|
|
""" |
|
|
Code is adapted from https://github.com/CompVis/stable-diffusion/blob/21f890f9da3cfbeaba8e2ac3c425ee9e998d5229/ldm/models/diffusion/ddpm.py |
|
|
""" |
|
|
import warnings |
|
|
from typing import Sequence, Union, Dict, Any, Optional, Callable |
|
|
from copy import deepcopy |
|
|
from contextlib import contextmanager |
|
|
from functools import partial |
|
|
from tqdm import tqdm |
|
|
import torch |
|
|
import torch.nn as nn |
|
|
import numpy as np |
|
|
from einops import rearrange |
|
|
import lightning.pytorch as pl |
|
|
from lightning.pytorch.utilities.rank_zero import rank_zero_only |
|
|
from diffusers.models.autoencoder_kl import AutoencoderKLOutput, DecoderOutput |
|
|
|
|
|
from .utils import make_beta_schedule, extract_into_tensor, noise_like, default |
|
|
from models.model_utils.distributions import DiagonalGaussianDistribution |
|
|
from utils.ema import LitEma |
|
|
from utils.layout import parse_layout_shape |
|
|
from utils.optim import disabled_train |
|
|
|
|
|
|
|
|
class LatentDiffusion(pl.LightningModule): |
|
|
def __init__(self, |
|
|
torch_nn_module: nn.Module, |
|
|
layout: str = "NTHWC", |
|
|
data_shape: Sequence[int] = (10, 128, 128, 4), |
|
|
timesteps=1000, |
|
|
beta_schedule="linear", |
|
|
loss_type="l2", |
|
|
monitor="val/loss", |
|
|
use_ema=True, |
|
|
log_every_t=100, |
|
|
clip_denoised=False, |
|
|
linear_start=1e-4, |
|
|
linear_end=2e-2, |
|
|
cosine_s=8e-3, |
|
|
given_betas=None, |
|
|
original_elbo_weight=0., |
|
|
v_posterior=0., |
|
|
l_simple_weight=1., |
|
|
parameterization="eps", |
|
|
learn_logvar=False, |
|
|
logvar_init=0., |
|
|
|
|
|
latent_shape: Sequence[int] = (10, 16, 16, 4), |
|
|
first_stage_model: nn.Module = None, |
|
|
cond_stage_model: Union[str, nn.Module] = None, |
|
|
num_timesteps_cond=None, |
|
|
cond_stage_trainable=False, |
|
|
cond_stage_forward=None, |
|
|
scale_by_std=False, |
|
|
scale_factor=1.0, |
|
|
): |
|
|
r""" |
|
|
Parameters |
|
|
---------- |
|
|
|
|
|
torch_nn_module: nn.Module |
|
|
The `.forward()` method of model should have the following signature: |
|
|
`x_hat = model.forward(x, t, *args, **kwargs)` |
|
|
layout: str |
|
|
e.g., "NTHWC", "NHWC". |
|
|
data_shape: Sequence[int] |
|
|
The shape of each data entry. Corresponds to `layout` without the batch axis "N". |
|
|
timesteps: int |
|
|
1000 by default. |
|
|
beta_schedule: str |
|
|
one of ["linear", "cosine", "sqrt_linear", "sqrt"]. |
|
|
loss_type: str |
|
|
one of ["l2", "l1"]. |
|
|
monitor: str |
|
|
name of logged var for selecting best val model. |
|
|
use_ema: bool |
|
|
log_every_t: int |
|
|
log intermediate denoising steps. Should be <= `timesteps`. |
|
|
clip_denoised: bool |
|
|
linear_start: float |
|
|
linear_end: float |
|
|
cosine_s: float |
|
|
given_betas: Optional |
|
|
If provided, `linear_start`, `linear_end`, `cosine_s` take no effect. |
|
|
If None, `linear_start`, `linear_end`, `cosine_s` are used to generate betas via `make_beta_schedule()`. |
|
|
original_elbo_weight: float |
|
|
0. by default |
|
|
v_posterior: float |
|
|
0. by default |
|
|
weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta |
|
|
l_simple_weight: float |
|
|
1. by default |
|
|
parameterization: str |
|
|
"eps" by default, to predict the noise from `t` to `t-1`. |
|
|
"x0" to predict the `x_{t-1}` from `x_t`. |
|
|
all assuming fixed variance schedules. |
|
|
learn_logvar: bool |
|
|
use fixed var by default. |
|
|
logvar_init: float |
|
|
(initial) values of `logvar`. |
|
|
|
|
|
latent_shape: Sequence[int] |
|
|
The shape of downsampled data entry. Corresponds to `layout` without the batch axis "N". |
|
|
first_stage_model: nn.Module |
|
|
nn.Module : a model that has method ".encode()" to encode the inputs. |
|
|
cond_stage_model: str or nn.Module |
|
|
"__is_first_stage__": use the first_stage_model also for encoding conditionings. |
|
|
nn.Module : a model that has method ".encode()" or use `self()` to encodes the conditionings. |
|
|
cond_stage_trainable: bool |
|
|
Whether to train the cond_stage_model jointly |
|
|
num_timesteps_cond: int |
|
|
cond_stage_forward: str |
|
|
The name of the forward method of the cond_stage_model. |
|
|
scale_by_std |
|
|
scale_factor |
|
|
""" |
|
|
super(LatentDiffusion, self).__init__() |
|
|
assert parameterization in ["eps", "x0"], 'currently only supporting "eps" and "x0"' |
|
|
self.parameterization = parameterization |
|
|
print(f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode") |
|
|
self.clip_denoised = clip_denoised |
|
|
self.log_every_t = log_every_t |
|
|
self.torch_nn_module = torch_nn_module |
|
|
self.layout = layout |
|
|
self.data_shape = data_shape |
|
|
self.parse_layout_shape(layout=layout) |
|
|
self.use_ema = use_ema |
|
|
if self.use_ema: |
|
|
self.model_ema = LitEma(self.torch_nn_module) |
|
|
print(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 |
|
|
|
|
|
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) |
|
|
|
|
|
self.loss_type = loss_type |
|
|
|
|
|
self.learn_logvar = learn_logvar |
|
|
logvar = torch.full(fill_value=logvar_init, size=(self.num_timesteps,)) |
|
|
if self.learn_logvar: |
|
|
self.logvar = nn.Parameter(logvar, requires_grad=True) |
|
|
else: |
|
|
self.register_buffer('logvar', logvar) |
|
|
|
|
|
self.latent_shape = latent_shape |
|
|
self.num_timesteps_cond = default(num_timesteps_cond, 1) |
|
|
assert self.num_timesteps_cond <= timesteps |
|
|
self.shorten_cond_schedule = self.num_timesteps_cond > 1 |
|
|
if self.shorten_cond_schedule: |
|
|
self.make_cond_schedule() |
|
|
|
|
|
self.cond_stage_trainable = cond_stage_trainable |
|
|
self.scale_by_std = scale_by_std |
|
|
if not scale_by_std: |
|
|
self.scale_factor = scale_factor |
|
|
else: |
|
|
self.register_buffer('scale_factor', torch.tensor(scale_factor)) |
|
|
|
|
|
self.instantiate_first_stage(first_stage_model) |
|
|
self.instantiate_cond_stage(cond_stage_model, cond_stage_forward) |
|
|
|
|
|
def set_alignment(self, alignment_fn: Callable = None): |
|
|
r""" |
|
|
Call this method to set alignment after __init__ of LatentDiffusion, |
|
|
to avoid error "cannot assign module before Module.__init__() call" |
|
|
when assigning alignment model to the LatentDiffusion before its __init__. |
|
|
|
|
|
Parameters |
|
|
---------- |
|
|
alignment_fn: Callable |
|
|
Should have signature `alignment_fn(zt, t, zc=None, y=None, xt=None, **kwargs)`. |
|
|
""" |
|
|
self.alignment_fn = alignment_fn |
|
|
|
|
|
def parse_layout_shape(self, layout): |
|
|
parsed_dict = parse_layout_shape(layout=layout) |
|
|
self.batch_axis = parsed_dict["batch_axis"] |
|
|
self.t_axis = parsed_dict["t_axis"] |
|
|
self.h_axis = parsed_dict["h_axis"] |
|
|
self.w_axis = parsed_dict["w_axis"] |
|
|
self.c_axis = parsed_dict["c_axis"] |
|
|
self.all_slice = [slice(None, None), ] * len(layout) |
|
|
|
|
|
def extract_into_tensor(self, a, t, x_shape): |
|
|
return extract_into_tensor(a=a, t=t, x_shape=x_shape, |
|
|
batch_axis=self.batch_axis) |
|
|
|
|
|
@property |
|
|
def loss_mean_dim(self): |
|
|
|
|
|
if not hasattr(self, "_loss_mean_dim"): |
|
|
_loss_mean_dim = list(range(len(self.layout))) |
|
|
_loss_mean_dim.pop(self.batch_axis) |
|
|
self._loss_mean_dim = tuple(_loss_mean_dim) |
|
|
return self._loss_mean_dim |
|
|
|
|
|
def get_batch_data_shape(self, batch_size=1): |
|
|
if not hasattr(self, "batch_data_shape"): |
|
|
_batch_data_shape = deepcopy(list(self.data_shape)) |
|
|
_batch_data_shape.insert(self.batch_axis, batch_size) |
|
|
elif self.batch_data_shape[self.batch_axis] != batch_size: |
|
|
_batch_data_shape = deepcopy(list(self.batch_data_shape)) |
|
|
_batch_data_shape[self.batch_axis] = batch_size |
|
|
else: |
|
|
return self.batch_data_shape |
|
|
self.batch_data_shape = tuple(_batch_data_shape) |
|
|
return self.batch_data_shape |
|
|
|
|
|
def get_batch_latent_shape(self, batch_size=1): |
|
|
if not hasattr(self, "batch_latent_shape"): |
|
|
_batch_latent_shape = deepcopy(list(self.latent_shape)) |
|
|
_batch_latent_shape.insert(self.batch_axis, batch_size) |
|
|
elif self.batch_latent_shape[self.batch_axis] != batch_size: |
|
|
_batch_latent_shape = deepcopy(list(self.batch_latent_shape)) |
|
|
_batch_latent_shape[self.batch_axis] = batch_size |
|
|
else: |
|
|
return self.batch_latent_shape |
|
|
self.batch_latent_shape = tuple(_batch_latent_shape) |
|
|
return self.batch_latent_shape |
|
|
|
|
|
def register_schedule(self, |
|
|
given_betas=None, beta_schedule="linear", timesteps=1000, |
|
|
linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): |
|
|
if given_betas is not None: |
|
|
betas = given_betas |
|
|
else: |
|
|
betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end, |
|
|
cosine_s=cosine_s) |
|
|
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))) |
|
|
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))) |
|
|
|
|
|
|
|
|
posterior_variance = (1 - self.v_posterior) * betas * (1. - alphas_cumprod_prev) / (1. - alphas_cumprod) + self.v_posterior * betas |
|
|
|
|
|
self.register_buffer('posterior_variance', to_torch(posterior_variance)) |
|
|
|
|
|
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)) |
|
|
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=None): |
|
|
if self.use_ema: |
|
|
self.model_ema.store(self.torch_nn_module.parameters()) |
|
|
self.model_ema.copy_to(self.torch_nn_module) |
|
|
if context is not None: |
|
|
print(f"{context}: Switched to EMA weights") |
|
|
try: |
|
|
yield None |
|
|
finally: |
|
|
if self.use_ema: |
|
|
self.model_ema.restore(self.torch_nn_module.parameters()) |
|
|
if context is not None: |
|
|
print(f"{context}: Restored training weights") |
|
|
|
|
|
def make_cond_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() |
|
|
cond_ids[:self.num_timesteps_cond] = ids |
|
|
self.register_buffer('cond_ids', cond_ids) |
|
|
|
|
|
@rank_zero_only |
|
|
@torch.no_grad() |
|
|
def on_train_batch_start(self, batch, batch_idx): |
|
|
|
|
|
|
|
|
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' |
|
|
|
|
|
print("### USING STD-RESCALING ###") |
|
|
x, _ = self.get_input(batch) |
|
|
x = x.to(self.device) |
|
|
x = rearrange(x, f"{self.einops_layout} -> {self.einops_spatial_layout}") |
|
|
z = self.encode_first_stage(x) |
|
|
del self.scale_factor |
|
|
self.register_buffer('scale_factor', 1. / z.flatten().std()) |
|
|
print(f"setting self.scale_factor to {self.scale_factor}") |
|
|
print("### USING STD-RESCALING ###") |
|
|
|
|
|
def instantiate_first_stage(self, first_stage_model): |
|
|
if isinstance(first_stage_model, nn.Module): |
|
|
model = first_stage_model |
|
|
else: |
|
|
assert first_stage_model is None |
|
|
raise NotImplementedError("No default first_stage_model supported yet!") |
|
|
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, cond_stage_model, cond_stage_forward): |
|
|
if cond_stage_model is None: |
|
|
self.cond_stage_model = None |
|
|
self.cond_stage_forward = None |
|
|
return |
|
|
|
|
|
is_first_stage_flag = cond_stage_model == "__is_first_stage__" |
|
|
if cond_stage_model == "__is_first_stage__": |
|
|
model = self.first_stage_model |
|
|
if self.cond_stage_trainable: |
|
|
warnings.warn("`cond_stage_trainable` is True while `cond_stage_model` is '__is_first_stage__'. " |
|
|
"force `cond_stage_trainable` to be False") |
|
|
self.cond_stage_trainable = False |
|
|
elif isinstance(cond_stage_model, nn.Module): |
|
|
model = cond_stage_model |
|
|
else: |
|
|
raise NotImplementedError |
|
|
self.cond_stage_model = model |
|
|
if (self.cond_stage_model is not None) and (not self.cond_stage_trainable): |
|
|
for param in self.cond_stage_model.parameters(): |
|
|
param.requires_grad = False |
|
|
|
|
|
if cond_stage_forward is None: |
|
|
if hasattr(self.cond_stage_model, 'encode') and callable(self.cond_stage_model.encode): |
|
|
cond_stage_forward = self.cond_stage_model.encode |
|
|
else: |
|
|
cond_stage_forward = self.cond_stage_model.__call__ |
|
|
else: |
|
|
assert hasattr(self.cond_stage_model, cond_stage_forward) |
|
|
cond_stage_forward = getattr(self.cond_stage_model, cond_stage_forward) |
|
|
|
|
|
def wrapper(cond_stage_forward: Callable, is_first_stage_flag=False): |
|
|
def func(c: Dict[str, Any]): |
|
|
if is_first_stage_flag: |
|
|
|
|
|
|
|
|
c = c.get("y") |
|
|
batch_size = c.shape[self.batch_axis] |
|
|
c = rearrange(c, f"{self.einops_layout} -> {self.einops_spatial_layout}") |
|
|
c = cond_stage_forward(c) |
|
|
if isinstance(c, DiagonalGaussianDistribution): |
|
|
c = c.mode() |
|
|
elif isinstance(c, AutoencoderKLOutput): |
|
|
c = c.latent_dist.mode() |
|
|
else: |
|
|
pass |
|
|
if is_first_stage_flag: |
|
|
c = rearrange(c, f"{self.einops_spatial_layout} -> {self.einops_layout}", N=batch_size) |
|
|
return c |
|
|
return func |
|
|
self.cond_stage_forward = wrapper(cond_stage_forward, is_first_stage_flag) |
|
|
|
|
|
def get_first_stage_encoding(self, encoder_posterior): |
|
|
if isinstance(encoder_posterior, DiagonalGaussianDistribution): |
|
|
z = encoder_posterior.sample() |
|
|
elif isinstance(encoder_posterior, torch.Tensor): |
|
|
z = encoder_posterior |
|
|
elif isinstance(encoder_posterior, AutoencoderKLOutput): |
|
|
z = encoder_posterior.latent_dist.sample() |
|
|
else: |
|
|
raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented") |
|
|
return self.scale_factor * z |
|
|
|
|
|
@property |
|
|
def einops_layout(self): |
|
|
return " ".join(self.layout) |
|
|
|
|
|
@property |
|
|
def einops_spatial_layout(self): |
|
|
if not hasattr(self, "_einops_spatial_layout"): |
|
|
assert len(self.layout) == 4 or len(self.layout) == 5 |
|
|
self._einops_spatial_layout = "(N T) C H W" if self.layout.find("T") else "N C H W" |
|
|
return self._einops_spatial_layout |
|
|
|
|
|
@torch.no_grad() |
|
|
def get_input(self, batch, **kwargs): |
|
|
r""" |
|
|
dataset dependent |
|
|
re-implement it for each specific dataset |
|
|
|
|
|
Parameters |
|
|
---------- |
|
|
batch: Any |
|
|
raw data batch from specific dataloader |
|
|
|
|
|
Returns |
|
|
------- |
|
|
out: Sequence[torch.Tensor, Dict[str, Any]] |
|
|
out[0] should be a torch.Tensor which is the target to generate |
|
|
out[1] should be a dict consists of several key-value pairs for conditioning |
|
|
""" |
|
|
return batch |
|
|
|
|
|
@torch.no_grad() |
|
|
def decode_first_stage(self, z): |
|
|
z = 1. / self.scale_factor * z |
|
|
batch_size = z.shape[self.batch_axis] |
|
|
z = rearrange(z, f"{self.einops_layout} -> {self.einops_spatial_layout}") |
|
|
output = self.first_stage_model.decode(z) |
|
|
if isinstance(output, DecoderOutput): |
|
|
output = output.sample |
|
|
output = rearrange(output, f"{self.einops_spatial_layout} -> {self.einops_layout}", N=batch_size) |
|
|
return output |
|
|
|
|
|
@torch.no_grad() |
|
|
def encode_first_stage(self, x): |
|
|
encoder_posterior = self.first_stage_model.encode(x) |
|
|
output = self.get_first_stage_encoding(encoder_posterior).detach() |
|
|
return output |
|
|
|
|
|
def apply_model(self, x_noisy, t, cond): |
|
|
x_recon = self.torch_nn_module(x_noisy, t, cond) |
|
|
if isinstance(x_recon, tuple): |
|
|
return x_recon[0] |
|
|
else: |
|
|
return x_recon |
|
|
|
|
|
def forward(self, batch, verbose=False): |
|
|
x, c = self.get_input(batch) |
|
|
if verbose: |
|
|
print("inputs:") |
|
|
print(f"x.shape = {x.shape}") |
|
|
for key, val in c.items(): |
|
|
if hasattr(val, "shape"): |
|
|
print(f"{key}.shape = {val.shape}") |
|
|
batch_size = x.shape[self.batch_axis] |
|
|
x = x.to(self.device) |
|
|
x = rearrange(x, f"{self.einops_layout} -> {self.einops_spatial_layout}") |
|
|
z = self.encode_first_stage(x) |
|
|
if verbose: |
|
|
print("after first stage:") |
|
|
print(f"z.shape = {z.shape}") |
|
|
|
|
|
z = rearrange(z, f"{self.einops_spatial_layout} -> {self.einops_layout}", N=batch_size) |
|
|
|
|
|
t = torch.randint(0, self.num_timesteps, (batch_size,), device=self.device).long() |
|
|
if self.cond_stage_model is not None: |
|
|
assert c is not None |
|
|
zc = self.cond_stage_forward(c) |
|
|
if self.shorten_cond_schedule: |
|
|
tc = self.cond_ids[t] |
|
|
zc = self.q_sample(x_start=zc, t=tc, noise=torch.randn_like(c.float())) |
|
|
if verbose and hasattr(zc, "shape"): |
|
|
print(f"zc.shape = {zc.shape}") |
|
|
else: |
|
|
zc = c if isinstance(c, torch.Tensor) else c.get("y", None) |
|
|
return self.p_losses(z, zc, t, noise=None) |
|
|
|
|
|
def training_step(self, batch, batch_idx): |
|
|
loss, loss_dict = self(batch) |
|
|
|
|
|
self.log_dict(loss_dict, prog_bar=True, logger=True, on_step=True, on_epoch=True) |
|
|
return loss |
|
|
|
|
|
def on_train_batch_end(self, *args, **kwargs): |
|
|
if self.use_ema: |
|
|
self.model_ema(self.torch_nn_module) |
|
|
|
|
|
@torch.no_grad() |
|
|
def validation_step(self, batch, batch_idx): |
|
|
_, loss_dict_no_ema = self(batch) |
|
|
with self.ema_scope(): |
|
|
_, loss_dict_ema = self(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 q_sample(self, x_start, t, noise=None): |
|
|
noise = default(noise, lambda: torch.randn_like(x_start)) |
|
|
return (self.extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start + |
|
|
self.extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise) |
|
|
|
|
|
def get_loss(self, pred, target, mean=True): |
|
|
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, cond, t, noise=None): |
|
|
noise = default(noise, lambda: torch.randn_like(x_start)) |
|
|
x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise) |
|
|
model_output = self.apply_model(x_noisy, t, cond) |
|
|
|
|
|
loss_dict = {} |
|
|
prefix = 'train' if self.training else 'val' |
|
|
|
|
|
|
|
|
if self.parameterization == "x0": |
|
|
target = x_start |
|
|
elif self.parameterization == "eps": |
|
|
target = noise |
|
|
else: |
|
|
raise NotImplementedError() |
|
|
|
|
|
loss_simple = self.get_loss(model_output, target, mean=False).mean(dim=self.loss_mean_dim) |
|
|
loss_dict.update({f'{prefix}/loss_simple': loss_simple.mean()}) |
|
|
|
|
|
logvar_t = self.logvar[t] |
|
|
loss = loss_simple / torch.exp(logvar_t) + logvar_t |
|
|
|
|
|
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=self.loss_mean_dim) |
|
|
loss_vlb = (self.lvlb_weights[t] * loss_vlb).mean() |
|
|
loss_dict.update({f'{prefix}/loss_vlb': loss_vlb}) |
|
|
loss += (self.original_elbo_weight * loss_vlb) |
|
|
loss_dict.update({f'{prefix}/loss': loss}) |
|
|
|
|
|
return loss, loss_dict |
|
|
|
|
|
def predict_start_from_noise(self, x_t, t, noise): |
|
|
return ( |
|
|
self.extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - |
|
|
self.extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise |
|
|
) |
|
|
|
|
|
def q_posterior(self, x_start, x_t, t): |
|
|
posterior_mean = ( |
|
|
self.extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * x_start + |
|
|
self.extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t |
|
|
) |
|
|
posterior_variance = self.extract_into_tensor(self.posterior_variance, t, x_t.shape) |
|
|
posterior_log_variance_clipped = self.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, zt, zc, t, clip_denoised: bool, |
|
|
return_x0=False, score_corrector=None, corrector_kwargs=None): |
|
|
t_in = t |
|
|
model_out = self.apply_model(zt, t_in, zc) |
|
|
|
|
|
if score_corrector is not None: |
|
|
assert self.parameterization == "eps" |
|
|
model_out = score_corrector.modify_score(self, model_out, zt, t, zc, **corrector_kwargs) |
|
|
|
|
|
if self.parameterization == "eps": |
|
|
z_recon = self.predict_start_from_noise(zt, t=t, noise=model_out) |
|
|
elif self.parameterization == "x0": |
|
|
z_recon = model_out |
|
|
else: |
|
|
raise NotImplementedError() |
|
|
|
|
|
if clip_denoised: |
|
|
z_recon.clamp_(-1., 1.) |
|
|
model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=z_recon, x_t=zt, t=t) |
|
|
if return_x0: |
|
|
return model_mean, posterior_variance, posterior_log_variance, z_recon |
|
|
else: |
|
|
return model_mean, posterior_variance, posterior_log_variance |
|
|
|
|
|
def aligned_mean(self, zt, t, zc, y, |
|
|
orig_mean, orig_log_var, **kwargs): |
|
|
align_gradient = self.alignment_fn(zt, t, zc=zc, y=y, **kwargs) |
|
|
new_mean = orig_mean - (0.5 * orig_log_var).exp() * align_gradient |
|
|
return new_mean |
|
|
|
|
|
@torch.no_grad() |
|
|
def p_sample(self, zt, zc, t, y=None, use_alignment=False, alignment_kwargs=None, |
|
|
clip_denoised=False, return_x0=False, |
|
|
temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None): |
|
|
batch_size = zt.shape[self.batch_axis] |
|
|
device = zt.device |
|
|
outputs = self.p_mean_variance(zt=zt, zc=zc, t=t, clip_denoised=clip_denoised, |
|
|
return_x0=return_x0, |
|
|
score_corrector=score_corrector, corrector_kwargs=corrector_kwargs) |
|
|
if use_alignment: |
|
|
if alignment_kwargs is None: |
|
|
alignment_kwargs = {} |
|
|
model_mean, posterior_variance, model_log_variance, *_ = outputs |
|
|
model_mean = self.aligned_mean(zt=zt, t=t, zc=zc, y=y, |
|
|
orig_mean=model_mean, orig_log_var=model_log_variance, |
|
|
**alignment_kwargs) |
|
|
outputs = (model_mean, posterior_variance, model_log_variance, *outputs[3:]) |
|
|
if return_x0: |
|
|
model_mean, _, model_log_variance, x0 = outputs |
|
|
else: |
|
|
model_mean, _, model_log_variance = outputs |
|
|
|
|
|
noise = noise_like(zt.shape, device) * temperature |
|
|
if noise_dropout > 0.: |
|
|
noise = torch.nn.functional.dropout(noise, p=noise_dropout) |
|
|
|
|
|
nonzero_mask_shape = [1, ] * len(zt.shape) |
|
|
nonzero_mask_shape[self.batch_axis] = batch_size |
|
|
nonzero_mask = (1 - (t == 0).float()).reshape(*nonzero_mask_shape) |
|
|
|
|
|
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, shape, y=None, |
|
|
use_alignment=False, alignment_kwargs=None, |
|
|
return_intermediates=False, x_T=None, |
|
|
verbose=False, callback=None, timesteps=None, |
|
|
mask=None, x0=None, img_callback=None, start_T=None, |
|
|
log_every_t=None): |
|
|
|
|
|
if not log_every_t: |
|
|
log_every_t = self.log_every_t |
|
|
device = self.betas.device |
|
|
batch_size = shape[self.batch_axis] |
|
|
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((batch_size,), i, device=device, dtype=torch.long) |
|
|
if self.shorten_cond_schedule: |
|
|
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(zt=img, zc=cond, t=ts, y=y, |
|
|
use_alignment=use_alignment, |
|
|
alignment_kwargs=alignment_kwargs, |
|
|
clip_denoised=self.clip_denoised, ) |
|
|
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=16, |
|
|
use_alignment=False, alignment_kwargs=None, |
|
|
return_intermediates=False, x_T=None, |
|
|
verbose=False, timesteps=None, |
|
|
mask=None, x0=None, shape=None, return_decoded=True, **kwargs): |
|
|
if use_alignment: |
|
|
assert self.alignment_fn is not None, "Alignment function not set." |
|
|
if shape is None: |
|
|
shape = self.get_batch_latent_shape(batch_size=batch_size) |
|
|
if self.cond_stage_model is not None: |
|
|
assert cond is not None |
|
|
cond_tensor_slice = [slice(None, None), ] * len(self.data_shape) |
|
|
cond_tensor_slice[self.batch_axis] = slice(0, batch_size) |
|
|
if isinstance(cond, dict): |
|
|
zc = {key: cond[key][cond_tensor_slice] if not isinstance(cond[key], list) else |
|
|
list(map(lambda x: x[cond_tensor_slice], cond[key])) for key in cond} |
|
|
else: |
|
|
zc = [c[cond_tensor_slice] for c in cond] if isinstance(cond, list) else cond[cond_tensor_slice] |
|
|
zc = self.cond_stage_forward(zc) |
|
|
else: |
|
|
zc = cond if isinstance(cond, torch.Tensor) else cond.get("y", None) |
|
|
y = cond if isinstance(cond, torch.Tensor) else cond.get("y", None) |
|
|
output = self.p_sample_loop( |
|
|
cond=zc, shape=shape, y=y, |
|
|
use_alignment=use_alignment, alignment_kwargs=alignment_kwargs, |
|
|
return_intermediates=return_intermediates, x_T=x_T, |
|
|
verbose=verbose, timesteps=timesteps, |
|
|
mask=mask, x0=x0) |
|
|
|
|
|
if return_decoded: |
|
|
if return_intermediates: |
|
|
samples, intermediates = output |
|
|
decoded_samples = self.decode_first_stage(samples) |
|
|
decoded_intermediates = [self.decode_first_stage(ele) for ele in intermediates] |
|
|
output = [decoded_samples, decoded_intermediates] |
|
|
else: |
|
|
output = self.decode_first_stage(output) |
|
|
return output |
|
|
|
|
|
def configure_optimizers(self): |
|
|
lr = self.learning_rate |
|
|
params = list(self.torch_nn_module.parameters()) |
|
|
if self.cond_stage_trainable: |
|
|
print(f"{self.__class__.__name__}: Also optimizing conditioner params!") |
|
|
params = params + list(self.cond_stage_model.parameters()) |
|
|
if self.learn_logvar: |
|
|
print('Diffusion model optimizing logvar') |
|
|
params.append(self.logvar) |
|
|
opt = torch.optim.AdamW(params, lr=lr) |
|
|
return opt |
|
|
|