/NVAE

The Official PyTorch Implementation of "NVAE: A Deep Hierarchical Variational Autoencoder" (NeurIPS 2020 spotlight paper)

Primary LanguagePythonOtherNOASSERTION

The Official PyTorch Implementation of "NVAE: A Deep Hierarchical Variational Autoencoder" (NeurIPS 2020 Spotlight Paper)

Arash Vahdat   ·   Jan Kautz


NVAE is a deep hierarchical variational autoencoder that enables training SOTA likelihood-based generative models on several image datasets.

Requirements

NVAE is built in Python 3.7 using PyTorch 1.6.0. Use the following command to install the requirements:

pip install -r requirements.txt

Set up file paths and data

We have examined NVAE on several datasets. For large datasets, we store the data in LMDB datasets for I/O efficiency. Click below on each dataset to see how you can prepare your data. Below, $DATA_DIR indicates the path to a data directory that will contain all the datasets and $CODE_DIR refers to the code directory:

MNIST and CIFAR-10

These datasets will be downloaded automatically, when you run the main training for NVAE using train.py for the first time. You can use --data=$DATA_DIR/mnist or --data=$DATA_DIR/cifar10, so that the datasets are downloaded to the corresponding directories.

CelebA 64 Run the following commands to download the CelebA images and store them in an LMDB dataset: 
cd $CODE_DIR/scripts
python create_celeba64_lmdb.py --split train --img_path $DATA_DIR/celeba_org --lmdb_path $DATA_DIR/celeba64_lmdb
python create_celeba64_lmdb.py --split valid --img_path $DATA_DIR/celeba_org --lmdb_path $DATA_DIR/celeba64_lmdb
python create_celeba64_lmdb.py --split test  --img_path $DATA_DIR/celeba_org --lmdb_path $DATA_DIR/celeba64_lmdb

Above, the images will be downloaded to $DATA_DIR/celeba_org automatically and then then LMDB datasets are created at $DATA_DIR/celeba64_lmdb.

ImageNet 32x32

Run the following commands to download tfrecord files from GLOW and to convert them to LMDB datasets

mkdir -p $DATA_DIR/imagenet-oord
cd $DATA_DIR/imagenet-oord
wget https://storage.googleapis.com/glow-demo/data/imagenet-oord-tfr.tar
tar -xvf imagenet-oord-tfr.tar
cd $CODE_DIR/scripts
python convert_tfrecord_to_lmdb.py --dataset=imagenet-oord_32 --tfr_path=$DATA_DIR/imagenet-oord/mnt/host/imagenet-oord-tfr --lmdb_path=$DATA_DIR/imagenet-oord/imagenet-oord-lmdb_32 --split=train
python convert_tfrecord_to_lmdb.py --dataset=imagenet-oord_32 --tfr_path=$DATA_DIR/imagenet-oord/mnt/host/imagenet-oord-tfr --lmdb_path=$DATA_DIR/imagenet-oord/imagenet-oord-lmdb_32 --split=validation
CelebA HQ 256

Run the following commands to download tfrecord files from GLOW and to convert them to LMDB datasets

mkdir -p $DATA_DIR/celeba
cd $DATA_DIR/celeba
wget https://storage.googleapis.com/glow-demo/data/celeba-tfr.tar
tar -xvf celeba-tfr.tar
cd $CODE_DIR/scripts
python convert_tfrecord_to_lmdb.py --dataset=celeba --tfr_path=$DATA_DIR/celeba/celeba-tfr --lmdb_path=$DATA_DIR/celeba/celeba-lmdb --split=train
python convert_tfrecord_to_lmdb.py --dataset=celeba --tfr_path=$DATA_DIR/celeba/celeba-tfr --lmdb_path=$DATA_DIR/celeba/celeba-lmdb --split=validation
FFHQ 256

Visit this Google drive location and download images1024x1024.zip. Run the following commands to unzip the images and to store them in LMDB datasets:

mkdir -p $DATA_DIR/ffhq
unzip images1024x1024.zip -d $DATA_DIR/ffhq/
cd $CODE_DIR/scripts
python create_ffhq_lmdb.py --ffhq_img_path=$DATA_DIR/ffhq/images1024x1024/ --ffhq_lmdb_path=$DATA_DIR/ffhq/ffhq-lmdb --split=train
python create_ffhq_lmdb.py --ffhq_img_path=$DATA_DIR/ffhq/images1024x1024/ --ffhq_lmdb_path=$DATA_DIR/ffhq/ffhq-lmdb --split=validation
LSUN

We use LSUN datasets in our follow-up works. Visit LSUN for instructions on how to download this dataset. Since the LSUN scene datasets come in the LMDB format, they are ready to be loaded using torchvision data loaders.

Running the main NVAE training and evaluation scripts

We use the following commands on each dataset for training NVAEs on each dataset for Table 1 in the paper. In all the datasets but MNIST normalizing flows are enabled. Check Table 6 in the paper for more information on training details. Note that for the multinode training (more than 8-GPU experiments), we use the mpirun command to run the training scripts on multiple nodes. Please adjust the commands below according to your setup. Below IP_ADDR is the IP address of the machine that will host the process with rank 0 (see here). NODE_RANK is the index of each node among all the nodes that are running the job.

MNIST

Two 16-GB V100 GPUs are used for training NVAE on dynamically binarized MNIST. Training takes about 21 hours.

export EXPR_ID=UNIQUE_EXPR_ID
export DATA_DIR=PATH_TO_DATA_DIR
export CHECKPOINT_DIR=PATH_TO_CHECKPOINT_DIR
export CODE_DIR=PATH_TO_CODE_DIR
cd $CODE_DIR
python train.py --data $DATA_DIR/mnist --root $CHECKPOINT_DIR --save $EXPR_ID --dataset mnist --batch_size 200 \
        --epochs 400 --num_latent_scales 2 --num_groups_per_scale 10 --num_postprocess_cells 3 --num_preprocess_cells 3 \
        --num_cell_per_cond_enc 2 --num_cell_per_cond_dec 2 --num_latent_per_group 20 --num_preprocess_blocks 2 \
        --num_postprocess_blocks 2 --weight_decay_norm 1e-2 --num_channels_enc 32 --num_channels_dec 32 --num_nf 0 \
        --ada_groups --num_process_per_node 2 --use_se --res_dist --fast_adamax
CIFAR-10

Eight 16-GB V100 GPUs are used for training NVAE on CIFAR-10. Training takes about 55 hours.

export EXPR_ID=UNIQUE_EXPR_ID
export DATA_DIR=PATH_TO_DATA_DIR
export CHECKPOINT_DIR=PATH_TO_CHECKPOINT_DIR
export CODE_DIR=PATH_TO_CODE_DIR
cd $CODE_DIR
python train.py --data $DATA_DIR/cifar10 --root $CHECKPOINT_DIR --save $EXPR_ID --dataset cifar10 \
        --num_channels_enc 128 --num_channels_dec 128 --epochs 400 --num_postprocess_cells 2 --num_preprocess_cells 2 \
        --num_latent_scales 1 --num_latent_per_group 20 --num_cell_per_cond_enc 2 --num_cell_per_cond_dec 2 \
        --num_preprocess_blocks 1 --num_postprocess_blocks 1 --num_groups_per_scale 30 --batch_size 32 \
        --weight_decay_norm 1e-2 --num_nf 1 --num_process_per_node 8 --use_se --res_dist --fast_adamax
CelebA 64

Eight 16-GB V100 GPUs are used for training NVAE on CelebA 64. Training takes about 92 hours.

export EXPR_ID=UNIQUE_EXPR_ID
export DATA_DIR=PATH_TO_DATA_DIR
export CHECKPOINT_DIR=PATH_TO_CHECKPOINT_DIR
export CODE_DIR=PATH_TO_CODE_DIR
cd $CODE_DIR
python train.py --data $DATA_DIR/celeba64_lmdb --root $CHECKPOINT_DIR --save $EXPR_ID --dataset celeba_64 \
        --num_channels_enc 64 --num_channels_dec 64 --epochs 90 --num_postprocess_cells 2 --num_preprocess_cells 2 \
        --num_latent_scales 3 --num_latent_per_group 20 --num_cell_per_cond_enc 2 --num_cell_per_cond_dec 2 \
        --num_preprocess_blocks 1 --num_postprocess_blocks 1 --weight_decay_norm 1e-1 --num_groups_per_scale 20 \
        --batch_size 16 --num_nf 1 --ada_groups --num_process_per_node 8 --use_se --res_dist --fast_adamax
ImageNet 32x32

24 16-GB V100 GPUs are used for training NVAE on ImageNet 32x32. Training takes about 70 hours.

export EXPR_ID=UNIQUE_EXPR_ID
export DATA_DIR=PATH_TO_DATA_DIR
export CHECKPOINT_DIR=PATH_TO_CHECKPOINT_DIR
export CODE_DIR=PATH_TO_CODE_DIR
export IP_ADDR=IP_ADDRESS
export NODE_RANK=NODE_RANK_BETWEEN_0_TO_2
cd $CODE_DIR
mpirun --allow-run-as-root -np 3 -npernode 1 bash -c \
        'python train.py --data $DATA_DIR/imagenet-oord/imagenet-oord-lmdb_32 --root $CHECKPOINT_DIR --save $EXPR_ID --dataset imagenet_32 \
        --num_channels_enc 192 --num_channels_dec 192 --epochs 45 --num_postprocess_cells 2 --num_preprocess_cells 2 \
        --num_latent_scales 1 --num_latent_per_group 20 --num_cell_per_cond_enc 2 --num_cell_per_cond_dec 2 \
        --num_preprocess_blocks 1 --num_postprocess_blocks 1 --num_groups_per_scale 28 \
        --batch_size 24 --num_nf 1 --warmup_epochs 1 \
        --weight_decay_norm 1e-2 --weight_decay_norm_anneal --weight_decay_norm_init 1e0 \
        --num_process_per_node 8 --use_se --res_dist \
        --fast_adamax --node_rank $NODE_RANK --num_proc_node 3 --master_address $IP_ADDR '
CelebA HQ 256

24 32-GB V100 GPUs are used for training NVAE on CelebA HQ 256. Training takes about 94 hours.

export EXPR_ID=UNIQUE_EXPR_ID
export DATA_DIR=PATH_TO_DATA_DIR
export CHECKPOINT_DIR=PATH_TO_CHECKPOINT_DIR
export CODE_DIR=PATH_TO_CODE_DIR
export IP_ADDR=IP_ADDRESS
export NODE_RANK=NODE_RANK_BETWEEN_0_TO_2
cd $CODE_DIR
mpirun --allow-run-as-root -np 3 -npernode 1 bash -c \
        'python train.py --data $DATA_DIR/celeba/celeba-lmdb --root $CHECKPOINT_DIR --save $EXPR_ID --dataset celeba_256 \
        --num_channels_enc 30 --num_channels_dec 30 --epochs 300 --num_postprocess_cells 2 --num_preprocess_cells 2 \
        --num_latent_scales 5 --num_latent_per_group 20 --num_cell_per_cond_enc 2 --num_cell_per_cond_dec 2 \
        --num_preprocess_blocks 1 --num_postprocess_blocks 1 --weight_decay_norm 1e-2 --num_groups_per_scale 16 \
        --batch_size 4 --num_nf 2 --ada_groups --min_groups_per_scale 4 \
        --weight_decay_norm_anneal --weight_decay_norm_init 1. --num_process_per_node 8 --use_se --res_dist \
        --fast_adamax --num_x_bits 5 --node_rank $NODE_RANK --num_proc_node 3 --master_address $IP_ADDR '

In our early experiments, a smaller model with 24 channels instead of 30, could be trained on only 8 GPUs in the same time (with the batch size of 6). The smaller models obtain only 0.01 bpd higher negative log-likelihood.

FFHQ 256

24 32-GB V100 GPUs are used for training NVAE on FFHQ 256. Training takes about 160 hours.

export EXPR_ID=UNIQUE_EXPR_ID
export DATA_DIR=PATH_TO_DATA_DIR
export CHECKPOINT_DIR=PATH_TO_CHECKPOINT_DIR
export CODE_DIR=PATH_TO_CODE_DIR
export IP_ADDR=IP_ADDRESS
export NODE_RANK=NODE_RANK_BETWEEN_0_TO_2
cd $CODE_DIR
mpirun --allow-run-as-root -np 3 -npernode 1 bash -c \
        'python train.py --data $DATA_DIR/ffhq/ffhq-lmdb --root $CHECKPOINT_DIR --save $EXPR_ID --dataset ffhq \
        --num_channels_enc 30 --num_channels_dec 30 --epochs 200 --num_postprocess_cells 2 --num_preprocess_cells 2 \
        --num_latent_scales 5 --num_latent_per_group 20 --num_cell_per_cond_enc 2 --num_cell_per_cond_dec 2 \
        --num_preprocess_blocks 1 --num_postprocess_blocks 1 --weight_decay_norm 1e-1  --num_groups_per_scale 16 \
        --batch_size 4 --num_nf 2  --ada_groups --min_groups_per_scale 4 \
        --weight_decay_norm_anneal --weight_decay_norm_init 1. --num_process_per_node 8 --use_se --res_dist \
        --fast_adamax --num_x_bits 5 --learning_rate 8e-3 --node_rank $NODE_RANK --num_proc_node 3 --master_address $IP_ADDR '

In our early experiments, a smaller model with 24 channels instead of 30, could be trained on only 8 GPUs in the same time (with the batch size of 6). The smaller models obtain only 0.01 bpd higher negative log-likelihood.

If for any reason your training is stopped, use the exact same commend with the addition of --cont_training to continue training from the last saved checkpoint. If you observe NaN, continuing the training using this flag usually will not fix the NaN issue.

Known Issues

Cannot build CelebA 64 or training gives NaN right at the beginning on this dataset

Several users have reported issues building CelebA 64 or have encountered NaN at the beginning of training on this dataset. If you face similar issues on this dataset, you can download this dataset manually and build LMDBs using instructions on this issue NVlabs#2 .

Getting NaN after a few epochs of training

One of the main challenges in training very deep hierarchical VAEs is training instability that we discussed in the paper. We have verified that the settings in the commands above can be trained in a stable way. If you modify the settings above and you encounter NaN after a few epochs of training, you can use these tricks to stabilize your training: i) increase the spectral regularization coefficient, --weight_decay_norm. ii) Use exponential decay on --weight_decay_norm using --weight_decay_norm_anneal and --weight_decay_norm_init. iii) Decrease learning rate.

Training freezes with no NaN

In some very rare cases, we observed that training freezes after 2-3 days of training. We believe the root cause of this is because of a racing condition that is happening in one of the low-level libraries. If for any reason the training is stopped, kill your current run, and use the exact same commend with the addition of --cont_training to continue training from the last saved checkpoint.

Monitoring the training progress

While running any of the commands above, you can monitor the training progress using Tensorboard:

Click here 
tensorboard --logdir $CHECKPOINT_DIR/eval-$EXPR_ID/

Above, $CHECKPOINT_DIR and $EXPR_ID are the same variables used for running the main training script.

Post-training sampling, evaluation, and checkpoints

Evaluating Log-Likelihood

You can use the following command to load a trained model and evaluate it on the test datasets:

cd $CODE_DIR
python evaluate.py --checkpoint $CHECKPOINT_DIR/eval-$EXPR_ID/checkpoint.pt --data $DATA_DIR/mnist --eval_mode=evaluate --num_iw_samples=1000

Above, --num_iw_samples indicates the number of importance weighted samples used in evaluation. $CHECKPOINT_DIR and $EXPR_ID are the same variables used for running the main training script. Set --data to the same argument that was used when training NVAE (our example is for MNIST).

Sampling

You can also use the following command to generate samples from a trained model:

cd $CODE_DIR
python evaluate.py --checkpoint $CHECKPOINT_DIR/eval-$EXPR_ID/checkpoint.pt --eval_mode=sample --temp=0.6 --readjust_bn

where --temp sets the temperature used for sampling and --readjust_bn enables readjustment of the BN statistics as described in the paper. If you remove --readjust_bn, the sampling will proceed with BN layer in the eval mode (i.e., BN layers will use running mean and variances extracted during training).

Computing FID

You can compute the FID score using 50K samples. To do so, you will need to create a mean and covariance statistics file on the training data using a command like:

cd $CODE_DIR
python scripts/precompute_fid_statistics.py --data $DATA_DIR/cifar10 --dataset cifar10 --fid_dir /tmp/fid-stats/

The command above computes the references statistics on the CIFAR-10 dataset and stores them in the --fid_dir durectory. Given the reference statistics file, we can run the following command to compute the FID score:

cd $CODE_DIR
python evaluate.py --checkpoint $CHECKPOINT_DIR/eval-$EXPR_ID/checkpoint.pt --data $DATA_DIR/cifar10 --eval_mode=evaluate_fid  --fid_dir /tmp/fid-stats/ --temp=0.6 --readjust_bn

where --temp sets the temperature used for sampling and --readjust_bn enables readjustment of the BN statistics as described in the paper. If you remove --readjust_bn, the sampling will proceed with BN layer in the eval mode (i.e., BN layers will use running mean and variances extracted during training). Above, $CHECKPOINT_DIR and $EXPR_ID are the same variables used for running the main training script. Set --data to the same argument that was used when training NVAE (our example is for MNIST).

Checkpoints

We provide checkpoints on MNIST, CIFAR-10, CelebA 64, CelebA HQ 256, FFHQ in this Google drive directory. For CIFAR10, we provide two checkpoints as we observed that a multiscale NVAE provides better qualitative results than a single scale model on this dataset. The multiscale model is only slightly worse in terms of log-likelihood (0.01 bpd). We also observe that one of our early models on CelebA HQ 256 with 0.01 bpd worse likelihood generates much better images in low temperature on this dataset.

You can use the commands above to evaluate or sample from these checkpoints.

How to construct smaller NVAE models

In the commands above, we are constructing big NVAE models that require several days of training in most cases. If you'd like to construct smaller NVAEs, you can use these tricks:

  • Reduce the network width: --num_channels_enc and --num_channels_dec are controlling the number of initial channels in the bottom-up and top-down networks respectively. Recall that we halve the number of channels with every spatial downsampling layer in the bottom-up network, and we double the number of channels with every upsampling layer in the top-down network. By reducing --num_channels_enc and --num_channels_dec, you can reduce the overall width of the networks.

  • Reduce the number of residual cells in the hierarchy: --num_cell_per_cond_enc and --num_cell_per_cond_dec control the number of residual cells used between every latent variable group in the bottom-up and top-down networks respectively. In most of our experiments, we are using two cells per group for both networks. You can reduce the number of residual cells to one to make the model smaller.

  • Reduce the number of epochs: You can reduce the training time by reducing --epochs.

  • Reduce the number of groups: You can make NVAE smaller by using a smaller number of latent variable groups. We use two schemes for setting the number of groups:

    1. An equal number of groups: This is set by --num_groups_per_scale which indicates the number of groups in each scale of latent variables. Reduce this number to have a small NVAE.

    2. An adaptive number of groups: This is enabled by --ada_groups. In this case, the highest resolution of latent variables will have --num_groups_per_scale groups and the smaller scales will get half the number of groups successively (see groups_per_scale in utils.py). We don't let the number of groups go below --min_groups_per_scale. You can reduce the total number of groups by reducing --num_groups_per_scale and --min_groups_per_scale when --ada_groups is enabled.

Understanding the implementation

If you are modifying the code, you can use the following figure to map the code to the paper.

Traversing the latent space

We can generate images by traversing in the latent space of NVAE. This sequence is generated using our model trained on CelebA HQ, by interpolating between samples generated with temperature 0.6. Some artifacts are due to color quantization in GIFs.

License

Please check the LICENSE file. NVAE may be used non-commercially, meaning for research or evaluation purposes only. For business inquiries, please contact researchinquiries@nvidia.com.

You should take into consideration that VAEs are trained to mimic the training data distribution, and, any bias introduced in data collection will make VAEs generate samples with a similar bias. Additional bias could be introduced during model design, training, or when VAEs are sampled using small temperatures. Bias correction in generative learning is an active area of research, and we recommend interested readers to check this area before building applications using NVAE.

Bibtex:

Please cite our paper, if you happen to use this codebase:

@inproceedings{vahdat2020NVAE,
  title={{NVAE}: A Deep Hierarchical Variational Autoencoder},
  author={Vahdat, Arash and Kautz, Jan},
  booktitle={Neural Information Processing Systems (NeurIPS)},
  year={2020}
}