mindformers.models.vit.ViTModel

class mindformers.models.vit.ViTModel(config=None)

Vision Transformer with support for patch or hybrid CNN input stage. The supported model name could be selected from ViTConfig.show_support_list().

Args:

config (ViTConfig): the config of Vit model.

Examples:

>>> # input model name, load model and weights
>>> model_a = ViTModel.from_pretrained('vit_base_p16')
>>> # input config, load model without weights
>>> from mindformers import AutoConfig
>>> config = AutoConfig.from_pretrained('vit_base_p16')
>>> model_b = ViTModel(config)

construct_without_pool(image, mask=None)

construct of vit without pool

fix_init_weight()

fix init weight

init_weights_vit()

init weights vit ViT weight initialization, original timm impl (for reproducibility)