Skip to content

Latest commit

 

History

History
122 lines (98 loc) · 5.93 KB

README.md

File metadata and controls

122 lines (98 loc) · 5.93 KB
Raw

Multi-modal Cycle-consistent Generalized Zero-Shot Learning

This repository contains a PyTorch implementation of the ECCV'18 paper "Multi-modal Cycle-consistent Generalized Zero-Shot Learning" - link to paper.

The original TensorFlow implementation can be found by clicking this link. This repository also contains some utility code originally published here.

Citation

@inproceedings{felix2018multi,
  title={Multi-modal Cycle-Consistent Generalized Zero-Shot Learning},
  author={Felix, Rafael and Kumar, BG Vijay and Reid, Ian and Carneiro, Gustavo},
  booktitle={European Conference on Computer Vision},
  pages={21--37},
  year={2018},
  organization={Springer}
}

Set-up

The following set-up assumes you are using an Ubuntu system.

Anaconda Set-up

We use a conda virtual environment for this implementation. If you do not have conda or Anaconda installed, enter the following commands into your terminal:

# The version of Anaconda may be different depending on when you are installing`

$ curl -O https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
$ sh Miniconda3-latest-Linux-x86_64.sh

# and follow the prompts. Make sure to select yes for running conda init, otherwise the defaults are generally good.`

You may have to open a new terminal or re-source your ~/.bashrc to get access to the conda command.

If don't want conda's base environment to be activated on startup, set the auto_activate_base parameter to false:

$ conda config --set auto_activate_base false

Clone repository

Enter the following commands to clone this repository:

$ git clone https://github.com/Best-of-ACRV/cycle_consistent_GZSL.git
$ cd cycle_consistent_GZSL/

Virtual Environment

The conda virtual environment is created from the .yml file virtual_environment/pytorch_gzsl.yml - enter the following command to create the environment:

$ conda env create -f virtual_environment/pytorch_gzsl.yml

This will create a conda environment named pytorch_gzsl. Check that the environment was created successfully:

$ conda info --envs

Datasets

The four datasets analysed in the paper (CUB, FLO, SUN and AWA1) are available in h5 format. We provide a script to download and unzip to the expected directory (data/). The download is ~1.3GB and the unzipped data will take up about ~3.1GB. To download and unpack the data, run the following from the cycle_consistent_GZSL root directory:

$ bash data/download_datasets.sh

Alternatively, the datasets can be downloaded via this link.

Usage

Ensure your working directory is set to repository base directory (cycle_consistent_GZSL) and activate the virtual environment:

$ conda activate pytorch_gzsl

You should now see (pytorch_gzsl) in your terminal prompt:

(pytorch_gzsl) $

Basic Usage

Training the cycle-consistent GZSL method consists of the following steps:

  1. GAN training.
    1. Pre-train validation classifier for monitoring generator training.
    2. Pre-train regressor (used for cycle loss component).
    3. Train generator/discriminator.
  2. Generating fake visual features from unseen (and optionally, seen) classes.
  3. Training a GZSL classifier on the fake visual features (or a combination of fake/real).

Model classes (including training/validation/testing routines) are found in models.py - classes include Classifier (for the validation classifier and GZSL classifier), Regressor, Generator, Discriminator and GAN (which has an instantiation of the other four classes as attributes). The model classes expect a dictionary of training/model options. We provide configuration .json files for each of the four datasets in the configs/ directory. Full details on the model/training configuration options can be found by entering the following commands:

(pytorch_gzsl) $ python
>>> import models
>>> help(models.GAN.__init__)
>>> help(models.Generator.__init__)
>>> help(models.Discriminator.__init__)
>>> help(models.Regressor.__init__)
>>> help(models.Classifier.__init__)

In general, experiments are run by calling the run.py file. To view the expected arguments, enter:

(pytorch_gzsl) $ python run.py --help

The example command below runs all training steps outlined above (with default settings), where CONFIG_JSON is a model/training configuration .json file and GPU_ID is the device ID to be used.

(pytorch_gzsl) $ python run.py --config CONFIG_JSON --gpu GPU_ID --train-gan --gen-fake --train-cls

Pre-defined Experiments

We provide bash scripts to run experiments on each of the CUB, FLO, SUN and AWA datasets with the configurations/settings used to produce the results in the paper. For example, the scripts/cub directory contains the following files:

  1. run_all.sh: Runs a complete GZSL experiment, including training the GAN, generating a fake dataset, training a GZSL classifier on the real and/or fake data and evaluating the GZSL classifier on the test data.
  2. run_train_gan.sh: Run GAN training only.
  3. run_gen_fake.sh: Generates a fake dataset using a trained generator - change the WORKDIR to the correct directory.
  4. run_train_cls.sh: Trains a GZSL on real and/or fake data - change the WORKDIR to the correct directory.
  5. run_test_cls.sh: Evaluate a trained GZSL classifier on the test data - change the WORKDIR to the correct directory.

Each script takes an optional argument to specify the GPU ID - if not ID is provided, GPU 0 is used.

To run the complete pre-defined experiment on the CUB dataset enter the following command:

(pytorch_gzsl) $ bash scripts/cub/run_all.sh GPU_ID #Replace GPU_ID with desired device ID