pipeline_ddim.py 2.72 KB
Newer Older
Patrick von Platen's avatar
Patrick von Platen committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
# Copyright 2022 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and

# limitations under the License.


Pedro Cuenca's avatar
Pedro Cuenca committed
17
import warnings
18
from typing import Tuple, Union
Pedro Cuenca's avatar
Pedro Cuenca committed
19

Patrick von Platen's avatar
Patrick von Platen committed
20
21
import torch

22
from ...pipeline_utils import DiffusionPipeline, ImagePipelineOutput
Patrick von Platen's avatar
Patrick von Platen committed
23
24


Patrick von Platen's avatar
Patrick von Platen committed
25
class DDIMPipeline(DiffusionPipeline):
26
    def __init__(self, unet, scheduler):
Patrick von Platen's avatar
Patrick von Platen committed
27
        super().__init__()
28
29
        scheduler = scheduler.set_format("pt")
        self.register_modules(unet=unet, scheduler=scheduler)
Patrick von Platen's avatar
Patrick von Platen committed
30

Patrick von Platen's avatar
Patrick von Platen committed
31
    @torch.no_grad()
32
33
34
35
36
37
38
39
40
41
    def __call__(
        self,
        batch_size=1,
        generator=None,
        eta=0.0,
        num_inference_steps=50,
        output_type="pil",
        return_dict: bool = True,
        **kwargs,
    ) -> Union[ImagePipelineOutput, Tuple]:
Pedro Cuenca's avatar
Pedro Cuenca committed
42
43
44
45
46
47
48

        if "torch_device" in kwargs:
            device = kwargs.pop("torch_device")
            warnings.warn(
                "`torch_device` is deprecated as an input argument to `__call__` and will be removed in v0.3.0."
                " Consider using `pipe.to(torch_device)` instead."
            )
Patrick von Platen's avatar
Patrick von Platen committed
49

Pedro Cuenca's avatar
Pedro Cuenca committed
50
51
52
53
54
55
            # Set device as before (to be removed in 0.3.0)
            if device is None:
                device = "cuda" if torch.cuda.is_available() else "cpu"
            self.to(device)

        # eta corresponds to η in paper and should be between [0, 1]
Patrick von Platen's avatar
Patrick von Platen committed
56
57

        # Sample gaussian noise to begin loop
Patrick von Platen's avatar
Patrick von Platen committed
58
        image = torch.randn(
Patrick von Platen's avatar
Patrick von Platen committed
59
            (batch_size, self.unet.in_channels, self.unet.sample_size, self.unet.sample_size),
Patrick von Platen's avatar
Patrick von Platen committed
60
61
            generator=generator,
        )
Pedro Cuenca's avatar
Pedro Cuenca committed
62
        image = image.to(self.device)
Patrick von Platen's avatar
Patrick von Platen committed
63

64
65
        # set step values
        self.scheduler.set_timesteps(num_inference_steps)
Patrick von Platen's avatar
Patrick von Platen committed
66

hysts's avatar
hysts committed
67
        for t in self.progress_bar(self.scheduler.timesteps):
Patrick von Platen's avatar
Patrick von Platen committed
68
            # 1. predict noise model_output
69
            model_output = self.unet(image, t).sample
Patrick von Platen's avatar
Patrick von Platen committed
70

71
72
            # 2. predict previous mean of image x_t-1 and add variance depending on eta
            # do x_t -> x_t-1
73
            image = self.scheduler.step(model_output, t, image, eta).prev_sample
Patrick von Platen's avatar
Patrick von Platen committed
74

75
76
        image = (image / 2 + 0.5).clamp(0, 1)
        image = image.cpu().permute(0, 2, 3, 1).numpy()
anton-l's avatar
anton-l committed
77
78
        if output_type == "pil":
            image = self.numpy_to_pil(image)
79

80
81
82
83
        if not return_dict:
            return (image,)

        return ImagePipelineOutput(images=image)