consolver/denoise_ppo.py at master · G-U-N/consolver · GitHub

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
import torch
from typing import Union, List, Optional
from torch.utils.checkpoint import checkpoint
import contextlib

def denoise_diffusion(
    text_encoder,
    scheduler,
    unet,
    noise: torch.FloatTensor,
    text: Union[str, List[str]],
    tokenizer,
    cfg: float = 3,  # guidance scale
    num_inference_steps: int = 50,
    gradient_checkpointing = False,
) -> torch.FloatTensor:

    if isinstance(text, str):
        batch_size = 1
        text = [text]
    else:
        batch_size = len(text)
    device = noise.device

    with torch.no_grad():
        text_inputs = tokenizer(
            text,
            padding="max_length",
            max_length=tokenizer.model_max_length,
            truncation=True,
            return_tensors="pt",
        )
        text_input_ids = text_inputs.input_ids.to(device)
        prompt_embeds = text_encoder(text_input_ids)[0]
    prompt_embeds_txt = prompt_embeds

    do_classifier_free_guidance = cfg > 1.0
    if do_classifier_free_guidance:
        with torch.no_grad():
            uncond_tokens = [""] * batch_size
            uncond_inputs = tokenizer(
                uncond_tokens,
                padding="max_length",
                max_length=tokenizer.model_max_length,
                truncation=True,
                return_tensors="pt",
            )
            uncond_input_ids = uncond_inputs.input_ids.to(device)
            negative_prompt_embeds = text_encoder(uncond_input_ids)[0]
            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])

    latents = noise.clone()
    scheduler.set_timesteps(num_inference_steps, device=device)
    timesteps = scheduler.timesteps


    # conds_ = []
    conds_ = dict(x=[], epsilon=[])
    actions_ = []
    probs_ = []
    masks_ = []
    for i, t in enumerate(timesteps):
        unet_apply = unet
        ctx = contextlib.nullcontext()

        latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
        latent_model_input = scheduler.scale_model_input(latent_model_input, t)

        with ctx:
            if gradient_checkpointing:
                def unet_call(latent_model_input, t, encoder_hidden_states, *args, **kwargs):
                    return unet_apply(
                        latent_model_input,
                        t,
                        encoder_hidden_states=encoder_hidden_states,
                        return_dict=False,
                        *args,
                        **kwargs
                    )[0]

                noise_pred = checkpoint(
                    unet_call,
                    latent_model_input,
                    t,
                    prompt_embeds,
                    use_reentrant=False
                )
            else:
                noise_pred = unet_apply(
                    latent_model_input,
                    t,
                    encoder_hidden_states=prompt_embeds,
                    return_dict=False
                )[0]

        if do_classifier_free_guidance:
            noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
            noise_pred_uncond = noise_pred_uncond.detach()

            noise_pred = noise_pred_uncond + cfg * (noise_pred_text - noise_pred_uncond)


        latents, actions, probs, conds, masks = scheduler.step(noise_pred, t, latents, return_dict=False)
        # conds_.append(conds.unsqueeze(1))
        if i > 0:

            conds_["x"].append(conds["x"].unsqueeze(1))
            conds_["epsilon"].append(conds["epsilon"].unsqueeze(1))
            probs_.append(probs.unsqueeze(1))
            actions_.append(actions.unsqueeze(1))
            masks_.append(masks.unsqueeze(1))

        latents = latents.detach() # detach() is very important
    # conds_ = torch.cat(conds_, dim=1)
    conds_ = {k: torch.cat(v, dim=1) for k,v in conds_.items()}
    probs_ = torch.cat(probs_, dim=1)
    actions_ = torch.cat(actions_, dim=1)
    masks_ = torch.cat(masks_, dim=1)

    return latents, conds_, probs_, actions_, masks_, prompt_embeds_txt

if __name__ == "__main__":
    from transformers import CLIPTextModel, CLIPTokenizer
    from diffusers import UNet2DConditionModel, DDIMScheduler

    tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-base-patch32")
    text_encoder = CLIPTextModel.from_pretrained("openai/clip-vit-base-patch32")
    unet = UNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5", subfolder="unet")
    teacher_unet = UNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5", subfolder="unet")
    scheduler = DDIMScheduler.from_pretrained("runwayml/stable-diffusion-v1-5", subfolder="scheduler")

    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    text_encoder = text_encoder.to(device)
    unet = unet.to(device)
    teacher_unet = teacher_unet.to(device)
    scheduler = scheduler.to(device)

    noise = torch.randn(1, 4, 64, 64).to(device)  # [batch_size, channels, height, width]
    text = "a photo of a cat"

    latents = denoise_diffusion(
        text_encoder=text_encoder,
        scheduler=scheduler,
        teacher_unet=teacher_unet,
        unet=unet,
        noise=noise,
        text=text,
        tokenizer=tokenizer,
        cfg=3,
        num_inference_steps=50
    )
    print("Denoised latents shape:", latents.shape)