BlockFusion/repaint.py at main · Tencent/BlockFusion · 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
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
import tqdm
from diffusers import DiffusionPipeline,RePaintScheduler,RePaintPipeline
import torch
import numpy as np
import os
import matplotlib.pyplot as plt
from diffusers.utils.torch_utils import randn_tensor
from math import ceil
def get_views(panorama_height, panorama_width, window_size=64, stride=32):
    num_blocks_height = ceil((panorama_height - window_size) / stride) + 1
    num_blocks_width = ceil((panorama_width - window_size) / stride) + 1
    views = [[] for _ in range(num_blocks_height)]
    for i in range(num_blocks_height):
        for j in range(num_blocks_width):
            h_start = int(i  * stride)
            h_end = h_start + window_size
            w_start = int(j  * stride)
            w_end = w_start + window_size
            views[i].append((h_start, h_end, w_start, w_end))
    return views

pipeline = DiffusionPipeline.from_pretrained("conditioned", torch_dtype=torch.float32)
pipeline.to("cuda")
device = 'cuda'
generator = torch.Generator(device=pipeline.device).manual_seed(7)
for v in pipeline.unet.parameters():
    v.requires_grad=False

x = 32
z = 56
y = 32
stride = 24
target_vol_shape = (x,y,z)
views = get_views(x, z,window_size=32,stride=stride)
num_views = len(views)*len(views[0])

in_channel = 2
num_inference_steps = 1000
image_shape = (num_views, in_channel, *pipeline.unet.config.sample_size)
images = randn_tensor(image_shape, generator=generator, device=pipeline.device)

sourcedir = 'layout'
layoutlatent = np.zeros((num_views,9,32,96),dtype=np.float32)
for i in range(len(views)):
    for j in range(len(views[0])):
        layoutlatent[i * len(views[0]) + j,:,:,32:64]  = np.load(os.path.join(sourcedir,'{}_{}.npy'.format(i,j))).astype(np.float32)
layoutlatent = torch.from_numpy(layoutlatent).to(pipeline.device)

pipeline.scheduler.set_timesteps(num_inference_steps)
with torch.autocast('cuda'):
    for t in tqdm.tqdm(pipeline.scheduler.timesteps):
        images = torch.cat((images, layoutlatent), dim=1)
        model_output = pipeline.unet(images, t).sample
        images = pipeline.scheduler.step(model_output, t, images[:,:in_channel,:,:], generator=generator).prev_sample


# images = images.cpu().numpy()*1.42
images = images*1.42
# savedir = 'latentseeds'
# os.makedirs(savedir,exist_ok=True)
# for i in range(len(views)):
#     for j in range(len(views[0])):
#         image = images[i * len(views[0]) + j]
#         np.save(os.path.join(savedir, str(i) + '_' + str(j) + '.npy'), image[np.newaxis, :])

inchannel = 2
x = 32 + (len(views)-1)*stride
z = 32 + (len(views[0])-1)*stride
latent1 = torch.randn((len(views[0]), inchannel, x, y), device=device)
latent2 = torch.randn((1, inchannel, x, z), device=device)
latent3 = torch.randn((len(views), inchannel, z, y), device=device)
noise = torch.randn((num_views, inchannel, 32, 96), device=device)

original_image = torch.zeros_like(noise)
mask_image = torch.zeros_like(noise)

for i in range(len(views)):
    for j in range(len(views[0])):
        if i % 2 == 0 and j % 2 == 0:
            original_image[i * len(views[0]) + j, :, :, :] = images[i * len(views[0]) + j]
            mask_image[i*len(views[0])+j, :, :, :] = 1

del pipeline,images
torch.cuda.empty_cache()

generator = torch.Generator(device='cuda').manual_seed(7)
scheduler = RePaintScheduler.from_pretrained("conditioned/scheduler")
pipeline = RePaintPipeline.from_pretrained("conditioned", scheduler=scheduler)
for v in pipeline.unet.parameters():
    v.requires_grad=False
device = "cuda"
pipeline = pipeline.to(device)
superiority = 1000000

count1 = torch.zeros_like(latent1)
value1 = torch.zeros_like(latent1)
count2 = torch.zeros_like(latent2)
value2 = torch.zeros_like(latent2)
count3 = torch.zeros_like(latent3)
value3 = torch.zeros_like(latent3)

a = 0.5
num_inference_steps=1000
eta=0.0
jump_length=100
jump_n_sample=15
generator=generator
scheduler.set_timesteps(num_inference_steps, jump_length, jump_n_sample, device)
t_last = scheduler.timesteps[0] + 1

with torch.autocast('cuda'):
    for i, t in enumerate(tqdm.tqdm(scheduler.timesteps)):
        if t < t_last:
            noise = torch.cat((noise,layoutlatent),dim=1)
            count1.zero_()
            value1.zero_()
            count2.zero_()
            value2.zero_()
            count3.zero_()
            value3.zero_()
            latent1.zero_()
            latent2.zero_()
            latent3.zero_()
            model_output = pipeline.unet(noise, t).sample
            noise = scheduler.step(model_output, t, noise[:,:inchannel,:,:], original_image, mask_image, generator).prev_sample
            for i in range(len(views)):
                for j in range(len(views[0])):
                    # if known[i*len(views[0])+j]:
                    if i%2==0 and j%2==0:
                    # if i * len(views[0]) + j in seed:
                    # if i == 0 and j == 0:
                        x_start, x_end, z_start, z_end = views[i][j]
                        noise1, noise2, noise3 = noise[i * len(views[0]) + j].split(32, dim=-1)
                        value1[j, :, x_start:x_end, :] += superiority*noise1
                        count1[j, :, x_start:x_end, :] += superiority
                        value2[:, :, x_start:x_end, z_start:z_end] += superiority*noise2
                        count2[:, :, x_start:x_end, z_start:z_end] += superiority
                        value3[i, :, z_start:z_end, :] += superiority*noise3
                        count3[i, :, z_start:z_end, :] += superiority
                    else:
                        x_start, x_end, z_start, z_end = views[i][j]
                        noise1, noise2, noise3 = noise[i * len(views[0]) + j].split(32, dim=-1)
                        value1[j, :, x_start:x_end, :] += noise1
                        count1[j, :, x_start:x_end, :] += 1
                        value2[:, :, x_start:x_end, z_start:z_end] += noise2
                        count2[:, :, x_start:x_end, z_start:z_end] += 1
                        value3[i, :, z_start:z_end, :] += noise3
                        count3[i, :, z_start:z_end, :] += 1
            latent1 = torch.where(count1 > 0, value1 / count1, value1)
            latent2 = torch.where(count2 > 0, value2 / count2, value2)
            latent3 = torch.where(count3 > 0, value3 / count3, value3)
            for i in range(len(views)):
                for j in range(len(views[0])):
                    x_start, x_end, z_start, z_end = views[i][j]
                    noise[i*len(views[0])+j, :, :, :32] = latent1[j,:,x_start:x_end, :]
                    noise[i*len(views[0])+j, :, :,32:64] = latent2[:,:,x_start:x_end, z_start:z_end]
                    noise[i*len(views[0])+j, :, :,64:] = latent3[i,:,z_start:z_end,:]
        else:
            noise = scheduler.undo_step(noise, t_last, generator)
        # print(noise.min(),noise.max())
        t_last = t

images = noise.cpu().numpy()
images = images*1.42
savedir = 'latents'
os.makedirs(savedir,exist_ok=True)
for i in range(len(views)):
    for j in range(len(views[0])):
        image = images[i*len(views[0])+j]
        np.save(os.path.join(savedir,str(i)+'_'+str(j)+'.npy'),image[np.newaxis,:])