TensorFlow-ResNets

This repository contains TensorFlow Keras ResNet models. Below, you will find the supported variants of ResNet and what weights are supported.

The codebase takes inspiration from TensorFlow ResNets and PyTorch ResNets.

This repository is compatible with TF 2.2.0 even though grouped convolutions are only supported in TF Nightly. Grouped convolutions will be imported from convolutions.py which is the TF Nightly version of convolution layers. This allows the use of TPUs since they don't work with TF Nightly.

These models will not work without a GPU or TPU due to the use of grouped convolutions.

Installation

To install, you can use pip install tf2-resnets.
To use a model, you can do

from tf2_resnets import models

# Weights here are ImageNet.
# They can also be 'ssl' (semi-supervised)
# or 'swsl' (semi-weakly supervised)
# for a selection of models.
model = models.ResNeXt50(input_shape=(224, 224, 3), weights='imagenet')

Models

* ResNeSt models' Top-1 Accuracies were reported using different crop sizes. Crop sizes in order - 224, 256, 320, and 416.

Preprocessing

There are two different preprocessing functions which are meant to be used according to the task at hand. The first one, located in imagenet_preprocessing.py, is meant to be used for reproducing the ImageNet results. The second function is located inside resnet.py. This function is meant to be used for transfer learning. The first function includes an additional resize to 256 by 256 and a central crop to 224 by 224.

Original Implementations

The original implementions of these models are listed below.

• ResNet (PyTorch)
• ResNeSt (PyTorch)
• ResNeXt (PyTorch)
• Wide ResNet (PyTorch)