ameroyer / flextuning

Flextuning, WACV 2020

ReadMe and instructions: Work in progress

Introduction

Implementation of Flextuning described in A Flexible Selection Scheme for Minimum-Effort Transfer Learning, WACV 2020

@article{flextuning,
		        author = {Royer, Am\'{e}lie and Lampert, Christoph H.},
		        title = {A Flexible Selection Scheme for Minimum-Effort Transfer Learning},
		        journal = {Winter Conference on Applications of Computer Vision (WACV)},
		        year = {2020}
		      }

Quick usage

# Requires Tensorflow (tested with 1.12 / 1.14)

# Download Mnist source/target domains from 
wget ist.ac.at/~aroyer/Models/Data.zip
# Alternatively you can compute them (see input_pipeline.ipynb)


# Run only the first time
mkdir results
mkdir txt_logs
unzip Data.zip

# Main
python3 main.py mnist --all

Will run one repeat of all tuning experiments on MNIST -> target domains. Results of experiments are printed in terminal + saved in a dictionary (in results/). Additional logs are output in txt_logs/ (just text files) and log/ (Tensorboard summaries). See main.py.

Codebase

Datasets:

• input_pipeline.ipynb : Visualize the datasets + Convert them to TFRecords format for convenience. Also pretrain a "base model" on the source domain that will be used as initialization for tuning.
• setting.py : Defines the different source and target domains for convenience + if it is run at main, it will create susampled ratios of each datasets.
• include/dataset_utils.py: Functions to parse the datasets + image transformation for the target domains. Use low-level functions defined in include/tfrecords_utils.py and include/preprocess_utils.py.

Classifiers:

• net_utils.py : Architectures (Note: The second half of the file is only used for experiments described in section 5.4 in the paper. You can also ignore the file include/graph_utils.py, it is also only really useful for section 5.4 for optional preprcessing when feeding outputs from the preprocessing module to the classifier)
• include/graph_manager.py : It defines two things: First a "Setting" object, which essentially a wrapper to define source/target datasets. and a "Classifier" object which defines the training loop. it comes with "train" and "test" functions: This part is probably the most "Tensorflow specific" part of the code. For convenience, subclasses of the Classifier base class are then defined in include/classifiers.py

Run experiments:

• main.py: Basically gathers the whole experiment setups
• include/run_exp.py: Given a classfier, and a list of source/domain datasets, it takes care of training/evaluating all different finetuned/flextuned models and save logs.

Fast and Faster-flex variants

Which unit to "flextune" ? The two faster selectio criteria are explored in selection_criterion.ipynb