This repository contains training code for the CVPR 2018 spotlight
A Variational U-Net for Conditional Appearance and Shape Generation
The model learns to infer appearance from a single image and can synthesize images with that appearance in different poses.
Project page with more results
This is a slightly modified version of the code that was used to produce the
results in the paper. The original code was cleaned up, the data dependent
weight initialization was made compatible with tensorflow >= 1.3.0
and a
unified model between the datasets is used. You can find the original code and
checkpoints online (vunet/runs
) but if you want to use
them, please keep in mind that:
- the original checkpoints are not compatible with the graphs defined in this repository. You must use the original code distributed with the checkpoints.
- the original code uses a data dependent weight initialization scheme which
does not work with
tensorflow >= 1.3.0
. You should usetensorflow==1.2.1
. - the original code became a bit of a mess and we can no longer provide support for it.
The code was developed with Python 3. Dependencies can be installed with
pip install -r requirements.txt
These requirements correspond to the dependency versions used to generate the pretrained models but other versions might work as well.
Download and unpack the desired dataset.
This results in a folder containing an index.p
file. Either add a symbolic
link named data
pointing to the download directory or adjust the path to
the index.p
file in the <dataset>.yaml
config file.
For convenience, you can also run
./download_data.sh <dataset> <store_dir>
which will perform the above steps automatically. <dataset>
can be one of
coco
, deepfashion
or market
. To train the model, run
python main.py --config <dataset>.yaml
By default, images and checkpoints are saved to log/<current date>
. To
change the log directory and other options, see
python main.py -h
and the corresponding configuration file. To obtain images of optimal quality it is recommended to train for a second round with a loss based on Gram matrices. To do so run
python main.py --config <dataset>_retrain.yaml --retrain --checkpoint <path to checkpoint of first round>
You can find pretrained models
online (vunet/pretrained_checkpoints
).
To be able to train the model on your own dataset you must provide a pickled dictionary with the following keys:
joint_order
: list indicating the order of joints.imgs
: list of paths to images (relative to pickle file).train
: list of booleans indicating if this image belongs to training splitjoints
: list of[0,1]
normalized xy joint coordinates of shape(len(joint_jorder), 2)
. Use negative values for occluded joints.
joint_order
should contain
'rankle', 'rknee', 'rhip', 'rshoulder', 'relbow', 'rwrist', 'reye', 'lankle', 'lknee', 'lhip', 'lshoulder', 'lelbow', 'lwrist', 'leye', 'cnose'
and images without valid values for rhip, rshoulder, lhip, lshoulder
are
ignored.