Skip to content

Latest commit

 

History

History
79 lines (60 loc) · 2.71 KB

INSTRUCTIONS.md

File metadata and controls

79 lines (60 loc) · 2.71 KB
Raw

Instructions

Table of contents

Environment

The project is based on Python 3.6, to manage the dependencies contained in requirements.txt a virtual environment is recommended.

python3.6 -m venv venv
source venv/bin/activate
pip install -e .

Even better, use a Conda environment:

conda create -y -n koalarization python=3.6
conda activate koalarization
pip install -e .

For GPU-support, run:

pip install -e .[gpu]

If it does not work, manually install tensorflow-gpu library (must be compatible with your OS, python version and tensorflow library). More info here.

Dataset

Prior to training, the images from ImageNet need to be downloaded, resized and processed.

Extracting the embeddings in real time during training could potentially slow down the process, so the Inception embeddings are computed in advance, stored together with the images and fed directly to the fusion layer.

At training time, for each image, we need the greyscale and colored versions and the inception embeddings. Storing the images and the embedding as separate files on disk (jpeg/png and csv for instance) would impact the performances during training, so all the image-embedding pairs are stored in binary format in large continuous TFRecords.

Refer to DATASET for detailed instructions on data preparation.

Training and evaluation

Before training, ensure that the folder data contains:

  • training records as tfrecords/lab_images_*.tfrecord
  • validation records as tfrecords/val_lab_images_*.tfrecord (just rename some training records as validation, but do it before any training!)

The training script will train on all the training images, and regularly checkpoint the weights and save to disk some colored images from the validation set.

All training logs, metrics and checkpoints are saved in runs/run_id.

python -m koalarization.train \
  --run-id 'run1' \
  --train-steps 100 \
  --val-every 20 \
  'data/tfrecords' 'runs/'

The evaluation script will load the latest checkpoint, colorize images from the validation records and save them to disk. At the moment, it is not possible to operate on normal image files (e.g. jpeg or png), but the images must be processed as TFRecords first.

python -m koalarization.evaluate \
  --run-id 'run1' \
  'data/tfrecords' 'runs/'

Development and test the code

See here.