ConvNextBase.py

from transformers import PretrainedConfig, PreTrainedModel
from typing import List
from torchvision.models import convnext_base, ConvNeXt_Base_Weights


class ConvNextBaseConfig(PretrainedConfig):
    model_type = "ConvNext"

    def __init__(
        self,
        **kwargs,
    ):
        super().__init__(**kwargs)


class ConvNextBaseModel(PreTrainedModel):
    config_class = ConvNextBaseConfig

    def __init__(self, config):
        super().__init__(config)
        self.model = convnext_base()

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


class ConvNextBaseModelForImageClassification(PreTrainedModel):
    config_class = ConvNextBaseConfig

    def __init__(self, config):
        super().__init__(config)
        self.model = convnext_base()

    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}

ConvNextBaseConfig.register_for_auto_class()
ConvNextBaseModel.register_for_auto_class("AutoModel")
ConvNextBaseModelForImageClassification.register_for_auto_class("AutoModelForImageClassification")