File size: 7,216 Bytes
c19ca42
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
import inspect
from typing import Union, Optional, Callable, List, Any
import numpy as np
import torch
import diffusers
from diffusers.pipelines.onnx_utils import ORT_TO_NP_TYPE
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
from modules.onnx_impl.pipelines import CallablePipelineBase
from modules.onnx_impl.pipelines.utils import prepare_latents


class OnnxStableDiffusionPipeline(diffusers.OnnxStableDiffusionPipeline, CallablePipelineBase):
    __module__ = 'diffusers'
    __name__ = 'OnnxStableDiffusionPipeline'

    def __init__(
        self,
        vae_encoder: diffusers.OnnxRuntimeModel,
        vae_decoder: diffusers.OnnxRuntimeModel,
        text_encoder: diffusers.OnnxRuntimeModel,
        tokenizer: Any,
        unet: diffusers.OnnxRuntimeModel,
        scheduler: Any,
        safety_checker: diffusers.OnnxRuntimeModel,
        feature_extractor: Any,
        requires_safety_checker: bool = True
    ):
        super().__init__(vae_encoder, vae_decoder, text_encoder, tokenizer, unet, scheduler, safety_checker, feature_extractor, requires_safety_checker)

    def __call__(
        self,
        prompt: Union[str, List[str]] = None,
        height: Optional[int] = 512,
        width: Optional[int] = 512,
        num_inference_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[Union[torch.Generator, List[torch.Generator]]] = 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 = torch.Generator("cpu")

        # 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 = 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 = prepare_latents(
            self.scheduler.init_noise_sigma,
            batch_size * num_images_per_prompt,
            height,
            width,
            prompt_embeds.dtype,
            generator,
            latents
        )

        # set timesteps
        self.scheduler.set_timesteps(num_inference_steps)

        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
        # eta (ฮท) is only used with the DDIMScheduler, it will be ignored for other schedulers.
        # eta corresponds to ฮท 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:
                step_idx = i // getattr(self.scheduler, "order", 1)
                callback(step_idx, t, latents)

        has_nsfw_concept = None

        if output_type != "latent":
            latents /= self.vae_decoder.config.get("scaling_factor", 0.18215)

            # image = self.vae_decoder(latent_sample=latents)[0]
            # 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 self.safety_checker is not None:
                safety_checker_input = self.feature_extractor(
                    self.numpy_to_pil(image), return_tensors="np"
                ).pixel_values.astype(image.dtype)

                images, has_nsfw_concept = [], []
                for i in range(image.shape[0]):
                    image_i, has_nsfw_concept_i = self.safety_checker(
                        clip_input=safety_checker_input[i : i + 1], images=image[i : i + 1]
                    )
                    images.append(image_i)
                    has_nsfw_concept.append(has_nsfw_concept_i[0])
                image = np.concatenate(images)

            if output_type == "pil":
                image = self.numpy_to_pil(image)
        else:
            image = latents

        if not return_dict:
            return (image, has_nsfw_concept)

        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)