Compute-Features

Computes features for images using various pretrained Tensorflow models. For each model, this will output the final fully connected layer (e.g. fc7 for Alexnet, fc8 for VGG_19, etc.). These can be used within various applications such as classification, clustering, etc.

Usage

Download one of the checkpoint files below, for example Inception V1.

tar -xvf inception_v1_2016_08_28.tar.gz
python compute_features.py --data_dir=test_images/ --checkpoint_file=inception_v1.ckpt --model=inception_v1

The output will be in inception_v1_features.pkl, which contains a dictionary of the form {image_path:feature}.

Look at load_features.py for an example of how to use the features that were computed. For example,

python load_features.py features/inception_v1_features.pkl

Some of these aren't working, such as inception_v4 due to differences in the model checkpoint and the model defined in the nets/ directory. Still working on that.

Pre-trained Models

You can download all models here. Otherwise links for individual models can be found below.

These CNNs have been trained on the ILSVRC-2012-CLS image classification dataset.

In the table below, we list each model, the corresponding TensorFlow model file, the link to the model checkpoint, and the top 1 and top 5 accuracy (on the imagenet test set). Note that the VGG and ResNet V1 parameters have been converted from their original caffe formats (here and here), whereas the Inception and ResNet V2 parameters have been trained internally at Google. Also be aware that these accuracies were computed by evaluating using a single image crop. Some academic papers report higher accuracy by using multiple crops at multiple scales.