Pre-training model

[jimliu741523]

adding kernel to support pre-training model.

[CorcovadoMing]

• [x] Load the pre-trained weights for inference
• [ ] Resume the training state from the checkpoints
• [ ] Load the partial pre-trained weights for fine-tuning or being stacked with new layers on top

Resume from the training state needs to store not only the network weights, also the state of optimizer (i.e, current gradient).

Loading the partial pre-trained weights for fine-tuning needs to store the network weights by naming

which is not implemented so for

[CorcovadoMing]

The issue is closed. Welcome to re-open it when the last two features needed to be implemented