2023-02-11 21:02:27 +00:00
|
|
|
###
|
|
|
|
# Parts of this file are copied or derived from:
|
|
|
|
# https://github.com/huggingface/diffusers/blob/main/scripts/convert_stable_diffusion_checkpoint_to_onnx.py
|
|
|
|
#
|
|
|
|
# Originally by https://github.com/huggingface
|
|
|
|
# Those portions *are not* covered by the MIT licensed used for the rest of the onnx-web project.
|
|
|
|
#
|
|
|
|
# HuggingFace code used under the Apache License, Version 2.0
|
|
|
|
# https://github.com/huggingface/diffusers/blob/main/LICENSE
|
|
|
|
###
|
|
|
|
|
2023-02-11 04:41:24 +00:00
|
|
|
from logging import getLogger
|
|
|
|
from os import mkdir, path
|
|
|
|
from pathlib import Path
|
|
|
|
from shutil import rmtree
|
|
|
|
from typing import Dict
|
|
|
|
|
|
|
|
import torch
|
2023-01-28 23:09:19 +00:00
|
|
|
from diffusers import (
|
2023-02-17 04:52:25 +00:00
|
|
|
AutoencoderKL,
|
2023-01-28 23:09:19 +00:00
|
|
|
OnnxRuntimeModel,
|
|
|
|
OnnxStableDiffusionPipeline,
|
|
|
|
StableDiffusionPipeline,
|
|
|
|
)
|
2023-02-09 04:35:54 +00:00
|
|
|
from onnx import load, save_model
|
2023-02-11 04:41:24 +00:00
|
|
|
from torch.onnx import export
|
2023-01-16 15:57:59 +00:00
|
|
|
|
2023-02-21 05:07:16 +00:00
|
|
|
from ...diffusion.load import optimize_pipeline
|
|
|
|
from ...diffusion.pipeline_onnx_stable_diffusion_upscale import (
|
2023-02-11 04:41:24 +00:00
|
|
|
OnnxStableDiffusionUpscalePipeline,
|
|
|
|
)
|
2023-02-21 05:07:16 +00:00
|
|
|
from ..utils import ConversionContext
|
2023-01-21 14:52:35 +00:00
|
|
|
|
2023-01-28 23:09:19 +00:00
|
|
|
logger = getLogger(__name__)
|
|
|
|
|
|
|
|
|
2023-01-16 23:48:50 +00:00
|
|
|
def onnx_export(
|
|
|
|
model,
|
|
|
|
model_args: tuple,
|
|
|
|
output_path: Path,
|
|
|
|
ordered_input_names,
|
|
|
|
output_names,
|
|
|
|
dynamic_axes,
|
|
|
|
opset,
|
|
|
|
use_external_data_format=False,
|
|
|
|
):
|
2023-02-05 13:53:26 +00:00
|
|
|
"""
|
2023-01-16 23:48:50 +00:00
|
|
|
From https://github.com/huggingface/diffusers/blob/main/scripts/convert_stable_diffusion_checkpoint_to_onnx.py
|
2023-02-05 13:53:26 +00:00
|
|
|
"""
|
2023-01-16 23:48:50 +00:00
|
|
|
|
|
|
|
output_path.parent.mkdir(parents=True, exist_ok=True)
|
|
|
|
export(
|
|
|
|
model,
|
|
|
|
model_args,
|
|
|
|
f=output_path.as_posix(),
|
|
|
|
input_names=ordered_input_names,
|
|
|
|
output_names=output_names,
|
|
|
|
dynamic_axes=dynamic_axes,
|
|
|
|
do_constant_folding=True,
|
|
|
|
opset_version=opset,
|
|
|
|
)
|
|
|
|
|
|
|
|
|
|
|
|
@torch.no_grad()
|
2023-02-21 05:07:16 +00:00
|
|
|
def convert_diffusion_diffusers(
|
2023-02-11 04:41:24 +00:00
|
|
|
ctx: ConversionContext,
|
|
|
|
model: Dict,
|
|
|
|
source: str,
|
2023-02-05 13:53:26 +00:00
|
|
|
):
|
|
|
|
"""
|
2023-01-16 23:48:50 +00:00
|
|
|
From https://github.com/huggingface/diffusers/blob/main/scripts/convert_stable_diffusion_checkpoint_to_onnx.py
|
2023-02-05 13:53:26 +00:00
|
|
|
"""
|
2023-02-11 04:41:24 +00:00
|
|
|
name = model.get("name")
|
|
|
|
source = source or model.get("source")
|
|
|
|
single_vae = model.get("single_vae")
|
2023-02-17 04:52:25 +00:00
|
|
|
replace_vae = model.get("vae")
|
2023-02-11 04:41:24 +00:00
|
|
|
|
|
|
|
dtype = torch.float16 if ctx.half else torch.float32
|
2023-02-09 04:35:54 +00:00
|
|
|
dest_path = path.join(ctx.model_path, name)
|
2023-01-21 14:52:35 +00:00
|
|
|
|
|
|
|
# diffusers go into a directory rather than .onnx file
|
2023-02-11 05:32:16 +00:00
|
|
|
logger.info(
|
|
|
|
"converting Stable Diffusion model %s: %s -> %s/", name, source, dest_path
|
|
|
|
)
|
2023-01-16 23:48:50 +00:00
|
|
|
|
2023-01-29 21:23:01 +00:00
|
|
|
if single_vae:
|
2023-02-05 13:53:26 +00:00
|
|
|
logger.info("converting model with single VAE")
|
2023-01-29 21:23:01 +00:00
|
|
|
|
2023-02-09 04:35:54 +00:00
|
|
|
if path.exists(dest_path):
|
2023-02-17 00:42:05 +00:00
|
|
|
logger.info("ONNX model already exists, skipping")
|
2023-01-16 23:48:50 +00:00
|
|
|
return
|
2023-01-16 15:57:59 +00:00
|
|
|
|
2023-01-16 23:48:50 +00:00
|
|
|
pipeline = StableDiffusionPipeline.from_pretrained(
|
2023-02-11 04:41:24 +00:00
|
|
|
source,
|
|
|
|
torch_dtype=dtype,
|
|
|
|
use_auth_token=ctx.token,
|
2023-02-09 04:35:54 +00:00
|
|
|
).to(ctx.training_device)
|
2023-01-16 23:48:50 +00:00
|
|
|
output_path = Path(dest_path)
|
|
|
|
|
2023-02-24 00:43:49 +00:00
|
|
|
optimize_pipeline(ctx, pipeline)
|
|
|
|
|
2023-01-16 23:48:50 +00:00
|
|
|
# TEXT ENCODER
|
|
|
|
num_tokens = pipeline.text_encoder.config.max_position_embeddings
|
|
|
|
text_hidden_size = pipeline.text_encoder.config.hidden_size
|
|
|
|
text_input = pipeline.tokenizer(
|
|
|
|
"A sample prompt",
|
|
|
|
padding="max_length",
|
|
|
|
max_length=pipeline.tokenizer.model_max_length,
|
|
|
|
truncation=True,
|
|
|
|
return_tensors="pt",
|
|
|
|
)
|
|
|
|
onnx_export(
|
|
|
|
pipeline.text_encoder,
|
|
|
|
# casting to torch.int32 until the CLIP fix is released: https://github.com/huggingface/transformers/pull/18515/files
|
2023-02-11 04:41:24 +00:00
|
|
|
model_args=(
|
|
|
|
text_input.input_ids.to(device=ctx.training_device, dtype=torch.int32)
|
|
|
|
),
|
2023-01-16 23:48:50 +00:00
|
|
|
output_path=output_path / "text_encoder" / "model.onnx",
|
|
|
|
ordered_input_names=["input_ids"],
|
|
|
|
output_names=["last_hidden_state", "pooler_output"],
|
|
|
|
dynamic_axes={
|
|
|
|
"input_ids": {0: "batch", 1: "sequence"},
|
|
|
|
},
|
2023-02-11 04:41:24 +00:00
|
|
|
opset=ctx.opset,
|
2023-01-16 23:48:50 +00:00
|
|
|
)
|
|
|
|
del pipeline.text_encoder
|
|
|
|
|
2023-02-05 13:53:26 +00:00
|
|
|
logger.debug("UNET config: %s", pipeline.unet.config)
|
2023-01-30 00:42:05 +00:00
|
|
|
|
2023-01-16 23:48:50 +00:00
|
|
|
# UNET
|
2023-01-30 00:42:05 +00:00
|
|
|
if single_vae:
|
|
|
|
unet_inputs = ["sample", "timestep", "encoder_hidden_states", "class_labels"]
|
2023-02-16 02:50:12 +00:00
|
|
|
unet_scale = torch.tensor(4).to(device=ctx.training_device, dtype=torch.long)
|
2023-01-30 00:42:05 +00:00
|
|
|
else:
|
|
|
|
unet_inputs = ["sample", "timestep", "encoder_hidden_states", "return_dict"]
|
2023-02-11 04:41:24 +00:00
|
|
|
unet_scale = torch.tensor(False).to(
|
|
|
|
device=ctx.training_device, dtype=torch.bool
|
|
|
|
)
|
2023-01-30 00:42:05 +00:00
|
|
|
|
2023-01-16 23:48:50 +00:00
|
|
|
unet_in_channels = pipeline.unet.config.in_channels
|
|
|
|
unet_sample_size = pipeline.unet.config.sample_size
|
|
|
|
unet_path = output_path / "unet" / "model.onnx"
|
|
|
|
onnx_export(
|
|
|
|
pipeline.unet,
|
|
|
|
model_args=(
|
|
|
|
torch.randn(2, unet_in_channels, unet_sample_size, unet_sample_size).to(
|
2023-02-09 04:35:54 +00:00
|
|
|
device=ctx.training_device, dtype=dtype
|
2023-02-05 13:53:26 +00:00
|
|
|
),
|
2023-02-09 04:35:54 +00:00
|
|
|
torch.randn(2).to(device=ctx.training_device, dtype=dtype),
|
2023-01-16 23:48:50 +00:00
|
|
|
torch.randn(2, num_tokens, text_hidden_size).to(
|
2023-02-09 04:35:54 +00:00
|
|
|
device=ctx.training_device, dtype=dtype
|
2023-02-05 13:53:26 +00:00
|
|
|
),
|
2023-01-30 00:42:05 +00:00
|
|
|
unet_scale,
|
2023-01-16 23:48:50 +00:00
|
|
|
),
|
|
|
|
output_path=unet_path,
|
2023-01-30 00:42:05 +00:00
|
|
|
ordered_input_names=unet_inputs,
|
2023-01-16 23:48:50 +00:00
|
|
|
# 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"},
|
|
|
|
},
|
2023-02-11 04:41:24 +00:00
|
|
|
opset=ctx.opset,
|
2023-01-16 23:48:50 +00:00
|
|
|
use_external_data_format=True, # UNet is > 2GB, so the weights need to be split
|
|
|
|
)
|
|
|
|
unet_model_path = str(unet_path.absolute().as_posix())
|
|
|
|
unet_dir = path.dirname(unet_model_path)
|
|
|
|
unet = load(unet_model_path)
|
|
|
|
# clean up existing tensor files
|
|
|
|
rmtree(unet_dir)
|
|
|
|
mkdir(unet_dir)
|
|
|
|
# collate external tensor files into one
|
|
|
|
save_model(
|
|
|
|
unet,
|
|
|
|
unet_model_path,
|
|
|
|
save_as_external_data=True,
|
|
|
|
all_tensors_to_one_file=True,
|
|
|
|
location="weights.pb",
|
|
|
|
convert_attribute=False,
|
|
|
|
)
|
|
|
|
del pipeline.unet
|
|
|
|
|
2023-02-17 04:52:25 +00:00
|
|
|
if replace_vae is not None:
|
|
|
|
logger.debug("loading custom VAE: %s", replace_vae)
|
|
|
|
vae = AutoencoderKL.from_pretrained(replace_vae)
|
|
|
|
pipeline.vae = vae
|
|
|
|
|
2023-01-29 21:23:01 +00:00
|
|
|
if single_vae:
|
2023-02-05 13:53:26 +00:00
|
|
|
logger.debug("VAE config: %s", pipeline.vae.config)
|
2023-01-30 00:42:05 +00:00
|
|
|
|
2023-01-29 21:23:01 +00:00
|
|
|
# SINGLE VAE
|
|
|
|
vae_only = pipeline.vae
|
2023-01-30 00:42:05 +00:00
|
|
|
vae_latent_channels = vae_only.config.latent_channels
|
|
|
|
vae_out_channels = vae_only.config.out_channels
|
|
|
|
# forward only through the decoder part
|
|
|
|
vae_only.forward = vae_only.decode
|
2023-01-29 21:23:01 +00:00
|
|
|
onnx_export(
|
|
|
|
vae_only,
|
|
|
|
model_args=(
|
2023-02-05 13:53:26 +00:00
|
|
|
torch.randn(
|
|
|
|
1, vae_latent_channels, unet_sample_size, unet_sample_size
|
2023-02-09 04:35:54 +00:00
|
|
|
).to(device=ctx.training_device, dtype=dtype),
|
2023-01-29 21:23:01 +00:00
|
|
|
False,
|
|
|
|
),
|
|
|
|
output_path=output_path / "vae" / "model.onnx",
|
2023-01-30 00:42:05 +00:00
|
|
|
ordered_input_names=["latent_sample", "return_dict"],
|
|
|
|
output_names=["sample"],
|
2023-01-29 21:23:01 +00:00
|
|
|
dynamic_axes={
|
2023-01-30 00:42:05 +00:00
|
|
|
"latent_sample": {0: "batch", 1: "channels", 2: "height", 3: "width"},
|
2023-01-29 21:23:01 +00:00
|
|
|
},
|
2023-02-11 04:41:24 +00:00
|
|
|
opset=ctx.opset,
|
2023-01-29 21:23:01 +00:00
|
|
|
)
|
|
|
|
else:
|
|
|
|
# VAE ENCODER
|
|
|
|
vae_encoder = pipeline.vae
|
|
|
|
vae_in_channels = vae_encoder.config.in_channels
|
|
|
|
vae_sample_size = vae_encoder.config.sample_size
|
|
|
|
# need to get the raw tensor output (sample) from the encoder
|
|
|
|
vae_encoder.forward = lambda sample, return_dict: vae_encoder.encode(
|
2023-02-05 13:53:26 +00:00
|
|
|
sample, return_dict
|
|
|
|
)[0].sample()
|
2023-01-29 21:23:01 +00:00
|
|
|
onnx_export(
|
|
|
|
vae_encoder,
|
|
|
|
model_args=(
|
|
|
|
torch.randn(1, vae_in_channels, vae_sample_size, vae_sample_size).to(
|
2023-02-09 04:35:54 +00:00
|
|
|
device=ctx.training_device, dtype=dtype
|
2023-02-05 13:53:26 +00:00
|
|
|
),
|
2023-01-29 21:23:01 +00:00
|
|
|
False,
|
|
|
|
),
|
|
|
|
output_path=output_path / "vae_encoder" / "model.onnx",
|
|
|
|
ordered_input_names=["sample", "return_dict"],
|
|
|
|
output_names=["latent_sample"],
|
|
|
|
dynamic_axes={
|
|
|
|
"sample": {0: "batch", 1: "channels", 2: "height", 3: "width"},
|
|
|
|
},
|
2023-02-11 04:41:24 +00:00
|
|
|
opset=ctx.opset,
|
2023-01-29 21:23:01 +00:00
|
|
|
)
|
|
|
|
|
|
|
|
# VAE DECODER
|
|
|
|
vae_decoder = pipeline.vae
|
|
|
|
vae_latent_channels = vae_decoder.config.latent_channels
|
|
|
|
vae_out_channels = vae_decoder.config.out_channels
|
|
|
|
# forward only through the decoder part
|
|
|
|
vae_decoder.forward = vae_encoder.decode
|
|
|
|
onnx_export(
|
|
|
|
vae_decoder,
|
|
|
|
model_args=(
|
2023-02-05 13:53:26 +00:00
|
|
|
torch.randn(
|
|
|
|
1, vae_latent_channels, unet_sample_size, unet_sample_size
|
2023-02-09 04:35:54 +00:00
|
|
|
).to(device=ctx.training_device, dtype=dtype),
|
2023-01-29 21:23:01 +00:00
|
|
|
False,
|
|
|
|
),
|
|
|
|
output_path=output_path / "vae_decoder" / "model.onnx",
|
|
|
|
ordered_input_names=["latent_sample", "return_dict"],
|
|
|
|
output_names=["sample"],
|
|
|
|
dynamic_axes={
|
|
|
|
"latent_sample": {0: "batch", 1: "channels", 2: "height", 3: "width"},
|
|
|
|
},
|
2023-02-11 04:41:24 +00:00
|
|
|
opset=ctx.opset,
|
2023-01-29 21:23:01 +00:00
|
|
|
)
|
2023-01-16 23:48:50 +00:00
|
|
|
|
|
|
|
del pipeline.vae
|
|
|
|
|
|
|
|
# SAFETY CHECKER
|
|
|
|
if pipeline.safety_checker is not None:
|
|
|
|
safety_checker = pipeline.safety_checker
|
|
|
|
clip_num_channels = safety_checker.config.vision_config.num_channels
|
|
|
|
clip_image_size = safety_checker.config.vision_config.image_size
|
|
|
|
safety_checker.forward = safety_checker.forward_onnx
|
|
|
|
onnx_export(
|
|
|
|
pipeline.safety_checker,
|
|
|
|
model_args=(
|
|
|
|
torch.randn(
|
|
|
|
1,
|
|
|
|
clip_num_channels,
|
|
|
|
clip_image_size,
|
|
|
|
clip_image_size,
|
2023-02-09 04:35:54 +00:00
|
|
|
).to(device=ctx.training_device, dtype=dtype),
|
2023-01-16 23:48:50 +00:00
|
|
|
torch.randn(1, vae_sample_size, vae_sample_size, vae_out_channels).to(
|
2023-02-09 04:35:54 +00:00
|
|
|
device=ctx.training_device, dtype=dtype
|
2023-02-05 13:53:26 +00:00
|
|
|
),
|
2023-01-16 23:48:50 +00:00
|
|
|
),
|
|
|
|
output_path=output_path / "safety_checker" / "model.onnx",
|
|
|
|
ordered_input_names=["clip_input", "images"],
|
|
|
|
output_names=["out_images", "has_nsfw_concepts"],
|
|
|
|
dynamic_axes={
|
|
|
|
"clip_input": {0: "batch", 1: "channels", 2: "height", 3: "width"},
|
|
|
|
"images": {0: "batch", 1: "height", 2: "width", 3: "channels"},
|
|
|
|
},
|
2023-02-11 04:41:24 +00:00
|
|
|
opset=ctx.opset,
|
2023-01-16 23:48:50 +00:00
|
|
|
)
|
|
|
|
del pipeline.safety_checker
|
|
|
|
safety_checker = OnnxRuntimeModel.from_pretrained(
|
2023-02-05 13:53:26 +00:00
|
|
|
output_path / "safety_checker"
|
|
|
|
)
|
2023-01-16 23:48:50 +00:00
|
|
|
feature_extractor = pipeline.feature_extractor
|
|
|
|
else:
|
|
|
|
safety_checker = None
|
|
|
|
feature_extractor = None
|
|
|
|
|
2023-01-29 21:23:01 +00:00
|
|
|
if single_vae:
|
2023-02-09 04:35:54 +00:00
|
|
|
onnx_pipeline = OnnxStableDiffusionUpscalePipeline(
|
2023-02-05 13:53:26 +00:00
|
|
|
vae=OnnxRuntimeModel.from_pretrained(output_path / "vae"),
|
|
|
|
text_encoder=OnnxRuntimeModel.from_pretrained(output_path / "text_encoder"),
|
2023-01-29 21:23:01 +00:00
|
|
|
tokenizer=pipeline.tokenizer,
|
|
|
|
low_res_scheduler=pipeline.scheduler,
|
|
|
|
unet=OnnxRuntimeModel.from_pretrained(output_path / "unet"),
|
|
|
|
scheduler=pipeline.scheduler,
|
|
|
|
)
|
|
|
|
else:
|
|
|
|
onnx_pipeline = OnnxStableDiffusionPipeline(
|
2023-02-05 13:53:26 +00:00
|
|
|
vae_encoder=OnnxRuntimeModel.from_pretrained(output_path / "vae_encoder"),
|
|
|
|
vae_decoder=OnnxRuntimeModel.from_pretrained(output_path / "vae_decoder"),
|
|
|
|
text_encoder=OnnxRuntimeModel.from_pretrained(output_path / "text_encoder"),
|
2023-01-29 21:23:01 +00:00
|
|
|
tokenizer=pipeline.tokenizer,
|
|
|
|
unet=OnnxRuntimeModel.from_pretrained(output_path / "unet"),
|
|
|
|
scheduler=pipeline.scheduler,
|
|
|
|
safety_checker=safety_checker,
|
|
|
|
feature_extractor=feature_extractor,
|
|
|
|
requires_safety_checker=safety_checker is not None,
|
|
|
|
)
|
2023-01-16 23:48:50 +00:00
|
|
|
|
2023-02-05 13:53:26 +00:00
|
|
|
logger.info("exporting ONNX model")
|
2023-01-21 14:52:35 +00:00
|
|
|
|
2023-01-16 23:48:50 +00:00
|
|
|
onnx_pipeline.save_pretrained(output_path)
|
2023-01-30 00:42:05 +00:00
|
|
|
logger.info("ONNX pipeline saved to %s", output_path)
|
2023-01-16 23:48:50 +00:00
|
|
|
|
|
|
|
del pipeline
|
|
|
|
del onnx_pipeline
|
2023-01-29 21:23:01 +00:00
|
|
|
|
|
|
|
if single_vae:
|
2023-02-09 04:35:54 +00:00
|
|
|
_ = OnnxStableDiffusionUpscalePipeline.from_pretrained(
|
2023-01-29 21:23:01 +00:00
|
|
|
output_path, provider="CPUExecutionProvider"
|
|
|
|
)
|
|
|
|
else:
|
|
|
|
_ = OnnxStableDiffusionPipeline.from_pretrained(
|
2023-02-05 13:53:26 +00:00
|
|
|
output_path, provider="CPUExecutionProvider"
|
|
|
|
)
|
2023-01-29 21:23:01 +00:00
|
|
|
|
2023-01-28 23:09:19 +00:00
|
|
|
logger.info("ONNX pipeline is loadable")
|