type

 type : enum, defines how the initializer behaves when the graph is used in training mode. It determines whether the tensor is treated as a constant, as trainable weights, or as frozen weights. During inference, it makes no difference whether the type is Constant, Train Weights, or Frozen Weights.

Constant

The initializer is treated as a fixed, non-trainable variable (e.g., a constant coefficient in a multiplication).

Train Weight

The initializer is treated as trainable weights during training (it will be updated by the optimizer).

Frozen Weight

The initializer is treated as weights, but they remain frozen (not updated) during training.