be that Pytorch is more sensitive to this. I could not find any sources why, but I am all for adding feature scaling. Maybe that is a good thing. In fact, I recall one of the learners, picking up on this "why don't we do any feature scaling"?

from sklearn.preprocessing import StandardScaler

# Feature scaling for better convergence
scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train.values)
X_test_scaled = scaler.transform(X_test.values)

# Convert one-hot encoded target dataframes to numpy arrays
y_train_array = y_train.values
y_test_array = y_test.values

...

history = model.fit(X_train_scaled, y_train_array, epochs=500, validation_split=0.2, verbose=1)

...

def plot_history(history, metrics):
    """
    Plot the training history

    Args:
        history (keras History object that is returned by model.fit())
        metrics (str, list): Metric or a list of metrics to plot
    """
    history_df = pd.DataFrame.from_dict(history.history)
    sns.lineplot(data=history_df[metrics])
    plt.xlabel("epochs")
    plt.ylabel("metric")

plot_history(history, 'loss')

A side effect of this is Pytorch converges slowly, but the model is rather good. I assumed it was perhaps because no overfitting happened yet.

I tried to increase training to 500 epochs to force some overfitting and the result remains the same:

Now the task is to do some "bad" scaling. So we can explain the confusion matrix.