TensorFlow Feature Extractor

This is a convenient wrapper for feature extraction or classification in TensorFlow. Given well known pre-trained models on ImageNet, the extractor runs over a list or directory of images. Optionally, features can be saved as HDF5 file. It supports all the pre-trained models listed on the official page.

TensorFlow models tested:

1. Inception v1-v4
2. ResNet v1 and v2
3. VGG 16-19

Requirements

• TensorFlow (tested with version 1.3)
• TensorFlow Models
• The usual suspects: numpy, scipy.
• Optionally h5py for saving features to HDF5 file

Setup

1. Checkout the TensorFlow models repository somewhere on your machine. The path where you checkout the repository will be denoted <checkout_dir>/models

git clone https://github.com/tensorflow/models/

2. Add the directory <checkout_dir>/research/slim to the$PYTHONPATH variable. Or add a line to your .bashrc file.

export PYTHONPATH="<checkout_dir>/research/slim:$PYTHONPATH"

3. Download the model checkpoints from the official page.

Usage

There are two example files, one for classification and one for feature extraction.

Feature Extraction

ResNet-v1-101

example_feat_extract.py --network resnet_v1_101 --checkpoint ./checkpoints/resnet_v1_101.ckpt --image_path ./images_dir/ --out_file ./features.h5
--num_classes 1000 --layer_names resnet_v1_101/logits

ResNet-v2-101

example_feat_extract.py --network resnet_v2_101 --checkpoint ./checkpoints/resnet_v2_101.ckpt --image_path ./images_dir/
--out_file ./features.h5 --layer_names resnet_v2_101/logits --preproc_func inception

Inception-v4

example_feat_extract.py --network inception_v4 --checkpoint ./checkpoints/inception_v4.ckpt --image_path ./images_dir/
--out_file ./features.h5 --layer_names Logits

Image Classification

example_classification.py
--network resnet_v1_101 --checkpoint ./checkpoints/resnet_v1_101.ckpt --image_path ./images_dir/
--num_classes 1000 --logits_name resnet_v1_101/logits

Work in Progress

1. Save image file names to HDF5 file
2. Support for multi-threaded preprocessing