First commit

LICENSE.md

@@ -0,0 +1,7 @@
+Copyright © 2020 OpenAI
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

README.md

@@ -0,0 +1,103 @@
+# Very Deep VAEs
+
+Repository for the paper "Very Deep VAEs Generalize Autoregressive Models and Can Outperform Them on Images," submitted to ICLR 2021 (https://openreview.net/forum?id=RLRXCV6DbEJ)
+
+Some model samples and a visualization of how it generates them:
+![image](header-image.png)
+
+This repository is tested with PyTorch 1.6, CUDA 10.1, Numpy 1.16, Ubuntu 18.04, and V100 GPUs.
+
+# Setup
+Several additional packages are required, including NVIDIA Apex:
+```
+pip install imageio
+pip install mpi4py
+pip install sklearn
+git clone https://github.com/NVIDIA/apex
+cd apex
+pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./
+cd ..
+```
+
+Also, you'll have to download the data, depending on which one you want to run:
+```
+./setup_cifar10.sh
+./setup_imagenet.sh imagenet32
+./setup_imagenet.sh imagenet64
+./setup_ffhq256.sh
+./setup_ffhq1024.sh  /path/to/images1024x1024  # this one depends on you first downloading the subfolder `images_1024x1024` from https://github.com/NVlabs/ffhq-dataset on your own
+```
+
+# Training models
+Hyperparameters all reside in `hps.py`. We use 2 gpus for our CIFAR-10 runs, and 32 for the rest of the models. (Using a lower batch size is also possible and results in slower learning, and may also require a lower learning rate).
+
+The `mpiexec` arguments you use for runs with more than 1 node depend on the configuration of your system, so please adapt accordingly.
+
+```bash
+mpiexec -n 2 python train.py --hps cifar10
+mpiexec -n 32 python train.py --hps imagenet32
+mpiexec -n 32 python train.py --hps imagenet64
+mpiexec -n 32 python train.py --hps ffhq256
+mpiexec -n 32 python train.py --hps ffhq1024
+```
+
+# Restoring saved models
+For convenience, we have included training checkpoints which can be restored in order to confirm performance, continue training, or generate samples.
+
+### ImageNet 32
+```bash
+# 119M parameter model, trained for 1.7M iters (about 2.5 weeks on 32 V100)
+wget https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets/imagenet32-iter-1700000-log.jsonl
+wget https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets/imagenet32-iter-1700000-model.th
+wget https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets/imagenet32-iter-1700000-model-ema.th
+wget https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets/imagenet32-iter-1700000-opt.th
+python train.py --hps imagenet32 --restore_path imagenet32-iter-1700000-model.th --restore_ema_path imagenet32-iter-1700000-model-ema.th --restore_log_path imagenet32-iter-1700000-log.jsonl --restore_optimizer_path imagenet32-iter-1700000-opt.th --test_eval
+# should give 2.6364 nats per dim, which is 3.80 bpd
+```
+
+### ImageNet 64
+```bash
+# 125M parameter model, trained for 1.6M iters (about 2.5 weeks on 32 V100)
+wget https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets-2/imagenet64-iter-1600000-log.jsonl
+wget https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets-2/imagenet64-iter-1600000-model.th
+wget https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets-2/imagenet64-iter-1600000-model-ema.th
+wget https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets-2/imagenet64-iter-1600000-opt.th
+python train.py --hps imagenet64 --restore_path imagenet64-iter-1600000-model.th --restore_ema_path imagenet64-iter-1600000-model-ema.th --restore_log_path imagenet64-iter-1600000-log.jsonl --restore_optimizer_path imagenet64-iter-1600000-opt.th --test_eval
+# should be 2.44 nats, or 3.52 bits per dim
+```
+
+### FFHQ-256
+```bash
+# 115M parameters, trained for 1.7M iterations (or about 2.5 weeks) on 32 V100
+wget https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets/ffhq256-iter-1700000-log.jsonl
+wget https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets/ffhq256-iter-1700000-model.th
+wget https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets/ffhq256-iter-1700000-model-ema.th
+wget https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets/ffhq256-iter-1700000-opt.th
+python train.py --hps ffhq256 --restore_path ffhq256-iter-1700000-model.th --restore_ema_path ffhq256-iter-1700000-model-ema.th --restore_log_path ffhq256-iter-1700000-log.jsonl --restore_optimizer_path ffhq256-iter-1700000-opt.th --test_eval
+# should be 0.4232 nats, or 0.61 bits per dim
+```
+
+### FFHQ-1024
+```bash
+# 115M parameters, trained for 1.7M iterations (or about 2.5 weeks) on 32 V100
+wget https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets/ffhq1024-iter-1700000-log.jsonl
+wget https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets/ffhq1024-iter-1700000-model.th
+wget https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets/ffhq1024-iter-1700000-model-ema.th
+wget https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets/ffhq1024-iter-1700000-opt.th
+python train.py --hps ffhq1024 --restore_path ffhq1024-iter-1700000-model.th --restore_ema_path ffhq1024-iter-1700000-model-ema.th --restore_log_path ffhq1024-iter-1700000-log.jsonl --restore_optimizer_path ffhq1024-iter-1700000-opt.th --test_eval
+# should be 1.678 nats, or 2.42 bits per dim
+```
+
+### CIFAR-10
+```bash
+# 39M parameters, trained for ~1M iterations with early stopping (a little less than a week on 2 GPUs)
+wget https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets-2/cifar10-seed0-iter-900000-model-ema.th
+wget https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets-2/cifar10-seed1-iter-1050000-model-ema.th
+wget https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets-2/cifar10-seed2-iter-650000-model-ema.th
+wget https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets-2/cifar10-seed3-iter-1050000-model-ema.th
+python train.py --hps cifar10 --restore_ema_path cifar10-seed0-iter-900000-model-ema.th --test_eval
+python train.py --hps cifar10 --restore_ema_path cifar10-seed1-iter-1050000-model-ema.th --test_eval
+python train.py --hps cifar10 --restore_ema_path cifar10-seed2-iter-650000-model-ema.th --test_eval
+python train.py --hps cifar10 --restore_ema_path cifar10-seed3-iter-1050000-model-ema.th --test_eval
+# seeds 0, 1, 2, 3 should give 2.879, 2.842, 2.898, 2.864 bits per dim, for an average of 2.87 bits per dim.
+```

data.py

@@ -0,0 +1,163 @@
+import numpy as np
+import pickle
+import os
+import torch
+from torch.utils.data import TensorDataset
+from torchvision.datasets import ImageFolder
+import torchvision.transforms as transforms
+from sklearn.model_selection import train_test_split
+
+
+def set_up_data(H):
+    shift_loss = -127.5
+    scale_loss = 1. / 127.5
+    if H.dataset == 'imagenet32':
+        trX, vaX, teX = imagenet32(H.data_root)
+        H.image_size = 32
+        H.image_channels = 3
+        shift = -116.2373
+        scale = 1. / 69.37404
+    elif H.dataset == 'imagenet64':
+        trX, vaX, teX = imagenet64(H.data_root)
+        H.image_size = 64
+        H.image_channels = 3
+        shift = -115.92961967
+        scale = 1. / 69.37404
+    elif H.dataset == 'ffhq_256':
+        trX, vaX, teX = ffhq256(H.data_root)
+        H.image_size = 256
+        H.image_channels = 3
+        shift = -112.8666757481
+        scale = 1. / 69.84780273
+    elif H.dataset == 'ffhq_1024':
+        trX, vaX, teX = ffhq1024(H.data_root)
+        H.image_size = 1024
+        H.image_channels = 3
+        shift = -0.4387
+        scale = 1.0 / 0.2743
+        shift_loss = -0.5
+        scale_loss = 2.0
+    elif H.dataset == 'cifar10':
+        (trX, _), (vaX, _), (teX, _) = cifar10(H.data_root, one_hot=False)
+        H.image_size = 32
+        H.image_channels = 3
+        shift = -120.63838
+        scale = 1. / 64.16736
+    else:
+        raise ValueError('unknown dataset: ', H.dataset)
+
+    do_low_bit = H.dataset in ['ffhq_256']
+
+    if H.test_eval:
+        print('DOING TEST')
+        eval_dataset = teX
+    else:
+        eval_dataset = vaX
+
+    shift = torch.tensor([shift]).cuda().view(1, 1, 1, 1)
+    scale = torch.tensor([scale]).cuda().view(1, 1, 1, 1)
+    shift_loss = torch.tensor([shift_loss]).cuda().view(1, 1, 1, 1)
+    scale_loss = torch.tensor([scale_loss]).cuda().view(1, 1, 1, 1)
+
+    if H.dataset == 'ffhq_1024':
+        train_data = ImageFolder(trX, transforms.ToTensor())
+        valid_data = ImageFolder(eval_dataset, transforms.ToTensor())
+        untranspose = True
+    else:
+        train_data = TensorDataset(torch.as_tensor(trX))
+        valid_data = TensorDataset(torch.as_tensor(eval_dataset))
+        untranspose = False
+
+    def preprocess_func(x):
+        nonlocal shift
+        nonlocal scale
+        nonlocal shift_loss
+        nonlocal scale_loss
+        nonlocal do_low_bit
+        nonlocal untranspose
+        'takes in a data example and returns the preprocessed input'
+        'as well as the input processed for the loss'
+        if untranspose:
+            x[0] = x[0].permute(0, 2, 3, 1)
+        inp = x[0].cuda(non_blocking=True).float()
+        out = inp.clone()
+        inp.add_(shift).mul_(scale)
+        if do_low_bit:
+            # 5 bits of precision
+            out.mul_(1. / 8.).floor_().mul_(8.)
+        out.add_(shift_loss).mul_(scale_loss)
+        return inp, out
+
+    return H, train_data, valid_data, preprocess_func
+
+
+def mkdir_p(path):
+    os.makedirs(path, exist_ok=True)
+
+
+def flatten(outer):
+    return [el for inner in outer for el in inner]
+
+
+def unpickle_cifar10(file):
+    fo = open(file, 'rb')
+    data = pickle.load(fo, encoding='bytes')
+    fo.close()
+    data = dict(zip([k.decode() for k in data.keys()], data.values()))
+    return data
+
+
+def imagenet32(data_root):
+    trX = np.load(os.path.join(data_root, 'imagenet32-train.npy'), mmap_mode='r')
+    np.random.seed(42)
+    tr_va_split_indices = np.random.permutation(trX.shape[0])
+    train = trX[tr_va_split_indices[:-5000]]
+    valid = trX[tr_va_split_indices[-5000:]]
+    test = np.load(os.path.join(data_root, 'imagenet32-valid.npy'), mmap_mode='r')
+    return train, valid, test
+
+
+def imagenet64(data_root):
+    trX = np.load(os.path.join(data_root, 'imagenet64-train.npy'), mmap_mode='r')
+    np.random.seed(42)
+    tr_va_split_indices = np.random.permutation(trX.shape[0])
+    train = trX[tr_va_split_indices[:-5000]]
+    valid = trX[tr_va_split_indices[-5000:]]
+    test = np.load(os.path.join(data_root, 'imagenet64-valid.npy'), mmap_mode='r')  # this is test.
+    return train, valid, test
+
+
+def ffhq1024(data_root):
+    # we did not significantly tune hyperparameters on ffhq-1024, and so simply evaluate on the test set
+    return os.path.join(data_root, 'ffhq1024/train'), os.path.join(data_root, 'ffhq1024/valid'), os.path.join(data_root, 'ffhq1024/valid')
+
+
+def ffhq256(data_root):
+    trX = np.load(os.path.join(data_root, 'ffhq-256.npy'), mmap_mode='r')
+    np.random.seed(5)
+    tr_va_split_indices = np.random.permutation(trX.shape[0])
+    train = trX[tr_va_split_indices[:-7000]]
+    valid = trX[tr_va_split_indices[-7000:]]
+    # we did not significantly tune hyperparameters on ffhq-256, and so simply evaluate on the test set
+    return train, valid, valid
+
+
+def cifar10(data_root, one_hot=True):
+    tr_data = [unpickle_cifar10(os.path.join(data_root, 'cifar-10-batches-py/', 'data_batch_%d' % i)) for i in range(1, 6)]
+    trX = np.vstack(data['data'] for data in tr_data)
+    trY = np.asarray(flatten([data['labels'] for data in tr_data]))
+    te_data = unpickle_cifar10(os.path.join(data_root, 'cifar-10-batches-py/', 'test_batch'))
+    teX = np.asarray(te_data['data'])
+    teY = np.asarray(te_data['labels'])
+    trX = trX.reshape(-1, 3, 32, 32).transpose(0, 2, 3, 1)
+    teX = teX.reshape(-1, 3, 32, 32).transpose(0, 2, 3, 1)
+    trX, vaX, trY, vaY = train_test_split(trX, trY, test_size=5000, random_state=11172018)
+    if one_hot:
+        trY = np.eye(10, dtype=np.float32)[trY]
+        vaY = np.eye(10, dtype=np.float32)[vaY]
+        teY = np.eye(10, dtype=np.float32)[teY]
+    else:
+        trY = np.reshape(trY, [-1, 1])
+        vaY = np.reshape(vaY, [-1, 1])
+        teY = np.reshape(teY, [-1, 1])
+    return (trX, trY), (vaX, vaY), (teX, teY)

files_to_npy.py

@@ -0,0 +1,14 @@
+import sys
+import numpy as np
+import imageio
+import glob
+import os
+
+if __name__ == "__main__":
+    print("moving images in", sys.argv[1], "to", sys.argv[2])
+    files = glob.glob(os.path.join(sys.argv[1], "*.png"))
+    shape = imageio.imread(files[0]).shape
+    data = np.zeros(shape=(len(files), *shape), dtype=np.uint8)
+    for idx, f in enumerate(files):
+        data[idx] = imageio.imread(f)
+    np.save(sys.argv[2], data)