NVAE from scratch

Pytorch implementation (using Distributed Data Parallel and Automatic Mixed Precision) of Nouveau-VAE: paper link

The implementation uses FFCV for fast data loading and WandB for logging.

Note on the implementation and useful resources

The repository is based on the official github repo of NVAE, implementing all the main features and details, while trying to keep the code well organized and clean.

Further resources that have been useful in the implementation include:

• Unofficial NVAE implementations by mgp123 and GlassyWing.
• The amazing Blog Post by M.G. Portnoy, which well explains the main architecture and some tricks.

Install and Run

Installation

# Dependencies Install 
conda env create --file environment.yml
conda activate nvae_from_scratch

# package install (after cloning)
pip install .

Note: Check the pytorch-cuda version in environment.yml to ensure it is compatible with your cuda version.

Dataset Creation

You need to create the .beton files to correctly load your datasets. You can use the provided script ./data/make_ffcv.py, that will take as input one or more image folders (train, validation, test) and output the corresponding train.beton, validation.beton, test.beton files in the provided output folder.

Please refer to the official ffcv docs for more information.

Training

To train your own NVAE model, you need to launch the train.py script. The script is compatible with multi-gpu and multi-node setups using torchrun or mpirun.

For example:
On a multinode cluster I use mpirun (enables to run only from master node)

mpirun -np world_size 
       -H ip_node_0:n_gpus,ip_node_1:n_gpus ... 
       -x MASTER_ADDR=ip_master 
       -x MASTER_PORT=1234
       -bind-to none -map-by slot -mca pml ob1 -mca btl ^openib
       python train.py --args

In a single node, multi-gpu, local environment I use torchrun

torchrun --nproc_per_node=ngpus 
         --nnodes=1 
         --node_rank=0 
         --master_addr='localhost'
         --master_port=1234 
         train.py --args

For debugging on 1 GPU I simply run the script

python train.py --args

Regarding the --args, these are:

• --run_name xxx represents the unique name of the training run, used to save checkpoints and logging.
• --conf_file ./example.yaml path to a .yaml configuration file, like the one in ./configuration_files/base_conf.yaml, where you can find more information on the single parameters.
• --data_path ./dataset/ path to a directory containing the train.beton and validation.beton files previously created.
• --checkpoint_base_path ./checkpoints/ path to the base directory where checkpoints are saved.
• --resume_from ./checkpoints/last.pt path to a specific checkpoint file, in case of resuming training.
• --logging will attempt to log results using wandb.
• --wandb_id z3c4z5r3 unique id of a specific wandb run, useful when resuming.

Evaluation Scripts

The ./scripts/ directory contains three scripts:

1. update_bn_stats.py: it is beneficial to readjust the batch norm statistics of a trained model. This script does it, for a specific sampling temperature. The updated model is saved as a copy on the same checkpoint file passed as input. This is the first script to launch after training.
2. evaluate.py: computes L2 reconstruction error, FID and IS scores for generation ability measures.
3. sample_random_images.py: samples and save some images for qualitative evaluations.

Note: check the single scripts for a description of the arguments to be passed.

Results

Some pretrained models and their configuration files can be downloaded from: https://huggingface.co/SerezD/NVAE-from-scratch

CIFAR 10 (Unconditioned)

The trained models have 3 scales, each one with 8 groups (24 latents in total). One version has additional normalizing flows cells.

Run Name	Temperature	L2 (1e-5)	FID	IS	# (trainable) params
NVAE 8x3	0.6	10.097	19.73	6.00 +- 0.24	705.678 M
NVAE 8x3	0.8	03.918	23.63	6.36 +- 0.16	705.678 M
NVAE 8x3	1.0	01.243	32.50	5.77 +- 0.19	705.678 M
NVAE 8x3 NF	0.6	07.814	20.66	6.94 +- 0.09	707.403 M
NVAE 8x3 NF	0.8	03.223	24.73	6.53 +- 0.22	707.403 M
NVAE 8x3 NF	1.0	00.997	33.28	5.80 +- 0.13	707.403 M

Random Samples from NVAE 8x3 NF (temp 0.6):

CELEBA 64x64

The trained model has 3 scales, each one with 8 groups (24 latents in total).

Run Name	Temperature	L2 (1e-5)	FID	IS	# (trainable) params
NVAE 8x3	0.6	02.751	14.86	2.20 +- 0.03	705.678 M
NVAE 8x3	0.8	01.996	12.82	2.18 +- 0.06	705.678 M
NVAE 8x3	1.0	01.746	15.10	2.16 +- 0.05	705.678 M

Random Samples (temp 0.6):

Random Samples (temp 0.8):

Random Samples (temp 1.0):