16 KiB
Repo & Config Structure
Repo Structure
Open-Sora
├── README.md
├── docs
│ ├── acceleration.md -> Acceleration & Speed benchmark
│ ├── command.md -> Commands for training & inference
│ ├── datasets.md -> Datasets used in this project
│ ├── structure.md -> This file
│ └── report_v1.md -> Report for Open-Sora v1
├── scripts
│ ├── train.py -> diffusion training script
│ └── inference.py -> Report for Open-Sora v1
├── configs -> Configs for training & inference
├── opensora
│ ├── __init__.py
│ ├── registry.py -> Registry helper
│ ├── acceleration -> Acceleration related code
│ ├── dataset -> Dataset related code
│ ├── models
│ │ ├── layers -> Common layers
│ │ ├── vae -> VAE as image encoder
│ │ ├── text_encoder -> Text encoder
│ │ │ ├── classes.py -> Class id encoder (inference only)
│ │ │ ├── clip.py -> CLIP encoder
│ │ │ └── t5.py -> T5 encoder
│ │ ├── dit
│ │ ├── latte
│ │ ├── pixart
│ │ └── stdit -> Our STDiT related code
│ ├── schedulers -> Diffusion schedulers
│ │ ├── iddpm -> IDDPM for training and inference
│ │ └── dpms -> DPM-Solver for fast inference
│ └── utils
└── tools -> Tools for data processing and more
Configs
Our config files follows MMEgine. MMEngine will reads the config file (a .py file) and parse it into a dictionary-like object.
Open-Sora
└── configs -> Configs for training & inference
├── opensora -> STDiT related configs
│ ├── inference
│ │ ├── 16x256x256.py -> Sample videos 16 frames 256x256
│ │ ├── 16x512x512.py -> Sample videos 16 frames 512x512
│ │ └── 64x512x512.py -> Sample videos 64 frames 512x512
│ └── train
│ ├── 16x256x256.py -> Train on videos 16 frames 256x256
│ ├── 16x256x256.py -> Train on videos 16 frames 256x256
│ └── 64x512x512.py -> Train on videos 64 frames 512x512
├── dit -> DiT related configs
│ ├── inference
│ │ ├── 1x256x256-class.py -> Sample images with ckpts from DiT
│ │ ├── 1x256x256.py -> Sample images with clip condition
│ │ └── 16x256x256.py -> Sample videos
│ └── train
│ ├── 1x256x256.py -> Train on images with clip condition
│ └── 16x256x256.py -> Train on videos
├── latte -> Latte related configs
└── pixart -> PixArt related configs
Inference config demos
To change the inference settings, you can directly modify the corresponding config file. Or you can pass arguments to overwrite the config file (config_utils.py). To change sampling prompts, you should modify the .txt file passed to the --prompt_path argument.
--prompt_path ./assets/texts/t2v_samples.txt -> prompt_path
--ckpt-path ./path/to/your/ckpt.pth -> model["from_pretrained"]
The explanation of each field is provided below.
# Define sampling size
num_frames = 64 # number of frames
fps = 24 // 2 # frames per second (divided by 2 for frame_interval=2)
image_size = (512, 512) # image size (height, width)
# Define model
model = dict(
type="STDiT-XL/2", # Select model type (STDiT-XL/2, DiT-XL/2, etc.)
space_scale=1.0, # (Optional) Space positional encoding scale (new height / old height)
time_scale=2 / 3, # (Optional) Time positional encoding scale (new frame_interval / old frame_interval)
enable_flashattn=True, # (Optional) Speed up training and inference with flash attention
enable_layernorm_kernel=True, # (Optional) Speed up training and inference with fused kernel
from_pretrained="PRETRAINED_MODEL", # (Optional) Load from pretrained model
no_temporal_pos_emb=True, # (Optional) Disable temporal positional encoding (for image)
)
vae = dict(
type="VideoAutoencoderKL", # Select VAE type
from_pretrained="stabilityai/sd-vae-ft-ema", # Load from pretrained VAE
micro_batch_size=128, # VAE with micro batch size to save memory
)
text_encoder = dict(
type="t5", # Select text encoder type (t5, clip)
from_pretrained="DeepFloyd/t5-v1_1-xxl", # Load from pretrained text encoder
model_max_length=120, # Maximum length of input text
)
scheduler = dict(
type="iddpm", # Select scheduler type (iddpm, dpm-solver)
num_sampling_steps=100, # Number of sampling steps
cfg_scale=7.0, # hyper-parameter for classifier-free diffusion
cfg_channel=3, # how many channels to use for classifier-free diffusion, if None, use all channels
)
dtype = "fp16" # Computation type (fp16, fp32, bf16)
# Other settings
batch_size = 1 # batch size
seed = 42 # random seed
prompt_path = "./assets/texts/t2v_samples.txt" # path to prompt file
save_dir = "./samples" # path to save samples
Training config demos
# Define sampling size
num_frames = 64
frame_interval = 2 # sample every 2 frames
image_size = (512, 512)
# Define dataset
root = None # root path to the dataset
data_path = "CSV_PATH" # path to the csv file
use_image_transform = False # True if training on images
num_workers = 4 # number of workers for dataloader
# Define acceleration
dtype = "bf16" # Computation type (fp16, bf16)
grad_checkpoint = True # Use gradient checkpointing
plugin = "zero2" # Plugin for distributed training (zero2, zero2-seq)
sp_size = 1 # Sequence parallelism size (1 for no sequence parallelism)
# Define model
model = dict(
type="STDiT-XL/2",
space_scale=1.0,
time_scale=2 / 3,
from_pretrained="YOUR_PRETRAINED_MODEL",
enable_flashattn=True, # Enable flash attention
enable_layernorm_kernel=True, # Enable layernorm kernel
)
vae = dict(
type="VideoAutoencoderKL",
from_pretrained="stabilityai/sd-vae-ft-ema",
micro_batch_size=128,
)
text_encoder = dict(
type="t5",
from_pretrained="DeepFloyd/t5-v1_1-xxl",
model_max_length=120,
shardformer=True, # Enable shardformer for T5 acceleration
)
scheduler = dict(
type="iddpm",
timestep_respacing="", # Default 1000 timesteps
)
# Others
seed = 42
outputs = "outputs" # path to save checkpoints
wandb = False # Use wandb for logging
epochs = 1000 # number of epochs (just large enough, kill when satisfied)
log_every = 10
ckpt_every = 250
load = None # path to resume training
batch_size = 4
lr = 2e-5
grad_clip = 1.0 # gradient clipping
Inference-long specific arguments
The inference-long.py script is used to generate long videos, and it also provides all functions of the inference.py script. The following arguments are specific to the inference-long.py script.
loop = 10
condition_frame_length = 4
reference_path = ["one.png;two.mp4"]
mask_strategy = ["0,0,0,1,0;0,0,0,1,-1"]
To generate a long video of any time, our strategy is to generate a video with a fixed length first, and then use the last condition_frame_length number of frames for the next video generation. This will loop for loop times. Thus, the total length of the video is loop * (num_frames - condition_frame_length) + condition_frame_length.
To condition the generation on images or videos, we introduce the mask_strategy. It is 5 number tuples separated by ;. Each tuple indicate an insertion of the condition image or video to the target generation. The meaning of each number is:
- First number: the index of the condition image or video in the
reference_path. (0 means one.png, and 1 means two.mp4) - Second number: the loop index of the condition image or video. (0 means the first loop, 1 means the second loop, etc.)
- Third number: the start frame of the condition image or video. (0 means the first frame, and images only have one frame)
- Fourth number: the number of frames to insert. (1 means insert one frame, and images only have one frame)
- Fifth number: the location to insert. (0 means insert at the beginning, 1 means insert at the end, and -1 means insert at the end of the video)
Thus, "0,0,0,1,-1" means insert the first frame of one.png at the end of the video at the first loop.
Bucket Configs
To enable dynamic training (for STDiT2), use VariableVideoText dataset, and set the bucket_config in the config. An example is:
bucket_config = {
"240p": {16: (1.0, 16), 32: (1.0, 8), 64: (1.0, 4), 128: (1.0, 2)},
"256": {1: (1.0, 256)},
"512": {1: (1.0, 80)},
"480p": {1: (1.0, 52), 16: (0.5, 4), 32: (0.0, None)},
"720p": {16: (1.0, 2), 32: (0.0, None)},
"1024": {1: (1.0, 20)},
"1080p": {1: (1.0, 8)},
}
This looks a bit difficult to understand at the first glance. Let's understand this config step by step.
Why bucket?
Dynamic training needs to support training on different resolution (HxW), different aspect ratio (H/W), and different frame length. There are several possible ways to achieve this:
- NaViT: support dynamic size within the same batch by masking, without efficiency loss. However, the system is a bit complex to implement, and may not benefit from optimized kernels such as flash attention.
- FiT: support dynamic size within the same batch by padding. However, padding different resolutions to the same size is not efficient.
- PixArt: support dynamic size in different batches by bucketing, but the size must be the same within the same batch, and only a fixed number of size can be applied. With the same size in a batch, we do not need to implement complex masking or padding.
Bucketing means we pre-define some fixed resolution, and allocate different samples to different bucket. The concern for bucketing is listed below. But we can see that the concern is not a big issue in our case.
- The bucket size is limited to a fixed number: First, in real-world applications, only a few aspect ratios (9:16, 3:4) and resolutions (240p, 1080p) are commonly used. Second, we find trained models can generalize well to unseen resolutions.
- The size in each batch is the same, breaks the i.i.d. assumption: Since we are using multiple GPUs, the local batches on different GPUs have different sizes. We did not see a significant performance drop due to this issue.
- The may not be enough samples to fill each bucket and the distribution may be biased: First, our dataset is large enough to fill each bucket when local batch size is not too large. Second, we should analyze the data's distribution on sizes and define the bucket size accordingly. Third, an unbalanced distribution did not affect the training process significantly.
- Different resolutions and frame lengths may have different processing speed: Different from PixArt, which only deals with aspect ratios of similar resolutions (similar token numbers), we need to consider the processing speed of different resolutions and frame lengths. We can use the
bucket_configto define the batch size for each bucket to ensure the processing speed is similar.
Three-level bucket
We design a three-level bucket: (resolution, num_frames, aspect_ratios). The resolution and aspect ratios is predefined in aspect.py. Commonly used resolutions (e.g., 240p, 1080p) are supported, and the name represents the number of pixels (e.g., 240p is 240x426, however, we define 240p to represent any size with HxW approximately 240x426=102240 pixels). The aspect ratios are defined for each resolution. You do not need to define the aspect ratios in the bucket_config.
The num_frames is the number of frames in each sample, with num_frames=1 especially for images. If frame_intervals is not 1, a bucket with num_frames=k will contain videos with k*frame_intervals frames except for images. Only a video with more than num_frames and more than resolution pixels will be likely to be put into the bucket.
The two number defined in the bucket config is (keep_prob, batch_size). Since the memory and speed of samples from different buckets may be different, we use batch_size to balance the processing speed. Since our computation is limited, we cannot process videos with their original resolution as stated in OpenAI's sora's report. Thus, we give a keep_prob to control the number of samples in each bucket. The keep_prob is the probability to keep a sample in the bucket. Let's take the following config as an example:
bucket_config = {
"480p": {16: (1.0, 8),},
"720p": {16: (0.5, 4),},
"1080p": {16: (0.2, 2)},
"4K", {16: (0.1, 1)},
}
Given a 2K video with more than 16 frames, the program will first try to put it into bucket "1080p" since it has a larger resolution than 1080p but less than 4K. Since the keep_prob for 1080p is 20%, a random number is generated, and if it is less than 0.2, the video will be put into the bucket. If the video is not put into the bucket, the program will try to put it into the "720p" bucket. Since the keep_prob for 720p is 50%, the video has a 50% chance to be put into the bucket. If the video is not put into the bucket, the program will try to put it into the "480p" bucket directly as it is the smallest resolution.
Examples
Let's see some simple examples to understand the bucket config. First, the aspect ratio bucket is compulsory, if you want to modify this you need to add your own resolution definition in aspect.py. Then, to keep only 256x256 resolution and 16 frames as OpenSora 1.0, you can use the following config:
bucket_config = {
"256": {16: (1.0, 8)},
}
If you want to train a model supporting different resolutions of images, you can use the following config:
bucket_config = {
"256": {1: (1.0, 256)},
"512": {1: (1.0, 80)},
"480p": {1: (1.0, 52)},
"1024": {1: (1.0, 20)},
"1080p": {1: (1.0, 8)},
}
Or if you find the number of high-resolution images is too large, you can modify the keep_prob to reduce the number of samples in the bucket:
bucket_config = {
"256": {1: (1.0, 256)},
"512": {1: (0.8, 80)},
"480p": {1: (0.5, 52)},
"1024": {1: (0.5, 20)},
"1080p": {1: (0.2, 8)},
}
And similarly for videos:
bucket_config = {
"240p": {16: (1.0, 16), 32: (1.0, 8), 64: (1.0, 4), 128: (1.0, 2)},
"480p": {16: (1.0, 4)},
"720p": {16: (0.5, 2)},
}
Note that in the above case, a video with 480p resolution and more than 16 frames will all go into bucket ("480p", 16), since they all satisfy this bucket's requirement. But training long videos with 480p resolution may be slow, so you can modify the config as follows to enforce the video with more than 32 frames to go into the 240p bucket.
bucket_config = {
"240p": {16: (1.0, 16), 32: (1.0, 8), 64: (1.0, 4), 128: (1.0, 2)},
"480p": {16: (1.0, 4), 32: (0.0, None)},
"720p": {16: (0.5, 2)},
}
Combine the above examples together, we think you can understand the bucket config provided at the beginning of this section and in the config files.