|
| 1 | +# Copyright 2022 The HuggingFace Team. All rights reserved. |
| 2 | +# |
| 3 | +# Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | +# you may not use this file except in compliance with the License. |
| 5 | +# You may obtain a copy of the License at |
| 6 | +# |
| 7 | +# http://www.apache.org/licenses/LICENSE-2.0 |
| 8 | +# |
| 9 | +# Unless required by applicable law or agreed to in writing, software |
| 10 | +# distributed under the License is distributed on an "AS IS" BASIS, |
| 11 | +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 | +# See the License for the specific language governing permissions and |
| 13 | +# limitations under the License. |
| 14 | + |
| 15 | +import argparse |
| 16 | +import os |
| 17 | +import shutil |
| 18 | +from pathlib import Path |
| 19 | + |
| 20 | +import torch |
| 21 | +from torch.onnx import export |
| 22 | + |
| 23 | +import onnx |
| 24 | +from diffusers import OnnxRuntimeModel, OnnxStableDiffusionPipeline, StableDiffusionPipeline, AutoencoderKL |
| 25 | +from packaging import version |
| 26 | + |
| 27 | + |
| 28 | +is_torch_less_than_1_11 = version.parse(version.parse(torch.__version__).base_version) < version.parse("1.11") |
| 29 | + |
| 30 | + |
| 31 | +def onnx_export( |
| 32 | + model, |
| 33 | + model_args: tuple, |
| 34 | + output_path: Path, |
| 35 | + ordered_input_names, |
| 36 | + output_names, |
| 37 | + dynamic_axes, |
| 38 | + opset, |
| 39 | + use_external_data_format=False, |
| 40 | +): |
| 41 | + output_path.parent.mkdir(parents=True, exist_ok=True) |
| 42 | + # PyTorch deprecated the `enable_onnx_checker` and `use_external_data_format` arguments in v1.11, |
| 43 | + # so we check the torch version for backwards compatibility |
| 44 | + if is_torch_less_than_1_11: |
| 45 | + export( |
| 46 | + model, |
| 47 | + model_args, |
| 48 | + f=output_path.as_posix(), |
| 49 | + input_names=ordered_input_names, |
| 50 | + output_names=output_names, |
| 51 | + dynamic_axes=dynamic_axes, |
| 52 | + do_constant_folding=True, |
| 53 | + use_external_data_format=use_external_data_format, |
| 54 | + enable_onnx_checker=True, |
| 55 | + opset_version=opset, |
| 56 | + ) |
| 57 | + else: |
| 58 | + export( |
| 59 | + model, |
| 60 | + model_args, |
| 61 | + f=output_path.as_posix(), |
| 62 | + input_names=ordered_input_names, |
| 63 | + output_names=output_names, |
| 64 | + dynamic_axes=dynamic_axes, |
| 65 | + do_constant_folding=True, |
| 66 | + opset_version=opset, |
| 67 | + ) |
| 68 | + |
| 69 | + |
| 70 | +@torch.no_grad() |
| 71 | +def convert_models(model_path: str, output_path: str, opset: int, fp16: bool = False): |
| 72 | + dtype = torch.float16 if fp16 else torch.float32 |
| 73 | + if fp16 and torch.cuda.is_available(): |
| 74 | + device = "cuda" |
| 75 | + elif fp16 and not torch.cuda.is_available(): |
| 76 | + raise ValueError("`float16` model export is only supported on GPUs with CUDA") |
| 77 | + else: |
| 78 | + device = "cpu" |
| 79 | + output_path = Path(output_path) |
| 80 | + |
| 81 | + # VAE DECODER |
| 82 | + vae_decoder = AutoencoderKL.from_pretrained(model_path + "/vae") |
| 83 | + vae_latent_channels = vae_decoder.config.latent_channels |
| 84 | + vae_out_channels = vae_decoder.config.out_channels |
| 85 | + # forward only through the decoder part |
| 86 | + vae_decoder.forward = vae_decoder.decode |
| 87 | + onnx_export( |
| 88 | + vae_decoder, |
| 89 | + model_args=( |
| 90 | + torch.randn(1, vae_latent_channels, 25, 25).to(device=device, dtype=dtype), |
| 91 | + False, |
| 92 | + ), |
| 93 | + output_path=output_path / "vae_decoder" / "model.onnx", |
| 94 | + ordered_input_names=["latent_sample", "return_dict"], |
| 95 | + output_names=["sample"], |
| 96 | + dynamic_axes={ |
| 97 | + "latent_sample": {0: "batch", 1: "channels", 2: "height", 3: "width"}, |
| 98 | + }, |
| 99 | + opset=opset, |
| 100 | + ) |
| 101 | + del vae_decoder |
| 102 | + |
| 103 | + |
| 104 | +if __name__ == "__main__": |
| 105 | + parser = argparse.ArgumentParser() |
| 106 | + |
| 107 | + parser.add_argument( |
| 108 | + "--model_path", |
| 109 | + type=str, |
| 110 | + required=True, |
| 111 | + help="Path to the `diffusers` checkpoint to convert (either a local directory or on the Hub).", |
| 112 | + ) |
| 113 | + |
| 114 | + parser.add_argument("--output_path", type=str, required=True, help="Path to the output model.") |
| 115 | + parser.add_argument( |
| 116 | + "--opset", |
| 117 | + default=14, |
| 118 | + type=int, |
| 119 | + help="The version of the ONNX operator set to use.", |
| 120 | + ) |
| 121 | + parser.add_argument("--fp16", action="store_true", default=False, help="Export the models in `float16` mode") |
| 122 | + |
| 123 | + args = parser.parse_args() |
| 124 | + print(args.output_path) |
| 125 | + convert_models(args.model_path, args.output_path, args.opset, args.fp16) |
| 126 | + print("SD: Done: ONNX") |
0 commit comments