ViTMAETimm.py

from transformers import PreTrainedModel
from timm.models.resnet import BasicBlock, Bottleneck, ResNet

from transformers import PretrainedConfig
from typing import List
import torch
import timm


class ViTMAEConfig(PretrainedConfig):
    model_type = "vit_mae_custom"

    def __init__( self,  model_name='timm/vit_base_patch16_224.mae', num_classes: int = 1000, **kwargs ):
        self.model_name = model_name
        self.num_classes = num_classes
        super().__init__(**kwargs)
        
# 'timm/vit_huge_patch14_224.mae'
# class ViTMAEModel(PreTrainedModel):
#     config_class = ViTMAEConfig

#     def __init__(self, config):
#         super().__init__(config)
#         self.model = timm.create_model(config.model_name, num_classes=config.num_classes, pretrained=True)

#     def forward(self, tensor):
#         return self.model.forward_features(tensor)

class ViTMAEModelForImageClassification(PreTrainedModel):
    config_class = ViTMAEConfig

    def __init__(self, config):
        super().__init__(config)
  
        self.model = timm.create_model(config.model_name, num_classes=config.num_classes, pretrained=True)

    def forward(self, tensor, labels=None):
        logits = self.model(tensor)
        if labels is not None:
            loss = torch.nn.cross_entropy(logits, labels)
            return {"loss": loss, "logits": logits}
        return {"logits": logits}