modeling_efficientnet.py

from torch import nn, Tensor, tensor
from typing import Union, List, Optional

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

from .configuration_efficientnet import EfficientNetConfig


class EfficientNetModel(PreTrainedModel):
    """
    EfficientNet model wrapper using Hugging Face's PreTrainedModel.

    This class initializes an EfficientNet model from `timm` library 
    and defines a forward method that extracts feature representations.

    Attributes
    ----------
        config:  
            Configuration object containing model parameters.
        model:   
            Instantiated EfficientNet model.
    """
    config_class = EfficientNetConfig

    def __init__(self, config):
        super().__init__(config)
        self.config = config
        self.model = create_model(
            config.model_name, 
            pretrained = config.pretrained, 
            num_classes = config.num_classes, 
            global_pool = config.global_pool,
        )
        
    def forward(self, pixel_values: Tensor) -> BaseModelOutputWithPoolingAndNoAttention:
        """
        Parameters
        ----------
            pixel_values : torch.Tensor
                Input tensor representing image pixel values.

        Returns
        -------
            BaseModelOutputWithPoolingAndNoAttention
                Object containing the `last_hidden_state` and `pooled_output`.
        """
        last_hidden_state = self.model.forward_features(pixel_values)
        pooler_output = self.model.forward_head(last_hidden_state, pre_logits=True)
        
        return BaseModelOutputWithPoolingAndNoAttention(
            last_hidden_state = last_hidden_state,
            pooler_output=pooler_output
        )


class EfficientNetModelForImageClassification(PreTrainedModel):
    """
    EfficientNet model wrapper using Hugging Face's PreTrainedModel.

    This class initializes an EfficientNet model from `timm` library 
    and defines a forward method that return logits.

    It supports training when labels are provided

    Attributes
    ----------
        config :  
            Configuration object containing model parameters.
        model :   
            Instantiated EfficientNet model.
    """
    config_class = EfficientNetConfig

    def __init__(self, config):
        super().__init__(config)
        self.config = config
        self.model = create_model(
            config.model_name, 
            pretrained = config.pretrained, 
            num_classes = config.num_classes, 
            global_pool = config.global_pool,
        )
    
    def forward(
        self, 
        pixel_values: Tensor, 
        labels: Optional[Union[List[int], Tensor]] = None
    ) -> ImageClassifierOutputWithNoAttention:
        """
        Parameters
        ----------
            pixel_values : torch.Tensor
                Input tensor representing image pixel values.
            labels : Optional[Union[List[int], torch.Tensor]]
                Ground truth labels for training and computing loss. 
                List of integers/tensor representing class IDs.

        Returns
        -------
            ImageClassifierOutputWithNoAttention
                Object containing `logits` and `loss`.
        """
        self.model.training = False if labels is None else True

        logits = self.model(pixel_values)

        loss = None
        if self.model.training:
            labels = tensor(labels)
            ce_loss = nn.CrossEntropyLoss()
            loss = ce_loss(logits, labels)

        return ImageClassifierOutputWithNoAttention(
            loss = loss,
            logits = logits,
        )
    
__all__ = [
    "EfficientNetModel",
    "EfficientNetModelForImageClassification"
]