418 lines
15 KiB
Python
418 lines
15 KiB
Python
import inspect
|
|
from logging import getLogger
|
|
from typing import Callable, List, Optional, Union
|
|
|
|
import numpy as np
|
|
import torch
|
|
import torch.nn.functional as F
|
|
from diffusers.models.unet_2d_condition import UNet2DConditionOutput
|
|
from diffusers.pipelines.onnx_utils import ORT_TO_NP_TYPE, OnnxRuntimeModel
|
|
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
|
|
from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
|
|
from transformers import CLIPImageProcessor, CLIPTokenizer
|
|
|
|
from ...constants import LATENT_CHANNELS, LATENT_FACTOR, ONNX_MODEL
|
|
from ...convert.utils import onnx_export
|
|
from .base import OnnxStableDiffusionBasePipeline
|
|
|
|
logger = getLogger(__name__)
|
|
|
|
|
|
class OnnxStableDiffusionHighresPipeline(OnnxStableDiffusionBasePipeline):
|
|
upscaler: OnnxRuntimeModel
|
|
|
|
def __init__(
|
|
self,
|
|
vae_encoder: OnnxRuntimeModel,
|
|
vae_decoder: OnnxRuntimeModel,
|
|
text_encoder: OnnxRuntimeModel,
|
|
tokenizer: CLIPTokenizer,
|
|
unet: OnnxRuntimeModel,
|
|
scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
|
|
safety_checker: OnnxRuntimeModel,
|
|
feature_extractor: CLIPImageProcessor,
|
|
requires_safety_checker: bool = True,
|
|
upscaler: OnnxRuntimeModel = None,
|
|
):
|
|
super().__init__(
|
|
vae_encoder=vae_encoder,
|
|
vae_decoder=vae_decoder,
|
|
text_encoder=text_encoder,
|
|
tokenizer=tokenizer,
|
|
unet=unet,
|
|
scheduler=scheduler,
|
|
safety_checker=safety_checker,
|
|
feature_extractor=feature_extractor,
|
|
requires_safety_checker=requires_safety_checker,
|
|
)
|
|
|
|
self.upscaler = upscaler
|
|
|
|
@torch.no_grad()
|
|
def text2img(
|
|
self,
|
|
prompt: Union[str, List[str]] = None,
|
|
height: Optional[int] = 512,
|
|
width: Optional[int] = 512,
|
|
num_inference_steps: Optional[int] = 50,
|
|
num_upscale_steps: Optional[int] = 50,
|
|
guidance_scale: Optional[float] = 7.5,
|
|
negative_prompt: Optional[Union[str, List[str]]] = None,
|
|
num_images_per_prompt: Optional[int] = 1,
|
|
eta: Optional[float] = 0.0,
|
|
generator: Optional[np.random.RandomState] = None,
|
|
latents: Optional[np.ndarray] = None,
|
|
prompt_embeds: Optional[np.ndarray] = None,
|
|
negative_prompt_embeds: Optional[np.ndarray] = None,
|
|
output_type: Optional[str] = "pil",
|
|
return_dict: bool = True,
|
|
callback: Optional[Callable[[int, int, np.ndarray], None]] = None,
|
|
callback_steps: int = 1,
|
|
):
|
|
# check inputs. Raise error if not correct
|
|
self.check_inputs(
|
|
prompt,
|
|
height,
|
|
width,
|
|
callback_steps,
|
|
negative_prompt,
|
|
prompt_embeds,
|
|
negative_prompt_embeds,
|
|
)
|
|
|
|
# define call parameters
|
|
if prompt is not None and isinstance(prompt, str):
|
|
batch_size = 1
|
|
elif prompt is not None and isinstance(prompt, list):
|
|
batch_size = len(prompt)
|
|
else:
|
|
batch_size = prompt_embeds.shape[0]
|
|
|
|
if generator is None:
|
|
generator = np.random
|
|
|
|
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
|
|
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
|
|
# corresponds to doing no classifier free guidance.
|
|
do_classifier_free_guidance = guidance_scale > 1.0
|
|
|
|
prompt_embeds, text_pooler_out = self._encode_prompt(
|
|
prompt,
|
|
num_images_per_prompt,
|
|
do_classifier_free_guidance,
|
|
negative_prompt,
|
|
prompt_embeds=prompt_embeds,
|
|
negative_prompt_embeds=negative_prompt_embeds,
|
|
)
|
|
|
|
# get the initial random noise unless the user supplied it
|
|
latents_dtype = prompt_embeds.dtype
|
|
latents_shape = (
|
|
batch_size * num_images_per_prompt,
|
|
LATENT_CHANNELS,
|
|
height // LATENT_FACTOR,
|
|
width // LATENT_FACTOR,
|
|
)
|
|
if latents is None:
|
|
latents = generator.randn(*latents_shape).astype(latents_dtype)
|
|
elif latents.shape != latents_shape:
|
|
raise ValueError(
|
|
f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}"
|
|
)
|
|
|
|
# set timesteps
|
|
self.scheduler.set_timesteps(num_inference_steps)
|
|
|
|
latents = latents * np.float64(self.scheduler.init_noise_sigma)
|
|
|
|
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
|
|
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
|
|
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
|
|
# and should be between [0, 1]
|
|
accepts_eta = "eta" in set(
|
|
inspect.signature(self.scheduler.step).parameters.keys()
|
|
)
|
|
extra_step_kwargs = {}
|
|
if accepts_eta:
|
|
extra_step_kwargs["eta"] = eta
|
|
|
|
timestep_dtype = next(
|
|
(
|
|
input.type
|
|
for input in self.unet.model.get_inputs()
|
|
if input.name == "timestep"
|
|
),
|
|
"tensor(float)",
|
|
)
|
|
timestep_dtype = ORT_TO_NP_TYPE[timestep_dtype]
|
|
|
|
for i, t in enumerate(self.progress_bar(self.scheduler.timesteps)):
|
|
# expand the latents if we are doing classifier free guidance
|
|
latent_model_input = (
|
|
np.concatenate([latents] * 2)
|
|
if do_classifier_free_guidance
|
|
else latents
|
|
)
|
|
latent_model_input = self.scheduler.scale_model_input(
|
|
torch.from_numpy(latent_model_input), t
|
|
)
|
|
latent_model_input = latent_model_input.cpu().numpy()
|
|
|
|
# predict the noise residual
|
|
timestep = np.array([t], dtype=timestep_dtype)
|
|
noise_pred = self.unet(
|
|
sample=latent_model_input,
|
|
timestep=timestep,
|
|
encoder_hidden_states=prompt_embeds,
|
|
)
|
|
noise_pred = noise_pred[0]
|
|
|
|
# perform guidance
|
|
if do_classifier_free_guidance:
|
|
noise_pred_uncond, noise_pred_text = np.split(noise_pred, 2)
|
|
noise_pred = noise_pred_uncond + guidance_scale * (
|
|
noise_pred_text - noise_pred_uncond
|
|
)
|
|
|
|
# compute the previous noisy sample x_t -> x_t-1
|
|
scheduler_output = self.scheduler.step(
|
|
torch.from_numpy(noise_pred),
|
|
t,
|
|
torch.from_numpy(latents),
|
|
**extra_step_kwargs,
|
|
)
|
|
latents = scheduler_output.prev_sample.numpy()
|
|
|
|
# call the callback, if provided
|
|
if callback is not None and i % callback_steps == 0:
|
|
callback(i, t, latents)
|
|
|
|
if self.upscaler is not None:
|
|
# 5. set upscale timesteps
|
|
self.scheduler.set_timesteps(num_upscale_steps)
|
|
timesteps = self.scheduler.timesteps
|
|
|
|
batch_multiplier = 2 if do_classifier_free_guidance else 1
|
|
image = np.concatenate([latents] * batch_multiplier)
|
|
|
|
# 5. Add noise to image (set to be 0):
|
|
# (see below notes from the author):
|
|
# "the This step theoretically can make the model work better on out-of-distribution inputs, but mostly
|
|
# just seems to make it match the input less, so it's turned off by default."
|
|
noise_level = np.array([0.0], dtype=np.float32)
|
|
noise_level = np.concatenate([noise_level] * image.shape[0])
|
|
inv_noise_level = (noise_level**2 + 1) ** (-0.5)
|
|
|
|
image_cond = (
|
|
F.interpolate(torch.tensor(image), scale_factor=2, mode="nearest")
|
|
* inv_noise_level[:, None, None, None]
|
|
)
|
|
image_cond = image_cond.numpy().astype(prompt_embeds.dtype)
|
|
|
|
noise_level_embed = np.concatenate(
|
|
[
|
|
np.ones(
|
|
(text_pooler_out.shape[0], 64), dtype=text_pooler_out.dtype
|
|
),
|
|
np.zeros(
|
|
(text_pooler_out.shape[0], 64), dtype=text_pooler_out.dtype
|
|
),
|
|
],
|
|
axis=1,
|
|
)
|
|
|
|
# upscaling latents
|
|
latents_shape = (
|
|
batch_size * num_images_per_prompt,
|
|
LATENT_CHANNELS,
|
|
height * 2 // LATENT_FACTOR,
|
|
width * 2 // LATENT_FACTOR,
|
|
)
|
|
latents = generator.randn(*latents_shape).astype(latents_dtype)
|
|
|
|
timestep_condition = np.concatenate(
|
|
[noise_level_embed, text_pooler_out], axis=1
|
|
)
|
|
|
|
num_warmup_steps = 0
|
|
|
|
with self.progress_bar(total=num_upscale_steps) as progress_bar:
|
|
for i, t in enumerate(timesteps):
|
|
sigma = self.scheduler.sigmas[i]
|
|
# expand the latents if we are doing classifier free guidance
|
|
latent_model_input = (
|
|
np.concatenate([latents] * 2)
|
|
if do_classifier_free_guidance
|
|
else latents
|
|
)
|
|
scaled_model_input = self.scheduler.scale_model_input(
|
|
latent_model_input, t
|
|
)
|
|
|
|
scaled_model_input = np.concatenate(
|
|
[scaled_model_input, image_cond], axis=1
|
|
)
|
|
# preconditioning parameter based on Karras et al. (2022) (table 1)
|
|
timestep = np.log(sigma) * 0.25
|
|
|
|
noise_pred = self.upscaler(
|
|
sample=scaled_model_input,
|
|
timestep=timestep,
|
|
encoder_hidden_states=prompt_embeds,
|
|
timestep_cond=timestep_condition,
|
|
).sample
|
|
|
|
# in original repo, the output contains a variance channel that's not used
|
|
noise_pred = noise_pred[:, :-1]
|
|
|
|
# apply preconditioning, based on table 1 in Karras et al. (2022)
|
|
inv_sigma = 1 / (sigma**2 + 1)
|
|
noise_pred = (
|
|
inv_sigma * latent_model_input
|
|
+ self.scheduler.scale_model_input(sigma, t) * noise_pred
|
|
)
|
|
|
|
# perform guidance
|
|
if do_classifier_free_guidance:
|
|
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
|
noise_pred = noise_pred_uncond + guidance_scale * (
|
|
noise_pred_text - noise_pred_uncond
|
|
)
|
|
|
|
# compute the previous noisy sample x_t -> x_t-1
|
|
scheduler_output = self.scheduler.step(
|
|
noise_pred, t, torch.from_numpy(latents)
|
|
)
|
|
latents = scheduler_output.prev_sample.numpy()
|
|
|
|
# call the callback, if provided
|
|
if i == len(timesteps) - 1 or (
|
|
(i + 1) > num_warmup_steps
|
|
and (i + 1) % self.scheduler.order == 0
|
|
):
|
|
progress_bar.update()
|
|
if callback is not None and i % callback_steps == 0:
|
|
callback(i, t, latents)
|
|
else:
|
|
logger.debug("skipping latent upscaler, no model provided")
|
|
|
|
# decode image
|
|
latents = 1 / 0.18215 * latents
|
|
|
|
# it seems likes there is a strange result for using half-precision vae decoder if batchsize>1
|
|
image = np.concatenate(
|
|
[
|
|
self.vae_decoder(latent_sample=latents[i : i + 1])[0]
|
|
for i in range(latents.shape[0])
|
|
]
|
|
)
|
|
|
|
image = np.clip(image / 2 + 0.5, 0, 1)
|
|
image = image.transpose((0, 2, 3, 1))
|
|
|
|
if output_type == "pil":
|
|
image = self.numpy_to_pil(image)
|
|
|
|
if not return_dict:
|
|
return (image, None)
|
|
|
|
return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=None)
|
|
|
|
|
|
def export_unet(pipeline, output_path, unet_sample_size=1024):
|
|
device = torch.device("cpu")
|
|
dtype = torch.float32
|
|
|
|
num_tokens = pipeline.text_encoder.config.max_position_embeddings
|
|
text_hidden_size = pipeline.text_encoder.config.hidden_size
|
|
|
|
unet_inputs = ["sample", "timestep", "encoder_hidden_states", "timestep_cond"]
|
|
unet_in_channels = pipeline.unet.config.in_channels
|
|
unet_path = output_path / "unet" / ONNX_MODEL
|
|
|
|
logger.info("exporting UNet to %s", unet_path)
|
|
onnx_export(
|
|
pipeline.unet,
|
|
model_args=(
|
|
torch.randn(
|
|
2,
|
|
unet_in_channels,
|
|
unet_sample_size // LATENT_FACTOR,
|
|
unet_sample_size // LATENT_FACTOR,
|
|
).to(device=device, dtype=dtype),
|
|
torch.randn(2).to(device=device, dtype=dtype),
|
|
torch.randn(2, num_tokens, text_hidden_size).to(device=device, dtype=dtype),
|
|
torch.randn(2, 64, 64, 2).to(
|
|
device=device, dtype=dtype
|
|
), # TODO: not the right shape
|
|
),
|
|
output_path=unet_path,
|
|
ordered_input_names=unet_inputs,
|
|
# has to be different from "sample" for correct tracing
|
|
output_names=["out_sample"],
|
|
dynamic_axes={
|
|
"sample": {0: "batch"}, # , 1: "channels", 2: "height", 3: "width"},
|
|
"timestep": {0: "batch"},
|
|
"encoder_hidden_states": {0: "batch", 1: "sequence"},
|
|
},
|
|
opset=14,
|
|
half=False,
|
|
external_data=True,
|
|
v2=False,
|
|
)
|
|
|
|
|
|
def load_and_export(output, source="stabilityai/sd-x2-latent-upscaler"):
|
|
from pathlib import Path
|
|
|
|
from diffusers import StableDiffusionLatentUpscalePipeline
|
|
|
|
upscaler = StableDiffusionLatentUpscalePipeline.from_pretrained(
|
|
source, torch_dtype=torch.float32
|
|
)
|
|
export_unet(upscaler, Path(output))
|
|
|
|
|
|
def load_and_run(
|
|
prompt,
|
|
output,
|
|
source="stabilityai/sd-x2-latent-upscaler",
|
|
checkpoint="../models/stable-diffusion-onnx-v1-5",
|
|
):
|
|
from diffusers import (
|
|
EulerAncestralDiscreteScheduler,
|
|
StableDiffusionLatentUpscalePipeline,
|
|
)
|
|
|
|
upscaler = StableDiffusionLatentUpscalePipeline.from_pretrained(source)
|
|
highres = OnnxStableDiffusionHighresPipeline.from_pretrained(checkpoint)
|
|
scheduler = EulerAncestralDiscreteScheduler.from_pretrained(
|
|
f"{checkpoint}/scheduler"
|
|
)
|
|
|
|
# combine them
|
|
highres.scheduler = scheduler
|
|
highres.upscaler = RetorchModel(upscaler.unet)
|
|
|
|
# run
|
|
result = highres.text2img(prompt, num_inference_steps=25, num_upscale_steps=25)
|
|
image = result.images[0]
|
|
image.save(output)
|
|
|
|
|
|
class RetorchModel:
|
|
"""
|
|
Shim back from ONNX to PyTorch
|
|
"""
|
|
|
|
def __init__(self, model) -> None:
|
|
self.model = model
|
|
|
|
def __call__(self, **kwargs):
|
|
inputs = {
|
|
k: torch.from_numpy(v) if isinstance(v, np.ndarray) else v
|
|
for k, v in kwargs.items()
|
|
}
|
|
outputs = self.model(**inputs)
|
|
return UNet2DConditionOutput(sample=outputs.sample.numpy())
|