modeling.py

import torch
import torchvision
from torch import nn
from transformers import PreTrainedModel
from configuration import EffNetPlantDiseaseConfig

# create model class
class EffNetPlantDiseaseClassification(PreTrainedModel):
  config_class = EffNetPlantDiseaseConfig
  def __init__(self, config):
    super().__init__(config)
    # get the model architecture from torchvision
    self.model = torchvision.models.efficientnet_v2_s()
    # modify the classifier head according to the config 
    self.model.classifier = nn.Sequential(
        nn.Dropout(p=config.dropout_rate, inplace=True),
        nn.Linear(in_features=self.model.classifier[-1].in_features,
                  out_features=config.num_classes)
      )
    self.num_classes = config.num_classes
    self.loss_fn =  nn.CrossEntropyLoss()
  # define forward method to be similar to hugging face model standards
  def forward(self, image, label=None):
    logits = self.model(image)
    loss = None
    if label is not None:
      loss = self.loss_fn(logits, label)

    return {"logits":logits,
            "loss": loss}