tx = optax.adamw(learning_rate=2e-5, weight_decay=0.01)

        __
        __