Hello, I want to train a model from scratch, but the training according to the configuration file in the code does not reach the accuracy of the official model. May I ask if the encoder in the model, i.e. hybrid vit, uses a pre-trained model on ImageNet? If so, how can I download it?