Files
olivame a90efc6718 VAE 也做 BF16
这个权重不做修改更好精度
2026-01-18 21:14:55 +08:00

2563 lines
102 KiB
Python

"""
wild mixture of
https://github.com/openai/improved-diffusion/blob/e94489283bb876ac1477d5dd7709bbbd2d9902ce/improved_diffusion/gaussian_diffusion.py
https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
https://github.com/CompVis/taming-transformers
-- merci
"""
import random
import torch
import torch.nn as nn
import copy
import numpy as np
import pytorch_lightning as pl
import torch.nn.functional as F
import logging
mainlogger = logging.getLogger('mainlogger')
from functools import partial
from contextlib import contextmanager
from tqdm import tqdm
from einops import rearrange, repeat, reduce
from torch.optim.lr_scheduler import LambdaLR, CosineAnnealingLR
from torchvision.utils import make_grid
from pytorch_lightning.utilities import rank_zero_only
from omegaconf import OmegaConf
from typing import Optional, Sequence, Any, Tuple, Union, List, Dict
from collections.abc import Mapping, Iterable, Callable
from torch import Tensor
from unifolm_wma.utils.utils import instantiate_from_config
from unifolm_wma.utils.ema import LitEma
from unifolm_wma.utils.distributions import DiagonalGaussianDistribution
from unifolm_wma.utils.diffusion import make_beta_schedule, rescale_zero_terminal_snr
from unifolm_wma.utils.basics import disabled_train
from unifolm_wma.utils.common import (extract_into_tensor, noise_like, exists,
default)
from unifolm_wma.models.samplers.ddim import DDIMSampler
from unifolm_wma.models.diffusion_head.common.lr_scheduler import get_scheduler, SelectiveLRScheduler
from unifolm_wma.models.diffusion_head.ema_model import EMAModel
from unifolm_wma.models.diffusion_head.positional_embedding import SinusoidalPosEmb
from unifolm_wma.modules.encoders.condition import MLPProjector
from unifolm_wma.data.normolize import Normalize, Unnormalize
__conditioning_keys__ = {
'concat': 'c_concat',
'crossattn': 'c_crossattn',
'adm': 'y'
}
class DDPM(pl.LightningModule):
"""
Denoising Diffusion Probabilistic Model (DDPM) LightningModule.
"""
def __init__(
self,
wma_config: OmegaConf,
timesteps: int = 1000,
beta_schedule: str = "linear",
loss_type: str = "l2",
ckpt_path: Optional[str] = None,
ignore_keys: Optional[Sequence[str]] = [],
load_only_unet: bool = False,
monitor: str = None,
use_ema: bool = True,
first_stage_key: str = "image",
image_size: int = 256,
channels: int = 3,
log_every_t: int = 100,
clip_denoised: bool = True,
linear_start: float = 1e-4,
linear_end: float = 2e-2,
cosine_s: float = 8e-3,
given_betas: Optional[np.ndarray] = None,
original_elbo_weight: float = 0.0,
v_posterior: float = 0.0,
l_simple_weight: float = 1.0,
conditioning_key: Optional[str] = None,
parameterization: str = "eps",
scheduler_config: Optional[Mapping[str, Any]] = None,
use_positional_encodings: bool = False,
learn_logvar: bool = False,
logvar_init: float = 0.0,
rescale_betas_zero_snr: bool = False,
):
"""
wma_config: Config object used to build the underlying model.
timesteps: Number of diffusion steps.
beta_schedule: Schedule type for betas (e.g., 'linear', 'cosine').
loss_type: Loss type.
ckpt_path: Optional checkpoint path to restore weights.
ignore_keys: Keys to ignore when loading the checkpoint.
load_only_unet: If True, load the backbone into self.model only.
monitor: Metric key for monitoring.
use_ema: If True, maintain EMA weights.
first_stage_key: Key in batch dict for inputs.
image_size: Image size.
channels: Number of channels.
log_every_t: Interval of timesteps to log intermediates during sampling.
clip_denoised: Clamp x_0 predictions or not.
linear_start: Linear schedule start.
linear_end: Linear schedule end.
cosine_s: Cosine schedule s parameter.
given_betas: Externally provided betas; overrides schedule if set.
original_elbo_weight: Weight for VLB term.
v_posterior: Interpolation weight for posterior variance.
l_simple_weight: Weight for simple loss term.
conditioning_key: Conditioning mechanism key (if used by wrapper).
parameterization: One of {'eps','x0','v'}.
scheduler_config: Optional LR scheduler config.
use_positional_encodings: Whether to inject positional encodings.
learn_logvar: If True, learn per-timestep log-variance.
logvar_init: Initial value for log-variance.
rescale_betas_zero_snr: If True, apply zero-SNR rescaling to betas.
"""
super().__init__()
assert parameterization in [
"eps", "x0", "v"
], 'currently only supporting "eps" and "x0" and "v"'
self.parameterization = parameterization
mainlogger.info(
f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode"
)
self.cond_stage_model = None
self.clip_denoised = clip_denoised
self.log_every_t = log_every_t
self.first_stage_key = first_stage_key
self.channels = channels
self.temporal_length = wma_config.params.temporal_length
self.image_size = image_size
if isinstance(self.image_size, int):
self.image_size = [self.image_size, self.image_size]
self.use_positional_encodings = use_positional_encodings
self.model = DiffusionWrapper(wma_config, conditioning_key)
self.use_ema = use_ema
self.rescale_betas_zero_snr = rescale_betas_zero_snr
if self.use_ema:
self.model_ema = LitEma(self.model)
mainlogger.info(
f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
self.v_posterior = v_posterior
self.original_elbo_weight = original_elbo_weight
self.l_simple_weight = l_simple_weight
if monitor is not None:
self.monitor = monitor
if ckpt_path is not None:
self.init_from_ckpt(ckpt_path,
ignore_keys=ignore_keys,
only_model=load_only_unet)
self.register_schedule(given_betas=given_betas,
beta_schedule=beta_schedule,
timesteps=timesteps,
linear_start=linear_start,
linear_end=linear_end,
cosine_s=cosine_s)
# For reschedule
self.given_betas = given_betas
self.beta_schedule = beta_schedule
self.timesteps = timesteps
self.cosine_s = cosine_s
self.loss_type = loss_type
self.learn_logvar = learn_logvar
self.logvar = torch.full(fill_value=logvar_init,
size=(self.num_timesteps, ))
if self.learn_logvar:
self.logvar = nn.Parameter(self.logvar, requires_grad=True)
def register_schedule(self,
given_betas: Optional[np.ndarray] = None,
beta_schedule: str = "linear",
timesteps: int = 1000,
linear_start: float = 1e-4,
linear_end: float = 2e-2,
cosine_s: float = 8e-3) -> None:
"""
Create and register diffusion buffers (betas, alphas, posteriors, weights).
Args:
given_betas: If provided, use these instead of building a schedule.
beta_schedule: Name of schedule to create if betas not given.
timesteps: Number of diffusion steps.
linear_start: Linear schedule start.
linear_end: Linear schedule end.
cosine_s: Cosine schedule parameter
"""
if exists(given_betas):
betas = given_betas
else:
betas = make_beta_schedule(beta_schedule,
timesteps,
linear_start=linear_start,
linear_end=linear_end,
cosine_s=cosine_s)
if self.rescale_betas_zero_snr:
betas = rescale_zero_terminal_snr(betas)
alphas = 1. - betas
alphas_cumprod = np.cumprod(alphas, axis=0)
alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
timesteps, = betas.shape
self.num_timesteps = int(timesteps)
self.linear_start = linear_start
self.linear_end = linear_end
assert alphas_cumprod.shape[
0] == self.num_timesteps, 'alphas have to be defined for each timestep'
to_torch = partial(torch.tensor, dtype=torch.float32)
self.register_buffer('betas', to_torch(betas))
self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
self.register_buffer('alphas_cumprod_prev',
to_torch(alphas_cumprod_prev))
self.register_buffer('sqrt_alphas_cumprod',
to_torch(np.sqrt(alphas_cumprod)))
self.register_buffer('sqrt_one_minus_alphas_cumprod',
to_torch(np.sqrt(1. - alphas_cumprod)))
self.register_buffer('log_one_minus_alphas_cumprod',
to_torch(np.log(1. - alphas_cumprod)))
if self.parameterization != 'v':
self.register_buffer('sqrt_recip_alphas_cumprod',
to_torch(np.sqrt(1. / alphas_cumprod)))
self.register_buffer('sqrt_recipm1_alphas_cumprod',
to_torch(np.sqrt(1. / alphas_cumprod - 1)))
else:
self.register_buffer('sqrt_recip_alphas_cumprod',
torch.zeros_like(to_torch(alphas_cumprod)))
self.register_buffer('sqrt_recipm1_alphas_cumprod',
torch.zeros_like(to_torch(alphas_cumprod)))
posterior_variance = (1 - self.v_posterior) * betas * (
1. - alphas_cumprod_prev) / (
1. - alphas_cumprod) + self.v_posterior * betas
# Above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t)
self.register_buffer('posterior_variance',
to_torch(posterior_variance))
# Below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain
self.register_buffer(
'posterior_log_variance_clipped',
to_torch(np.log(np.maximum(posterior_variance, 1e-20))))
self.register_buffer(
'posterior_mean_coef1',
to_torch(betas * np.sqrt(alphas_cumprod_prev) /
(1. - alphas_cumprod)))
self.register_buffer(
'posterior_mean_coef2',
to_torch((1. - alphas_cumprod_prev) * np.sqrt(alphas) /
(1. - alphas_cumprod)))
if self.parameterization == "eps":
lvlb_weights = self.betas**2 / (2 * self.posterior_variance *
to_torch(alphas) *
(1 - self.alphas_cumprod))
elif self.parameterization == "x0":
lvlb_weights = 0.5 * np.sqrt(torch.Tensor(alphas_cumprod)) / (
2. * 1 - torch.Tensor(alphas_cumprod))
elif self.parameterization == "v":
lvlb_weights = torch.ones_like(
self.betas**2 /
(2 * self.posterior_variance * to_torch(alphas) *
(1 - self.alphas_cumprod)))
else:
raise NotImplementedError("mu not supported")
lvlb_weights[0] = lvlb_weights[1]
self.register_buffer('lvlb_weights', lvlb_weights, persistent=False)
assert not torch.isnan(self.lvlb_weights).all()
@contextmanager
def ema_scope(self, context: Optional[str] = None) -> Iterable[None]:
"""
Context manager that temporarily swaps to EMA weights (if enabled).
Args:
context: Optional label for logging.
"""
if self.use_ema:
self.model_ema.store(self.model.parameters())
self.model_ema.copy_to(self.model)
if context is not None:
mainlogger.info(f"{context}: Switched to EMA weights")
try:
yield None
finally:
if self.use_ema:
self.model_ema.restore(self.model.parameters())
if context is not None:
mainlogger.info(f"{context}: Restored training weights")
def init_from_ckpt(self,
path: str,
ignore_keys: Sequence[str] = tuple(),
only_model: bool = False) -> None:
"""
Load a checkpoint, optionally filtering keys or loading only the inner model.
Args:
path: Path to checkpoint.
ignore_keys: State-dict keys (prefix match) to drop.
only_model: If True, load only into self.model.
"""
sd = torch.load(path, map_location="cpu")
if "state_dict" in list(sd.keys()):
sd = sd["state_dict"]
keys = list(sd.keys())
for k in keys:
for ik in ignore_keys:
if k.startswith(ik):
mainlogger.info(
"Deleting key {} from state_dict.".format(k))
del sd[k]
missing, unexpected = self.load_state_dict(
sd,
strict=False) if not only_model else self.model.load_state_dict(
sd, strict=False)
mainlogger.info(
f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys"
)
if len(missing) > 0:
mainlogger.info(f"Missing Keys: {missing}")
if len(unexpected) > 0:
mainlogger.info(f"Unexpected Keys: {unexpected}")
def q_mean_variance(self, x_start: Tensor,
t: Tensor) -> Tuple[Tensor, Tensor, Tensor]:
"""
Compute q(x_t | x_0): mean, variance, and log-variance.
Args:
x_start: the [N x C x ...] tensor of noiseless inputs..
t: the number of diffusion steps (minus 1). Here, 0 means one step..
Returns:
(mean, variance, log_variance), each shaped like x_start.
"""
mean = (
extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) *
x_start)
variance = extract_into_tensor(1.0 - self.alphas_cumprod, t,
x_start.shape)
log_variance = extract_into_tensor(self.log_one_minus_alphas_cumprod,
t, x_start.shape)
return mean, variance, log_variance
def predict_start_from_noise(self, x_t: Tensor, t: Tensor,
noise: Tensor) -> Tensor:
"""
Predict x_0 from x_t and noise.
"""
return (
extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) *
x_t - extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t,
x_t.shape) * noise)
def predict_start_from_z_and_v(self, x_t: Tensor, t: Tensor,
v: Tensor) -> Tensor:
"""
Predict x_0 from z and v (v-parameterization).
"""
return (
extract_into_tensor(self.sqrt_alphas_cumprod, t, x_t.shape) * x_t -
extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t,
x_t.shape) * v)
def predict_eps_from_z_and_v(self, x_t: Tensor, t: Tensor,
v: Tensor) -> Tensor:
"""
Predict epsilon from z and v (v-parameterization).
"""
return (
extract_into_tensor(self.sqrt_alphas_cumprod, t, x_t.shape) * v +
extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t,
x_t.shape) * x_t)
def q_posterior(self, x_start: Tensor, x_t: Tensor,
t: Tensor) -> Tuple[Tensor, Tensor, Tensor]:
"""
Compute posterior q(x_{t-1} | x_t, x_0): mean and (log-)variance.
"""
posterior_mean = (
extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) *
x_start +
extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t)
posterior_variance = extract_into_tensor(self.posterior_variance, t,
x_t.shape)
posterior_log_variance_clipped = extract_into_tensor(
self.posterior_log_variance_clipped, t, x_t.shape)
return posterior_mean, posterior_variance, posterior_log_variance_clipped
def p_mean_variance(self, x: Tensor, t: Tensor,
clip_denoised: bool) -> Tuple[Tensor, Tensor, Tensor]:
"""
Predict mean and variance of p(x_{t-1} | x_t) using the model.
"""
model_out = self.model(x, t)
if self.parameterization == "eps":
x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
elif self.parameterization == "x0":
x_recon = model_out
if clip_denoised:
x_recon.clamp_(-1., 1.)
model_mean, posterior_variance, posterior_log_variance = self.q_posterior(
x_start=x_recon, x_t=x, t=t)
return model_mean, posterior_variance, posterior_log_variance
@torch.no_grad()
def p_sample(self,
x: Tensor,
t: Tensor,
clip_denoised: bool = True,
repeat_noise: bool = False) -> Tensor:
"""
Draw a single reverse-diffusion sample step: x_{t-1} from x_t.
Args:
x: Current noisy sample (B, C, ...).
t: Current timestep indices (B,).
clip_denoised: Clamp x_0 predictions or not.
repeat_noise: Reuse the same noise across the batch.
Returns:
Next sample x_{t-1}.
"""
b, *_, device = *x.shape, x.device
model_mean, _, model_log_variance = self.p_mean_variance(
x=x, t=t, clip_denoised=clip_denoised)
noise = noise_like(x.shape, device, repeat_noise)
# No noise when t == 0
nonzero_mask = (1 - (t == 0).float()).reshape(
b, *((1, ) * (len(x.shape) - 1)))
return model_mean + nonzero_mask * (0.5 *
model_log_variance).exp() * noise
@torch.no_grad()
def p_sample_loop(
self,
shape: Sequence[int],
return_intermediates: bool = False
) -> Union[Tensor, Tuple[Tensor, List[Tensor]]]:
"""
Run the full reverse process starting from Gaussian noise.
Args:
shape: Output tensor shape (B, C, ...).
return_intermediates: If True, also return intermediate frames.
Returns:
Final sample, and optionally the intermediate list.
"""
device = self.betas.device
b = shape[0]
img = torch.randn(shape, device=device)
intermediates = [img]
for i in tqdm(reversed(range(0, self.num_timesteps)),
desc='Sampling t',
total=self.num_timesteps):
img = self.p_sample(img,
torch.full((b, ),
i,
device=device,
dtype=torch.long),
clip_denoised=self.clip_denoised)
if i % self.log_every_t == 0 or i == self.num_timesteps - 1:
intermediates.append(img)
if return_intermediates:
return img, intermediates
return img
@torch.no_grad()
def sample(
self,
batch_size: int = 16,
return_intermediates: bool = False
) -> Union[Tensor, Tuple[Tensor, List[Tensor]]]:
"""
Convenience wrapper to sample square images of configured size.
Args:
batch_size: Number of samples.
return_intermediates: If True, also return intermediate frames.
Returns:
Final sample (and optionally intermediates).
"""
image_size = self.image_size
channels = self.channels
return self.p_sample_loop(
(batch_size, channels, image_size, image_size),
return_intermediates=return_intermediates)
def q_sample(self,
x_start: Tensor,
t: Tensor,
noise: Optional[Tensor] = None) -> Tensor:
"""
Forward noising step: sample x_t ~ q(x_t | x_0).
"""
noise = default(noise, lambda: torch.randn_like(x_start))
return (
extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) *
x_start + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod,
t, x_start.shape) * noise)
def get_v(self, x: Tensor, noise: Tensor, t: Tensor) -> Tensor:
"""
Compute v-target given x and epsilon.
"""
return (
extract_into_tensor(self.sqrt_alphas_cumprod, t, x.shape) * noise -
extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t,
x.shape) * x)
def get_loss(self,
pred: Tensor,
target: Tensor,
mean: bool = True) -> Tensor:
"""
Compute training loss between prediction and target.
Args:
pred: Model output.
target: Target tensor.
mean: If True, reduce to mean.
Returns:
Loss tensor (scalar if reduced).
"""
if self.loss_type == 'l1':
loss = (target - pred).abs()
if mean:
loss = loss.mean()
elif self.loss_type == 'l2':
if mean:
loss = torch.nn.functional.mse_loss(target, pred)
else:
loss = torch.nn.functional.mse_loss(target,
pred,
reduction='none')
else:
raise NotImplementedError("unknown loss type '{loss_type}'")
return loss
def p_losses(self,
x_start: Tensor,
t: Tensor,
noise: Optional[Tensor] = None
) -> Tuple[Tensor, Dict[str, Tensor]]:
"""
Compute the per-step training loss for a batch.
Args:
x_start: Clean inputs (B, C, ...).
t: Timesteps (B,).
noise: Optional pre-sampled epsilon.
Returns:
(loss, log_dict)
"""
noise = default(noise, lambda: torch.randn_like(x_start))
x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
model_out = self.model(x_noisy, t)
loss_dict = {}
if self.parameterization == "eps":
target = noise
elif self.parameterization == "x0":
target = x_start
elif self.parameterization == "v":
target = self.get_v(x_start, noise, t)
else:
raise NotImplementedError(
f"Paramterization {self.parameterization} not yet supported")
loss = self.get_loss(model_out, target, mean=False).mean(dim=[1, 2, 3])
log_prefix = 'train' if self.training else 'val'
loss_dict.update({f'{log_prefix}/loss_simple': loss.mean()})
loss_simple = loss.mean() * self.l_simple_weight
loss_vlb = (self.lvlb_weights[t] * loss).mean()
loss_dict.update({f'{log_prefix}/loss_vlb': loss_vlb})
loss = loss_simple + self.original_elbo_weight * loss_vlb
loss_dict.update({f'{log_prefix}/loss': loss})
return loss, loss_dict
def forward(self, x: Tensor, *args: Any,
**kwargs: Any) -> Tuple[Tensor, Dict[str, Tensor]]:
"""
Lightning forward: sample random timesteps and compute losses.
Args:
x: Clean batch (B, C, ...).
Returns:
(loss, log_dict)
"""
t = torch.randint(0,
self.num_timesteps, (x.shape[0], ),
device=self.device).long()
return self.p_losses(x, t, *args, **kwargs)
def get_input(self, batch: Mapping[str, Tensor], k: str) -> Tensor:
"""
Fetch and format the network input from batch.
Args:
batch: Batch mapping.
k: Key for the tensor to use.
Returns:
(B, C, ...) float32 contiguous tensor.
"""
x = batch[k]
'''
if len(x.shape) == 3:
x = x[..., None]
x = rearrange(x, 'b h w c -> b c h w')
'''
x = x.to(memory_format=torch.contiguous_format).float()
return x
def shared_step(
self, batch: Mapping[str,
Tensor]) -> Tuple[Tensor, Dict[str, Tensor]]:
"""
Common train/val step computing loss and logs.
"""
x = self.get_input(batch, self.first_stage_key)
loss, loss_dict = self(x)
return loss, loss_dict
def training_step(self, batch: Mapping[str, Tensor],
batch_idx: int) -> Tensor:
"""
PyTorch Lightning training step.
"""
loss, loss_dict = self.shared_step(batch)
self.log_dict(loss_dict,
prog_bar=True,
logger=True,
on_step=True,
on_epoch=True)
self.log("global_step",
self.global_step,
prog_bar=True,
logger=True,
on_step=True,
on_epoch=False)
if self.use_scheduler:
lr = self.optimizers().param_groups[0]['lr']
self.log('lr_abs',
lr,
prog_bar=True,
logger=True,
on_step=True,
on_epoch=False)
return loss
@torch.no_grad()
def validation_step(self, batch: Mapping[str, Tensor],
batch_idx: int) -> None:
"""
PyTorch Lightning validation step with and without EMA.
"""
_, loss_dict_no_ema = self.shared_step(batch)
with self.ema_scope():
_, loss_dict_ema = self.shared_step(batch)
loss_dict_ema = {
key + '_ema': loss_dict_ema[key]
for key in loss_dict_ema
}
self.log_dict(loss_dict_no_ema,
prog_bar=False,
logger=True,
on_step=False,
on_epoch=True)
self.log_dict(loss_dict_ema,
prog_bar=False,
logger=True,
on_step=False,
on_epoch=True)
def on_train_batch_end(self, *args: Any, **kwargs: Any) -> None:
"""
Update EMA after each train batch (if enabled).
"""
if self.use_ema:
self.model_ema(self.model)
def _get_rows_from_list(self, samples: List[Tensor]) -> Tensor:
"""
Arrange a list of (B, C, ...) tensors into a grid for logging.
Args:
samples: List of tensors at different timesteps.
Returns:
Grid image tensor suitable for visualization.
"""
n_imgs_per_row = len(samples)
denoise_grid = rearrange(samples, 'n b c h w -> b n c h w')
denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
return denoise_grid
@torch.no_grad()
def log_images(
self,
batch: Mapping[str, Tensor],
N: int = 8,
n_row: int = 2,
sample: bool = True,
return_keys: Optional[Sequence[str]] = None,
**kwargs: Any,
) -> Dict[str, Tensor]:
"""
Create tensors for image logging: inputs, diffusion row, (optional) samples.
Args:
batch: Batch mapping.
N: Number of examples to visualize.
n_row: Number of examples for diffusion-row visualization.
sample: If True, also run reverse diffusion to produce samples.
return_keys: If provided, filter the returned dict to these keys.
Returns:
Dict of image tensors.
"""
log = dict()
x = self.get_input(batch, self.first_stage_key)
N = min(x.shape[0], N)
n_row = min(x.shape[0], n_row)
x = x.to(self.device)[:N]
log["inputs"] = x
# Get diffusion row
diffusion_row = list()
x_start = x[:n_row]
for t in range(self.num_timesteps):
if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
t = t.to(self.device).long()
noise = torch.randn_like(x_start)
x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
diffusion_row.append(x_noisy)
log["diffusion_row"] = self._get_rows_from_list(diffusion_row)
if sample:
# Get denoise row
with self.ema_scope("Plotting"):
samples, denoise_row = self.sample(batch_size=N,
return_intermediates=True)
log["samples"] = samples
log["denoise_row"] = self._get_rows_from_list(denoise_row)
if return_keys:
if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
return log
else:
return {key: log[key] for key in return_keys}
return log
def configure_optimizers(self) -> torch.optim.Optimizer:
"""
Build the optimizer (AdamW) over model parameters (+ logvar if learned).
"""
lr = self.learning_rate
params = list(self.model.parameters())
if self.learn_logvar:
params = params + [self.logvar]
opt = torch.optim.AdamW(params, lr=lr)
return opt
class LatentDiffusion(DDPM):
"""
Main Class: Latent-diffusion model on top of DDPM (first/cond stages + guidance).
"""
def __init__(self,
first_stage_config: OmegaConf,
cond_stage_config: OmegaConf,
num_timesteps_cond: int | None = None,
cond_stage_key: str = "instruction",
cond_stage_trainable: bool = False,
cond_stage_forward: str | None = None,
conditioning_key: str | None = None,
uncond_prob: float = 0.2,
uncond_type: str = "empty_seq",
scale_factor: str = 1.0,
scale_by_std: bool = False,
encoder_type: str = "2d",
only_model: bool = False,
noise_strength: float = 0.0,
use_dynamic_rescale: bool = False,
base_scale: float = 0.7,
turning_step: int = 400,
interp_mode: bool = False,
fps_condition_type: str = 'fs',
perframe_ae: bool = False,
logdir: str | None = None,
rand_cond_frame: bool = False,
en_and_decode_n_samples_a_time: int | None = None,
*args,
**kwargs):
"""
Args:
first_stage_config: OmegaConf config for the first-stage autoencoder.
cond_stage_config: OmegaConf config for the conditioning encoder/module.
num_timesteps_cond: Number of condition timesteps used for schedule shortening.
cond_stage_key: Batch key for conditioning input (e.g., "instruction").
cond_stage_trainable: Whether the conditioning module is trainable.
cond_stage_forward: Optional method name to call on cond model instead of default.
conditioning_key: Conditioning mode (e.g., "crossattn", "concat").
uncond_prob: Probability to drop/zero the condition for classifier-free guidance.
uncond_type: Strategy for unconditional condition ("zero_embed" or "empty_seq").
scale_factor: Fixed latent scale multiplier if not using std-scaling.
scale_by_std: If True, compute scale as 1/std of latents at first batch.
encoder_type: "2d" (per-frame) or "3d" (volumetric) first-stage behavior.
only_model: If True, load only inner model weights when restoring from ckpt.
noise_strength: Extra offset noise strength for inputs (when > 0).
use_dynamic_rescale: If True, apply time-dependent rescaling array.
base_scale: Target base scale used by dynamic rescaling after turning_step.
turning_step: Steps to transition from 1.0 to base_scale in dynamic rescaling.
interp_mode: Flag for interpolation-specific behaviors (reserved).
fps_condition_type: Frame-per-second conditioning mode label.
perframe_ae: If True, encode/decode one frame at a time.
logdir: Optional directory for logs.
rand_cond_frame: If True, randomly select conditioning frames.
en_and_decode_n_samples_a_time: Optional per-step batch size for (en|de)code loops.
"""
self.num_timesteps_cond = default(num_timesteps_cond, 1)
self.scale_by_std = scale_by_std
assert self.num_timesteps_cond <= kwargs['timesteps']
# For backwards compatibility after implementation of DiffusionWrapper
ckpt_path = kwargs.pop("ckpt_path", None)
ignore_keys = kwargs.pop("ignore_keys", [])
conditioning_key = default(conditioning_key, 'crossattn')
super().__init__(conditioning_key=conditioning_key, *args, **kwargs)
self.cond_stage_trainable = cond_stage_trainable
self.cond_stage_key = cond_stage_key
self.noise_strength = noise_strength
self.use_dynamic_rescale = use_dynamic_rescale
self.interp_mode = interp_mode
self.fps_condition_type = fps_condition_type
self.perframe_ae = perframe_ae
self.logdir = logdir
self.rand_cond_frame = rand_cond_frame
self.en_and_decode_n_samples_a_time = en_and_decode_n_samples_a_time
try:
self.num_downs = len(
first_stage_config.params.ddconfig.ch_mult) - 1
except:
self.num_downs = 0
if not scale_by_std:
self.scale_factor = scale_factor
else:
self.register_buffer('scale_factor', torch.tensor(scale_factor))
if use_dynamic_rescale:
scale_arr1 = np.linspace(1.0, base_scale, turning_step)
scale_arr2 = np.full(self.num_timesteps, base_scale)
scale_arr = np.concatenate((scale_arr1, scale_arr2))
to_torch = partial(torch.tensor, dtype=torch.float32)
self.register_buffer('scale_arr', to_torch(scale_arr))
self.instantiate_first_stage(first_stage_config)
self.instantiate_cond_stage(cond_stage_config)
self.first_stage_config = first_stage_config
self.cond_stage_config = cond_stage_config
self.clip_denoised = False
self.cond_stage_forward = cond_stage_forward
self.encoder_type = encoder_type
assert (encoder_type in ["2d", "3d"])
self.uncond_prob = uncond_prob
self.classifier_free_guidance = True if uncond_prob > 0 else False
assert (uncond_type in ["zero_embed", "empty_seq"])
self.uncond_type = uncond_type
self.restarted_from_ckpt = False
if ckpt_path is not None:
self.init_from_ckpt(ckpt_path, ignore_keys, only_model=only_model)
self.restarted_from_ckpt = True
def make_cond_schedule(self) -> None:
"""
Build the condition timestep schedule.
"""
self.cond_ids = torch.full(size=(self.num_timesteps, ),
fill_value=self.num_timesteps - 1,
dtype=torch.long)
ids = torch.round(
torch.linspace(0, self.num_timesteps - 1,
self.num_timesteps_cond)).long()
self.cond_ids[:self.num_timesteps_cond] = ids
@rank_zero_only
@torch.no_grad()
def on_train_batch_start(self,
batch: Mapping[str, Any],
batch_idx: int,
dataloader_idx: int | None = None) -> None:
"""
Args:
batch: Current batch mapping.
batch_idx: Index of the batch within the epoch.
dataloader_idx: Optional dataloader index in multi-loader setups.
"""
# Only for very first batch, reset the self.scale_factor
if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 and batch_idx == 0 and \
not self.restarted_from_ckpt:
assert self.scale_factor == 1., 'rather not use custom rescaling and std-rescaling simultaneously'
# set rescale weight to 1./std of encodings
mainlogger.info("### USING STD-RESCALING ###")
x = super().get_input(batch, self.first_stage_key)
x = x.to(self.device)
encoder_posterior = self.encode_first_stage(x)
z = self.get_first_stage_encoding(encoder_posterior).detach()
del self.scale_factor
self.register_buffer('scale_factor', 1. / z.flatten().std())
mainlogger.info(
f"setting self.scale_factor to {self.scale_factor}")
mainlogger.info("### USING STD-RESCALING ###")
mainlogger.info(f"std={z.flatten().std()}")
def register_schedule(self,
given_betas: np.ndarray | None = None,
beta_schedule: str = "linear",
timesteps: int = 1000,
linear_start: float = 1e-4,
linear_end: float = 2e-2,
cosine_s: float = 8e-3) -> None:
"""
Extend base schedule registration and optionally shorten conditioning schedule.
Args:
given_betas: Optional precomputed beta schedule.
beta_schedule: Name of schedule function ("linear", "cosine", etc.).
timesteps: Number of diffusion steps.
linear_start: Linear schedule start beta (if used).
linear_end: Linear schedule end beta (if used).
cosine_s: Cosine schedule parameter (if used).
"""
super().register_schedule(given_betas, beta_schedule, timesteps,
linear_start, linear_end, cosine_s)
self.shorten_cond_schedule = self.num_timesteps_cond > 1
if self.shorten_cond_schedule:
self.make_cond_schedule()
def instantiate_first_stage(self, config: OmegaConf) -> None:
"""
Build and freeze the first-stage (autoencoder) model.
Args:
config: OmegaConf config describing the first-stage model to instantiate.
"""
model = instantiate_from_config(config)
self.first_stage_model = model.eval()
self.first_stage_model.train = disabled_train
for param in self.first_stage_model.parameters():
param.requires_grad = False
def instantiate_cond_stage(self, config: OmegaConf) -> None:
"""
Build the conditioning stage model.
Args:
config: OmegaConf config describing the conditioning model to instantiate.
"""
if not self.cond_stage_trainable:
model = instantiate_from_config(config)
self.cond_stage_model = model.eval()
self.cond_stage_model.train = disabled_train
for param in self.cond_stage_model.parameters():
param.requires_grad = False
else:
model = instantiate_from_config(config)
self.cond_stage_model = model
def get_learned_conditioning(self, c: Any) -> Tensor:
"""
Encode conditioning input into an embedding tensor.
Args:
c: Raw conditioning input (tensor, list/dict of strings, tokens, etc.).
Returns:
Conditioning embedding as a tensor (shape depends on cond model).
"""
if self.cond_stage_forward is None:
if hasattr(self.cond_stage_model, 'encode') and callable(
self.cond_stage_model.encode):
c = self.cond_stage_model.encode(c)
if isinstance(c, DiagonalGaussianDistribution):
c = c.mode()
else:
c = self.cond_stage_model(c)
else:
assert hasattr(self.cond_stage_model, self.cond_stage_forward)
c = getattr(self.cond_stage_model, self.cond_stage_forward)(c)
return c
def get_first_stage_encoding(
self,
encoder_posterior: DiagonalGaussianDistribution | Tensor,
noise: Tensor | None = None) -> Tensor:
"""
Convert encoder posterior to latent z and apply scaling.
Args:
encoder_posterior: First-stage output; either a Gaussian posterior or a latent tensor.
noise: Optional noise for sampling if posterior is Gaussian.
Returns:
Scaled latent tensor z.
"""
if isinstance(encoder_posterior, DiagonalGaussianDistribution):
z = encoder_posterior.sample(noise=noise)
elif isinstance(encoder_posterior, torch.Tensor):
z = encoder_posterior
else:
raise NotImplementedError(
f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented"
)
return self.scale_factor * z
@torch.no_grad()
def encode_first_stage(self, x: Tensor) -> Tensor:
"""
Encode input frames/images into latent space.
Args:
x: Input tensor, either (B, C, ...).
Returns:
Latent tensor with shape matched to input.
"""
if self.encoder_type == "2d" and x.dim() == 5:
b, _, t, _, _ = x.shape
x = rearrange(x, 'b c t h w -> (b t) c h w')
reshape_back = True
else:
reshape_back = False
## Consume more GPU memory but faster
if not self.perframe_ae:
encoder_posterior = self.first_stage_model.encode(x)
results = self.get_first_stage_encoding(encoder_posterior).detach()
else: ## Consume less GPU memory but slower
results = []
for index in range(x.shape[0]):
frame_batch = self.first_stage_model.encode(x[index:index +
1, :, :, :])
frame_result = self.get_first_stage_encoding(
frame_batch).detach()
results.append(frame_result)
results = torch.cat(results, dim=0)
if reshape_back:
results = rearrange(results, '(b t) c h w -> b c t h w', b=b, t=t)
return results
def decode_core(self, z: Tensor, **kwargs: Any) -> Tensor:
"""
Decode latent z back to pixel space (2D or per-frame).
Args:
z: Latent tensor (B, C, ...).
Returns:
Decoded tensor in pixel space with shape matching the input layout.
"""
if self.encoder_type == "2d" and z.dim() == 5:
b, _, t, _, _ = z.shape
z = rearrange(z, 'b c t h w -> (b t) c h w')
reshape_back = True
else:
reshape_back = False
if not self.perframe_ae:
z = 1. / self.scale_factor * z
results = self.first_stage_model.decode(z, **kwargs)
else:
results = []
for index in range(z.shape[0]):
frame_z = 1. / self.scale_factor * z[index:index + 1, :, :, :]
frame_result = self.first_stage_model.decode(frame_z, **kwargs)
results.append(frame_result)
results = torch.cat(results, dim=0)
if reshape_back:
results = rearrange(results, '(b t) c h w -> b c t h w', b=b, t=t)
return results
@torch.no_grad()
def decode_first_stage(self, z: Tensor, **kwargs: Any) -> Tensor:
"""
Decode latent with no gradient.
Args:
z: Latent tensor to decode.
**kwargs: Extra args for the decoder.
Returns:
Decoded tensor in pixel space.
"""
return self.decode_core(z, **kwargs)
# Same as above but without decorator
def differentiable_decode_first_stage(self, z: Tensor,
**kwargs: Any) -> Tensor:
"""
Decode latent with gradients enabled.
Args:
z: Latent tensor to decode.
Returns:
ecoded tensor in pixel space.
"""
return self.decode_core(z, **kwargs)
@torch.no_grad()
def get_batch_input(self,
batch: Mapping[str, Any],
random_uncond: bool,
return_first_stage_outputs: bool = False,
return_original_cond: bool = False) -> list[Any]:
"""
Prepare batch: encode inputs to latents and produce conditioning embeddings.
Args:
batch: Batch mapping containing first-stage inputs and conditioning.
random_uncond: If True and `uncond_type` allows, randomly drop conditions.
return_first_stage_outputs: If True, also decode z to xrec for logging.
return_original_cond: If True, also return the raw condition object.
"""
x = super().get_input(batch, self.first_stage_key)
# Encode video frames x to z via a 2D encoder
z = self.encode_first_stage(x)
# Get instruction condition
cond = batch[self.cond_stage_key]
if random_uncond and self.uncond_type == 'empty_seq':
for i, ci in enumerate(cond):
if random.random() < self.uncond_prob:
cond[i] = ""
if isinstance(cond, dict) or isinstance(cond, list):
cond_emb = self.get_learned_conditioning(cond)
else:
cond_emb = self.get_learned_conditioning(cond.to(self.device))
if random_uncond and self.uncond_type == 'zero_embed':
for i, ci in enumerate(cond):
if random.random() < self.uncond_prob:
cond_emb[i] = torch.zeros_like(cond_emb[i])
out = [z, cond_emb]
if return_first_stage_outputs:
xrec = self.decode_first_stage(z)
out.extend([xrec])
if return_original_cond:
out.append(cond)
return out
def forward(
self,
x: Tensor,
x_action: Tensor,
x_state: Tensor,
c: Any,
**kwargs: Any,
) -> tuple[Tensor, dict[str, Tensor]]:
"""
Args:
x: Input latent (or pixel) tensor for the primary stream.
x_action: Action tensor associated with the batch.
x_state: State tensor associated with the batch.
c: Conditioning object (tensor/list/dict) consumed by `apply_model`.
Returns:
(loss, loss_dict) tuple.
"""
t = torch.randint(0,
self.num_timesteps, (x.shape[0], ),
device=self.device).long()
if self.use_dynamic_rescale:
x = x * extract_into_tensor(self.scale_arr, t, x.shape)
return self.p_losses(x, x_action, x_state, c, t, **kwargs)
def shared_step(self, batch: Mapping[str, Any], random_uncond: bool,
**kwargs: Any) -> tuple[Tensor, dict[str, Tensor]]:
"""
Common train/val step: build inputs, run forward, return loss/logs.
Args:
batch: Input batch mapping.
random_uncond: Whether to apply classifier-free guidance dropout to cond.
**kwargs: Extra args forwarded to `forward`.
Returns:
(loss, loss_dict) tuple.
"""
x, c = self.get_batch_input(batch, random_uncond=random_uncond)
loss, loss_dict = self(x, c, **kwargs)
return loss, loss_dict
def apply_model(self, x_noisy: Tensor, x_action_noisy: Tensor,
x_state_noisy: Tensor, t: Tensor, cond: Any,
**kwargs: Any) -> Tensor | tuple[Tensor, Tensor, Tensor]:
"""
Apply inner diffusion model with standardized conditioning dict.
Args:
x_noisy: Noisy latent input for the primary stream.
x_action_noisy: Noisy action tensor aligned with t.
x_state_noisy: Noisy state tensor aligned with t.
t: Timestep indices (B,).
cond: Raw conditioning; will be wrapped into the proper key if not a dict.
**kwargs: Extra args forwarded to the inner model call.
Returns:
Either a single tensor or a tuple of tensors (x, action, state) depending on model.
"""
if isinstance(cond, dict):
pass
else:
if not isinstance(cond, list):
cond = [cond]
key = 'c_concat' if self.model.conditioning_key == 'concat' else 'c_crossattn'
cond = {key: cond}
x_recon, x_action_recon, x_state_recon = self.model(
x_noisy, x_action_noisy, x_state_noisy, t, **cond, **kwargs)
if isinstance(x_recon, tuple):
return x_recon[0]
else:
return x_recon, x_action_recon, x_state_recon
def p_losses(
self,
x_start: Tensor,
x_action_start: Tensor,
x_state_start: Tensor,
cond: Any,
t: Tensor,
noise: Tensor | None = None,
action_noise: Tensor | None = None,
**kwargs: Any,
) -> tuple[Tensor, dict[str, Tensor]]:
"""
Compute the per-step training losses for latent diffusion.
Args:
x_start: Clean primary latent (or pixel) tensor.
x_action_start: Clean action tensor aligned with x_start.
x_state_start: Clean state tensor aligned with x_start.
cond: Conditioning object; may include masks for action/state losses.
t: Timestep indices (B,).
noise: Optional epsilon noise for the primary stream (else sampled).
action_noise: Optional epsilon noise for the action stream (else sampled).
**kwargs: Extra args forwarded into `apply_model` (and logged if needed).
Returns:
(loss, loss_dict)
"""
if self.noise_strength > 0:
b, c, f, _, _ = x_start.shape
offset_noise = torch.randn(b, c, f, 1, 1, device=x_start.device)
noise = default(
noise, lambda: torch.randn_like(x_start) + self.noise_strength
* offset_noise)
else:
noise = default(noise, lambda: torch.randn_like(x_start))
action_noise = torch.randn(x_action_start.shape,
device=x_action_start.device)
action_noise_new = action_noise + self.input_pertub * torch.randn(
x_action_start.shape, device=x_action_start.device)
state_noise = torch.randn(x_state_start.shape,
device=x_state_start.device)
state_noise_new = state_noise + self.input_pertub * torch.randn(
x_state_start.shape, device=x_state_start.device)
x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
x_action_noisy = self.dp_noise_scheduler_action.add_noise(
x_action_start, action_noise_new, t[:x_action_start.shape[0]])
x_state_noisy = self.dp_noise_scheduler_state.add_noise(
x_state_start, state_noise_new, t[:x_state_start.shape[0]])
kwargs['x_start'] = x_start
model_output, model_action_output, model_state_output = self.apply_model(
x_noisy, x_action_noisy, x_state_noisy, t, cond, **kwargs)
loss_dict = {}
prefix = 'train' if self.training else 'val'
if self.parameterization == "x0":
target = x_start
elif self.parameterization == "eps":
target = noise
elif self.parameterization == "v":
target = self.get_v(x_start, noise, t)
else:
raise NotImplementedError()
target_action = action_noise
target_state = state_noise
loss_simple = self.get_loss(model_output, target,
mean=False).mean([1, 2, 3, 4])
loss_dict.update({f'{prefix}/loss_simple': loss_simple.mean()})
loss_action_simple = F.mse_loss(model_action_output,
target_action,
reduction='none')
action_mask = cond['c_crossattn_action'][-1]
loss_action_simple *= action_mask
loss_action_simple = loss_action_simple.type(loss_action_simple.dtype)
loss_action_simple = reduce(loss_action_simple, 'b ... -> b (...)',
'mean')
loss_action_simple = loss_action_simple.sum() / action_mask.sum()
loss_dict.update({f'{prefix}/loss_action_simple': loss_action_simple})
loss_state_simple = F.mse_loss(model_state_output,
target_state,
reduction='none')
state_mask = cond['c_crossattn_action'][-2]
loss_state_simple *= state_mask
loss_state_simple = loss_state_simple.type(loss_state_simple.dtype)
loss_state_simple = reduce(loss_state_simple, 'b ... -> b (...)',
'mean')
loss_state_simple = loss_state_simple.sum() / state_mask.sum()
loss_dict.update({f'{prefix}/loss_state_simple': loss_state_simple})
if self.logvar.device is not self.device:
self.logvar = self.logvar.to(self.device)
logvar_t = self.logvar[t]
loss = loss_simple / torch.exp(logvar_t) + logvar_t
loss_action = loss_action_simple
loss_state = loss_state_simple
if self.learn_logvar:
loss_dict.update({f'{prefix}/loss_gamma': loss.mean()})
loss_dict.update({'logvar': self.logvar.data.mean()})
loss = self.l_simple_weight * loss.mean()
loss_vlb = self.get_loss(model_output, target,
mean=False).mean(dim=(1, 2, 3, 4))
loss_vlb = (self.lvlb_weights[t] * loss_vlb).mean()
loss_dict.update({f'{prefix}/loss_vlb': loss_vlb})
loss_dict.update({f'{prefix}/loss_action_vlb': loss_action})
loss_dict.update({f'{prefix}/loss_state_vlb': loss_state})
loss += (self.original_elbo_weight * loss_vlb)
loss_dict.update({f'{prefix}/loss': loss})
loss_dict.update({f'{prefix}/loss_action': loss_action})
loss_dict.update({f'{prefix}/loss_state': loss_state})
if cond['c_crossattn_action'][2]:
return loss + loss_state + loss_action * 0.0, loss_dict
else:
return loss + loss_action + loss_state * 0.0, loss_dict
def training_step(self, batch: Mapping[str, Any],
batch_idx: int) -> Tensor:
"""
Lightning training step: compute loss and log metrics.
Args:
batch: Training batch mapping.
batch_idx: Batch index within current epoch.
Returns:
Scalar loss tensor for optimization.
"""
loss, loss_dict = self.shared_step(
batch, random_uncond=self.classifier_free_guidance)
loss_dict.update(
{'lr': self.trainer.optimizers[0].param_groups[0]['lr']})
loss_dict.update({
'lr_action_unet':
self.trainer.optimizers[0].param_groups[1]['lr']
})
# Sync_dist | rank_zero_only
self.log_dict(loss_dict,
prog_bar=True,
logger=True,
on_step=True,
on_epoch=True,
sync_dist=False)
if (batch_idx + 1) % self.log_every_t == 0:
mainlogger.info(
f"batch:{batch_idx}|epoch:{self.current_epoch} [globalstep:{self.global_step}]: loss={loss}"
)
return loss
@torch.no_grad()
def validation_step(self, batch: Mapping[str, Any],
batch_idx: int) -> None:
"""
Lightning validation step: compute loss and log metrics.
Args:
batch: Validation batch mapping.
batch_idx: Batch index in validation loop.
"""
_, loss_dict_no_ema = self.shared_step(batch, random_uncond=False)
self.log_dict(loss_dict_no_ema,
prog_bar=False,
logger=True,
on_step=False,
on_epoch=True)
def _get_denoise_row_from_list(self,
samples: Sequence[Tensor],
desc: str = '') -> Tensor:
"""
Decode a list of latents and pack into a grid for visualization.
Args:
samples: Sequence of latent tensors to decode and tile.
desc: Optional tqdm description string.
Returns:
Grid image tensor suitable for logging.
"""
denoise_row = []
for zd in tqdm(samples, desc=desc):
denoise_row.append(self.decode_first_stage(zd.to(self.device)))
n_log_timesteps = len(denoise_row)
denoise_row = torch.stack(denoise_row)
if denoise_row.dim() == 5:
denoise_grid = rearrange(denoise_row, 'n b c h w -> b n c h w')
denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
denoise_grid = make_grid(denoise_grid, nrow=n_log_timesteps)
elif denoise_row.dim() == 6:
video_length = denoise_row.shape[3]
denoise_grid = rearrange(denoise_row, 'n b c t h w -> b n c t h w')
denoise_grid = rearrange(denoise_grid,
'b n c t h w -> (b n) c t h w')
denoise_grid = rearrange(denoise_grid, 'n c t h w -> (n t) c h w')
denoise_grid = make_grid(denoise_grid, nrow=video_length)
else:
raise ValueError
return denoise_grid
@torch.no_grad()
def log_images(self,
batch: Mapping[str, Any],
sample: bool = True,
ddim_steps: int = 200,
ddim_eta: float = 1.0,
plot_denoise_rows: bool = False,
unconditional_guidance_scale: float = 1.0,
**kwargs: Any) -> dict[str, Tensor]:
""" Log images for LatentDiffusion """
# Control sampled imgae for logging, larger value may cause OOM
sampled_img_num = 2
for key in batch.keys():
batch[key] = batch[key][:sampled_img_num]
# TBD: currently, classifier_free_guidance sampling is only supported by DDIM
use_ddim = ddim_steps is not None
log = dict()
z, c, xrec, xc = self.get_batch_input(batch,
random_uncond=False,
return_first_stage_outputs=True,
return_original_cond=True,
logging=True)
N = xrec.shape[0]
log["reconst"] = xrec
log["condition"] = xc
if sample:
uc = None
with self.ema_scope("Plotting"):
samples, z_denoise_row = self.sample_log(
cond=c,
batch_size=N,
ddim=use_ddim,
ddim_steps=ddim_steps,
eta=ddim_eta,
unconditional_guidance_scale=unconditional_guidance_scale,
unconditional_conditioning=uc,
x0=z,
**kwargs)
x_samples = self.decode_first_stage(samples)
log["samples"] = x_samples
if plot_denoise_rows:
denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
log["denoise_row"] = denoise_grid
return log
def p_mean_variance(
self,
x: Tensor,
c: Any,
t: Tensor,
clip_denoised: bool,
return_x0: bool = False,
score_corrector: Any = None,
corrector_kwargs: Mapping[str, Any] | None = None,
**kwargs: Any
) -> tuple[Tensor, Tensor, Tensor] | tuple[Tensor, Tensor, Tensor, Tensor]:
"""
Predict posterior parameters (and optionally x0) at timestep t.
Args:
x: Current latent at timestep t.
c: Conditioning object passed to the inner model/score corrector.
t: Timestep indices (B,).
clip_denoised: If True, clamp predicted x0 to [-1, 1].
return_x0: If True, also return predicted x0.
score_corrector: Optional score-corrector object with `modify_score`.
corrector_kwargs: Extra kwargs for the score corrector.
**kwargs: Forwarded to `apply_model`.
Returns:
(mean, var, log_var) or (mean, var, log_var, x0) tensors.
"""
t_in = t
model_out = self.apply_model(x, t_in, c, **kwargs)
if score_corrector is not None:
assert self.parameterization == "eps"
model_out = score_corrector.modify_score(self, model_out, x, t, c,
**corrector_kwargs)
if self.parameterization == "eps":
x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
elif self.parameterization == "x0":
x_recon = model_out
else:
raise NotImplementedError()
if clip_denoised:
x_recon.clamp_(-1., 1.)
model_mean, posterior_variance, posterior_log_variance = self.q_posterior(
x_start=x_recon, x_t=x, t=t)
if return_x0:
return model_mean, posterior_variance, posterior_log_variance, x_recon
else:
return model_mean, posterior_variance, posterior_log_variance
@torch.no_grad()
def p_sample(self,
x: Tensor,
c: Any,
t: Tensor,
clip_denoised: bool = False,
repeat_noise: bool = False,
return_x0: bool = False,
temperature: float = 1.0,
noise_dropout: float = 0.0,
score_corrector: Any = None,
corrector_kwargs: Mapping[str, Any] | None = None,
**kwargs: Any) -> Tensor | tuple[Tensor, Tensor]:
"""
Draw a single reverse-diffusion step (optionally return x0).
Args:
x: Current latent at timestep t.
c: Conditioning object for the model.
t: Timestep indices (B,).
clip_denoised: Clamp predicted x0 to [-1, 1] when forming the mean.
repeat_noise: If True, reuse the same noise across batch.
return_x0: If True, also return the predicted x0.
temperature: Temperature for sampling noise scale.
noise_dropout: Dropout probability applied to the sampled noise.
score_corrector: Optional score-corrector to adjust model outputs.
corrector_kwargs: Extra kwargs for the corrector.
**kwargs: Forwarded to `p_mean_variance`.
Returns:
Next latent (and optionally x0).
"""
b, *_, device = *x.shape, x.device
outputs = self.p_mean_variance(x=x, c=c, t=t, clip_denoised=clip_denoised, return_x0=return_x0, \
score_corrector=score_corrector, corrector_kwargs=corrector_kwargs, **kwargs)
if return_x0:
model_mean, _, model_log_variance, x0 = outputs
else:
model_mean, _, model_log_variance = outputs
noise = noise_like(x.shape, device, repeat_noise) * temperature
if noise_dropout > 0.:
noise = torch.nn.functional.dropout(noise, p=noise_dropout)
# No noise when t == 0
nonzero_mask = (1 - (t == 0).float()).reshape(
b, *((1, ) * (len(x.shape) - 1)))
if return_x0:
return model_mean + nonzero_mask * (
0.5 * model_log_variance).exp() * noise, x0
else:
return model_mean + nonzero_mask * (
0.5 * model_log_variance).exp() * noise
@torch.no_grad()
def p_sample_loop(self,
cond: Any,
shape: Sequence[int],
return_intermediates: bool = False,
x_T: Tensor | None = None,
verbose: bool = True,
callback: Callable[[int], Any] | None = None,
timesteps: int | None = None,
mask: Tensor | None = None,
x0: Tensor | None = None,
img_callback: Callable[[Tensor, int], Any] | None = None,
start_T: int | None = None,
log_every_t: int | None = None,
**kwargs: Any) -> Tensor | tuple[Tensor, list[Tensor]]:
"""
Run the full reverse process from noise to sample(s).
Args:
cond: Conditioning object (tensor/dict/list), optionally noised when cond schedule is shortened.
shape: Output latent shape (B, C, ...).
return_intermediates: If True, also return intermediate latents.
x_T: Optional starting noise latent (else sampled from N(0, I)).
verbose: If True, show tqdm progress.
callback: Optional function called with the current timestep i.
timesteps: Number of reverse steps to perform (default: self.num_timesteps).
mask: Optional inpainting mask; ones keep original x0 regions.
x0: Optional original latent for masked regions (when using `mask`).
img_callback: Optional function called with (img, i) every step.
start_T: Optional cap to limit starting step (min(timesteps, start_T)).
log_every_t: Logging frequency for collecting intermediates (defaults to self.log_every_t).
Returns:
Final latent sample (and optionally the list of intermediates).
"""
if not log_every_t:
log_every_t = self.log_every_t
device = self.betas.device
b = shape[0]
# Sample an initial noise
if x_T is None:
img = torch.randn(shape, device=device)
else:
img = x_T
intermediates = [img]
if timesteps is None:
timesteps = self.num_timesteps
if start_T is not None:
timesteps = min(timesteps, start_T)
iterator = tqdm(
reversed(range(0, timesteps)), desc='Sampling t',
total=timesteps) if verbose else reversed(range(0, timesteps))
if mask is not None:
assert x0 is not None
assert x0.shape[2:3] == mask.shape[2:3]
for i in iterator:
ts = torch.full((b, ), i, device=device, dtype=torch.long)
if self.shorten_cond_schedule:
assert self.model.conditioning_key != 'hybrid'
tc = self.cond_ids[ts].to(cond.device)
cond = self.q_sample(x_start=cond,
t=tc,
noise=torch.randn_like(cond))
img = self.p_sample(img,
cond,
ts,
clip_denoised=self.clip_denoised,
**kwargs)
if mask is not None:
img_orig = self.q_sample(x0, ts)
img = img_orig * mask + (1. - mask) * img
if i % log_every_t == 0 or i == timesteps - 1:
intermediates.append(img)
if callback: callback(i)
if img_callback: img_callback(img, i)
if return_intermediates:
return img, intermediates
return img
@torch.no_grad()
def sample(self,
cond,
batch_size: int = 16,
return_intermediates: bool = False,
x_T: Tensor | None = None,
verbose: bool = True,
timesteps: int | None = None,
mask: Tensor | None = None,
x0: Tensor | None = None,
shape: Sequence[int] | None = None,
**kwargs: Any) -> Tensor | tuple[Tensor, list[Tensor]]:
"""
Convenience wrapper to run `p_sample_loop` with a full batch.
Args:
cond: Conditioning object; dict/list items are truncated to batch_size.
batch_size: Number of samples to generate.
return_intermediates: If True, return intermediates as well.
x_T: Optional starting noise latent (else sampled).
verbose: Whether to print sampling progress.
timesteps: Number of reverse steps (default: self.num_timesteps).
mask: Optional mask for partial generation/inpainting.
x0: Optional original latent used with `mask` during sampling.
shape: Optional output shape; if None, uses (B, C, T, H, W) from model config.
Returns:
Final latent (and optionally intermediates).
"""
if shape is None:
shape = (batch_size, self.channels, self.temporal_length,
*self.image_size)
if cond is not None:
if isinstance(cond, dict):
cond = {
key:
cond[key][:batch_size] if not isinstance(cond[key], list)
else list(map(lambda x: x[:batch_size], cond[key]))
for key in cond
}
else:
cond = [c[:batch_size] for c in cond] if isinstance(
cond, list) else cond[:batch_size]
return self.p_sample_loop(cond,
shape,
return_intermediates=return_intermediates,
x_T=x_T,
verbose=verbose,
timesteps=timesteps,
mask=mask,
x0=x0,
**kwargs)
@torch.no_grad()
def sample_log(self, cond: Any, batch_size: int, ddim: bool,
ddim_steps: int,
**kwargs: Any) -> tuple[Any, Any, Any, Any]:
"""
Produce samples (and intermediates), optionally via DDIM sampler.
Args:
cond: Conditioning object passed to the sampler.
batch_size: Number of samples to generate.
ddim: If True, use DDIM sampler; otherwise use ancestral sampling.
ddim_steps: Number of DDIM steps when `ddim` is True.
"""
if ddim:
ddim_sampler = DDIMSampler(self)
shape = (self.channels, self.temporal_length, *self.image_size)
samples, actions, states, intermediates = ddim_sampler.sample(
ddim_steps, batch_size, shape, cond, verbose=False, **kwargs)
else:
samples, intermediates = self.sample(cond=cond,
batch_size=batch_size,
return_intermediates=True,
**kwargs)
return samples, actions, states, intermediates
def configure_schedulers(
self, optimizer: torch.optim.Optimizer) -> dict[str, Any]:
"""
Build LR scheduler dict compatible with PyTorch Lightning.
Args:
optimizer: Optimizer instance for which to build the scheduler dict.
Returns:
Dict with keys {'scheduler', 'interval', 'frequency'} per Lightning API.
"""
assert 'target' in self.scheduler_config
scheduler_name = self.scheduler_config.target.split('.')[-1]
interval = self.scheduler_config.interval
frequency = self.scheduler_config.frequency
if scheduler_name == "LambdaLRScheduler":
scheduler = instantiate_from_config(self.scheduler_config)
scheduler.start_step = self.global_step
lr_scheduler = {
'scheduler': LambdaLR(optimizer, lr_lambda=scheduler.schedule),
'interval': interval,
'frequency': frequency
}
elif scheduler_name == "CosineAnnealingLRScheduler":
scheduler = instantiate_from_config(self.scheduler_config)
decay_steps = scheduler.decay_steps
last_step = -1 if self.global_step == 0 else scheduler.start_step
lr_scheduler = {
'scheduler':
CosineAnnealingLR(optimizer,
T_max=decay_steps,
last_epoch=last_step),
'interval':
interval,
'frequency':
frequency
}
else:
raise NotImplementedError
return lr_scheduler
class LatentVisualDiffusion(LatentDiffusion):
"""
Visual-conditioned latent diffusion with action/state heads and schedulers.
"""
def __init__(self,
img_cond_stage_config: OmegaConf,
image_proj_stage_config: OmegaConf,
noise_scheduler_config: OmegaConf,
dp_optimizer_config: OmegaConf,
dp_ema_config: OmegaConf,
freeze_embedder: bool = True,
image_proj_model_trainable: bool = True,
n_obs_steps_imagen: int = 2,
n_obs_steps_acting: int = 2,
agent_state_dim: int = 14,
agent_action_dim: int = 14,
global_emb_dim: int = 1024,
input_pertub: float = 0.1,
lr_scheduler: str = 'cosine',
lr_warmup_steps: int = 500,
num_epochs: int = 15000,
gradient_accumulate_every: int = 1,
use_scheduler: bool = False,
dp_use_ema: bool = False,
pretrained_checkpoint: str | None = None,
decision_making_only: bool = True,
projector_bf16: bool = True,
*args,
**kwargs):
"""
Args:
img_cond_stage_config: OmegaConf for the *image* conditioning encoder.
image_proj_stage_config: OmegaConf for the image feature projector.
noise_scheduler_config: OmegaConf for DP noise schedulers (state/action).
dp_optimizer_config: OmegaConf for optimizer params of the UNet heads.
dp_ema_config: Optional EMA config for the action UNet.
freeze_embedder: If True, freeze the image embedder params.
image_proj_model_trainable: If True, train the image projector.
n_obs_steps_imagen: Number of observed steps for image conditions.
n_obs_steps_acting: Number of observed steps for acting head.
agent_state_dim: Dimension of agent state vector.
agent_action_dim: Dimension of agent action vector.
global_emb_dim: Embedding size for state/action/text/image fusion.
input_pertub: Perturbation scale added to action/state noises.
lr_scheduler: Name of LR scheduler (for SelectiveLRScheduler wrapper).
lr_warmup_steps: Warmup steps for scheduler creation.
num_epochs: Total training epochs.
gradient_accumulate_every: Gradient accumulation steps.
use_scheduler: If True, enable LR scheduling.
dp_use_ema: If True, maintain EMA for action UNet head.
pretrained_checkpoint: Optional path to a pretrained checkpoint.
decision_making_only: If True, use decision-only augmentation path.
projector_bf16: If True, run image/state/action projectors under BF16 autocast.
"""
super().__init__(*args, **kwargs)
self.image_proj_model_trainable = image_proj_model_trainable
self.agent_state_dim = agent_state_dim
self.agent_action_dim = agent_action_dim
self.global_emb_dim = global_emb_dim
self.n_obs_steps_imagen = n_obs_steps_imagen
self.n_obs_steps_acting = n_obs_steps_acting
self.decision_making_only = decision_making_only
self.projector_bf16 = projector_bf16
self._init_embedder(img_cond_stage_config, freeze_embedder)
self._init_img_ctx_projector(image_proj_stage_config,
image_proj_model_trainable)
self._init_dp_noise_scheduler(noise_scheduler_config)
self._init_projectors()
if dp_use_ema:
self._init_dp_ema(dp_ema_config)
# Create a pos_embedder for state and action info, our state and action have an unified vector space
self.pos_embedder = SinusoidalPosEmb(self.global_emb_dim)
self.register_buffer('cond_pos_emb',
self.pos_embedder(torch.arange(
0, 16))) #NOTE HAND-CODE 16
self.input_pertub = input_pertub
self.dp_optimizer_config = dp_optimizer_config
self.lr_scheduler = lr_scheduler
self.lr_warmup_steps = lr_warmup_steps
self.num_epochs = num_epochs
self.gradient_accumulate_every = gradient_accumulate_every
self.use_scheduler = use_scheduler
self.dp_use_ema = dp_use_ema
self.pretrained_checkpoint = pretrained_checkpoint
def _init_img_ctx_projector(self, config: OmegaConf,
trainable: bool) -> None:
"""
Instantiate image context projector; optionally freeze.
Args:
config: OmegaConf for the projector module to instantiate.
trainable: If False, freeze the projector.
"""
self.image_proj_model = instantiate_from_config(config)
if not trainable:
self.image_proj_model.eval()
self.image_proj_model.train = disabled_train
for param in self.image_proj_model.parameters():
param.requires_grad = False
def _init_embedder(self, config: OmegaConf, freeze: bool = True) -> None:
"""
Instantiate the image embedder; optionally freeze.
Args:
config: OmegaConf for the embedder to instantiate.
freeze: If True, set to eval/disable grads.
"""
self.embedder = instantiate_from_config(config)
if freeze:
self.embedder.eval()
self.embedder.train = disabled_train
for param in self.embedder.parameters():
param.requires_grad = False
def init_normalizers(self, normalize_config: OmegaConf,
dataset_stats: Mapping[str, Any]) -> None:
"""
Create normalization and unnormalization utilities.
Args:
normalize_config: Config with shapes and normalization modes.
dataset_stats: Statistics dict used to compute normalization.
"""
self.normalize_inputs = Normalize(
normalize_config.input_shapes,
normalize_config.input_normalization_modes, dataset_stats)
self.unnormalize_outputs = Unnormalize(
normalize_config.output_shapes,
normalize_config.output_normalization_modes, dataset_stats)
def _init_dp_noise_scheduler(self, config: OmegaConf) -> None:
"""
Instantiate separate DP noise schedulers for action and state.
Args:
config: OmegaConf used to create scheduler instances.
"""
self.dp_noise_scheduler_action = instantiate_from_config(config)
self.dp_noise_scheduler_state = instantiate_from_config(config)
def _init_dp_ema(self, config: OmegaConf | None) -> None:
"""
Initialize EMA for UNet head.
Args:
config: EMA config, must contain 'params' sub-dict.
"""
self.dp_ema_model = copy.deepcopy(
self.model.diffusion_model.action_unet)
self.dp_ema_model_on_device = False
self.dp_ema = EMAModel(**config['params'], model=self.dp_ema_model)
def _init_projectors(self):
"""
Build small MLP projectors and positional embeddings for state/action.
"""
self.state_projector = MLPProjector(self.agent_state_dim,
1024) # NOTE HAND CODE
self.action_projector = MLPProjector(self.agent_action_dim,
1024) # NOTE HAND CODE
self.agent_action_pos_emb = nn.Parameter(
torch.randn(1, 16, self.global_emb_dim))
self.agent_state_pos_emb = nn.Parameter(
torch.randn(1, self.n_obs_steps_imagen, self.global_emb_dim))
def _projector_forward(self, projector: nn.Module, x: Tensor,
target_dtype: torch.dtype | None) -> Tensor:
use_bf16 = (self.projector_bf16 and x.device.type == "cuda"
and torch.cuda.is_bf16_supported())
if not use_bf16:
weight_dtype = None
for param in projector.parameters():
weight_dtype = param.dtype
break
if weight_dtype is not None and x.dtype != weight_dtype:
x = x.to(dtype=weight_dtype)
if use_bf16:
with torch.autocast(device_type="cuda", dtype=torch.bfloat16):
out = projector(x)
else:
out = projector(x)
if not hasattr(self, "_printed_projector_bf16"):
print(
">>> projector bf16 autocast: "
f"enabled={self.projector_bf16} "
f"use_bf16={use_bf16} "
f"input={x.dtype} "
f"output={out.dtype} "
f"target={target_dtype}")
self._printed_projector_bf16 = True
if target_dtype is not None and out.dtype != target_dtype:
out = out.to(dtype=target_dtype)
return out
def _get_augmented_batch(
self,
z: Tensor,
state: Tensor,
obs_state: Tensor,
action: Tensor,
ins: Tensor,
null_ins: Tensor,
img: Tensor,
sim_mode: bool = False,
pre_action: Tensor | None = None,
logging: bool = False) -> tuple[Tensor, Tensor, list[Tensor]]:
"""
Construct augmented conditioning batch for decision/simulation modes.
Args:
z: Latent video tensor (B, C, ...).
state: Full state tensor (B, T, D_s).
obs_state: Observed state embeddings (B, T, E).
action: Action embeddings (B, T, E).
ins: Instruction/text embeddings (B, L, E) after projector.
null_ins: Null/empty instruction embedding for CFG.
img: Image conditioning embedding (B, E_img) or batched equivalent.
sim_mode: If True, build simulated-mode batch; else decision-making.
pre_action: Optional previous action(s). (unused here; reserved)
logging: If True, may include extra returns for logs. (unused)
Returns:
Tuple of (z, state, [mode_batch]) where mode_batch is a single tensor combining the selected conditioning streams.
"""
b, _, t, _, _ = z.shape
if self.decision_making_only:
mode_batch = torch.cat([obs_state, ins, img], dim=1)
return z, state, [mode_batch]
if not sim_mode:
zero_action = torch.zeros_like(action)
mode_batch = torch.cat([obs_state, zero_action, ins, img], dim=1)
else:
null_ins_batch = null_ins.repeat_interleave(repeats=ins.shape[0],
dim=0)
mode_batch = torch.cat([obs_state, action, null_ins_batch, img],
dim=1)
return z, state, [mode_batch]
def on_train_batch_end(self, outputs: Any, batch: Mapping[str, Any],
batch_idx: int) -> None:
"""
Update EMA for action UNet after each train batch (if enabled).
Args:
batch: Current training batch mapping.
batch_idx: Batch index within the epoch.
"""
if self.dp_use_ema:
if self.dp_ema_model is not None and not self.dp_ema_model_on_device:
device = self.model.device
self.dp_ema_model.to(device)
self.dp_ema_model_on_device = True
self.dp_ema.step(self.model.diffusion_model.action_unet)
def shared_step(self, batch: Mapping[str, Any], random_uncond: bool,
**kwargs: Any) -> tuple[Tensor, dict[str, Tensor]]:
"""
Common train/val step for visual diffusion.
Args:
batch: Input batch mapping.
random_uncond: Whether to apply classifier-free guidance dropout.
Returns:
(loss, loss_dict) tuple.
"""
x, x_action, x_state, c, fs = self.get_batch_input(
batch, random_uncond=random_uncond, return_fs=True)
kwargs.update({"fs": fs.long()})
loss, loss_dict = self(x, x_action, x_state, c, **kwargs)
return loss, loss_dict
def get_batch_input(self,
batch: Mapping[str, Any],
random_uncond: bool,
return_first_stage_outputs: bool = False,
return_original_cond: bool = False,
return_fs: bool = False,
return_cond_frame: bool = False,
return_original_input: bool = False,
logging: bool = False,
**kwargs: Any) -> list[Any]:
"""
Prepare model inputs & conditioning from a raw training batch.
Args:
batch: Batch mapping with keys like image/state/action/obs/etc.
random_uncond: Apply stochastic condition dropout for CFG.
return_first_stage_outputs: If True, also return xrec (decoded z).
return_original_cond: If True, also return raw instruction text.
return_fs: If True, return fps or frame_stride per config.
return_cond_frame: If True, return conditioning frames (obs images).
return_original_input: If True, return original x (pre-encoding).
logging: If True, append sim_mode flag at the end.
Returns:
A list of inputs
"""
# x: b c t h w
x = super().get_input(batch, self.first_stage_key)
b, _, t, _, _ = x.shape
# Get actions: b t d
action = super().get_input(batch, 'action')
# Get states: b t d
state = super().get_input(batch, 'next.state')
# Get observable states: b t d
obs_state = super().get_input(batch, 'observation.state')
# Get observable images: b c t h w
obs = super().get_input(batch, 'observation.image')
# Encode video frames x to z via a 2D encoder
z = self.encode_first_stage(x)
cond = {}
# Get instruction condition
cond_ins_input = batch[self.cond_stage_key]
if isinstance(cond_ins_input, dict) or isinstance(
cond_ins_input, list):
cond_ins_emb = self.get_learned_conditioning(cond_ins_input)
else:
cond_ins_emb = self.get_learned_conditioning(
cond_ins_input.to(self.device))
# To support classifier-free guidance, randomly drop out only text conditioning
# 5%, only image conditioning 5%, and both 5%.
if random_uncond:
random_num = torch.rand(b, device=x.device)
else:
random_num = torch.ones(b, device=x.device)
prompt_mask = rearrange(random_num < 2 * self.uncond_prob,
"n -> n 1 1")
null_prompt = self.get_learned_conditioning([""])
cond_ins_emb = torch.where(prompt_mask, null_prompt,
cond_ins_emb.detach())
target_dtype = cond_ins_emb.dtype
# Get conditioning frames
cond_frame_index = 0
img = obs[:, :, -1, ...]
input_mask = 1 - rearrange(
(random_num >= self.uncond_prob).float() *
(random_num < 3 * self.uncond_prob).float(), "n -> n 1 1 1")
cond_img = input_mask * img
cond_img_emb = self.embedder(cond_img)
cond_img_emb = self._projector_forward(self.image_proj_model,
cond_img_emb, target_dtype)
if self.model.conditioning_key == 'hybrid':
if self.interp_mode:
img_cat_cond = torch.zeros_like(z)
img_cat_cond[:, :, 0, :, :] = z[:, :, 0, :, :]
img_cat_cond[:, :, -1, :, :] = z[:, :, -1, :, :]
else:
img_cat_cond = z[:, :, cond_frame_index, :, :]
img_cat_cond = img_cat_cond.unsqueeze(2)
img_cat_cond = repeat(img_cat_cond,
'b c t h w -> b c (repeat t) h w',
repeat=z.shape[2])
cond["c_concat"] = [img_cat_cond]
cond_action = self._projector_forward(self.action_projector, action,
target_dtype)
cond_action_emb = self.agent_action_pos_emb.to(
dtype=target_dtype) + cond_action
# Get conditioning states
cond_state = self._projector_forward(self.state_projector, obs_state,
target_dtype)
cond_state_emb = self.agent_state_pos_emb.to(
dtype=target_dtype) + cond_state
if self.decision_making_only:
is_sim_mode = False
else:
is_sim_mode = torch.rand(1) < 0.5
z, state, cond["c_crossattn"] = self._get_augmented_batch(
z,
state,
cond_state_emb,
cond_action_emb,
cond_ins_emb,
null_prompt,
cond_img_emb,
sim_mode=is_sim_mode,
logging=logging)
cond["c_crossattn_action"] = [
obs[:, :, -self.n_obs_steps_acting:],
state[:, -self.n_obs_steps_acting:], is_sim_mode,
batch['state_mask'], batch['action_mask']
]
out = [z, action, state, cond]
if return_first_stage_outputs:
xrec = self.decode_first_stage(z)
out.extend([xrec])
if return_original_cond:
out.append(cond_ins_input)
if return_fs:
if self.fps_condition_type == 'fs':
fs = super().get_input(batch, 'frame_stride')
elif self.fps_condition_type == 'fps':
fs = super().get_input(batch, 'fps')
out.append(fs)
if return_cond_frame:
out.append(obs)
if return_original_input:
out.append(x)
if logging:
out.append(is_sim_mode)
return out
@torch.no_grad()
def log_images(self,
batch: Mapping[str, Any],
sample: bool = True,
ddim_steps: int = 50,
ddim_eta: float = 1.0,
plot_denoise_rows: bool = False,
unconditional_guidance_scale: float = 1.0,
mask: Tensor | None = None,
**kwargs) -> dict[str, Tensor]:
"""
Log images for LatentVisualDiffusion
Args:
batch: Batch mapping used to form inputs/conditions.
sample: If True, also run sampling for visualization.
ddim_steps: Number of DDIM steps when using DDIM.
ddim_eta: DDIM eta parameter (stochasticity).
plot_denoise_rows: If True, include denoise progression grid.
unconditional_guidance_scale: Guidance scale for CFG sampling.
mask: Optional mask for sampling-time inpainting.
Returns:
Dict of visualization tensors (images/actions/states/progress).
"""
##### sampled_img_num: control sampled imgae for logging, larger value may cause OOM
sampled_img_num = 1
for key in batch.keys():
batch[key] = batch[key][:sampled_img_num]
## TBD: currently, classifier_free_guidance sampling is only supported by DDIM
use_ddim = ddim_steps is not None
log = dict()
z, act, state, c, xrec, xc, fs, cond_x, is_sim_mode = self.get_batch_input(
batch,
random_uncond=False,
return_first_stage_outputs=True,
return_original_cond=True,
return_fs=True,
return_cond_frame=True,
logging=True)
kwargs['x_start'] = z
N = xrec.shape[0]
log["image_condition"] = cond_x
log["reconst"] = xrec
if is_sim_mode:
xc = ["NULL"]
xc_with_fs = []
for idx, content in enumerate(xc):
xc_with_fs.append(content + '_fs=' + str(fs[idx].item()))
log['instruction'] = xc
log["condition"] = xc_with_fs
kwargs.update({"fs": fs.long()})
if sample:
uc = None
with self.ema_scope("Plotting"):
samples, action_samples, state_samples, z_denoise_row = self.sample_log(
cond=c,
batch_size=N,
ddim=use_ddim,
ddim_steps=ddim_steps,
eta=ddim_eta,
unconditional_guidance_scale=unconditional_guidance_scale,
unconditional_conditioning=uc,
x0=z,
**kwargs)
x_samples = self.decode_first_stage(samples)
log["samples"] = x_samples
# Log actions
mb, mt, _ = batch['action_mask'].shape
act_mask = batch['action_mask'] == 1.0
action_target = act[act_mask].reshape(mb, mt, -1)
action_samples = action_samples[act_mask].reshape(mb, mt, -1)
log["action"] = torch.cat((action_target, action_samples), dim=0)
# Log states
mb, mt, _ = batch['state_mask'].shape
state_mask = batch['state_mask'] == 1.0
state_target = state[state_mask].reshape(mb, mt, -1)
state_samples = state_samples[state_mask].reshape(mb, mt, -1)
log["state"] = torch.cat((state_target, state_samples), dim=0)
if plot_denoise_rows:
denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
log["denoise_row"] = denoise_grid
log["video_idx"] = batch["path"][0].split('/')[-1][:-4]
return log
def configure_optimizers(self):
""" configure_optimizers for LatentDiffusion """
lr = self.learning_rate
params = [
param for name, param in self.model.named_parameters()
if not name.startswith("diffusion_model.action_unet")
and not name.startswith("diffusion_model.state_unet")
]
params_unet_head = list(
self.model.diffusion_model.action_unet.parameters()) + list(
self.model.diffusion_model.state_unet.parameters())
mainlogger.info(f"@Training [{len(params)}] Full Paramters.")
if self.cond_stage_trainable:
params_cond_stage = [
p for p in self.cond_stage_model.parameters()
if p.requires_grad == True
]
mainlogger.info(
f"@Training [{len(params_cond_stage)}] Paramters for Cond_stage_model."
)
params.extend(params_cond_stage)
if self.image_proj_model_trainable:
mainlogger.info(
f"@Training [{len(list(self.image_proj_model.parameters()))}] Paramters for Image_proj_model."
)
params.extend(list(self.image_proj_model.parameters()))
if self.learn_logvar:
mainlogger.info('Diffusion model optimizing logvar')
if isinstance(params[0], dict):
params.append({"params": [self.logvar]})
else:
params.append(self.logvar)
params_group = [{
'params': params,
'lr': lr
}, {
'params':
params_unet_head,
'lr':
self.dp_optimizer_config['params']['lr'],
'betas':
self.dp_optimizer_config['params']['betas'],
'eps':
self.dp_optimizer_config['params']['eps'],
'weight_decay':
self.dp_optimizer_config['params']['weight_decay']
}]
optimizer = torch.optim.AdamW(params_group, lr=lr)
if self.use_scheduler:
# mainlogger.info("Setting up scheduler...")
lr_scheduler = get_scheduler(
self.lr_scheduler,
optimizer=optimizer,
num_warmup_steps=self.lr_warmup_steps,
num_training_steps=(self.datasets_len * self.num_epochs) //
self.gradient_accumulate_every, # 50 is handcode
last_epoch=-1)
scheduler = SelectiveLRScheduler(
optimizer=optimizer,
base_scheduler=lr_scheduler,
group_indices=[1],
default_lr=[lr, self.dp_optimizer_config['params']['lr']])
return [optimizer], [{'scheduler': scheduler, 'interval': 'step'}]
return [optimizer]
class DiffusionWrapper(pl.LightningModule):
"""Thin wrapper that routes inputs/conditions to the underlying diffusion model."""
def __init__(self, diff_model_config: OmegaConf,
conditioning_key: str | None) -> None:
"""
Args:
diff_model_config: OmegaConf describing the inner diffusion model to instantiate.
conditioning_key: How conditioning is applied.
"""
super().__init__()
self.diffusion_model = instantiate_from_config(diff_model_config)
self.conditioning_key = conditioning_key
def forward(
self,
x: Tensor,
x_action: Tensor | None,
x_state: Tensor | None,
t: Tensor,
c_concat: Sequence[Tensor] | None = None,
c_crossattn: Sequence[Tensor] | None = None,
c_crossattn_action: list[Any] | None = None,
c_adm: Tensor | None = None,
s: Tensor | None = None,
mask: Tensor | None = None,
**kwargs: Any,
) -> Any:
"""
Route input(s) and condition(s) into the inner diffusion model based on `conditioning_key`.
Args:
x: Primary input tensor (e.g., latent/image) at timestep `t`.
x_action: Action stream tensor (used by 'hybrid' variants).
x_state: State stream tensor (used by 'hybrid' variants).
t: Timestep indices (B,).
c_concat: List of tensors to be concatenated channel-wise with `x` (for 'concat' / 'hybrid' modes).
c_crossattn: List of context tensors concatenated along sequence/channel dim for cross-attention.
c_crossattn_action: Mixed list used by action/state heads.
c_adm: Class/ADM conditioning (e.g., labels) when required by '*adm*' modes.
s: Optional additional time-like / scalar conditioning (e.g., fps/frame-stride) for '*time*' modes.
mask: Optional spatial/temporal mask (e.g., inpainting) for '*mask*' modes.
**kwargs: Extra keyword arguments forwarded to the inner diffusion model.
Returns:
Output from the inner diffusion model (tensor or tuple, depending on the model).
"""
if self.conditioning_key is None:
out = self.diffusion_model(x, t)
elif self.conditioning_key == 'concat':
xc = torch.cat([x] + c_concat, dim=1)
out = self.diffusion_model(xc, t, **kwargs)
elif self.conditioning_key == 'crossattn':
cc = torch.cat(c_crossattn, 1)
out = self.diffusion_model(x, t, context=cc, **kwargs)
elif self.conditioning_key == 'hybrid':
xc = torch.cat([x] + c_concat, dim=1)
cc = torch.cat(c_crossattn, 1)
cc_action = c_crossattn_action
out = self.diffusion_model(xc,
x_action,
x_state,
t,
context=cc,
context_action=cc_action,
**kwargs)
elif self.conditioning_key == 'resblockcond':
cc = c_crossattn[0]
out = self.diffusion_model(x, t, context=cc)
elif self.conditioning_key == 'adm':
cc = c_crossattn[0]
out = self.diffusion_model(x, t, y=cc)
elif self.conditioning_key == 'hybrid-adm':
assert c_adm is not None
xc = torch.cat([x] + c_concat, dim=1)
cc = torch.cat(c_crossattn, 1)
out = self.diffusion_model(xc, t, context=cc, y=c_adm, **kwargs)
elif self.conditioning_key == 'hybrid-time':
assert s is not None
xc = torch.cat([x] + c_concat, dim=1)
cc = torch.cat(c_crossattn, 1)
out = self.diffusion_model(xc, t, context=cc, s=s)
elif self.conditioning_key == 'concat-time-mask':
xc = torch.cat([x] + c_concat, dim=1)
out = self.diffusion_model(xc, t, context=None, s=s, mask=mask)
elif self.conditioning_key == 'concat-adm-mask':
if c_concat is not None:
xc = torch.cat([x] + c_concat, dim=1)
else:
xc = x
out = self.diffusion_model(xc, t, context=None, y=s, mask=mask)
elif self.conditioning_key == 'hybrid-adm-mask':
cc = torch.cat(c_crossattn, 1)
if c_concat is not None:
xc = torch.cat([x] + c_concat, dim=1)
else:
xc = x
out = self.diffusion_model(xc, t, context=cc, y=s, mask=mask)
elif self.conditioning_key == 'hybrid-time-adm':
assert c_adm is not None
xc = torch.cat([x] + c_concat, dim=1)
cc = torch.cat(c_crossattn, 1)
out = self.diffusion_model(xc, t, context=cc, s=s, y=c_adm)
elif self.conditioning_key == 'crossattn-adm':
assert c_adm is not None
cc = torch.cat(c_crossattn, 1)
out = self.diffusion_model(x, t, context=cc, y=c_adm)
else:
raise NotImplementedError()
return out