Implementation of NÜWA, state of the art attention network for text to video synthesis, in Pytorch. It also contain an extension into video and audio generation, using a dual decoder approach.
-
March 2022 - seeing signs of life with a difficult version of moving mnist
-
April 2022 - It seems as though a diffusion based method has taken the new throne for SOTA. However, I will continue on with NUWA, extending it to use multi-headed codes + hierarchical causal transformer. I think that direction is untapped for improving on this line of work.
$ pip install nuwa-pytorchFirst train the VAE
import torch
from nuwa_pytorch import VQGanVAE
vae = VQGanVAE(
dim = 512,
channels = 3, # default is 3, but can be changed to any value for the training of the segmentation masks (sketches)
image_size = 256, # image size
num_layers = 4, # number of downsampling layers
num_resnet_blocks = 2, # number of resnet blocks
vq_codebook_size = 8192, # codebook size
vq_decay = 0.8 # codebook exponential decay
)
imgs = torch.randn(10, 3, 256, 256)
# alternate learning for autoencoder ...
loss = vae(imgs, return_loss = True)
loss.backward()
# and the discriminator ...
discr_loss = vae(imgs, return_discr_loss = True)
discr_loss.backward()
# do above for many steps
# return reconstructed images and make sure they look ok
recon_imgs = vae(imgs)Then, with your learned VAE
import torch
from nuwa_pytorch import NUWA, VQGanVAE
# autoencoder
vae = VQGanVAE(
dim = 64,
num_layers = 4,
image_size = 256,
num_conv_blocks = 2,
vq_codebook_size = 8192
)
# NUWA transformer
nuwa = NUWA(
vae = vae,
dim = 512,
text_num_tokens = 20000, # number of text tokens
text_enc_depth = 12, # text encoder depth
text_enc_heads = 8, # number of attention heads for encoder
text_max_seq_len = 256, # max sequence length of text conditioning tokens (keep at 256 as in paper, or shorter, if your text is not that long)
max_video_frames = 10, # number of video frames
image_size = 256, # size of each frame of video
dec_depth = 64, # video decoder depth
dec_heads = 8, # number of attention heads in decoder
dec_reversible = True, # reversible networks - from reformer, decoupling memory usage from depth
enc_reversible = True, # reversible encoders, if you need it
attn_dropout = 0.05, # dropout for attention
ff_dropout = 0.05, # dropout for feedforward
sparse_3dna_kernel_size = (5, 3, 3), # kernel size of the sparse 3dna attention. can be a single value for frame, height, width, or different values (to simulate axial attention, etc)
sparse_3dna_dilation = (1, 2, 4), # cycle dilation of 3d conv attention in decoder, for more range
shift_video_tokens = True # cheap relative positions for sparse 3dna transformer, by shifting along spatial dimensions by one
).cuda()
# data
text = torch.randint(0, 20000, (1, 256)).cuda()
video = torch.randn(1, 10, 3, 256, 256).cuda() # (batch, frames, channels, height, width)
loss = nuwa(
text = text,
video = video,
return_loss = True # set this to True, only for training, to return cross entropy loss
)
loss.backward()
# do above with as much data as possible
# then you can generate a video from text
video = nuwa.generate(text = text, num_frames = 5) # (1, 5, 3, 256, 256)In the paper, they also present a way to condition the video generation based on segmentation mask(s). You can easily do this as well, given you train a VQGanVAE on the sketches before hand.
Then, you will use NUWASketch instead of NUWA, which can accept the sketch VAE as a reference
ex.
import torch
from nuwa_pytorch import NUWASketch, VQGanVAE
# autoencoder, one for main video, the other for the sketch
vae = VQGanVAE(
dim = 64,
num_layers = 4,
image_size = 256,
num_conv_blocks = 2,
vq_codebook_size = 8192
)
sketch_vae = VQGanVAE(
dim = 512,
channels = 5, # say the sketch has 5 classes
num_layers = 4,
image_size = 256,
num_conv_blocks = 2,
vq_codebook_size = 8192
)
# NUWA transformer for conditioning with sketches
nuwa = NUWASketch(
vae = vae,
sketch_vae = sketch_vae,
dim = 512, # model dimensions
sketch_enc_depth = 12, # sketch encoder depth
sketch_enc_heads = 8, # number of attention heads for sketch encoder
sketch_max_video_frames = 3, # max number of frames for sketches
sketch_enc_use_sparse_3dna = True, # whether to use 3d-nearby attention (of full attention if False) for sketch encoding transformer
max_video_frames = 10, # number of video frames
image_size = 256, # size of each frame of video
dec_depth = 64, # video decoder depth
dec_heads = 8, # number of attention heads in decoder
dec_reversible = True, # reversible networks - from reformer, decoupling memory usage from depth
enc_reversible = True, # reversible encoders, if you need it
attn_dropout = 0.05, # dropout for attention
ff_dropout = 0.05, # dropout for feedforward
sparse_3dna_kernel_size = (5, 3, 3), # kernel size of the sparse 3dna attention. can be a single value for frame, height, width, or different values (to simulate axial attention, etc)
sparse_3dna_dilation = (1, 2, 4), # cycle dilation of 3d conv attention in decoder, for more range
cross_2dna_kernel_size = 5, # 2d kernel size of spatial grouping of attention from video frames to sketches
cross_2dna_dilation = 1, # 2d dilation of spatial attention from video frames to sketches
shift_video_tokens = True # cheap relative positions for sparse 3dna transformer, by shifting along spatial dimensions by one
).cuda()
# data
sketch = torch.randn(2, 2, 5, 256, 256).cuda() # (batch, frames, segmentation classes, height, width)
sketch_mask = torch.ones(2, 2).bool().cuda() # (batch, frames) [Optional]
video = torch.randn(2, 10, 3, 256, 256).cuda() # (batch, frames, channels, height, width)
loss = nuwa(
sketch = sketch,
sketch_mask =sketch_mask,
video = video,
return_loss = True # set this to True, only for training, to return cross entropy loss
)
loss.backward()
# do above with as much data as possible
# then you can generate a video from sketch(es)
video = nuwa.generate(sketch = sketch, num_frames = 5) # (1, 5, 3, 256, 256)This repository will also offer a variant of NUWA that can produce both video and audio. For now, the audio will need to be encoded manually.
import torch
from nuwa_pytorch import NUWAVideoAudio, VQGanVAE
# autoencoder
vae = VQGanVAE(
dim = 64,
num_layers = 4,
image_size = 256,
num_conv_blocks = 2,
vq_codebook_size = 100
)
# NUWA transformer
nuwa = NUWAVideoAudio(
vae = vae,
dim = 512,
num_audio_tokens = 2048, # codebook size for audio tokens
num_audio_tokens_per_video_frame = 32, # number of audio tokens per video frame
cross_modality_attn_every = 3, # cross modality attention every N layers
text_num_tokens = 20000, # number of text tokens
text_enc_depth = 1, # text encoder depth
text_enc_heads = 8, # number of attention heads for encoder
text_max_seq_len = 256, # max sequence length of text conditioning tokens (keep at 256 as in paper, or shorter, if your text is not that long)
max_video_frames = 10, # number of video frames
image_size = 256, # size of each frame of video
dec_depth = 4, # video decoder depth
dec_heads = 8, # number of attention heads in decoder
enc_reversible = True, # reversible encoders, if you need it
dec_reversible = True, # quad-branched reversible network, for making depth of twin video / audio decoder independent of network depth. recommended to be turned on unless you have a ton of memory at your disposal
attn_dropout = 0.05, # dropout for attention
ff_dropout = 0.05, # dropout for feedforward
sparse_3dna_kernel_size = (5, 3, 3), # kernel size of the sparse 3dna attention. can be a single value for frame, height, width, or different values (to simulate axial attention, etc)
sparse_3dna_dilation = (1, 2, 4), # cycle dilation of 3d conv attention in decoder, for more range
shift_video_tokens = True # cheap relative positions for sparse 3dna transformer, by shifting along spatial dimensions by one
).cuda()
# data
text = torch.randint(0, 20000, (1, 256)).cuda()
audio = torch.randint(0, 2048, (1, 32 * 10)).cuda() # (batch, audio tokens per frame * max video frames)
video = torch.randn(1, 10, 3, 256, 256).cuda() # (batch, frames, channels, height, width)
loss = nuwa(
text = text,
video = video,
audio = audio,
return_loss = True # set this to True, only for training, to return cross entropy loss
)
loss.backward()
# do above with as much data as possible
# then you can generate a video from text
video, audio = nuwa.generate(text = text, num_frames = 5) # (1, 5, 3, 256, 256), (1, 32 * 5 == 160)This library will offer some utilities to make training easier. For starters, you can use the VQGanVAETrainer class to take care of training the VQGanVAE. Simply wrap the model and also pass in the image folder path as well as the various training hyperparameters.
import torch
from nuwa_pytorch import VQGanVAE, VQGanVAETrainer
vae = VQGanVAE(
dim = 64,
image_size = 256,
num_layers = 5,
vq_codebook_size = 1024,
vq_use_cosine_sim = True,
vq_codebook_dim = 32,
vq_orthogonal_reg_weight = 10,
vq_orthogonal_reg_max_codes = 128,
).cuda()
trainer = VQGanVAETrainer(
vae, # VAE defined above
folder ='/path/to/images', # path to images
lr = 3e-4, # learning rate
num_train_steps = 100000, # number of training steps
batch_size = 8, # batch size
grad_accum_every = 4 # gradient accumulation (effective batch size is (batch_size x grad_accum_every))
)
trainer.train()
# results and model checkpoints will be saved periodically to ./resultsTo train NUWA, first you need to organize a folder of .gif files with corresponding .txt files containing its caption. It should be organized as such.
ex.
📂video-and-text-data
┣ 📜cat.gif
┣ 📜cat.txt
┣ 📜dog.gif
┣ 📜dog.txt
┣ 📜turtle.gif
┗ 📜turtle.txt
Then you will load your previously trained VQGan-VAE and train NUWA with the GifVideoDataset and NUWATrainer classes.
import torch
from nuwa_pytorch import NUWA, VQGanVAE
from nuwa_pytorch.train_nuwa import GifVideoDataset, NUWATrainer
# dataset
ds = GifVideoDataset(
folder = './path/to/videos/',
channels = 1
)
# autoencoder
vae = VQGanVAE(
dim = 64,
image_size = 256,
num_layers = 5,
num_resnet_blocks = 2,
vq_codebook_size = 512,
attn_dropout = 0.1
)
vae.load_state_dict(torch.load('./path/to/trained/vae.pt'))
# NUWA transformer
nuwa = NUWA(
vae = vae,
dim = 512,
text_enc_depth = 6,
text_max_seq_len = 256,
max_video_frames = 10,
dec_depth = 12,
dec_reversible = True,
enc_reversible = True,
attn_dropout = 0.05,
ff_dropout = 0.05,
sparse_3dna_kernel_size = (5, 3, 3),
sparse_3dna_dilation = (1, 2, 4),
shift_video_tokens = True
).cuda()
# data
trainer = NUWATrainer(
nuwa = nuwa, # NUWA transformer
dataset = dataset, # video dataset class
num_train_steps = 1000000, # number of training steps
lr = 3e-4, # learning rate
wd = 0.01, # weight decay
batch_size = 8, # batch size
grad_accum_every = 4, # gradient accumulation
max_grad_norm = 0.5, # gradient clipping
num_sampled_frames = 10, # number of frames to sample
results_folder = './results' # folder to store checkpoints and samples
)
trainer.train()This library depends on this vector quantization library, which comes with a number of improvements (improved vqgan, orthogonal codebook regularization, etc). To use any of these improvements, you can configure the vector quantizer keyword params by prepending vq_ on VQGanVAE initialization.
ex. cosine sim proposed in improved vqgan
from nuwa_pytorch import VQGanVAE
vae = VQGanVAE(
dim = 64,
image_size = 256,
num_layers = 4,
vq_use_cosine_sim = True
# VectorQuantize will be initialized with use_cosine_sim = True
# https://github.com/lucidrains/vector-quantize-pytorch#cosine-similarity
).cuda()- complete 3dna causal attention in decoder
- write up easy generation functions
- make sure GAN portion of VQGan is correct, reread paper
- make sure adaptive weight in vqgan is correctly built
- offer new vqvae improvements (orthogonal reg and smaller codebook dimensions)
- batch video tokens -> vae during video generation, to prevent oom
- query chunking in 3dna attention, to put a cap on peak memory
- flesh out VAE resnet blocks, offer some choices
- add all stability tricks from cogview paper by default
- make VQGan able to accept custom VGG for LPAPs loss (audio)
- add feedforward chunking
- add shift token in decoder for cheap powerful RPE
- add reversible networks, to save on memory on depth
- support kernel sizes different along each dimension for sparse 3dna
- add some autotrainer that takes care of the alternating updates of discriminator and VQVAE generator
- segmentation mask encoder, make sure embeddings can undergo 3dna attention with decoder during cross attention
- finish 2d-nearby cross attention for sketches
- able to add convnext blocks to other layers in vqgan vae
- offer vqvae training script
- handle variable lengthed sketches, accept a mask on the sketch frames dimension
- take care of audio transformer and cross modality attention
- add audio transformer, and build audio / video nearby cross attention
- make dual decoder reversible
- rotary embeddings for encoder
- add cycle dilation to audio
- omit vgg from VAE state dict
- add cosine sim attention from swinv2 as an option
- add axial positional embedding to audio
- Triton kernel for 3dna attention
- offer a colab with moving mnist example, conditioned on present digits
- build NUWA controller class that can accept text or sketch
- key masking for 3dna attention - for variable sketch length masking
- figure out spec vqgan and fit it into the framework, take care of audio encoding / decoding automatically
- turn into CLI tool, like stylegan2-pytorch
- look into integrating https://github.com/lucidrains/RQ-Transformer for both video and audio
- inference caching
@misc{wu2021nuwa,
title = {N\"UWA: Visual Synthesis Pre-training for Neural visUal World creAtion},
author = {Chenfei Wu and Jian Liang and Lei Ji and Fan Yang and Yuejian Fang and Daxin Jiang and Nan Duan},
year = {2021},
eprint = {2111.12417},
archivePrefix = {arXiv},
primaryClass = {cs.CV}
}@misc{esser2021taming,
title = {Taming Transformers for High-Resolution Image Synthesis},
author = {Patrick Esser and Robin Rombach and Björn Ommer},
year = {2021},
eprint = {2012.09841},
archivePrefix = {arXiv},
primaryClass = {cs.CV}
}@misc{iashin2021taming,
title = {Taming Visually Guided Sound Generation},
author = {Vladimir Iashin and Esa Rahtu},
year = {2021},
eprint = {2110.08791},
archivePrefix = {arXiv},
primaryClass = {cs.CV}
}@misc{ding2021cogview,
title = {CogView: Mastering Text-to-Image Generation via Transformers},
author = {Ming Ding and Zhuoyi Yang and Wenyi Hong and Wendi Zheng and Chang Zhou and Da Yin and Junyang Lin and Xu Zou and Zhou Shao and Hongxia Yang and Jie Tang},
year = {2021},
eprint = {2105.13290},
archivePrefix = {arXiv},
primaryClass = {cs.CV}
}@misc{kitaev2020reformer,
title = {Reformer: The Efficient Transformer},
author = {Nikita Kitaev and Łukasz Kaiser and Anselm Levskaya},
year = {2020},
eprint = {2001.04451},
archivePrefix = {arXiv},
primaryClass = {cs.LG}
}@misc{shazeer2020talkingheads,
title = {Talking-Heads Attention},
author = {Noam Shazeer and Zhenzhong Lan and Youlong Cheng and Nan Ding and Le Hou},
year = {2020},
eprint = {2003.02436},
archivePrefix = {arXiv},
primaryClass = {cs.LG}
}@misc{shazeer2020glu,
title = {GLU Variants Improve Transformer},
author = {Noam Shazeer},
year = {2020},
url = {https://arxiv.org/abs/2002.05202}
}@misc{su2021roformer,
title = {RoFormer: Enhanced Transformer with Rotary Position Embedding},
author = {Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu},
year = {2021},
eprint = {2104.09864},
archivePrefix = {arXiv},
primaryClass = {cs.CL}
}@inproceedings{ho2021classifierfree,
title = {Classifier-Free Diffusion Guidance},
author = {Jonathan Ho and Tim Salimans},
booktitle = {NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications},
year = {2021},
url = {https://openreview.net/forum?id=qw8AKxfYbI}
}@misc{liu2021swin,
title = {Swin Transformer V2: Scaling Up Capacity and Resolution},
author = {Ze Liu and Han Hu and Yutong Lin and Zhuliang Yao and Zhenda Xie and Yixuan Wei and Jia Ning and Yue Cao and Zheng Zhang and Li Dong and Furu Wei and Baining Guo},
year = {2021},
eprint = {2111.09883},
archivePrefix = {arXiv},
primaryClass = {cs.CV}
}@misc{crowson2022,
author = {Katherine Crowson},
url = {https://twitter.com/RiversHaveWings/status/1478093658716966912}
}Attention is the rarest and purest form of generosity. - Simone Weil