Deep Automodulators

Code for the paper:

• Heljakka, A., Hou, Y., Kannala, J., and Solin, A. (2020). Deep Automodulators. In Advances in Neural Information Processing Systems (NeurIPS). [arXiv preprint]

Implementation by Ari Heljakka (code adapted from [2-5], metrics from [6-7], h5tool from [8]).

Quick start

a) Locally

We recommend to try out the models via PyTorch Hub, loading the pre-trained models automatically, with PyTorch v1.7+ and Python v3.6+ as the only dependencies.

For a quick test, just run:

python eval_example.py

The script eval_test.py is self-explanatory and lets you immediately start style-mixing, generating random samples, and reconstructing. For style-mixing and reconstruction, the FFHQ models are far superior to the others.

Alternatively, you can try the Jupyter Notebook at /src/pioneer/Automodulator%20Evaluator.ipynb

b) In Browser, via Colab

You can run the [Colab notebook] directly from browser.

c) In code

To style-mix images you have loaded, all you have to do is basically the following:

model = torch.hub.load('heljakka/automodulator', 'ffhq512', pretrained=True, source='github')
model.eval(useLN=False)
z = model.encode(images)
fused_image = model.decode(
    model.zbuilder().hi(z[0])
                    .mid(z[1])
                    .lo(z[1]))

To select exact decoder layers to modulate, switch to the .use() API, replacing the final line with, for example:

fused_image = model.decode(
    model.zbuilder().use(z[0], [0,3])
                    .use(z[1], [3,5])
                    .use(z[2], [5,-1]))

For scale-specific conditional sampling, where you fix e.g. only the coarse scales, all you nead to do is to leave those other scales undefined:

coarse_scale_conditioned_images = model.decode(
    model.zbuilder(16).hi(z[0]))

For all-scales random sampling, simply run

random_images = model.generate(16)

Pre-trained models

Pre-trained models in Zenodo are available for each dataset, either via Hub API or direct download. You can run them on command line with the usage examples below.

If you only use the pre-trained models, you do not need any datasets except for reproducing the exact reconstruction results shown on the paper.

Datasets

Supported datasets are:

• FFHQ
  • Download via NVIDIA repository and split identities 1-60k for training and 60001 - 70k for testing.
• CelebA-HQ
  • You need to put the dataset in H5 format for PyTorch, separately for training and testing, as follows:
    1. Download the CelebA dataset (originals, NOT the aligned & cropped one)
    2. Download the Delta files from the author https://github.com/tkarras/progressive_growing_of_gans (see "Data files needed to reconstruct the CelebA-HQ dataset")
    3. Run the dataset-tools/h5tool.py we provide (which contains the train/test split unlike the original CelebA-HQ script). For the syntax, run

       python h5tool.py -h
       

    Please run first data_split=train, then data_split=test. This will create a separate 27k/3k split for training and testing.
• LSUN Bedrooms
  • Used as normal. Please bear in mind the very long time on first run of any LSUN usage to just build the index (often many hours).
• LSUN Cars

Regular CelebA will likely also work but it has no scheduler setup (see Training) but you can try to replicate the CelebA-HQ scheduler. For CelebA, please choose the aligned & cropped version; also, you still need to do the train/test split into separate directories manually.

Dependencies

For reproducing the paper results, you need the full dependencies. For the paper, the models were evaluated with

• CUDA v10.2
• PyTorch v1.2.0

For all up-to-date dependencies, run

pip install -r requirements.txt

For logging (when training), either run the following or specify --no_TB:

pip install tensorboardx

Evaluation

For all command-line arguments, run

python train.py -h

You can evaluate a trained model by it's Hub model name in --hub_model as ffhq512|ffhq256|celebahq256|lsuncars256|lsunbedrooms256 or by giving its directory as save_dir. The checkpoints are saved under [save_dir]/checkpoint.

Use --testonly to designate the testing (evaluation) mode, as opposed to training. The examples below use the Hub.

1. Reconstructions, random samples and interpolations (FFHQ, 256x256 or 512x512):

python -m pioneer.train -d ffhq --hub_model=ffhq512 --train_path /data/FFHQ_train --test_path /data/FFHQ_test --sample_N=16 --reconstructions_N=8 --interpolate_N=3 --testonly

2. Reconstructions of your own input images (FFHQ, 256x256 or 512x512 - use the 512x512 model file unless your input resolution is lower): (Please remember to put your images under an extra subdirectory (e.g. /my_images/png/ for the below example. Face images must be cropped and aligned as in CelebA-HQ and FFHQ, respectively.)

python -m pioneer.train -d ffhq --hub_model=ffhq512 --sample_N=16 --reconstructions_N=8 --interpolate_N=3 --testonly --aux_inpath /my_images

For style-mixing examples, please see the eval_test.py. For not applying the layer noise, add --no_LN. (Also possible for models originally trained with layer noise.) For training or evaluating the CelebaHQ model, add --small_darch to indicate the smaller decoder architecture.

Training

All training uses the pre-configured scheduler. For new datasets, you can configure the schedule of phases, margin values and learning rates in makeTS(...) method of train.py. See the existing code for how to set them up. Pre-configured schedule exists for CelebA-HQ, FFHQ, LSUN Bedrooms and LSUN Cars.

The batch sizes have been selected to enable running on 12 GB of GPU memory. 1-GPU setup has been confirmed to work in all cases up to 9.6M steps, and 2-GPU setup on all other resolutions. Please ensure enough disk space for checkpoints.

To resume training from a checkpoint, it is sufficient to add --start_iteration=N where N is the step number of your latest state file (eg. for checkpoint/256000_state, N=256000). Use start_iteration=-1 to take the latest state.

The resolutions of each phase are defined in powers of 2, as follows: 0 = 4x4, ..., 3 = 32x32, 4 = 64x64, 5 = 128x128, 6 = 256x256, 7 = 512x512

Note that all loss function hyper-parameters in the code base are scaled by a factor of 0.1 in comparison to the paper.

Example of regular training FFHQ, up to 512x512:

python -m pioneer.train -d ffhq --save_dir FFHQ_quicktest --train_path /data/FFHQ_train --test_path /data/FFHQ_test --sample_N=16 --reconstructions_N=8 --interpolate_N=0 --total_kimg=45000

To enforce the flop-invariance on a specific layer (e.g. 2), add --flip_invariance_layer=2. This will both add an extra support layer in the decoder (as in the experiment in the paper) and also start using the training method which will horizontally flop 50% of the minibatch.

If the intermediate results are complete noise after ~10 Million training images for one of the supported datasets, then it is recommended to restart with changing the random seed with --manual_seed=N parameter. If you are working with a dataset of your own, then it is best to check whether the samples in lower-resolution stages looked correct (before breaking), and then either make the training stages in the last "stable" resolution longer, or make the KL margin tighter. You can adjust these in the make_ts() function.

Main Results

Dataset, resolution	FID-50k	PPL
FFHQ, 256x256	31.64	250.2
CelebaHQ, 256x256	29.13	203.8
LSUN Bedrooms, 256x256	25.53	-
LSUN Cars, 256x256	19.82	-

You can reproduce the samples for the table as follows.

FFHQ, 256x256:

python -m pioneer.train -d ffhq --hub_model=ffhq256 --sample_N=50100 --testonly

CelebaHQ, 256x256:

python -m pioneer.train -d celebaHQ --hub_model=celebahq256 --small_darch --sample_N=50100  --testonly

LSUN Bedrooms, 256x256:

python -m pioneer.train -d lsun --hub_model=lsunbedrooms256 --sample_N=50100  --testonly

LSUN Cars, 256x256:

python -m pioneer.train -d lsun --hub_model=lsuncars256 --sample_N=50100   --testonly

Support

For all correspondence, please contact heljakka@iki.fi

Support and email replies are not always guaranteed, but we will appreciate and evaluate all feedback.

References

[1] Heljakka, A., Hou, Y., Kannala, J., and Solin, A. (2020). Deep Automodulators. In Advances in Neural Information Processing Systems (NeurIPS). [arXiv preprint]

[2] https://github.com/AaltoVision/balanced-pioneer

[3] https://github.com/rosinality/progressive-gan-pytorch

[4] https://github.com/DmitryUlyanov/AGE

[5] Barron, J. T. (2019). A General and Adaptive Robust Loss Function In: CVPR. https://github.com/jonbarron/robust_loss_pytorch

[6] Zhang, R. and Isola, P. and Efros, A. A. and Shechtman, E. and Wang, O. (2018). The Unreasonable Effectiveness of Deep Features as a Perceptual Metric In: CVPR. https://github.com/richzhang/PerceptualSimilarity

[7] https://github.com/mseitzer/pytorch-fid

[8] Karras, T., Aila, T., Laine, S., and Lehtinen, J. (2018). Progressive growing of GANs for improved quality, stability, and variation. In: International Conference on Learning Representations (ICLR). https://github.com/tkarras/progressive_growing_of_gans

License

This software is distributed under the MIT License; please refer to the file LICENSE, included with the software, for details.