Pretrained VGG neural nets in TensorFlow

Set of VGG neural net models for TensorFlow. Weights converted from timm module (Pytorch).

Introduction

Sometimes pretrained nets like VGG useful for segmentation problems. Unfortunately there is only pretrained VGG16 and VGG19 without batchnorm available for tensorflow. I made converter from Pytorch to VGG which allows to use all family of VGG models with imagenet weights.

Available models

vgg11, vgg11_bn, vgg13, vgg13_bn, vgg16, vgg16_bn, vgg19, vgg19_bn

Notes

• Converted only feature extraction part (no classification part available)
• Models tested against PyTorch version to give the same result (max diff < 1e-05)
• Code for conversion of model weights available here: convert_model_weights.py

Usage

from vgg_tensorflow import vgg11_bn

model = vgg11_bn(inp_size=(224, 224, 3), pretrained=True)