Skip to content

ConvNeXt Models for Image Classification

This page documents the ConvNeXt image classification model pool in FusionBench. It wraps Hugging Face Transformers ConvNeXt models with convenient loading, processor management, dataset-aware head adaptation, and save utilities.

Implementation: ConvNextForImageClassificationPool

alt text

Quick start

Minimal Python usage with a single pretrained ConvNeXt model (e.g., base-224):

from fusion_bench.modelpool import ConvNextForImageClassificationPool

pool = ConvNextForImageClassificationPool(
        models={
                "_pretrained_": {
                        "config_path": "facebook/convnext-base-224",
                        "pretrained": True,
                        # set to a known dataset key (e.g., "cifar10") to resize classifier
                        # and populate id2label/label2id mappings
                        "dataset_name": None,
                }
        }
)

model = pool.load_model("_pretrained_")
processor = pool.load_processor()  # AutoImageProcessor

Low-level construction is available via helpers:

Ready-to-use config

Use the provided Hydra config to set up a pretrained ConvNeXt-base model:

config/modelpool/ConvNextForImageClassification/convnext-base-224.yaml
_target_: fusion_bench.modelpool.ConvNextForImageClassificationPool
_recursive_: False
models:
  _pretrained_:
    config_path: facebook/convnext-base-224
    pretrained: true
    dataset_name: null
train_datasets: null
val_datasets: null
test_datasets: null

Tip: set dataset_name to a supported dataset key (e.g., cifar10, svhn, gtsrb, …) to auto-resize the classifier and label mappings.

Implementation Details