modeling_efficientnet.py

from torch import nn
from functools import partial

from transformers import PreTrainedModel
from transformers.modeling_outputs import BaseModelOutputWithPoolingAndNoAttention, ImageClassifierOutputWithNoAttention
from timm import create_model

from configuration_efficientnet import EfficientNetConfig

class EfficientNetModel(PreTrainedModel):
    config_class = EfficientNetConfig

    def __init__(self, config):
        super().__init__(config)

        self.model = create_model(config.model_name, pretrained = config.pretrained)
    
    def forward(self, pixel_values):
        last_hidden_state = self.model.forward_features(pixel_values)
        return BaseModelOutputWithPoolingAndNoAttention(
            last_hidden_state = last_hidden_state
        )

class EfficientNetModelForImageClassification(PreTrainedModel):
    config_class = EfficientNetConfig

    def __init__(self, config):
        super().__init__(config)

        self.model = create_model(config.model_name, pretrained = config.pretrained)
    
    def forward(self, pixel_values, labels=None):
        logits = self.model(pixel_values)
        loss = None
        if labels is not None:
            loss = nn.CrossEntropyLoss(logits, labels)
        return ImageClassifierOutputWithNoAttention(
            loss = loss,
            logits = logits
        )
    
__all__ = [
    "EfficientNetModel",
    "EfficientNetModelForImageClassification"
]