This repository provides TensorFlow / Keras implementations of different ConvNeXt [1] variants. It also provides the TensorFlow / Keras models that have been populated with the original ConvNeXt pre-trained weights available from [2]. These models are not blackbox SavedModels i.e., they can be fully expanded into tf.keras.Model objects and one can call all the utility functions on them (example: .summary()).

As of today, all the TensorFlow / Keras variants of the models listed here are available in this repository except for the isotropic ones. This list includes the ImageNet-1k as well as ImageNet-21k models.

Refer to the "Using the models" section to get started. Additionally, here's a related blog post that jots down my experience.

TensorFlow / Keras implementations are available in models/convnext_tf.py. Conversion utilities are in convert.py.

The converted models are available on TF-Hub.

There should be a total of 15 different models each having two variants: classifier and feature extractor. You can load any model and get started like so:

import tensorflow as tf

model_gcs_path = "gs://tfhub-modules/sayakpaul/convnext_tiny_1k_224/1/uncompressed"
model = tf.keras.models.load_model(model_gcs_path)
print(model.summary(expand_nested=True))

The model names are interpreted as follows:

• convnext_large_21k_1k_384: This means that the model was first pre-trained on the ImageNet-21k dataset and was then fine-tuned on the ImageNet-1k dataset. Resolution used during pre-training and fine-tuning: 384x384. large denotes the topology of the underlying model.
• convnext_large_1k_224: Means that the model was pre-trained on the ImageNet-1k dataset with a resolution of 224x224.

Results are on ImageNet-1k validation set (top-1 accuracy).

Differences in the results are primarily because of the differences in the library implementations especially how image resizing is implemented in PyTorch and TensorFlow. Results can be verified with the code in i1k_eval. Logs are available at this URL.

Pre-trained models:

• Off-the-shelf classification: Colab Notebook
• Fine-tuning: Colab Notebook

Randomly initialized models:

from models.convnext_tf import get_convnext_model

convnext_tiny = get_convnext_model()
print(convnext_tiny.summary(expand_nested=True))

To view different model configurations, refer here.

• Align layer initializers (useful if someone wanted to train the models from scratch)
• Allow the models to accept arbitrary shapes (useful for downstream tasks)
• Convert the isotropic models as well
• Fine-tuning notebook (thanks to awsaf49)
• Off-the-shelf-classification notebook
• Publish models on TF-Hub

[1] ConvNeXt paper: https://arxiv.org/abs/2201.03545

[2] Official ConvNeXt code: https://github.com/facebookresearch/ConvNeXt

• Vasudev Gupta
• Gus
• Willi
• ML-GDE program