Spaces:
Sleeping
Sleeping
add SD
Browse filesThis view is limited to 50 files because it contains too many changes.
See raw diff
- stable-diffusion/configs/autoencoder/autoencoder_kl_16x16x16.yaml +54 -0
- stable-diffusion/configs/autoencoder/autoencoder_kl_32x32x4.yaml +53 -0
- stable-diffusion/configs/autoencoder/autoencoder_kl_64x64x3.yaml +54 -0
- stable-diffusion/configs/autoencoder/autoencoder_kl_8x8x64.yaml +53 -0
- stable-diffusion/configs/latent-diffusion/celebahq-ldm-vq-4.yaml +86 -0
- stable-diffusion/configs/latent-diffusion/cin-ldm-vq-f8.yaml +98 -0
- stable-diffusion/configs/latent-diffusion/cin256-v2.yaml +68 -0
- stable-diffusion/configs/latent-diffusion/ffhq-ldm-vq-4.yaml +85 -0
- stable-diffusion/configs/latent-diffusion/lsun_bedrooms-ldm-vq-4.yaml +85 -0
- stable-diffusion/configs/latent-diffusion/lsun_churches-ldm-kl-8.yaml +91 -0
- stable-diffusion/configs/latent-diffusion/txt2img-1p4B-eval.yaml +71 -0
- stable-diffusion/configs/retrieval-augmented-diffusion/768x768.yaml +68 -0
- stable-diffusion/configs/stable-diffusion/v1-inference.yaml +70 -0
- stable-diffusion/debug/generate.sh +14 -0
- stable-diffusion/debug/inverse.sh +13 -0
- stable-diffusion/debug/inverse_bip_ldm_laion.sh +13 -0
- stable-diffusion/debug/inverse_mb_ldm_laion.sh +13 -0
- stable-diffusion/ldm/__pycache__/util.cpython-38.pyc +0 -0
- stable-diffusion/ldm/__pycache__/util.cpython-39.pyc +0 -0
- stable-diffusion/ldm/data/__init__.py +0 -0
- stable-diffusion/ldm/data/base.py +23 -0
- stable-diffusion/ldm/data/imagenet.py +394 -0
- stable-diffusion/ldm/data/lsun.py +92 -0
- stable-diffusion/ldm/lr_scheduler.py +98 -0
- stable-diffusion/ldm/models/__pycache__/autoencoder.cpython-38.pyc +0 -0
- stable-diffusion/ldm/models/autoencoder.py +443 -0
- stable-diffusion/ldm/models/diffusion/__init__.py +0 -0
- stable-diffusion/ldm/models/diffusion/__pycache__/__init__.cpython-38.pyc +0 -0
- stable-diffusion/ldm/models/diffusion/__pycache__/__init__.cpython-39.pyc +0 -0
- stable-diffusion/ldm/models/diffusion/__pycache__/ddim.cpython-38.pyc +0 -0
- stable-diffusion/ldm/models/diffusion/__pycache__/ddim.cpython-39.pyc +0 -0
- stable-diffusion/ldm/models/diffusion/__pycache__/ddpm.cpython-38.pyc +0 -0
- stable-diffusion/ldm/models/diffusion/__pycache__/ddpm.cpython-39.pyc +0 -0
- stable-diffusion/ldm/models/diffusion/__pycache__/plms.cpython-38.pyc +0 -0
- stable-diffusion/ldm/models/diffusion/__pycache__/plms.cpython-39.pyc +0 -0
- stable-diffusion/ldm/models/diffusion/__pycache__/psld.cpython-38.pyc +0 -0
- stable-diffusion/ldm/models/diffusion/__pycache__/psld.cpython-39.pyc +0 -0
- stable-diffusion/ldm/models/diffusion/classifier.py +267 -0
- stable-diffusion/ldm/models/diffusion/ddim.py +241 -0
- stable-diffusion/ldm/models/diffusion/ddpm.py +1445 -0
- stable-diffusion/ldm/models/diffusion/dpm_solver/__init__.py +1 -0
- stable-diffusion/ldm/models/diffusion/dpm_solver/__pycache__/__init__.cpython-38.pyc +0 -0
- stable-diffusion/ldm/models/diffusion/dpm_solver/__pycache__/__init__.cpython-39.pyc +0 -0
- stable-diffusion/ldm/models/diffusion/dpm_solver/__pycache__/dpm_solver.cpython-38.pyc +0 -0
- stable-diffusion/ldm/models/diffusion/dpm_solver/__pycache__/dpm_solver.cpython-39.pyc +0 -0
- stable-diffusion/ldm/models/diffusion/dpm_solver/__pycache__/sampler.cpython-38.pyc +0 -0
- stable-diffusion/ldm/models/diffusion/dpm_solver/__pycache__/sampler.cpython-39.pyc +0 -0
- stable-diffusion/ldm/models/diffusion/dpm_solver/dpm_solver.py +1184 -0
- stable-diffusion/ldm/models/diffusion/dpm_solver/sampler.py +82 -0
- stable-diffusion/ldm/models/diffusion/plms.py +236 -0
stable-diffusion/configs/autoencoder/autoencoder_kl_16x16x16.yaml
ADDED
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
model:
|
2 |
+
base_learning_rate: 4.5e-6
|
3 |
+
target: ldm.models.autoencoder.AutoencoderKL
|
4 |
+
params:
|
5 |
+
monitor: "val/rec_loss"
|
6 |
+
embed_dim: 16
|
7 |
+
lossconfig:
|
8 |
+
target: ldm.modules.losses.LPIPSWithDiscriminator
|
9 |
+
params:
|
10 |
+
disc_start: 50001
|
11 |
+
kl_weight: 0.000001
|
12 |
+
disc_weight: 0.5
|
13 |
+
|
14 |
+
ddconfig:
|
15 |
+
double_z: True
|
16 |
+
z_channels: 16
|
17 |
+
resolution: 256
|
18 |
+
in_channels: 3
|
19 |
+
out_ch: 3
|
20 |
+
ch: 128
|
21 |
+
ch_mult: [ 1,1,2,2,4] # num_down = len(ch_mult)-1
|
22 |
+
num_res_blocks: 2
|
23 |
+
attn_resolutions: [16]
|
24 |
+
dropout: 0.0
|
25 |
+
|
26 |
+
|
27 |
+
data:
|
28 |
+
target: main.DataModuleFromConfig
|
29 |
+
params:
|
30 |
+
batch_size: 12
|
31 |
+
wrap: True
|
32 |
+
train:
|
33 |
+
target: ldm.data.imagenet.ImageNetSRTrain
|
34 |
+
params:
|
35 |
+
size: 256
|
36 |
+
degradation: pil_nearest
|
37 |
+
validation:
|
38 |
+
target: ldm.data.imagenet.ImageNetSRValidation
|
39 |
+
params:
|
40 |
+
size: 256
|
41 |
+
degradation: pil_nearest
|
42 |
+
|
43 |
+
lightning:
|
44 |
+
callbacks:
|
45 |
+
image_logger:
|
46 |
+
target: main.ImageLogger
|
47 |
+
params:
|
48 |
+
batch_frequency: 1000
|
49 |
+
max_images: 8
|
50 |
+
increase_log_steps: True
|
51 |
+
|
52 |
+
trainer:
|
53 |
+
benchmark: True
|
54 |
+
accumulate_grad_batches: 2
|
stable-diffusion/configs/autoencoder/autoencoder_kl_32x32x4.yaml
ADDED
@@ -0,0 +1,53 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
model:
|
2 |
+
base_learning_rate: 4.5e-6
|
3 |
+
target: ldm.models.autoencoder.AutoencoderKL
|
4 |
+
params:
|
5 |
+
monitor: "val/rec_loss"
|
6 |
+
embed_dim: 4
|
7 |
+
lossconfig:
|
8 |
+
target: ldm.modules.losses.LPIPSWithDiscriminator
|
9 |
+
params:
|
10 |
+
disc_start: 50001
|
11 |
+
kl_weight: 0.000001
|
12 |
+
disc_weight: 0.5
|
13 |
+
|
14 |
+
ddconfig:
|
15 |
+
double_z: True
|
16 |
+
z_channels: 4
|
17 |
+
resolution: 256
|
18 |
+
in_channels: 3
|
19 |
+
out_ch: 3
|
20 |
+
ch: 128
|
21 |
+
ch_mult: [ 1,2,4,4 ] # num_down = len(ch_mult)-1
|
22 |
+
num_res_blocks: 2
|
23 |
+
attn_resolutions: [ ]
|
24 |
+
dropout: 0.0
|
25 |
+
|
26 |
+
data:
|
27 |
+
target: main.DataModuleFromConfig
|
28 |
+
params:
|
29 |
+
batch_size: 12
|
30 |
+
wrap: True
|
31 |
+
train:
|
32 |
+
target: ldm.data.imagenet.ImageNetSRTrain
|
33 |
+
params:
|
34 |
+
size: 256
|
35 |
+
degradation: pil_nearest
|
36 |
+
validation:
|
37 |
+
target: ldm.data.imagenet.ImageNetSRValidation
|
38 |
+
params:
|
39 |
+
size: 256
|
40 |
+
degradation: pil_nearest
|
41 |
+
|
42 |
+
lightning:
|
43 |
+
callbacks:
|
44 |
+
image_logger:
|
45 |
+
target: main.ImageLogger
|
46 |
+
params:
|
47 |
+
batch_frequency: 1000
|
48 |
+
max_images: 8
|
49 |
+
increase_log_steps: True
|
50 |
+
|
51 |
+
trainer:
|
52 |
+
benchmark: True
|
53 |
+
accumulate_grad_batches: 2
|
stable-diffusion/configs/autoencoder/autoencoder_kl_64x64x3.yaml
ADDED
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
model:
|
2 |
+
base_learning_rate: 4.5e-6
|
3 |
+
target: ldm.models.autoencoder.AutoencoderKL
|
4 |
+
params:
|
5 |
+
monitor: "val/rec_loss"
|
6 |
+
embed_dim: 3
|
7 |
+
lossconfig:
|
8 |
+
target: ldm.modules.losses.LPIPSWithDiscriminator
|
9 |
+
params:
|
10 |
+
disc_start: 50001
|
11 |
+
kl_weight: 0.000001
|
12 |
+
disc_weight: 0.5
|
13 |
+
|
14 |
+
ddconfig:
|
15 |
+
double_z: True
|
16 |
+
z_channels: 3
|
17 |
+
resolution: 256
|
18 |
+
in_channels: 3
|
19 |
+
out_ch: 3
|
20 |
+
ch: 128
|
21 |
+
ch_mult: [ 1,2,4 ] # num_down = len(ch_mult)-1
|
22 |
+
num_res_blocks: 2
|
23 |
+
attn_resolutions: [ ]
|
24 |
+
dropout: 0.0
|
25 |
+
|
26 |
+
|
27 |
+
data:
|
28 |
+
target: main.DataModuleFromConfig
|
29 |
+
params:
|
30 |
+
batch_size: 12
|
31 |
+
wrap: True
|
32 |
+
train:
|
33 |
+
target: ldm.data.imagenet.ImageNetSRTrain
|
34 |
+
params:
|
35 |
+
size: 256
|
36 |
+
degradation: pil_nearest
|
37 |
+
validation:
|
38 |
+
target: ldm.data.imagenet.ImageNetSRValidation
|
39 |
+
params:
|
40 |
+
size: 256
|
41 |
+
degradation: pil_nearest
|
42 |
+
|
43 |
+
lightning:
|
44 |
+
callbacks:
|
45 |
+
image_logger:
|
46 |
+
target: main.ImageLogger
|
47 |
+
params:
|
48 |
+
batch_frequency: 1000
|
49 |
+
max_images: 8
|
50 |
+
increase_log_steps: True
|
51 |
+
|
52 |
+
trainer:
|
53 |
+
benchmark: True
|
54 |
+
accumulate_grad_batches: 2
|
stable-diffusion/configs/autoencoder/autoencoder_kl_8x8x64.yaml
ADDED
@@ -0,0 +1,53 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
model:
|
2 |
+
base_learning_rate: 4.5e-6
|
3 |
+
target: ldm.models.autoencoder.AutoencoderKL
|
4 |
+
params:
|
5 |
+
monitor: "val/rec_loss"
|
6 |
+
embed_dim: 64
|
7 |
+
lossconfig:
|
8 |
+
target: ldm.modules.losses.LPIPSWithDiscriminator
|
9 |
+
params:
|
10 |
+
disc_start: 50001
|
11 |
+
kl_weight: 0.000001
|
12 |
+
disc_weight: 0.5
|
13 |
+
|
14 |
+
ddconfig:
|
15 |
+
double_z: True
|
16 |
+
z_channels: 64
|
17 |
+
resolution: 256
|
18 |
+
in_channels: 3
|
19 |
+
out_ch: 3
|
20 |
+
ch: 128
|
21 |
+
ch_mult: [ 1,1,2,2,4,4] # num_down = len(ch_mult)-1
|
22 |
+
num_res_blocks: 2
|
23 |
+
attn_resolutions: [16,8]
|
24 |
+
dropout: 0.0
|
25 |
+
|
26 |
+
data:
|
27 |
+
target: main.DataModuleFromConfig
|
28 |
+
params:
|
29 |
+
batch_size: 12
|
30 |
+
wrap: True
|
31 |
+
train:
|
32 |
+
target: ldm.data.imagenet.ImageNetSRTrain
|
33 |
+
params:
|
34 |
+
size: 256
|
35 |
+
degradation: pil_nearest
|
36 |
+
validation:
|
37 |
+
target: ldm.data.imagenet.ImageNetSRValidation
|
38 |
+
params:
|
39 |
+
size: 256
|
40 |
+
degradation: pil_nearest
|
41 |
+
|
42 |
+
lightning:
|
43 |
+
callbacks:
|
44 |
+
image_logger:
|
45 |
+
target: main.ImageLogger
|
46 |
+
params:
|
47 |
+
batch_frequency: 1000
|
48 |
+
max_images: 8
|
49 |
+
increase_log_steps: True
|
50 |
+
|
51 |
+
trainer:
|
52 |
+
benchmark: True
|
53 |
+
accumulate_grad_batches: 2
|
stable-diffusion/configs/latent-diffusion/celebahq-ldm-vq-4.yaml
ADDED
@@ -0,0 +1,86 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
model:
|
2 |
+
base_learning_rate: 2.0e-06
|
3 |
+
target: ldm.models.diffusion.ddpm.LatentDiffusion
|
4 |
+
params:
|
5 |
+
linear_start: 0.0015
|
6 |
+
linear_end: 0.0195
|
7 |
+
num_timesteps_cond: 1
|
8 |
+
log_every_t: 200
|
9 |
+
timesteps: 1000
|
10 |
+
first_stage_key: image
|
11 |
+
image_size: 64
|
12 |
+
channels: 3
|
13 |
+
monitor: val/loss_simple_ema
|
14 |
+
|
15 |
+
unet_config:
|
16 |
+
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
|
17 |
+
params:
|
18 |
+
image_size: 64
|
19 |
+
in_channels: 3
|
20 |
+
out_channels: 3
|
21 |
+
model_channels: 224
|
22 |
+
attention_resolutions:
|
23 |
+
# note: this isn\t actually the resolution but
|
24 |
+
# the downsampling factor, i.e. this corresnponds to
|
25 |
+
# attention on spatial resolution 8,16,32, as the
|
26 |
+
# spatial reolution of the latents is 64 for f4
|
27 |
+
- 8
|
28 |
+
- 4
|
29 |
+
- 2
|
30 |
+
num_res_blocks: 2
|
31 |
+
channel_mult:
|
32 |
+
- 1
|
33 |
+
- 2
|
34 |
+
- 3
|
35 |
+
- 4
|
36 |
+
num_head_channels: 32
|
37 |
+
first_stage_config:
|
38 |
+
target: ldm.models.autoencoder.VQModelInterface
|
39 |
+
params:
|
40 |
+
embed_dim: 3
|
41 |
+
n_embed: 8192
|
42 |
+
ckpt_path: models/first_stage_models/vq-f4/model.ckpt
|
43 |
+
ddconfig:
|
44 |
+
double_z: false
|
45 |
+
z_channels: 3
|
46 |
+
resolution: 256
|
47 |
+
in_channels: 3
|
48 |
+
out_ch: 3
|
49 |
+
ch: 128
|
50 |
+
ch_mult:
|
51 |
+
- 1
|
52 |
+
- 2
|
53 |
+
- 4
|
54 |
+
num_res_blocks: 2
|
55 |
+
attn_resolutions: []
|
56 |
+
dropout: 0.0
|
57 |
+
lossconfig:
|
58 |
+
target: torch.nn.Identity
|
59 |
+
cond_stage_config: __is_unconditional__
|
60 |
+
data:
|
61 |
+
target: main.DataModuleFromConfig
|
62 |
+
params:
|
63 |
+
batch_size: 48
|
64 |
+
num_workers: 5
|
65 |
+
wrap: false
|
66 |
+
train:
|
67 |
+
target: taming.data.faceshq.CelebAHQTrain
|
68 |
+
params:
|
69 |
+
size: 256
|
70 |
+
validation:
|
71 |
+
target: taming.data.faceshq.CelebAHQValidation
|
72 |
+
params:
|
73 |
+
size: 256
|
74 |
+
|
75 |
+
|
76 |
+
lightning:
|
77 |
+
callbacks:
|
78 |
+
image_logger:
|
79 |
+
target: main.ImageLogger
|
80 |
+
params:
|
81 |
+
batch_frequency: 5000
|
82 |
+
max_images: 8
|
83 |
+
increase_log_steps: False
|
84 |
+
|
85 |
+
trainer:
|
86 |
+
benchmark: True
|
stable-diffusion/configs/latent-diffusion/cin-ldm-vq-f8.yaml
ADDED
@@ -0,0 +1,98 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
model:
|
2 |
+
base_learning_rate: 1.0e-06
|
3 |
+
target: ldm.models.diffusion.ddpm.LatentDiffusion
|
4 |
+
params:
|
5 |
+
linear_start: 0.0015
|
6 |
+
linear_end: 0.0195
|
7 |
+
num_timesteps_cond: 1
|
8 |
+
log_every_t: 200
|
9 |
+
timesteps: 1000
|
10 |
+
first_stage_key: image
|
11 |
+
cond_stage_key: class_label
|
12 |
+
image_size: 32
|
13 |
+
channels: 4
|
14 |
+
cond_stage_trainable: true
|
15 |
+
conditioning_key: crossattn
|
16 |
+
monitor: val/loss_simple_ema
|
17 |
+
unet_config:
|
18 |
+
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
|
19 |
+
params:
|
20 |
+
image_size: 32
|
21 |
+
in_channels: 4
|
22 |
+
out_channels: 4
|
23 |
+
model_channels: 256
|
24 |
+
attention_resolutions:
|
25 |
+
#note: this isn\t actually the resolution but
|
26 |
+
# the downsampling factor, i.e. this corresnponds to
|
27 |
+
# attention on spatial resolution 8,16,32, as the
|
28 |
+
# spatial reolution of the latents is 32 for f8
|
29 |
+
- 4
|
30 |
+
- 2
|
31 |
+
- 1
|
32 |
+
num_res_blocks: 2
|
33 |
+
channel_mult:
|
34 |
+
- 1
|
35 |
+
- 2
|
36 |
+
- 4
|
37 |
+
num_head_channels: 32
|
38 |
+
use_spatial_transformer: true
|
39 |
+
transformer_depth: 1
|
40 |
+
context_dim: 512
|
41 |
+
first_stage_config:
|
42 |
+
target: ldm.models.autoencoder.VQModelInterface
|
43 |
+
params:
|
44 |
+
embed_dim: 4
|
45 |
+
n_embed: 16384
|
46 |
+
ckpt_path: configs/first_stage_models/vq-f8/model.yaml
|
47 |
+
ddconfig:
|
48 |
+
double_z: false
|
49 |
+
z_channels: 4
|
50 |
+
resolution: 256
|
51 |
+
in_channels: 3
|
52 |
+
out_ch: 3
|
53 |
+
ch: 128
|
54 |
+
ch_mult:
|
55 |
+
- 1
|
56 |
+
- 2
|
57 |
+
- 2
|
58 |
+
- 4
|
59 |
+
num_res_blocks: 2
|
60 |
+
attn_resolutions:
|
61 |
+
- 32
|
62 |
+
dropout: 0.0
|
63 |
+
lossconfig:
|
64 |
+
target: torch.nn.Identity
|
65 |
+
cond_stage_config:
|
66 |
+
target: ldm.modules.encoders.modules.ClassEmbedder
|
67 |
+
params:
|
68 |
+
embed_dim: 512
|
69 |
+
key: class_label
|
70 |
+
data:
|
71 |
+
target: main.DataModuleFromConfig
|
72 |
+
params:
|
73 |
+
batch_size: 64
|
74 |
+
num_workers: 12
|
75 |
+
wrap: false
|
76 |
+
train:
|
77 |
+
target: ldm.data.imagenet.ImageNetTrain
|
78 |
+
params:
|
79 |
+
config:
|
80 |
+
size: 256
|
81 |
+
validation:
|
82 |
+
target: ldm.data.imagenet.ImageNetValidation
|
83 |
+
params:
|
84 |
+
config:
|
85 |
+
size: 256
|
86 |
+
|
87 |
+
|
88 |
+
lightning:
|
89 |
+
callbacks:
|
90 |
+
image_logger:
|
91 |
+
target: main.ImageLogger
|
92 |
+
params:
|
93 |
+
batch_frequency: 5000
|
94 |
+
max_images: 8
|
95 |
+
increase_log_steps: False
|
96 |
+
|
97 |
+
trainer:
|
98 |
+
benchmark: True
|
stable-diffusion/configs/latent-diffusion/cin256-v2.yaml
ADDED
@@ -0,0 +1,68 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
model:
|
2 |
+
base_learning_rate: 0.0001
|
3 |
+
target: ldm.models.diffusion.ddpm.LatentDiffusion
|
4 |
+
params:
|
5 |
+
linear_start: 0.0015
|
6 |
+
linear_end: 0.0195
|
7 |
+
num_timesteps_cond: 1
|
8 |
+
log_every_t: 200
|
9 |
+
timesteps: 1000
|
10 |
+
first_stage_key: image
|
11 |
+
cond_stage_key: class_label
|
12 |
+
image_size: 64
|
13 |
+
channels: 3
|
14 |
+
cond_stage_trainable: true
|
15 |
+
conditioning_key: crossattn
|
16 |
+
monitor: val/loss
|
17 |
+
use_ema: False
|
18 |
+
|
19 |
+
unet_config:
|
20 |
+
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
|
21 |
+
params:
|
22 |
+
image_size: 64
|
23 |
+
in_channels: 3
|
24 |
+
out_channels: 3
|
25 |
+
model_channels: 192
|
26 |
+
attention_resolutions:
|
27 |
+
- 8
|
28 |
+
- 4
|
29 |
+
- 2
|
30 |
+
num_res_blocks: 2
|
31 |
+
channel_mult:
|
32 |
+
- 1
|
33 |
+
- 2
|
34 |
+
- 3
|
35 |
+
- 5
|
36 |
+
num_heads: 1
|
37 |
+
use_spatial_transformer: true
|
38 |
+
transformer_depth: 1
|
39 |
+
context_dim: 512
|
40 |
+
|
41 |
+
first_stage_config:
|
42 |
+
target: ldm.models.autoencoder.VQModelInterface
|
43 |
+
params:
|
44 |
+
embed_dim: 3
|
45 |
+
n_embed: 8192
|
46 |
+
ddconfig:
|
47 |
+
double_z: false
|
48 |
+
z_channels: 3
|
49 |
+
resolution: 256
|
50 |
+
in_channels: 3
|
51 |
+
out_ch: 3
|
52 |
+
ch: 128
|
53 |
+
ch_mult:
|
54 |
+
- 1
|
55 |
+
- 2
|
56 |
+
- 4
|
57 |
+
num_res_blocks: 2
|
58 |
+
attn_resolutions: []
|
59 |
+
dropout: 0.0
|
60 |
+
lossconfig:
|
61 |
+
target: torch.nn.Identity
|
62 |
+
|
63 |
+
cond_stage_config:
|
64 |
+
target: ldm.modules.encoders.modules.ClassEmbedder
|
65 |
+
params:
|
66 |
+
n_classes: 1001
|
67 |
+
embed_dim: 512
|
68 |
+
key: class_label
|
stable-diffusion/configs/latent-diffusion/ffhq-ldm-vq-4.yaml
ADDED
@@ -0,0 +1,85 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
model:
|
2 |
+
base_learning_rate: 2.0e-06
|
3 |
+
target: ldm.models.diffusion.ddpm.LatentDiffusion
|
4 |
+
params:
|
5 |
+
linear_start: 0.0015
|
6 |
+
linear_end: 0.0195
|
7 |
+
num_timesteps_cond: 1
|
8 |
+
log_every_t: 200
|
9 |
+
timesteps: 1000
|
10 |
+
first_stage_key: image
|
11 |
+
image_size: 64
|
12 |
+
channels: 3
|
13 |
+
monitor: val/loss_simple_ema
|
14 |
+
unet_config:
|
15 |
+
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
|
16 |
+
params:
|
17 |
+
image_size: 64
|
18 |
+
in_channels: 3
|
19 |
+
out_channels: 3
|
20 |
+
model_channels: 224
|
21 |
+
attention_resolutions:
|
22 |
+
# note: this isn\t actually the resolution but
|
23 |
+
# the downsampling factor, i.e. this corresnponds to
|
24 |
+
# attention on spatial resolution 8,16,32, as the
|
25 |
+
# spatial reolution of the latents is 64 for f4
|
26 |
+
- 8
|
27 |
+
- 4
|
28 |
+
- 2
|
29 |
+
num_res_blocks: 2
|
30 |
+
channel_mult:
|
31 |
+
- 1
|
32 |
+
- 2
|
33 |
+
- 3
|
34 |
+
- 4
|
35 |
+
num_head_channels: 32
|
36 |
+
first_stage_config:
|
37 |
+
target: ldm.models.autoencoder.VQModelInterface
|
38 |
+
params:
|
39 |
+
embed_dim: 3
|
40 |
+
n_embed: 8192
|
41 |
+
ckpt_path: configs/first_stage_models/vq-f4/model.yaml
|
42 |
+
ddconfig:
|
43 |
+
double_z: false
|
44 |
+
z_channels: 3
|
45 |
+
resolution: 256
|
46 |
+
in_channels: 3
|
47 |
+
out_ch: 3
|
48 |
+
ch: 128
|
49 |
+
ch_mult:
|
50 |
+
- 1
|
51 |
+
- 2
|
52 |
+
- 4
|
53 |
+
num_res_blocks: 2
|
54 |
+
attn_resolutions: []
|
55 |
+
dropout: 0.0
|
56 |
+
lossconfig:
|
57 |
+
target: torch.nn.Identity
|
58 |
+
cond_stage_config: __is_unconditional__
|
59 |
+
data:
|
60 |
+
target: main.DataModuleFromConfig
|
61 |
+
params:
|
62 |
+
batch_size: 42
|
63 |
+
num_workers: 5
|
64 |
+
wrap: false
|
65 |
+
train:
|
66 |
+
target: taming.data.faceshq.FFHQTrain
|
67 |
+
params:
|
68 |
+
size: 256
|
69 |
+
validation:
|
70 |
+
target: taming.data.faceshq.FFHQValidation
|
71 |
+
params:
|
72 |
+
size: 256
|
73 |
+
|
74 |
+
|
75 |
+
lightning:
|
76 |
+
callbacks:
|
77 |
+
image_logger:
|
78 |
+
target: main.ImageLogger
|
79 |
+
params:
|
80 |
+
batch_frequency: 5000
|
81 |
+
max_images: 8
|
82 |
+
increase_log_steps: False
|
83 |
+
|
84 |
+
trainer:
|
85 |
+
benchmark: True
|
stable-diffusion/configs/latent-diffusion/lsun_bedrooms-ldm-vq-4.yaml
ADDED
@@ -0,0 +1,85 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
model:
|
2 |
+
base_learning_rate: 2.0e-06
|
3 |
+
target: ldm.models.diffusion.ddpm.LatentDiffusion
|
4 |
+
params:
|
5 |
+
linear_start: 0.0015
|
6 |
+
linear_end: 0.0195
|
7 |
+
num_timesteps_cond: 1
|
8 |
+
log_every_t: 200
|
9 |
+
timesteps: 1000
|
10 |
+
first_stage_key: image
|
11 |
+
image_size: 64
|
12 |
+
channels: 3
|
13 |
+
monitor: val/loss_simple_ema
|
14 |
+
unet_config:
|
15 |
+
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
|
16 |
+
params:
|
17 |
+
image_size: 64
|
18 |
+
in_channels: 3
|
19 |
+
out_channels: 3
|
20 |
+
model_channels: 224
|
21 |
+
attention_resolutions:
|
22 |
+
# note: this isn\t actually the resolution but
|
23 |
+
# the downsampling factor, i.e. this corresnponds to
|
24 |
+
# attention on spatial resolution 8,16,32, as the
|
25 |
+
# spatial reolution of the latents is 64 for f4
|
26 |
+
- 8
|
27 |
+
- 4
|
28 |
+
- 2
|
29 |
+
num_res_blocks: 2
|
30 |
+
channel_mult:
|
31 |
+
- 1
|
32 |
+
- 2
|
33 |
+
- 3
|
34 |
+
- 4
|
35 |
+
num_head_channels: 32
|
36 |
+
first_stage_config:
|
37 |
+
target: ldm.models.autoencoder.VQModelInterface
|
38 |
+
params:
|
39 |
+
ckpt_path: configs/first_stage_models/vq-f4/model.yaml
|
40 |
+
embed_dim: 3
|
41 |
+
n_embed: 8192
|
42 |
+
ddconfig:
|
43 |
+
double_z: false
|
44 |
+
z_channels: 3
|
45 |
+
resolution: 256
|
46 |
+
in_channels: 3
|
47 |
+
out_ch: 3
|
48 |
+
ch: 128
|
49 |
+
ch_mult:
|
50 |
+
- 1
|
51 |
+
- 2
|
52 |
+
- 4
|
53 |
+
num_res_blocks: 2
|
54 |
+
attn_resolutions: []
|
55 |
+
dropout: 0.0
|
56 |
+
lossconfig:
|
57 |
+
target: torch.nn.Identity
|
58 |
+
cond_stage_config: __is_unconditional__
|
59 |
+
data:
|
60 |
+
target: main.DataModuleFromConfig
|
61 |
+
params:
|
62 |
+
batch_size: 48
|
63 |
+
num_workers: 5
|
64 |
+
wrap: false
|
65 |
+
train:
|
66 |
+
target: ldm.data.lsun.LSUNBedroomsTrain
|
67 |
+
params:
|
68 |
+
size: 256
|
69 |
+
validation:
|
70 |
+
target: ldm.data.lsun.LSUNBedroomsValidation
|
71 |
+
params:
|
72 |
+
size: 256
|
73 |
+
|
74 |
+
|
75 |
+
lightning:
|
76 |
+
callbacks:
|
77 |
+
image_logger:
|
78 |
+
target: main.ImageLogger
|
79 |
+
params:
|
80 |
+
batch_frequency: 5000
|
81 |
+
max_images: 8
|
82 |
+
increase_log_steps: False
|
83 |
+
|
84 |
+
trainer:
|
85 |
+
benchmark: True
|
stable-diffusion/configs/latent-diffusion/lsun_churches-ldm-kl-8.yaml
ADDED
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
model:
|
2 |
+
base_learning_rate: 5.0e-5 # set to target_lr by starting main.py with '--scale_lr False'
|
3 |
+
target: ldm.models.diffusion.ddpm.LatentDiffusion
|
4 |
+
params:
|
5 |
+
linear_start: 0.0015
|
6 |
+
linear_end: 0.0155
|
7 |
+
num_timesteps_cond: 1
|
8 |
+
log_every_t: 200
|
9 |
+
timesteps: 1000
|
10 |
+
loss_type: l1
|
11 |
+
first_stage_key: "image"
|
12 |
+
cond_stage_key: "image"
|
13 |
+
image_size: 32
|
14 |
+
channels: 4
|
15 |
+
cond_stage_trainable: False
|
16 |
+
concat_mode: False
|
17 |
+
scale_by_std: True
|
18 |
+
monitor: 'val/loss_simple_ema'
|
19 |
+
|
20 |
+
scheduler_config: # 10000 warmup steps
|
21 |
+
target: ldm.lr_scheduler.LambdaLinearScheduler
|
22 |
+
params:
|
23 |
+
warm_up_steps: [10000]
|
24 |
+
cycle_lengths: [10000000000000]
|
25 |
+
f_start: [1.e-6]
|
26 |
+
f_max: [1.]
|
27 |
+
f_min: [ 1.]
|
28 |
+
|
29 |
+
unet_config:
|
30 |
+
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
|
31 |
+
params:
|
32 |
+
image_size: 32
|
33 |
+
in_channels: 4
|
34 |
+
out_channels: 4
|
35 |
+
model_channels: 192
|
36 |
+
attention_resolutions: [ 1, 2, 4, 8 ] # 32, 16, 8, 4
|
37 |
+
num_res_blocks: 2
|
38 |
+
channel_mult: [ 1,2,2,4,4 ] # 32, 16, 8, 4, 2
|
39 |
+
num_heads: 8
|
40 |
+
use_scale_shift_norm: True
|
41 |
+
resblock_updown: True
|
42 |
+
|
43 |
+
first_stage_config:
|
44 |
+
target: ldm.models.autoencoder.AutoencoderKL
|
45 |
+
params:
|
46 |
+
embed_dim: 4
|
47 |
+
monitor: "val/rec_loss"
|
48 |
+
ckpt_path: "models/first_stage_models/kl-f8/model.ckpt"
|
49 |
+
ddconfig:
|
50 |
+
double_z: True
|
51 |
+
z_channels: 4
|
52 |
+
resolution: 256
|
53 |
+
in_channels: 3
|
54 |
+
out_ch: 3
|
55 |
+
ch: 128
|
56 |
+
ch_mult: [ 1,2,4,4 ] # num_down = len(ch_mult)-1
|
57 |
+
num_res_blocks: 2
|
58 |
+
attn_resolutions: [ ]
|
59 |
+
dropout: 0.0
|
60 |
+
lossconfig:
|
61 |
+
target: torch.nn.Identity
|
62 |
+
|
63 |
+
cond_stage_config: "__is_unconditional__"
|
64 |
+
|
65 |
+
data:
|
66 |
+
target: main.DataModuleFromConfig
|
67 |
+
params:
|
68 |
+
batch_size: 96
|
69 |
+
num_workers: 5
|
70 |
+
wrap: False
|
71 |
+
train:
|
72 |
+
target: ldm.data.lsun.LSUNChurchesTrain
|
73 |
+
params:
|
74 |
+
size: 256
|
75 |
+
validation:
|
76 |
+
target: ldm.data.lsun.LSUNChurchesValidation
|
77 |
+
params:
|
78 |
+
size: 256
|
79 |
+
|
80 |
+
lightning:
|
81 |
+
callbacks:
|
82 |
+
image_logger:
|
83 |
+
target: main.ImageLogger
|
84 |
+
params:
|
85 |
+
batch_frequency: 5000
|
86 |
+
max_images: 8
|
87 |
+
increase_log_steps: False
|
88 |
+
|
89 |
+
|
90 |
+
trainer:
|
91 |
+
benchmark: True
|
stable-diffusion/configs/latent-diffusion/txt2img-1p4B-eval.yaml
ADDED
@@ -0,0 +1,71 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
model:
|
2 |
+
base_learning_rate: 5.0e-05
|
3 |
+
target: ldm.models.diffusion.ddpm.LatentDiffusion
|
4 |
+
params:
|
5 |
+
linear_start: 0.00085
|
6 |
+
linear_end: 0.012
|
7 |
+
num_timesteps_cond: 1
|
8 |
+
log_every_t: 200
|
9 |
+
timesteps: 1000
|
10 |
+
first_stage_key: image
|
11 |
+
cond_stage_key: caption
|
12 |
+
image_size: 32
|
13 |
+
channels: 4
|
14 |
+
cond_stage_trainable: true
|
15 |
+
conditioning_key: crossattn
|
16 |
+
monitor: val/loss_simple_ema
|
17 |
+
scale_factor: 0.18215
|
18 |
+
use_ema: False
|
19 |
+
|
20 |
+
unet_config:
|
21 |
+
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
|
22 |
+
params:
|
23 |
+
image_size: 32
|
24 |
+
in_channels: 4
|
25 |
+
out_channels: 4
|
26 |
+
model_channels: 320
|
27 |
+
attention_resolutions:
|
28 |
+
- 4
|
29 |
+
- 2
|
30 |
+
- 1
|
31 |
+
num_res_blocks: 2
|
32 |
+
channel_mult:
|
33 |
+
- 1
|
34 |
+
- 2
|
35 |
+
- 4
|
36 |
+
- 4
|
37 |
+
num_heads: 8
|
38 |
+
use_spatial_transformer: true
|
39 |
+
transformer_depth: 1
|
40 |
+
context_dim: 1280
|
41 |
+
use_checkpoint: true
|
42 |
+
legacy: False
|
43 |
+
|
44 |
+
first_stage_config:
|
45 |
+
target: ldm.models.autoencoder.AutoencoderKL
|
46 |
+
params:
|
47 |
+
embed_dim: 4
|
48 |
+
monitor: val/rec_loss
|
49 |
+
ddconfig:
|
50 |
+
double_z: true
|
51 |
+
z_channels: 4
|
52 |
+
resolution: 256
|
53 |
+
in_channels: 3
|
54 |
+
out_ch: 3
|
55 |
+
ch: 128
|
56 |
+
ch_mult:
|
57 |
+
- 1
|
58 |
+
- 2
|
59 |
+
- 4
|
60 |
+
- 4
|
61 |
+
num_res_blocks: 2
|
62 |
+
attn_resolutions: []
|
63 |
+
dropout: 0.0
|
64 |
+
lossconfig:
|
65 |
+
target: torch.nn.Identity
|
66 |
+
|
67 |
+
cond_stage_config:
|
68 |
+
target: ldm.modules.encoders.modules.BERTEmbedder
|
69 |
+
params:
|
70 |
+
n_embed: 1280
|
71 |
+
n_layer: 32
|
stable-diffusion/configs/retrieval-augmented-diffusion/768x768.yaml
ADDED
@@ -0,0 +1,68 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
model:
|
2 |
+
base_learning_rate: 0.0001
|
3 |
+
target: ldm.models.diffusion.ddpm.LatentDiffusion
|
4 |
+
params:
|
5 |
+
linear_start: 0.0015
|
6 |
+
linear_end: 0.015
|
7 |
+
num_timesteps_cond: 1
|
8 |
+
log_every_t: 200
|
9 |
+
timesteps: 1000
|
10 |
+
first_stage_key: jpg
|
11 |
+
cond_stage_key: nix
|
12 |
+
image_size: 48
|
13 |
+
channels: 16
|
14 |
+
cond_stage_trainable: false
|
15 |
+
conditioning_key: crossattn
|
16 |
+
monitor: val/loss_simple_ema
|
17 |
+
scale_by_std: false
|
18 |
+
scale_factor: 0.22765929
|
19 |
+
unet_config:
|
20 |
+
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
|
21 |
+
params:
|
22 |
+
image_size: 48
|
23 |
+
in_channels: 16
|
24 |
+
out_channels: 16
|
25 |
+
model_channels: 448
|
26 |
+
attention_resolutions:
|
27 |
+
- 4
|
28 |
+
- 2
|
29 |
+
- 1
|
30 |
+
num_res_blocks: 2
|
31 |
+
channel_mult:
|
32 |
+
- 1
|
33 |
+
- 2
|
34 |
+
- 3
|
35 |
+
- 4
|
36 |
+
use_scale_shift_norm: false
|
37 |
+
resblock_updown: false
|
38 |
+
num_head_channels: 32
|
39 |
+
use_spatial_transformer: true
|
40 |
+
transformer_depth: 1
|
41 |
+
context_dim: 768
|
42 |
+
use_checkpoint: true
|
43 |
+
first_stage_config:
|
44 |
+
target: ldm.models.autoencoder.AutoencoderKL
|
45 |
+
params:
|
46 |
+
monitor: val/rec_loss
|
47 |
+
embed_dim: 16
|
48 |
+
ddconfig:
|
49 |
+
double_z: true
|
50 |
+
z_channels: 16
|
51 |
+
resolution: 256
|
52 |
+
in_channels: 3
|
53 |
+
out_ch: 3
|
54 |
+
ch: 128
|
55 |
+
ch_mult:
|
56 |
+
- 1
|
57 |
+
- 1
|
58 |
+
- 2
|
59 |
+
- 2
|
60 |
+
- 4
|
61 |
+
num_res_blocks: 2
|
62 |
+
attn_resolutions:
|
63 |
+
- 16
|
64 |
+
dropout: 0.0
|
65 |
+
lossconfig:
|
66 |
+
target: torch.nn.Identity
|
67 |
+
cond_stage_config:
|
68 |
+
target: torch.nn.Identity
|
stable-diffusion/configs/stable-diffusion/v1-inference.yaml
ADDED
@@ -0,0 +1,70 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
model:
|
2 |
+
base_learning_rate: 1.0e-04
|
3 |
+
target: ldm.models.diffusion.ddpm.LatentDiffusion
|
4 |
+
params:
|
5 |
+
linear_start: 0.00085
|
6 |
+
linear_end: 0.0120
|
7 |
+
num_timesteps_cond: 1
|
8 |
+
log_every_t: 200
|
9 |
+
timesteps: 1000
|
10 |
+
first_stage_key: "jpg"
|
11 |
+
cond_stage_key: "txt"
|
12 |
+
image_size: 64
|
13 |
+
channels: 4
|
14 |
+
cond_stage_trainable: false # Note: different from the one we trained before
|
15 |
+
conditioning_key: crossattn
|
16 |
+
monitor: val/loss_simple_ema
|
17 |
+
scale_factor: 0.18215
|
18 |
+
use_ema: False
|
19 |
+
|
20 |
+
scheduler_config: # 10000 warmup steps
|
21 |
+
target: ldm.lr_scheduler.LambdaLinearScheduler
|
22 |
+
params:
|
23 |
+
warm_up_steps: [ 10000 ]
|
24 |
+
cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
|
25 |
+
f_start: [ 1.e-6 ]
|
26 |
+
f_max: [ 1. ]
|
27 |
+
f_min: [ 1. ]
|
28 |
+
|
29 |
+
unet_config:
|
30 |
+
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
|
31 |
+
params:
|
32 |
+
image_size: 32 # unused
|
33 |
+
in_channels: 4
|
34 |
+
out_channels: 4
|
35 |
+
model_channels: 320
|
36 |
+
attention_resolutions: [ 4, 2, 1 ]
|
37 |
+
num_res_blocks: 2
|
38 |
+
channel_mult: [ 1, 2, 4, 4 ]
|
39 |
+
num_heads: 8
|
40 |
+
use_spatial_transformer: True
|
41 |
+
transformer_depth: 1
|
42 |
+
context_dim: 768
|
43 |
+
use_checkpoint: True
|
44 |
+
legacy: False
|
45 |
+
|
46 |
+
first_stage_config:
|
47 |
+
target: ldm.models.autoencoder.AutoencoderKL
|
48 |
+
params:
|
49 |
+
embed_dim: 4
|
50 |
+
monitor: val/rec_loss
|
51 |
+
ddconfig:
|
52 |
+
double_z: true
|
53 |
+
z_channels: 4
|
54 |
+
resolution: 256
|
55 |
+
in_channels: 3
|
56 |
+
out_ch: 3
|
57 |
+
ch: 128
|
58 |
+
ch_mult:
|
59 |
+
- 1
|
60 |
+
- 2
|
61 |
+
- 4
|
62 |
+
- 4
|
63 |
+
num_res_blocks: 2
|
64 |
+
attn_resolutions: []
|
65 |
+
dropout: 0.0
|
66 |
+
lossconfig:
|
67 |
+
target: torch.nn.Identity
|
68 |
+
|
69 |
+
cond_stage_config:
|
70 |
+
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder
|
stable-diffusion/debug/generate.sh
ADDED
@@ -0,0 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
export CUDA_VISIBLE_DEVICES='2'
|
2 |
+
python scripts/inverse.py \
|
3 |
+
--file_id='00019.png' \
|
4 |
+
--task_config='configs/motion_deblur_config.yaml' \
|
5 |
+
--inpainting=0 \
|
6 |
+
--general_inverse=0 \
|
7 |
+
--gamma=1e-1 \
|
8 |
+
--omega=1e-1 \
|
9 |
+
--W=256 \
|
10 |
+
--H=256 \
|
11 |
+
--scale=5.0 \
|
12 |
+
--laion400m \
|
13 |
+
--prompt="a photograph of fantasy landscape trending in art station" \
|
14 |
+
--outdir="outputs/txt2img-samples-laion400m"
|
stable-diffusion/debug/inverse.sh
ADDED
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
export CUDA_VISIBLE_DEVICES='1'
|
2 |
+
python scripts/inverse.py \
|
3 |
+
--file_id='00019.png' \
|
4 |
+
--task_config='configs/motion_deblur_config.yaml' \
|
5 |
+
--inpainting=0 \
|
6 |
+
--general_inverse=0 \
|
7 |
+
--gamma=1e-1 \
|
8 |
+
--omega=1e-1 \
|
9 |
+
--W=256 \
|
10 |
+
--H=256 \
|
11 |
+
--scale=5.0 \
|
12 |
+
--laion400m \
|
13 |
+
--prompt="a virus monster is playing guitar, oil on canvas"
|
stable-diffusion/debug/inverse_bip_ldm_laion.sh
ADDED
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
export CUDA_VISIBLE_DEVICES='1'
|
2 |
+
python scripts/inverse.py \
|
3 |
+
--file_id='00019.png' \
|
4 |
+
--task_config='configs/box_inpainting_config.yaml' \
|
5 |
+
--inpainting=1 \
|
6 |
+
--general_inverse=0 \
|
7 |
+
--gamma=1e-1 \
|
8 |
+
--omega=1 \
|
9 |
+
--W=256 \
|
10 |
+
--H=256 \
|
11 |
+
--scale=5.0 \
|
12 |
+
--laion400m \
|
13 |
+
--outdir="outputs/psld-ldm-laion400m-bip"
|
stable-diffusion/debug/inverse_mb_ldm_laion.sh
ADDED
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
export CUDA_VISIBLE_DEVICES='1'
|
2 |
+
python scripts/inverse.py \
|
3 |
+
--file_id='00019.png' \
|
4 |
+
--task_config='configs/motion_deblur_config.yaml' \
|
5 |
+
--inpainting=0 \
|
6 |
+
--general_inverse=1 \
|
7 |
+
--gamma=1e-1 \
|
8 |
+
--omega=1 \
|
9 |
+
--W=256 \
|
10 |
+
--H=256 \
|
11 |
+
--scale=5.0 \
|
12 |
+
--laion400m \
|
13 |
+
--outdir="outputs/psld-ldm-laion400m-mb"
|
stable-diffusion/ldm/__pycache__/util.cpython-38.pyc
ADDED
Binary file (6.09 kB). View file
|
|
stable-diffusion/ldm/__pycache__/util.cpython-39.pyc
ADDED
Binary file (6.12 kB). View file
|
|
stable-diffusion/ldm/data/__init__.py
ADDED
File without changes
|
stable-diffusion/ldm/data/base.py
ADDED
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from abc import abstractmethod
|
2 |
+
from torch.utils.data import Dataset, ConcatDataset, ChainDataset, IterableDataset
|
3 |
+
|
4 |
+
|
5 |
+
class Txt2ImgIterableBaseDataset(IterableDataset):
|
6 |
+
'''
|
7 |
+
Define an interface to make the IterableDatasets for text2img data chainable
|
8 |
+
'''
|
9 |
+
def __init__(self, num_records=0, valid_ids=None, size=256):
|
10 |
+
super().__init__()
|
11 |
+
self.num_records = num_records
|
12 |
+
self.valid_ids = valid_ids
|
13 |
+
self.sample_ids = valid_ids
|
14 |
+
self.size = size
|
15 |
+
|
16 |
+
print(f'{self.__class__.__name__} dataset contains {self.__len__()} examples.')
|
17 |
+
|
18 |
+
def __len__(self):
|
19 |
+
return self.num_records
|
20 |
+
|
21 |
+
@abstractmethod
|
22 |
+
def __iter__(self):
|
23 |
+
pass
|
stable-diffusion/ldm/data/imagenet.py
ADDED
@@ -0,0 +1,394 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os, yaml, pickle, shutil, tarfile, glob
|
2 |
+
import cv2
|
3 |
+
import albumentations
|
4 |
+
import PIL
|
5 |
+
import numpy as np
|
6 |
+
import torchvision.transforms.functional as TF
|
7 |
+
from omegaconf import OmegaConf
|
8 |
+
from functools import partial
|
9 |
+
from PIL import Image
|
10 |
+
from tqdm import tqdm
|
11 |
+
from torch.utils.data import Dataset, Subset
|
12 |
+
|
13 |
+
import taming.data.utils as tdu
|
14 |
+
from taming.data.imagenet import str_to_indices, give_synsets_from_indices, download, retrieve
|
15 |
+
from taming.data.imagenet import ImagePaths
|
16 |
+
|
17 |
+
from ldm.modules.image_degradation import degradation_fn_bsr, degradation_fn_bsr_light
|
18 |
+
|
19 |
+
|
20 |
+
def synset2idx(path_to_yaml="data/index_synset.yaml"):
|
21 |
+
with open(path_to_yaml) as f:
|
22 |
+
di2s = yaml.load(f)
|
23 |
+
return dict((v,k) for k,v in di2s.items())
|
24 |
+
|
25 |
+
|
26 |
+
class ImageNetBase(Dataset):
|
27 |
+
def __init__(self, config=None):
|
28 |
+
self.config = config or OmegaConf.create()
|
29 |
+
if not type(self.config)==dict:
|
30 |
+
self.config = OmegaConf.to_container(self.config)
|
31 |
+
self.keep_orig_class_label = self.config.get("keep_orig_class_label", False)
|
32 |
+
self.process_images = True # if False we skip loading & processing images and self.data contains filepaths
|
33 |
+
self._prepare()
|
34 |
+
self._prepare_synset_to_human()
|
35 |
+
self._prepare_idx_to_synset()
|
36 |
+
self._prepare_human_to_integer_label()
|
37 |
+
self._load()
|
38 |
+
|
39 |
+
def __len__(self):
|
40 |
+
return len(self.data)
|
41 |
+
|
42 |
+
def __getitem__(self, i):
|
43 |
+
return self.data[i]
|
44 |
+
|
45 |
+
def _prepare(self):
|
46 |
+
raise NotImplementedError()
|
47 |
+
|
48 |
+
def _filter_relpaths(self, relpaths):
|
49 |
+
ignore = set([
|
50 |
+
"n06596364_9591.JPEG",
|
51 |
+
])
|
52 |
+
relpaths = [rpath for rpath in relpaths if not rpath.split("/")[-1] in ignore]
|
53 |
+
if "sub_indices" in self.config:
|
54 |
+
indices = str_to_indices(self.config["sub_indices"])
|
55 |
+
synsets = give_synsets_from_indices(indices, path_to_yaml=self.idx2syn) # returns a list of strings
|
56 |
+
self.synset2idx = synset2idx(path_to_yaml=self.idx2syn)
|
57 |
+
files = []
|
58 |
+
for rpath in relpaths:
|
59 |
+
syn = rpath.split("/")[0]
|
60 |
+
if syn in synsets:
|
61 |
+
files.append(rpath)
|
62 |
+
return files
|
63 |
+
else:
|
64 |
+
return relpaths
|
65 |
+
|
66 |
+
def _prepare_synset_to_human(self):
|
67 |
+
SIZE = 2655750
|
68 |
+
URL = "https://heibox.uni-heidelberg.de/f/9f28e956cd304264bb82/?dl=1"
|
69 |
+
self.human_dict = os.path.join(self.root, "synset_human.txt")
|
70 |
+
if (not os.path.exists(self.human_dict) or
|
71 |
+
not os.path.getsize(self.human_dict)==SIZE):
|
72 |
+
download(URL, self.human_dict)
|
73 |
+
|
74 |
+
def _prepare_idx_to_synset(self):
|
75 |
+
URL = "https://heibox.uni-heidelberg.de/f/d835d5b6ceda4d3aa910/?dl=1"
|
76 |
+
self.idx2syn = os.path.join(self.root, "index_synset.yaml")
|
77 |
+
if (not os.path.exists(self.idx2syn)):
|
78 |
+
download(URL, self.idx2syn)
|
79 |
+
|
80 |
+
def _prepare_human_to_integer_label(self):
|
81 |
+
URL = "https://heibox.uni-heidelberg.de/f/2362b797d5be43b883f6/?dl=1"
|
82 |
+
self.human2integer = os.path.join(self.root, "imagenet1000_clsidx_to_labels.txt")
|
83 |
+
if (not os.path.exists(self.human2integer)):
|
84 |
+
download(URL, self.human2integer)
|
85 |
+
with open(self.human2integer, "r") as f:
|
86 |
+
lines = f.read().splitlines()
|
87 |
+
assert len(lines) == 1000
|
88 |
+
self.human2integer_dict = dict()
|
89 |
+
for line in lines:
|
90 |
+
value, key = line.split(":")
|
91 |
+
self.human2integer_dict[key] = int(value)
|
92 |
+
|
93 |
+
def _load(self):
|
94 |
+
with open(self.txt_filelist, "r") as f:
|
95 |
+
self.relpaths = f.read().splitlines()
|
96 |
+
l1 = len(self.relpaths)
|
97 |
+
self.relpaths = self._filter_relpaths(self.relpaths)
|
98 |
+
print("Removed {} files from filelist during filtering.".format(l1 - len(self.relpaths)))
|
99 |
+
|
100 |
+
self.synsets = [p.split("/")[0] for p in self.relpaths]
|
101 |
+
self.abspaths = [os.path.join(self.datadir, p) for p in self.relpaths]
|
102 |
+
|
103 |
+
unique_synsets = np.unique(self.synsets)
|
104 |
+
class_dict = dict((synset, i) for i, synset in enumerate(unique_synsets))
|
105 |
+
if not self.keep_orig_class_label:
|
106 |
+
self.class_labels = [class_dict[s] for s in self.synsets]
|
107 |
+
else:
|
108 |
+
self.class_labels = [self.synset2idx[s] for s in self.synsets]
|
109 |
+
|
110 |
+
with open(self.human_dict, "r") as f:
|
111 |
+
human_dict = f.read().splitlines()
|
112 |
+
human_dict = dict(line.split(maxsplit=1) for line in human_dict)
|
113 |
+
|
114 |
+
self.human_labels = [human_dict[s] for s in self.synsets]
|
115 |
+
|
116 |
+
labels = {
|
117 |
+
"relpath": np.array(self.relpaths),
|
118 |
+
"synsets": np.array(self.synsets),
|
119 |
+
"class_label": np.array(self.class_labels),
|
120 |
+
"human_label": np.array(self.human_labels),
|
121 |
+
}
|
122 |
+
|
123 |
+
if self.process_images:
|
124 |
+
self.size = retrieve(self.config, "size", default=256)
|
125 |
+
self.data = ImagePaths(self.abspaths,
|
126 |
+
labels=labels,
|
127 |
+
size=self.size,
|
128 |
+
random_crop=self.random_crop,
|
129 |
+
)
|
130 |
+
else:
|
131 |
+
self.data = self.abspaths
|
132 |
+
|
133 |
+
|
134 |
+
class ImageNetTrain(ImageNetBase):
|
135 |
+
NAME = "ILSVRC2012_train"
|
136 |
+
URL = "http://www.image-net.org/challenges/LSVRC/2012/"
|
137 |
+
AT_HASH = "a306397ccf9c2ead27155983c254227c0fd938e2"
|
138 |
+
FILES = [
|
139 |
+
"ILSVRC2012_img_train.tar",
|
140 |
+
]
|
141 |
+
SIZES = [
|
142 |
+
147897477120,
|
143 |
+
]
|
144 |
+
|
145 |
+
def __init__(self, process_images=True, data_root=None, **kwargs):
|
146 |
+
self.process_images = process_images
|
147 |
+
self.data_root = data_root
|
148 |
+
super().__init__(**kwargs)
|
149 |
+
|
150 |
+
def _prepare(self):
|
151 |
+
if self.data_root:
|
152 |
+
self.root = os.path.join(self.data_root, self.NAME)
|
153 |
+
else:
|
154 |
+
cachedir = os.environ.get("XDG_CACHE_HOME", os.path.expanduser("~/.cache"))
|
155 |
+
self.root = os.path.join(cachedir, "autoencoders/data", self.NAME)
|
156 |
+
|
157 |
+
self.datadir = os.path.join(self.root, "data")
|
158 |
+
self.txt_filelist = os.path.join(self.root, "filelist.txt")
|
159 |
+
self.expected_length = 1281167
|
160 |
+
self.random_crop = retrieve(self.config, "ImageNetTrain/random_crop",
|
161 |
+
default=True)
|
162 |
+
if not tdu.is_prepared(self.root):
|
163 |
+
# prep
|
164 |
+
print("Preparing dataset {} in {}".format(self.NAME, self.root))
|
165 |
+
|
166 |
+
datadir = self.datadir
|
167 |
+
if not os.path.exists(datadir):
|
168 |
+
path = os.path.join(self.root, self.FILES[0])
|
169 |
+
if not os.path.exists(path) or not os.path.getsize(path)==self.SIZES[0]:
|
170 |
+
import academictorrents as at
|
171 |
+
atpath = at.get(self.AT_HASH, datastore=self.root)
|
172 |
+
assert atpath == path
|
173 |
+
|
174 |
+
print("Extracting {} to {}".format(path, datadir))
|
175 |
+
os.makedirs(datadir, exist_ok=True)
|
176 |
+
with tarfile.open(path, "r:") as tar:
|
177 |
+
tar.extractall(path=datadir)
|
178 |
+
|
179 |
+
print("Extracting sub-tars.")
|
180 |
+
subpaths = sorted(glob.glob(os.path.join(datadir, "*.tar")))
|
181 |
+
for subpath in tqdm(subpaths):
|
182 |
+
subdir = subpath[:-len(".tar")]
|
183 |
+
os.makedirs(subdir, exist_ok=True)
|
184 |
+
with tarfile.open(subpath, "r:") as tar:
|
185 |
+
tar.extractall(path=subdir)
|
186 |
+
|
187 |
+
filelist = glob.glob(os.path.join(datadir, "**", "*.JPEG"))
|
188 |
+
filelist = [os.path.relpath(p, start=datadir) for p in filelist]
|
189 |
+
filelist = sorted(filelist)
|
190 |
+
filelist = "\n".join(filelist)+"\n"
|
191 |
+
with open(self.txt_filelist, "w") as f:
|
192 |
+
f.write(filelist)
|
193 |
+
|
194 |
+
tdu.mark_prepared(self.root)
|
195 |
+
|
196 |
+
|
197 |
+
class ImageNetValidation(ImageNetBase):
|
198 |
+
NAME = "ILSVRC2012_validation"
|
199 |
+
URL = "http://www.image-net.org/challenges/LSVRC/2012/"
|
200 |
+
AT_HASH = "5d6d0df7ed81efd49ca99ea4737e0ae5e3a5f2e5"
|
201 |
+
VS_URL = "https://heibox.uni-heidelberg.de/f/3e0f6e9c624e45f2bd73/?dl=1"
|
202 |
+
FILES = [
|
203 |
+
"ILSVRC2012_img_val.tar",
|
204 |
+
"validation_synset.txt",
|
205 |
+
]
|
206 |
+
SIZES = [
|
207 |
+
6744924160,
|
208 |
+
1950000,
|
209 |
+
]
|
210 |
+
|
211 |
+
def __init__(self, process_images=True, data_root=None, **kwargs):
|
212 |
+
self.data_root = data_root
|
213 |
+
self.process_images = process_images
|
214 |
+
super().__init__(**kwargs)
|
215 |
+
|
216 |
+
def _prepare(self):
|
217 |
+
if self.data_root:
|
218 |
+
self.root = os.path.join(self.data_root, self.NAME)
|
219 |
+
else:
|
220 |
+
cachedir = os.environ.get("XDG_CACHE_HOME", os.path.expanduser("~/.cache"))
|
221 |
+
self.root = os.path.join(cachedir, "autoencoders/data", self.NAME)
|
222 |
+
self.datadir = os.path.join(self.root, "data")
|
223 |
+
self.txt_filelist = os.path.join(self.root, "filelist.txt")
|
224 |
+
self.expected_length = 50000
|
225 |
+
self.random_crop = retrieve(self.config, "ImageNetValidation/random_crop",
|
226 |
+
default=False)
|
227 |
+
if not tdu.is_prepared(self.root):
|
228 |
+
# prep
|
229 |
+
print("Preparing dataset {} in {}".format(self.NAME, self.root))
|
230 |
+
|
231 |
+
datadir = self.datadir
|
232 |
+
if not os.path.exists(datadir):
|
233 |
+
path = os.path.join(self.root, self.FILES[0])
|
234 |
+
if not os.path.exists(path) or not os.path.getsize(path)==self.SIZES[0]:
|
235 |
+
import academictorrents as at
|
236 |
+
atpath = at.get(self.AT_HASH, datastore=self.root)
|
237 |
+
assert atpath == path
|
238 |
+
|
239 |
+
print("Extracting {} to {}".format(path, datadir))
|
240 |
+
os.makedirs(datadir, exist_ok=True)
|
241 |
+
with tarfile.open(path, "r:") as tar:
|
242 |
+
tar.extractall(path=datadir)
|
243 |
+
|
244 |
+
vspath = os.path.join(self.root, self.FILES[1])
|
245 |
+
if not os.path.exists(vspath) or not os.path.getsize(vspath)==self.SIZES[1]:
|
246 |
+
download(self.VS_URL, vspath)
|
247 |
+
|
248 |
+
with open(vspath, "r") as f:
|
249 |
+
synset_dict = f.read().splitlines()
|
250 |
+
synset_dict = dict(line.split() for line in synset_dict)
|
251 |
+
|
252 |
+
print("Reorganizing into synset folders")
|
253 |
+
synsets = np.unique(list(synset_dict.values()))
|
254 |
+
for s in synsets:
|
255 |
+
os.makedirs(os.path.join(datadir, s), exist_ok=True)
|
256 |
+
for k, v in synset_dict.items():
|
257 |
+
src = os.path.join(datadir, k)
|
258 |
+
dst = os.path.join(datadir, v)
|
259 |
+
shutil.move(src, dst)
|
260 |
+
|
261 |
+
filelist = glob.glob(os.path.join(datadir, "**", "*.JPEG"))
|
262 |
+
filelist = [os.path.relpath(p, start=datadir) for p in filelist]
|
263 |
+
filelist = sorted(filelist)
|
264 |
+
filelist = "\n".join(filelist)+"\n"
|
265 |
+
with open(self.txt_filelist, "w") as f:
|
266 |
+
f.write(filelist)
|
267 |
+
|
268 |
+
tdu.mark_prepared(self.root)
|
269 |
+
|
270 |
+
|
271 |
+
|
272 |
+
class ImageNetSR(Dataset):
|
273 |
+
def __init__(self, size=None,
|
274 |
+
degradation=None, downscale_f=4, min_crop_f=0.5, max_crop_f=1.,
|
275 |
+
random_crop=True):
|
276 |
+
"""
|
277 |
+
Imagenet Superresolution Dataloader
|
278 |
+
Performs following ops in order:
|
279 |
+
1. crops a crop of size s from image either as random or center crop
|
280 |
+
2. resizes crop to size with cv2.area_interpolation
|
281 |
+
3. degrades resized crop with degradation_fn
|
282 |
+
|
283 |
+
:param size: resizing to size after cropping
|
284 |
+
:param degradation: degradation_fn, e.g. cv_bicubic or bsrgan_light
|
285 |
+
:param downscale_f: Low Resolution Downsample factor
|
286 |
+
:param min_crop_f: determines crop size s,
|
287 |
+
where s = c * min_img_side_len with c sampled from interval (min_crop_f, max_crop_f)
|
288 |
+
:param max_crop_f: ""
|
289 |
+
:param data_root:
|
290 |
+
:param random_crop:
|
291 |
+
"""
|
292 |
+
self.base = self.get_base()
|
293 |
+
assert size
|
294 |
+
assert (size / downscale_f).is_integer()
|
295 |
+
self.size = size
|
296 |
+
self.LR_size = int(size / downscale_f)
|
297 |
+
self.min_crop_f = min_crop_f
|
298 |
+
self.max_crop_f = max_crop_f
|
299 |
+
assert(max_crop_f <= 1.)
|
300 |
+
self.center_crop = not random_crop
|
301 |
+
|
302 |
+
self.image_rescaler = albumentations.SmallestMaxSize(max_size=size, interpolation=cv2.INTER_AREA)
|
303 |
+
|
304 |
+
self.pil_interpolation = False # gets reset later if incase interp_op is from pillow
|
305 |
+
|
306 |
+
if degradation == "bsrgan":
|
307 |
+
self.degradation_process = partial(degradation_fn_bsr, sf=downscale_f)
|
308 |
+
|
309 |
+
elif degradation == "bsrgan_light":
|
310 |
+
self.degradation_process = partial(degradation_fn_bsr_light, sf=downscale_f)
|
311 |
+
|
312 |
+
else:
|
313 |
+
interpolation_fn = {
|
314 |
+
"cv_nearest": cv2.INTER_NEAREST,
|
315 |
+
"cv_bilinear": cv2.INTER_LINEAR,
|
316 |
+
"cv_bicubic": cv2.INTER_CUBIC,
|
317 |
+
"cv_area": cv2.INTER_AREA,
|
318 |
+
"cv_lanczos": cv2.INTER_LANCZOS4,
|
319 |
+
"pil_nearest": PIL.Image.NEAREST,
|
320 |
+
"pil_bilinear": PIL.Image.BILINEAR,
|
321 |
+
"pil_bicubic": PIL.Image.BICUBIC,
|
322 |
+
"pil_box": PIL.Image.BOX,
|
323 |
+
"pil_hamming": PIL.Image.HAMMING,
|
324 |
+
"pil_lanczos": PIL.Image.LANCZOS,
|
325 |
+
}[degradation]
|
326 |
+
|
327 |
+
self.pil_interpolation = degradation.startswith("pil_")
|
328 |
+
|
329 |
+
if self.pil_interpolation:
|
330 |
+
self.degradation_process = partial(TF.resize, size=self.LR_size, interpolation=interpolation_fn)
|
331 |
+
|
332 |
+
else:
|
333 |
+
self.degradation_process = albumentations.SmallestMaxSize(max_size=self.LR_size,
|
334 |
+
interpolation=interpolation_fn)
|
335 |
+
|
336 |
+
def __len__(self):
|
337 |
+
return len(self.base)
|
338 |
+
|
339 |
+
def __getitem__(self, i):
|
340 |
+
example = self.base[i]
|
341 |
+
image = Image.open(example["file_path_"])
|
342 |
+
|
343 |
+
if not image.mode == "RGB":
|
344 |
+
image = image.convert("RGB")
|
345 |
+
|
346 |
+
image = np.array(image).astype(np.uint8)
|
347 |
+
|
348 |
+
min_side_len = min(image.shape[:2])
|
349 |
+
crop_side_len = min_side_len * np.random.uniform(self.min_crop_f, self.max_crop_f, size=None)
|
350 |
+
crop_side_len = int(crop_side_len)
|
351 |
+
|
352 |
+
if self.center_crop:
|
353 |
+
self.cropper = albumentations.CenterCrop(height=crop_side_len, width=crop_side_len)
|
354 |
+
|
355 |
+
else:
|
356 |
+
self.cropper = albumentations.RandomCrop(height=crop_side_len, width=crop_side_len)
|
357 |
+
|
358 |
+
image = self.cropper(image=image)["image"]
|
359 |
+
image = self.image_rescaler(image=image)["image"]
|
360 |
+
|
361 |
+
if self.pil_interpolation:
|
362 |
+
image_pil = PIL.Image.fromarray(image)
|
363 |
+
LR_image = self.degradation_process(image_pil)
|
364 |
+
LR_image = np.array(LR_image).astype(np.uint8)
|
365 |
+
|
366 |
+
else:
|
367 |
+
LR_image = self.degradation_process(image=image)["image"]
|
368 |
+
|
369 |
+
example["image"] = (image/127.5 - 1.0).astype(np.float32)
|
370 |
+
example["LR_image"] = (LR_image/127.5 - 1.0).astype(np.float32)
|
371 |
+
|
372 |
+
return example
|
373 |
+
|
374 |
+
|
375 |
+
class ImageNetSRTrain(ImageNetSR):
|
376 |
+
def __init__(self, **kwargs):
|
377 |
+
super().__init__(**kwargs)
|
378 |
+
|
379 |
+
def get_base(self):
|
380 |
+
with open("data/imagenet_train_hr_indices.p", "rb") as f:
|
381 |
+
indices = pickle.load(f)
|
382 |
+
dset = ImageNetTrain(process_images=False,)
|
383 |
+
return Subset(dset, indices)
|
384 |
+
|
385 |
+
|
386 |
+
class ImageNetSRValidation(ImageNetSR):
|
387 |
+
def __init__(self, **kwargs):
|
388 |
+
super().__init__(**kwargs)
|
389 |
+
|
390 |
+
def get_base(self):
|
391 |
+
with open("data/imagenet_val_hr_indices.p", "rb") as f:
|
392 |
+
indices = pickle.load(f)
|
393 |
+
dset = ImageNetValidation(process_images=False,)
|
394 |
+
return Subset(dset, indices)
|
stable-diffusion/ldm/data/lsun.py
ADDED
@@ -0,0 +1,92 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import numpy as np
|
3 |
+
import PIL
|
4 |
+
from PIL import Image
|
5 |
+
from torch.utils.data import Dataset
|
6 |
+
from torchvision import transforms
|
7 |
+
|
8 |
+
|
9 |
+
class LSUNBase(Dataset):
|
10 |
+
def __init__(self,
|
11 |
+
txt_file,
|
12 |
+
data_root,
|
13 |
+
size=None,
|
14 |
+
interpolation="bicubic",
|
15 |
+
flip_p=0.5
|
16 |
+
):
|
17 |
+
self.data_paths = txt_file
|
18 |
+
self.data_root = data_root
|
19 |
+
with open(self.data_paths, "r") as f:
|
20 |
+
self.image_paths = f.read().splitlines()
|
21 |
+
self._length = len(self.image_paths)
|
22 |
+
self.labels = {
|
23 |
+
"relative_file_path_": [l for l in self.image_paths],
|
24 |
+
"file_path_": [os.path.join(self.data_root, l)
|
25 |
+
for l in self.image_paths],
|
26 |
+
}
|
27 |
+
|
28 |
+
self.size = size
|
29 |
+
self.interpolation = {"linear": PIL.Image.LINEAR,
|
30 |
+
"bilinear": PIL.Image.BILINEAR,
|
31 |
+
"bicubic": PIL.Image.BICUBIC,
|
32 |
+
"lanczos": PIL.Image.LANCZOS,
|
33 |
+
}[interpolation]
|
34 |
+
self.flip = transforms.RandomHorizontalFlip(p=flip_p)
|
35 |
+
|
36 |
+
def __len__(self):
|
37 |
+
return self._length
|
38 |
+
|
39 |
+
def __getitem__(self, i):
|
40 |
+
example = dict((k, self.labels[k][i]) for k in self.labels)
|
41 |
+
image = Image.open(example["file_path_"])
|
42 |
+
if not image.mode == "RGB":
|
43 |
+
image = image.convert("RGB")
|
44 |
+
|
45 |
+
# default to score-sde preprocessing
|
46 |
+
img = np.array(image).astype(np.uint8)
|
47 |
+
crop = min(img.shape[0], img.shape[1])
|
48 |
+
h, w, = img.shape[0], img.shape[1]
|
49 |
+
img = img[(h - crop) // 2:(h + crop) // 2,
|
50 |
+
(w - crop) // 2:(w + crop) // 2]
|
51 |
+
|
52 |
+
image = Image.fromarray(img)
|
53 |
+
if self.size is not None:
|
54 |
+
image = image.resize((self.size, self.size), resample=self.interpolation)
|
55 |
+
|
56 |
+
image = self.flip(image)
|
57 |
+
image = np.array(image).astype(np.uint8)
|
58 |
+
example["image"] = (image / 127.5 - 1.0).astype(np.float32)
|
59 |
+
return example
|
60 |
+
|
61 |
+
|
62 |
+
class LSUNChurchesTrain(LSUNBase):
|
63 |
+
def __init__(self, **kwargs):
|
64 |
+
super().__init__(txt_file="data/lsun/church_outdoor_train.txt", data_root="data/lsun/churches", **kwargs)
|
65 |
+
|
66 |
+
|
67 |
+
class LSUNChurchesValidation(LSUNBase):
|
68 |
+
def __init__(self, flip_p=0., **kwargs):
|
69 |
+
super().__init__(txt_file="data/lsun/church_outdoor_val.txt", data_root="data/lsun/churches",
|
70 |
+
flip_p=flip_p, **kwargs)
|
71 |
+
|
72 |
+
|
73 |
+
class LSUNBedroomsTrain(LSUNBase):
|
74 |
+
def __init__(self, **kwargs):
|
75 |
+
super().__init__(txt_file="data/lsun/bedrooms_train.txt", data_root="data/lsun/bedrooms", **kwargs)
|
76 |
+
|
77 |
+
|
78 |
+
class LSUNBedroomsValidation(LSUNBase):
|
79 |
+
def __init__(self, flip_p=0.0, **kwargs):
|
80 |
+
super().__init__(txt_file="data/lsun/bedrooms_val.txt", data_root="data/lsun/bedrooms",
|
81 |
+
flip_p=flip_p, **kwargs)
|
82 |
+
|
83 |
+
|
84 |
+
class LSUNCatsTrain(LSUNBase):
|
85 |
+
def __init__(self, **kwargs):
|
86 |
+
super().__init__(txt_file="data/lsun/cat_train.txt", data_root="data/lsun/cats", **kwargs)
|
87 |
+
|
88 |
+
|
89 |
+
class LSUNCatsValidation(LSUNBase):
|
90 |
+
def __init__(self, flip_p=0., **kwargs):
|
91 |
+
super().__init__(txt_file="data/lsun/cat_val.txt", data_root="data/lsun/cats",
|
92 |
+
flip_p=flip_p, **kwargs)
|
stable-diffusion/ldm/lr_scheduler.py
ADDED
@@ -0,0 +1,98 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
|
3 |
+
|
4 |
+
class LambdaWarmUpCosineScheduler:
|
5 |
+
"""
|
6 |
+
note: use with a base_lr of 1.0
|
7 |
+
"""
|
8 |
+
def __init__(self, warm_up_steps, lr_min, lr_max, lr_start, max_decay_steps, verbosity_interval=0):
|
9 |
+
self.lr_warm_up_steps = warm_up_steps
|
10 |
+
self.lr_start = lr_start
|
11 |
+
self.lr_min = lr_min
|
12 |
+
self.lr_max = lr_max
|
13 |
+
self.lr_max_decay_steps = max_decay_steps
|
14 |
+
self.last_lr = 0.
|
15 |
+
self.verbosity_interval = verbosity_interval
|
16 |
+
|
17 |
+
def schedule(self, n, **kwargs):
|
18 |
+
if self.verbosity_interval > 0:
|
19 |
+
if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_lr}")
|
20 |
+
if n < self.lr_warm_up_steps:
|
21 |
+
lr = (self.lr_max - self.lr_start) / self.lr_warm_up_steps * n + self.lr_start
|
22 |
+
self.last_lr = lr
|
23 |
+
return lr
|
24 |
+
else:
|
25 |
+
t = (n - self.lr_warm_up_steps) / (self.lr_max_decay_steps - self.lr_warm_up_steps)
|
26 |
+
t = min(t, 1.0)
|
27 |
+
lr = self.lr_min + 0.5 * (self.lr_max - self.lr_min) * (
|
28 |
+
1 + np.cos(t * np.pi))
|
29 |
+
self.last_lr = lr
|
30 |
+
return lr
|
31 |
+
|
32 |
+
def __call__(self, n, **kwargs):
|
33 |
+
return self.schedule(n,**kwargs)
|
34 |
+
|
35 |
+
|
36 |
+
class LambdaWarmUpCosineScheduler2:
|
37 |
+
"""
|
38 |
+
supports repeated iterations, configurable via lists
|
39 |
+
note: use with a base_lr of 1.0.
|
40 |
+
"""
|
41 |
+
def __init__(self, warm_up_steps, f_min, f_max, f_start, cycle_lengths, verbosity_interval=0):
|
42 |
+
assert len(warm_up_steps) == len(f_min) == len(f_max) == len(f_start) == len(cycle_lengths)
|
43 |
+
self.lr_warm_up_steps = warm_up_steps
|
44 |
+
self.f_start = f_start
|
45 |
+
self.f_min = f_min
|
46 |
+
self.f_max = f_max
|
47 |
+
self.cycle_lengths = cycle_lengths
|
48 |
+
self.cum_cycles = np.cumsum([0] + list(self.cycle_lengths))
|
49 |
+
self.last_f = 0.
|
50 |
+
self.verbosity_interval = verbosity_interval
|
51 |
+
|
52 |
+
def find_in_interval(self, n):
|
53 |
+
interval = 0
|
54 |
+
for cl in self.cum_cycles[1:]:
|
55 |
+
if n <= cl:
|
56 |
+
return interval
|
57 |
+
interval += 1
|
58 |
+
|
59 |
+
def schedule(self, n, **kwargs):
|
60 |
+
cycle = self.find_in_interval(n)
|
61 |
+
n = n - self.cum_cycles[cycle]
|
62 |
+
if self.verbosity_interval > 0:
|
63 |
+
if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_f}, "
|
64 |
+
f"current cycle {cycle}")
|
65 |
+
if n < self.lr_warm_up_steps[cycle]:
|
66 |
+
f = (self.f_max[cycle] - self.f_start[cycle]) / self.lr_warm_up_steps[cycle] * n + self.f_start[cycle]
|
67 |
+
self.last_f = f
|
68 |
+
return f
|
69 |
+
else:
|
70 |
+
t = (n - self.lr_warm_up_steps[cycle]) / (self.cycle_lengths[cycle] - self.lr_warm_up_steps[cycle])
|
71 |
+
t = min(t, 1.0)
|
72 |
+
f = self.f_min[cycle] + 0.5 * (self.f_max[cycle] - self.f_min[cycle]) * (
|
73 |
+
1 + np.cos(t * np.pi))
|
74 |
+
self.last_f = f
|
75 |
+
return f
|
76 |
+
|
77 |
+
def __call__(self, n, **kwargs):
|
78 |
+
return self.schedule(n, **kwargs)
|
79 |
+
|
80 |
+
|
81 |
+
class LambdaLinearScheduler(LambdaWarmUpCosineScheduler2):
|
82 |
+
|
83 |
+
def schedule(self, n, **kwargs):
|
84 |
+
cycle = self.find_in_interval(n)
|
85 |
+
n = n - self.cum_cycles[cycle]
|
86 |
+
if self.verbosity_interval > 0:
|
87 |
+
if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_f}, "
|
88 |
+
f"current cycle {cycle}")
|
89 |
+
|
90 |
+
if n < self.lr_warm_up_steps[cycle]:
|
91 |
+
f = (self.f_max[cycle] - self.f_start[cycle]) / self.lr_warm_up_steps[cycle] * n + self.f_start[cycle]
|
92 |
+
self.last_f = f
|
93 |
+
return f
|
94 |
+
else:
|
95 |
+
f = self.f_min[cycle] + (self.f_max[cycle] - self.f_min[cycle]) * (self.cycle_lengths[cycle] - n) / (self.cycle_lengths[cycle])
|
96 |
+
self.last_f = f
|
97 |
+
return f
|
98 |
+
|
stable-diffusion/ldm/models/__pycache__/autoencoder.cpython-38.pyc
ADDED
Binary file (13.5 kB). View file
|
|
stable-diffusion/ldm/models/autoencoder.py
ADDED
@@ -0,0 +1,443 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import pytorch_lightning as pl
|
3 |
+
import torch.nn.functional as F
|
4 |
+
from contextlib import contextmanager
|
5 |
+
|
6 |
+
from taming.modules.vqvae.quantize import VectorQuantizer2 as VectorQuantizer
|
7 |
+
|
8 |
+
from ldm.modules.diffusionmodules.model import Encoder, Decoder
|
9 |
+
from ldm.modules.distributions.distributions import DiagonalGaussianDistribution
|
10 |
+
|
11 |
+
from ldm.util import instantiate_from_config
|
12 |
+
|
13 |
+
|
14 |
+
class VQModel(pl.LightningModule):
|
15 |
+
def __init__(self,
|
16 |
+
ddconfig,
|
17 |
+
lossconfig,
|
18 |
+
n_embed,
|
19 |
+
embed_dim,
|
20 |
+
ckpt_path=None,
|
21 |
+
ignore_keys=[],
|
22 |
+
image_key="image",
|
23 |
+
colorize_nlabels=None,
|
24 |
+
monitor=None,
|
25 |
+
batch_resize_range=None,
|
26 |
+
scheduler_config=None,
|
27 |
+
lr_g_factor=1.0,
|
28 |
+
remap=None,
|
29 |
+
sane_index_shape=False, # tell vector quantizer to return indices as bhw
|
30 |
+
use_ema=False
|
31 |
+
):
|
32 |
+
super().__init__()
|
33 |
+
self.embed_dim = embed_dim
|
34 |
+
self.n_embed = n_embed
|
35 |
+
self.image_key = image_key
|
36 |
+
self.encoder = Encoder(**ddconfig)
|
37 |
+
self.decoder = Decoder(**ddconfig)
|
38 |
+
self.loss = instantiate_from_config(lossconfig)
|
39 |
+
self.quantize = VectorQuantizer(n_embed, embed_dim, beta=0.25,
|
40 |
+
remap=remap,
|
41 |
+
sane_index_shape=sane_index_shape)
|
42 |
+
self.quant_conv = torch.nn.Conv2d(ddconfig["z_channels"], embed_dim, 1)
|
43 |
+
self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
|
44 |
+
if colorize_nlabels is not None:
|
45 |
+
assert type(colorize_nlabels)==int
|
46 |
+
self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1))
|
47 |
+
if monitor is not None:
|
48 |
+
self.monitor = monitor
|
49 |
+
self.batch_resize_range = batch_resize_range
|
50 |
+
if self.batch_resize_range is not None:
|
51 |
+
print(f"{self.__class__.__name__}: Using per-batch resizing in range {batch_resize_range}.")
|
52 |
+
|
53 |
+
self.use_ema = use_ema
|
54 |
+
if self.use_ema:
|
55 |
+
self.model_ema = LitEma(self)
|
56 |
+
print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
|
57 |
+
|
58 |
+
if ckpt_path is not None:
|
59 |
+
self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
|
60 |
+
self.scheduler_config = scheduler_config
|
61 |
+
self.lr_g_factor = lr_g_factor
|
62 |
+
|
63 |
+
@contextmanager
|
64 |
+
def ema_scope(self, context=None):
|
65 |
+
if self.use_ema:
|
66 |
+
self.model_ema.store(self.parameters())
|
67 |
+
self.model_ema.copy_to(self)
|
68 |
+
if context is not None:
|
69 |
+
print(f"{context}: Switched to EMA weights")
|
70 |
+
try:
|
71 |
+
yield None
|
72 |
+
finally:
|
73 |
+
if self.use_ema:
|
74 |
+
self.model_ema.restore(self.parameters())
|
75 |
+
if context is not None:
|
76 |
+
print(f"{context}: Restored training weights")
|
77 |
+
|
78 |
+
def init_from_ckpt(self, path, ignore_keys=list()):
|
79 |
+
sd = torch.load(path, map_location="cpu")["state_dict"]
|
80 |
+
keys = list(sd.keys())
|
81 |
+
for k in keys:
|
82 |
+
for ik in ignore_keys:
|
83 |
+
if k.startswith(ik):
|
84 |
+
print("Deleting key {} from state_dict.".format(k))
|
85 |
+
del sd[k]
|
86 |
+
missing, unexpected = self.load_state_dict(sd, strict=False)
|
87 |
+
print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
|
88 |
+
if len(missing) > 0:
|
89 |
+
print(f"Missing Keys: {missing}")
|
90 |
+
print(f"Unexpected Keys: {unexpected}")
|
91 |
+
|
92 |
+
def on_train_batch_end(self, *args, **kwargs):
|
93 |
+
if self.use_ema:
|
94 |
+
self.model_ema(self)
|
95 |
+
|
96 |
+
def encode(self, x):
|
97 |
+
h = self.encoder(x)
|
98 |
+
h = self.quant_conv(h)
|
99 |
+
quant, emb_loss, info = self.quantize(h)
|
100 |
+
return quant, emb_loss, info
|
101 |
+
|
102 |
+
def encode_to_prequant(self, x):
|
103 |
+
h = self.encoder(x)
|
104 |
+
h = self.quant_conv(h)
|
105 |
+
return h
|
106 |
+
|
107 |
+
def decode(self, quant):
|
108 |
+
quant = self.post_quant_conv(quant)
|
109 |
+
dec = self.decoder(quant)
|
110 |
+
return dec
|
111 |
+
|
112 |
+
def decode_code(self, code_b):
|
113 |
+
quant_b = self.quantize.embed_code(code_b)
|
114 |
+
dec = self.decode(quant_b)
|
115 |
+
return dec
|
116 |
+
|
117 |
+
def forward(self, input, return_pred_indices=False):
|
118 |
+
quant, diff, (_,_,ind) = self.encode(input)
|
119 |
+
dec = self.decode(quant)
|
120 |
+
if return_pred_indices:
|
121 |
+
return dec, diff, ind
|
122 |
+
return dec, diff
|
123 |
+
|
124 |
+
def get_input(self, batch, k):
|
125 |
+
x = batch[k]
|
126 |
+
if len(x.shape) == 3:
|
127 |
+
x = x[..., None]
|
128 |
+
x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float()
|
129 |
+
if self.batch_resize_range is not None:
|
130 |
+
lower_size = self.batch_resize_range[0]
|
131 |
+
upper_size = self.batch_resize_range[1]
|
132 |
+
if self.global_step <= 4:
|
133 |
+
# do the first few batches with max size to avoid later oom
|
134 |
+
new_resize = upper_size
|
135 |
+
else:
|
136 |
+
new_resize = np.random.choice(np.arange(lower_size, upper_size+16, 16))
|
137 |
+
if new_resize != x.shape[2]:
|
138 |
+
x = F.interpolate(x, size=new_resize, mode="bicubic")
|
139 |
+
x = x.detach()
|
140 |
+
return x
|
141 |
+
|
142 |
+
def training_step(self, batch, batch_idx, optimizer_idx):
|
143 |
+
# https://github.com/pytorch/pytorch/issues/37142
|
144 |
+
# try not to fool the heuristics
|
145 |
+
x = self.get_input(batch, self.image_key)
|
146 |
+
xrec, qloss, ind = self(x, return_pred_indices=True)
|
147 |
+
|
148 |
+
if optimizer_idx == 0:
|
149 |
+
# autoencode
|
150 |
+
aeloss, log_dict_ae = self.loss(qloss, x, xrec, optimizer_idx, self.global_step,
|
151 |
+
last_layer=self.get_last_layer(), split="train",
|
152 |
+
predicted_indices=ind)
|
153 |
+
|
154 |
+
self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=True)
|
155 |
+
return aeloss
|
156 |
+
|
157 |
+
if optimizer_idx == 1:
|
158 |
+
# discriminator
|
159 |
+
discloss, log_dict_disc = self.loss(qloss, x, xrec, optimizer_idx, self.global_step,
|
160 |
+
last_layer=self.get_last_layer(), split="train")
|
161 |
+
self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=True)
|
162 |
+
return discloss
|
163 |
+
|
164 |
+
def validation_step(self, batch, batch_idx):
|
165 |
+
log_dict = self._validation_step(batch, batch_idx)
|
166 |
+
with self.ema_scope():
|
167 |
+
log_dict_ema = self._validation_step(batch, batch_idx, suffix="_ema")
|
168 |
+
return log_dict
|
169 |
+
|
170 |
+
def _validation_step(self, batch, batch_idx, suffix=""):
|
171 |
+
x = self.get_input(batch, self.image_key)
|
172 |
+
xrec, qloss, ind = self(x, return_pred_indices=True)
|
173 |
+
aeloss, log_dict_ae = self.loss(qloss, x, xrec, 0,
|
174 |
+
self.global_step,
|
175 |
+
last_layer=self.get_last_layer(),
|
176 |
+
split="val"+suffix,
|
177 |
+
predicted_indices=ind
|
178 |
+
)
|
179 |
+
|
180 |
+
discloss, log_dict_disc = self.loss(qloss, x, xrec, 1,
|
181 |
+
self.global_step,
|
182 |
+
last_layer=self.get_last_layer(),
|
183 |
+
split="val"+suffix,
|
184 |
+
predicted_indices=ind
|
185 |
+
)
|
186 |
+
rec_loss = log_dict_ae[f"val{suffix}/rec_loss"]
|
187 |
+
self.log(f"val{suffix}/rec_loss", rec_loss,
|
188 |
+
prog_bar=True, logger=True, on_step=False, on_epoch=True, sync_dist=True)
|
189 |
+
self.log(f"val{suffix}/aeloss", aeloss,
|
190 |
+
prog_bar=True, logger=True, on_step=False, on_epoch=True, sync_dist=True)
|
191 |
+
if version.parse(pl.__version__) >= version.parse('1.4.0'):
|
192 |
+
del log_dict_ae[f"val{suffix}/rec_loss"]
|
193 |
+
self.log_dict(log_dict_ae)
|
194 |
+
self.log_dict(log_dict_disc)
|
195 |
+
return self.log_dict
|
196 |
+
|
197 |
+
def configure_optimizers(self):
|
198 |
+
lr_d = self.learning_rate
|
199 |
+
lr_g = self.lr_g_factor*self.learning_rate
|
200 |
+
print("lr_d", lr_d)
|
201 |
+
print("lr_g", lr_g)
|
202 |
+
opt_ae = torch.optim.Adam(list(self.encoder.parameters())+
|
203 |
+
list(self.decoder.parameters())+
|
204 |
+
list(self.quantize.parameters())+
|
205 |
+
list(self.quant_conv.parameters())+
|
206 |
+
list(self.post_quant_conv.parameters()),
|
207 |
+
lr=lr_g, betas=(0.5, 0.9))
|
208 |
+
opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(),
|
209 |
+
lr=lr_d, betas=(0.5, 0.9))
|
210 |
+
|
211 |
+
if self.scheduler_config is not None:
|
212 |
+
scheduler = instantiate_from_config(self.scheduler_config)
|
213 |
+
|
214 |
+
print("Setting up LambdaLR scheduler...")
|
215 |
+
scheduler = [
|
216 |
+
{
|
217 |
+
'scheduler': LambdaLR(opt_ae, lr_lambda=scheduler.schedule),
|
218 |
+
'interval': 'step',
|
219 |
+
'frequency': 1
|
220 |
+
},
|
221 |
+
{
|
222 |
+
'scheduler': LambdaLR(opt_disc, lr_lambda=scheduler.schedule),
|
223 |
+
'interval': 'step',
|
224 |
+
'frequency': 1
|
225 |
+
},
|
226 |
+
]
|
227 |
+
return [opt_ae, opt_disc], scheduler
|
228 |
+
return [opt_ae, opt_disc], []
|
229 |
+
|
230 |
+
def get_last_layer(self):
|
231 |
+
return self.decoder.conv_out.weight
|
232 |
+
|
233 |
+
def log_images(self, batch, only_inputs=False, plot_ema=False, **kwargs):
|
234 |
+
log = dict()
|
235 |
+
x = self.get_input(batch, self.image_key)
|
236 |
+
x = x.to(self.device)
|
237 |
+
if only_inputs:
|
238 |
+
log["inputs"] = x
|
239 |
+
return log
|
240 |
+
xrec, _ = self(x)
|
241 |
+
if x.shape[1] > 3:
|
242 |
+
# colorize with random projection
|
243 |
+
assert xrec.shape[1] > 3
|
244 |
+
x = self.to_rgb(x)
|
245 |
+
xrec = self.to_rgb(xrec)
|
246 |
+
log["inputs"] = x
|
247 |
+
log["reconstructions"] = xrec
|
248 |
+
if plot_ema:
|
249 |
+
with self.ema_scope():
|
250 |
+
xrec_ema, _ = self(x)
|
251 |
+
if x.shape[1] > 3: xrec_ema = self.to_rgb(xrec_ema)
|
252 |
+
log["reconstructions_ema"] = xrec_ema
|
253 |
+
return log
|
254 |
+
|
255 |
+
def to_rgb(self, x):
|
256 |
+
assert self.image_key == "segmentation"
|
257 |
+
if not hasattr(self, "colorize"):
|
258 |
+
self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x))
|
259 |
+
x = F.conv2d(x, weight=self.colorize)
|
260 |
+
x = 2.*(x-x.min())/(x.max()-x.min()) - 1.
|
261 |
+
return x
|
262 |
+
|
263 |
+
|
264 |
+
class VQModelInterface(VQModel):
|
265 |
+
def __init__(self, embed_dim, *args, **kwargs):
|
266 |
+
super().__init__(embed_dim=embed_dim, *args, **kwargs)
|
267 |
+
self.embed_dim = embed_dim
|
268 |
+
|
269 |
+
def encode(self, x):
|
270 |
+
h = self.encoder(x)
|
271 |
+
h = self.quant_conv(h)
|
272 |
+
return h
|
273 |
+
|
274 |
+
def decode(self, h, force_not_quantize=False):
|
275 |
+
# also go through quantization layer
|
276 |
+
if not force_not_quantize:
|
277 |
+
quant, emb_loss, info = self.quantize(h)
|
278 |
+
else:
|
279 |
+
quant = h
|
280 |
+
quant = self.post_quant_conv(quant)
|
281 |
+
dec = self.decoder(quant)
|
282 |
+
return dec
|
283 |
+
|
284 |
+
|
285 |
+
class AutoencoderKL(pl.LightningModule):
|
286 |
+
def __init__(self,
|
287 |
+
ddconfig,
|
288 |
+
lossconfig,
|
289 |
+
embed_dim,
|
290 |
+
ckpt_path=None,
|
291 |
+
ignore_keys=[],
|
292 |
+
image_key="image",
|
293 |
+
colorize_nlabels=None,
|
294 |
+
monitor=None,
|
295 |
+
):
|
296 |
+
super().__init__()
|
297 |
+
self.image_key = image_key
|
298 |
+
self.encoder = Encoder(**ddconfig)
|
299 |
+
self.decoder = Decoder(**ddconfig)
|
300 |
+
self.loss = instantiate_from_config(lossconfig)
|
301 |
+
assert ddconfig["double_z"]
|
302 |
+
self.quant_conv = torch.nn.Conv2d(2*ddconfig["z_channels"], 2*embed_dim, 1)
|
303 |
+
self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
|
304 |
+
self.embed_dim = embed_dim
|
305 |
+
if colorize_nlabels is not None:
|
306 |
+
assert type(colorize_nlabels)==int
|
307 |
+
self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1))
|
308 |
+
if monitor is not None:
|
309 |
+
self.monitor = monitor
|
310 |
+
if ckpt_path is not None:
|
311 |
+
self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
|
312 |
+
|
313 |
+
def init_from_ckpt(self, path, ignore_keys=list()):
|
314 |
+
sd = torch.load(path, map_location="cpu")["state_dict"]
|
315 |
+
keys = list(sd.keys())
|
316 |
+
for k in keys:
|
317 |
+
for ik in ignore_keys:
|
318 |
+
if k.startswith(ik):
|
319 |
+
print("Deleting key {} from state_dict.".format(k))
|
320 |
+
del sd[k]
|
321 |
+
self.load_state_dict(sd, strict=False)
|
322 |
+
print(f"Restored from {path}")
|
323 |
+
|
324 |
+
def encode(self, x):
|
325 |
+
h = self.encoder(x)
|
326 |
+
moments = self.quant_conv(h)
|
327 |
+
posterior = DiagonalGaussianDistribution(moments)
|
328 |
+
return posterior
|
329 |
+
|
330 |
+
def decode(self, z):
|
331 |
+
z = self.post_quant_conv(z)
|
332 |
+
dec = self.decoder(z)
|
333 |
+
return dec
|
334 |
+
|
335 |
+
def forward(self, input, sample_posterior=True):
|
336 |
+
posterior = self.encode(input)
|
337 |
+
if sample_posterior:
|
338 |
+
z = posterior.sample()
|
339 |
+
else:
|
340 |
+
z = posterior.mode()
|
341 |
+
dec = self.decode(z)
|
342 |
+
return dec, posterior
|
343 |
+
|
344 |
+
def get_input(self, batch, k):
|
345 |
+
x = batch[k]
|
346 |
+
if len(x.shape) == 3:
|
347 |
+
x = x[..., None]
|
348 |
+
x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float()
|
349 |
+
return x
|
350 |
+
|
351 |
+
def training_step(self, batch, batch_idx, optimizer_idx):
|
352 |
+
inputs = self.get_input(batch, self.image_key)
|
353 |
+
reconstructions, posterior = self(inputs)
|
354 |
+
|
355 |
+
if optimizer_idx == 0:
|
356 |
+
# train encoder+decoder+logvar
|
357 |
+
aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step,
|
358 |
+
last_layer=self.get_last_layer(), split="train")
|
359 |
+
self.log("aeloss", aeloss, prog_bar=True, logger=True, on_step=True, on_epoch=True)
|
360 |
+
self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=False)
|
361 |
+
return aeloss
|
362 |
+
|
363 |
+
if optimizer_idx == 1:
|
364 |
+
# train the discriminator
|
365 |
+
discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step,
|
366 |
+
last_layer=self.get_last_layer(), split="train")
|
367 |
+
|
368 |
+
self.log("discloss", discloss, prog_bar=True, logger=True, on_step=True, on_epoch=True)
|
369 |
+
self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=False)
|
370 |
+
return discloss
|
371 |
+
|
372 |
+
def validation_step(self, batch, batch_idx):
|
373 |
+
inputs = self.get_input(batch, self.image_key)
|
374 |
+
reconstructions, posterior = self(inputs)
|
375 |
+
aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, 0, self.global_step,
|
376 |
+
last_layer=self.get_last_layer(), split="val")
|
377 |
+
|
378 |
+
discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, 1, self.global_step,
|
379 |
+
last_layer=self.get_last_layer(), split="val")
|
380 |
+
|
381 |
+
self.log("val/rec_loss", log_dict_ae["val/rec_loss"])
|
382 |
+
self.log_dict(log_dict_ae)
|
383 |
+
self.log_dict(log_dict_disc)
|
384 |
+
return self.log_dict
|
385 |
+
|
386 |
+
def configure_optimizers(self):
|
387 |
+
lr = self.learning_rate
|
388 |
+
opt_ae = torch.optim.Adam(list(self.encoder.parameters())+
|
389 |
+
list(self.decoder.parameters())+
|
390 |
+
list(self.quant_conv.parameters())+
|
391 |
+
list(self.post_quant_conv.parameters()),
|
392 |
+
lr=lr, betas=(0.5, 0.9))
|
393 |
+
opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(),
|
394 |
+
lr=lr, betas=(0.5, 0.9))
|
395 |
+
return [opt_ae, opt_disc], []
|
396 |
+
|
397 |
+
def get_last_layer(self):
|
398 |
+
return self.decoder.conv_out.weight
|
399 |
+
|
400 |
+
@torch.no_grad()
|
401 |
+
def log_images(self, batch, only_inputs=False, **kwargs):
|
402 |
+
log = dict()
|
403 |
+
x = self.get_input(batch, self.image_key)
|
404 |
+
x = x.to(self.device)
|
405 |
+
if not only_inputs:
|
406 |
+
xrec, posterior = self(x)
|
407 |
+
if x.shape[1] > 3:
|
408 |
+
# colorize with random projection
|
409 |
+
assert xrec.shape[1] > 3
|
410 |
+
x = self.to_rgb(x)
|
411 |
+
xrec = self.to_rgb(xrec)
|
412 |
+
log["samples"] = self.decode(torch.randn_like(posterior.sample()))
|
413 |
+
log["reconstructions"] = xrec
|
414 |
+
log["inputs"] = x
|
415 |
+
return log
|
416 |
+
|
417 |
+
def to_rgb(self, x):
|
418 |
+
assert self.image_key == "segmentation"
|
419 |
+
if not hasattr(self, "colorize"):
|
420 |
+
self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x))
|
421 |
+
x = F.conv2d(x, weight=self.colorize)
|
422 |
+
x = 2.*(x-x.min())/(x.max()-x.min()) - 1.
|
423 |
+
return x
|
424 |
+
|
425 |
+
|
426 |
+
class IdentityFirstStage(torch.nn.Module):
|
427 |
+
def __init__(self, *args, vq_interface=False, **kwargs):
|
428 |
+
self.vq_interface = vq_interface # TODO: Should be true by default but check to not break older stuff
|
429 |
+
super().__init__()
|
430 |
+
|
431 |
+
def encode(self, x, *args, **kwargs):
|
432 |
+
return x
|
433 |
+
|
434 |
+
def decode(self, x, *args, **kwargs):
|
435 |
+
return x
|
436 |
+
|
437 |
+
def quantize(self, x, *args, **kwargs):
|
438 |
+
if self.vq_interface:
|
439 |
+
return x, None, [None, None, None]
|
440 |
+
return x
|
441 |
+
|
442 |
+
def forward(self, x, *args, **kwargs):
|
443 |
+
return x
|
stable-diffusion/ldm/models/diffusion/__init__.py
ADDED
File without changes
|
stable-diffusion/ldm/models/diffusion/__pycache__/__init__.cpython-38.pyc
ADDED
Binary file (169 Bytes). View file
|
|
stable-diffusion/ldm/models/diffusion/__pycache__/__init__.cpython-39.pyc
ADDED
Binary file (169 Bytes). View file
|
|
stable-diffusion/ldm/models/diffusion/__pycache__/ddim.cpython-38.pyc
ADDED
Binary file (7.18 kB). View file
|
|
stable-diffusion/ldm/models/diffusion/__pycache__/ddim.cpython-39.pyc
ADDED
Binary file (7.12 kB). View file
|
|
stable-diffusion/ldm/models/diffusion/__pycache__/ddpm.cpython-38.pyc
ADDED
Binary file (43.8 kB). View file
|
|
stable-diffusion/ldm/models/diffusion/__pycache__/ddpm.cpython-39.pyc
ADDED
Binary file (43.8 kB). View file
|
|
stable-diffusion/ldm/models/diffusion/__pycache__/plms.cpython-38.pyc
ADDED
Binary file (7.37 kB). View file
|
|
stable-diffusion/ldm/models/diffusion/__pycache__/plms.cpython-39.pyc
ADDED
Binary file (7.31 kB). View file
|
|
stable-diffusion/ldm/models/diffusion/__pycache__/psld.cpython-38.pyc
ADDED
Binary file (9.8 kB). View file
|
|
stable-diffusion/ldm/models/diffusion/__pycache__/psld.cpython-39.pyc
ADDED
Binary file (9.79 kB). View file
|
|
stable-diffusion/ldm/models/diffusion/classifier.py
ADDED
@@ -0,0 +1,267 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import torch
|
3 |
+
import pytorch_lightning as pl
|
4 |
+
from omegaconf import OmegaConf
|
5 |
+
from torch.nn import functional as F
|
6 |
+
from torch.optim import AdamW
|
7 |
+
from torch.optim.lr_scheduler import LambdaLR
|
8 |
+
from copy import deepcopy
|
9 |
+
from einops import rearrange
|
10 |
+
from glob import glob
|
11 |
+
from natsort import natsorted
|
12 |
+
|
13 |
+
from ldm.modules.diffusionmodules.openaimodel import EncoderUNetModel, UNetModel
|
14 |
+
from ldm.util import log_txt_as_img, default, ismap, instantiate_from_config
|
15 |
+
|
16 |
+
__models__ = {
|
17 |
+
'class_label': EncoderUNetModel,
|
18 |
+
'segmentation': UNetModel
|
19 |
+
}
|
20 |
+
|
21 |
+
|
22 |
+
def disabled_train(self, mode=True):
|
23 |
+
"""Overwrite model.train with this function to make sure train/eval mode
|
24 |
+
does not change anymore."""
|
25 |
+
return self
|
26 |
+
|
27 |
+
|
28 |
+
class NoisyLatentImageClassifier(pl.LightningModule):
|
29 |
+
|
30 |
+
def __init__(self,
|
31 |
+
diffusion_path,
|
32 |
+
num_classes,
|
33 |
+
ckpt_path=None,
|
34 |
+
pool='attention',
|
35 |
+
label_key=None,
|
36 |
+
diffusion_ckpt_path=None,
|
37 |
+
scheduler_config=None,
|
38 |
+
weight_decay=1.e-2,
|
39 |
+
log_steps=10,
|
40 |
+
monitor='val/loss',
|
41 |
+
*args,
|
42 |
+
**kwargs):
|
43 |
+
super().__init__(*args, **kwargs)
|
44 |
+
self.num_classes = num_classes
|
45 |
+
# get latest config of diffusion model
|
46 |
+
diffusion_config = natsorted(glob(os.path.join(diffusion_path, 'configs', '*-project.yaml')))[-1]
|
47 |
+
self.diffusion_config = OmegaConf.load(diffusion_config).model
|
48 |
+
self.diffusion_config.params.ckpt_path = diffusion_ckpt_path
|
49 |
+
self.load_diffusion()
|
50 |
+
|
51 |
+
self.monitor = monitor
|
52 |
+
self.numd = self.diffusion_model.first_stage_model.encoder.num_resolutions - 1
|
53 |
+
self.log_time_interval = self.diffusion_model.num_timesteps // log_steps
|
54 |
+
self.log_steps = log_steps
|
55 |
+
|
56 |
+
self.label_key = label_key if not hasattr(self.diffusion_model, 'cond_stage_key') \
|
57 |
+
else self.diffusion_model.cond_stage_key
|
58 |
+
|
59 |
+
assert self.label_key is not None, 'label_key neither in diffusion model nor in model.params'
|
60 |
+
|
61 |
+
if self.label_key not in __models__:
|
62 |
+
raise NotImplementedError()
|
63 |
+
|
64 |
+
self.load_classifier(ckpt_path, pool)
|
65 |
+
|
66 |
+
self.scheduler_config = scheduler_config
|
67 |
+
self.use_scheduler = self.scheduler_config is not None
|
68 |
+
self.weight_decay = weight_decay
|
69 |
+
|
70 |
+
def init_from_ckpt(self, path, ignore_keys=list(), only_model=False):
|
71 |
+
sd = torch.load(path, map_location="cpu")
|
72 |
+
if "state_dict" in list(sd.keys()):
|
73 |
+
sd = sd["state_dict"]
|
74 |
+
keys = list(sd.keys())
|
75 |
+
for k in keys:
|
76 |
+
for ik in ignore_keys:
|
77 |
+
if k.startswith(ik):
|
78 |
+
print("Deleting key {} from state_dict.".format(k))
|
79 |
+
del sd[k]
|
80 |
+
missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(
|
81 |
+
sd, strict=False)
|
82 |
+
print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
|
83 |
+
if len(missing) > 0:
|
84 |
+
print(f"Missing Keys: {missing}")
|
85 |
+
if len(unexpected) > 0:
|
86 |
+
print(f"Unexpected Keys: {unexpected}")
|
87 |
+
|
88 |
+
def load_diffusion(self):
|
89 |
+
model = instantiate_from_config(self.diffusion_config)
|
90 |
+
self.diffusion_model = model.eval()
|
91 |
+
self.diffusion_model.train = disabled_train
|
92 |
+
for param in self.diffusion_model.parameters():
|
93 |
+
param.requires_grad = False
|
94 |
+
|
95 |
+
def load_classifier(self, ckpt_path, pool):
|
96 |
+
model_config = deepcopy(self.diffusion_config.params.unet_config.params)
|
97 |
+
model_config.in_channels = self.diffusion_config.params.unet_config.params.out_channels
|
98 |
+
model_config.out_channels = self.num_classes
|
99 |
+
if self.label_key == 'class_label':
|
100 |
+
model_config.pool = pool
|
101 |
+
|
102 |
+
self.model = __models__[self.label_key](**model_config)
|
103 |
+
if ckpt_path is not None:
|
104 |
+
print('#####################################################################')
|
105 |
+
print(f'load from ckpt "{ckpt_path}"')
|
106 |
+
print('#####################################################################')
|
107 |
+
self.init_from_ckpt(ckpt_path)
|
108 |
+
|
109 |
+
@torch.no_grad()
|
110 |
+
def get_x_noisy(self, x, t, noise=None):
|
111 |
+
noise = default(noise, lambda: torch.randn_like(x))
|
112 |
+
continuous_sqrt_alpha_cumprod = None
|
113 |
+
if self.diffusion_model.use_continuous_noise:
|
114 |
+
continuous_sqrt_alpha_cumprod = self.diffusion_model.sample_continuous_noise_level(x.shape[0], t + 1)
|
115 |
+
# todo: make sure t+1 is correct here
|
116 |
+
|
117 |
+
return self.diffusion_model.q_sample(x_start=x, t=t, noise=noise,
|
118 |
+
continuous_sqrt_alpha_cumprod=continuous_sqrt_alpha_cumprod)
|
119 |
+
|
120 |
+
def forward(self, x_noisy, t, *args, **kwargs):
|
121 |
+
return self.model(x_noisy, t)
|
122 |
+
|
123 |
+
@torch.no_grad()
|
124 |
+
def get_input(self, batch, k):
|
125 |
+
x = batch[k]
|
126 |
+
if len(x.shape) == 3:
|
127 |
+
x = x[..., None]
|
128 |
+
x = rearrange(x, 'b h w c -> b c h w')
|
129 |
+
x = x.to(memory_format=torch.contiguous_format).float()
|
130 |
+
return x
|
131 |
+
|
132 |
+
@torch.no_grad()
|
133 |
+
def get_conditioning(self, batch, k=None):
|
134 |
+
if k is None:
|
135 |
+
k = self.label_key
|
136 |
+
assert k is not None, 'Needs to provide label key'
|
137 |
+
|
138 |
+
targets = batch[k].to(self.device)
|
139 |
+
|
140 |
+
if self.label_key == 'segmentation':
|
141 |
+
targets = rearrange(targets, 'b h w c -> b c h w')
|
142 |
+
for down in range(self.numd):
|
143 |
+
h, w = targets.shape[-2:]
|
144 |
+
targets = F.interpolate(targets, size=(h // 2, w // 2), mode='nearest')
|
145 |
+
|
146 |
+
# targets = rearrange(targets,'b c h w -> b h w c')
|
147 |
+
|
148 |
+
return targets
|
149 |
+
|
150 |
+
def compute_top_k(self, logits, labels, k, reduction="mean"):
|
151 |
+
_, top_ks = torch.topk(logits, k, dim=1)
|
152 |
+
if reduction == "mean":
|
153 |
+
return (top_ks == labels[:, None]).float().sum(dim=-1).mean().item()
|
154 |
+
elif reduction == "none":
|
155 |
+
return (top_ks == labels[:, None]).float().sum(dim=-1)
|
156 |
+
|
157 |
+
def on_train_epoch_start(self):
|
158 |
+
# save some memory
|
159 |
+
self.diffusion_model.model.to('cpu')
|
160 |
+
|
161 |
+
@torch.no_grad()
|
162 |
+
def write_logs(self, loss, logits, targets):
|
163 |
+
log_prefix = 'train' if self.training else 'val'
|
164 |
+
log = {}
|
165 |
+
log[f"{log_prefix}/loss"] = loss.mean()
|
166 |
+
log[f"{log_prefix}/acc@1"] = self.compute_top_k(
|
167 |
+
logits, targets, k=1, reduction="mean"
|
168 |
+
)
|
169 |
+
log[f"{log_prefix}/acc@5"] = self.compute_top_k(
|
170 |
+
logits, targets, k=5, reduction="mean"
|
171 |
+
)
|
172 |
+
|
173 |
+
self.log_dict(log, prog_bar=False, logger=True, on_step=self.training, on_epoch=True)
|
174 |
+
self.log('loss', log[f"{log_prefix}/loss"], prog_bar=True, logger=False)
|
175 |
+
self.log('global_step', self.global_step, logger=False, on_epoch=False, prog_bar=True)
|
176 |
+
lr = self.optimizers().param_groups[0]['lr']
|
177 |
+
self.log('lr_abs', lr, on_step=True, logger=True, on_epoch=False, prog_bar=True)
|
178 |
+
|
179 |
+
def shared_step(self, batch, t=None):
|
180 |
+
x, *_ = self.diffusion_model.get_input(batch, k=self.diffusion_model.first_stage_key)
|
181 |
+
targets = self.get_conditioning(batch)
|
182 |
+
if targets.dim() == 4:
|
183 |
+
targets = targets.argmax(dim=1)
|
184 |
+
if t is None:
|
185 |
+
t = torch.randint(0, self.diffusion_model.num_timesteps, (x.shape[0],), device=self.device).long()
|
186 |
+
else:
|
187 |
+
t = torch.full(size=(x.shape[0],), fill_value=t, device=self.device).long()
|
188 |
+
x_noisy = self.get_x_noisy(x, t)
|
189 |
+
logits = self(x_noisy, t)
|
190 |
+
|
191 |
+
loss = F.cross_entropy(logits, targets, reduction='none')
|
192 |
+
|
193 |
+
self.write_logs(loss.detach(), logits.detach(), targets.detach())
|
194 |
+
|
195 |
+
loss = loss.mean()
|
196 |
+
return loss, logits, x_noisy, targets
|
197 |
+
|
198 |
+
def training_step(self, batch, batch_idx):
|
199 |
+
loss, *_ = self.shared_step(batch)
|
200 |
+
return loss
|
201 |
+
|
202 |
+
def reset_noise_accs(self):
|
203 |
+
self.noisy_acc = {t: {'acc@1': [], 'acc@5': []} for t in
|
204 |
+
range(0, self.diffusion_model.num_timesteps, self.diffusion_model.log_every_t)}
|
205 |
+
|
206 |
+
def on_validation_start(self):
|
207 |
+
self.reset_noise_accs()
|
208 |
+
|
209 |
+
@torch.no_grad()
|
210 |
+
def validation_step(self, batch, batch_idx):
|
211 |
+
loss, *_ = self.shared_step(batch)
|
212 |
+
|
213 |
+
for t in self.noisy_acc:
|
214 |
+
_, logits, _, targets = self.shared_step(batch, t)
|
215 |
+
self.noisy_acc[t]['acc@1'].append(self.compute_top_k(logits, targets, k=1, reduction='mean'))
|
216 |
+
self.noisy_acc[t]['acc@5'].append(self.compute_top_k(logits, targets, k=5, reduction='mean'))
|
217 |
+
|
218 |
+
return loss
|
219 |
+
|
220 |
+
def configure_optimizers(self):
|
221 |
+
optimizer = AdamW(self.model.parameters(), lr=self.learning_rate, weight_decay=self.weight_decay)
|
222 |
+
|
223 |
+
if self.use_scheduler:
|
224 |
+
scheduler = instantiate_from_config(self.scheduler_config)
|
225 |
+
|
226 |
+
print("Setting up LambdaLR scheduler...")
|
227 |
+
scheduler = [
|
228 |
+
{
|
229 |
+
'scheduler': LambdaLR(optimizer, lr_lambda=scheduler.schedule),
|
230 |
+
'interval': 'step',
|
231 |
+
'frequency': 1
|
232 |
+
}]
|
233 |
+
return [optimizer], scheduler
|
234 |
+
|
235 |
+
return optimizer
|
236 |
+
|
237 |
+
@torch.no_grad()
|
238 |
+
def log_images(self, batch, N=8, *args, **kwargs):
|
239 |
+
log = dict()
|
240 |
+
x = self.get_input(batch, self.diffusion_model.first_stage_key)
|
241 |
+
log['inputs'] = x
|
242 |
+
|
243 |
+
y = self.get_conditioning(batch)
|
244 |
+
|
245 |
+
if self.label_key == 'class_label':
|
246 |
+
y = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"])
|
247 |
+
log['labels'] = y
|
248 |
+
|
249 |
+
if ismap(y):
|
250 |
+
log['labels'] = self.diffusion_model.to_rgb(y)
|
251 |
+
|
252 |
+
for step in range(self.log_steps):
|
253 |
+
current_time = step * self.log_time_interval
|
254 |
+
|
255 |
+
_, logits, x_noisy, _ = self.shared_step(batch, t=current_time)
|
256 |
+
|
257 |
+
log[f'inputs@t{current_time}'] = x_noisy
|
258 |
+
|
259 |
+
pred = F.one_hot(logits.argmax(dim=1), num_classes=self.num_classes)
|
260 |
+
pred = rearrange(pred, 'b h w c -> b c h w')
|
261 |
+
|
262 |
+
log[f'pred@t{current_time}'] = self.diffusion_model.to_rgb(pred)
|
263 |
+
|
264 |
+
for key in log:
|
265 |
+
log[key] = log[key][:N]
|
266 |
+
|
267 |
+
return log
|
stable-diffusion/ldm/models/diffusion/ddim.py
ADDED
@@ -0,0 +1,241 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""SAMPLING ONLY."""
|
2 |
+
|
3 |
+
import torch
|
4 |
+
import numpy as np
|
5 |
+
from tqdm import tqdm
|
6 |
+
from functools import partial
|
7 |
+
|
8 |
+
from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like, \
|
9 |
+
extract_into_tensor
|
10 |
+
|
11 |
+
|
12 |
+
class DDIMSampler(object):
|
13 |
+
def __init__(self, model, schedule="linear", **kwargs):
|
14 |
+
super().__init__()
|
15 |
+
self.model = model
|
16 |
+
self.ddpm_num_timesteps = model.num_timesteps
|
17 |
+
self.schedule = schedule
|
18 |
+
|
19 |
+
def register_buffer(self, name, attr):
|
20 |
+
if type(attr) == torch.Tensor:
|
21 |
+
if attr.device != torch.device("cuda"):
|
22 |
+
attr = attr.to(torch.device("cuda"))
|
23 |
+
setattr(self, name, attr)
|
24 |
+
|
25 |
+
def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
|
26 |
+
self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
|
27 |
+
num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)
|
28 |
+
alphas_cumprod = self.model.alphas_cumprod
|
29 |
+
assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
|
30 |
+
to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
|
31 |
+
|
32 |
+
self.register_buffer('betas', to_torch(self.model.betas))
|
33 |
+
self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
|
34 |
+
self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev))
|
35 |
+
|
36 |
+
# calculations for diffusion q(x_t | x_{t-1}) and others
|
37 |
+
self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))
|
38 |
+
self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
|
39 |
+
self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))
|
40 |
+
self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))
|
41 |
+
self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))
|
42 |
+
|
43 |
+
# ddim sampling parameters
|
44 |
+
ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),
|
45 |
+
ddim_timesteps=self.ddim_timesteps,
|
46 |
+
eta=ddim_eta,verbose=verbose)
|
47 |
+
self.register_buffer('ddim_sigmas', ddim_sigmas)
|
48 |
+
self.register_buffer('ddim_alphas', ddim_alphas)
|
49 |
+
self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
|
50 |
+
self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
|
51 |
+
sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
|
52 |
+
(1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
|
53 |
+
1 - self.alphas_cumprod / self.alphas_cumprod_prev))
|
54 |
+
self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
|
55 |
+
|
56 |
+
@torch.no_grad()
|
57 |
+
def sample(self,
|
58 |
+
S,
|
59 |
+
batch_size,
|
60 |
+
shape,
|
61 |
+
conditioning=None,
|
62 |
+
callback=None,
|
63 |
+
normals_sequence=None,
|
64 |
+
img_callback=None,
|
65 |
+
quantize_x0=False,
|
66 |
+
eta=0.,
|
67 |
+
mask=None,
|
68 |
+
x0=None,
|
69 |
+
temperature=1.,
|
70 |
+
noise_dropout=0.,
|
71 |
+
score_corrector=None,
|
72 |
+
corrector_kwargs=None,
|
73 |
+
verbose=True,
|
74 |
+
x_T=None,
|
75 |
+
log_every_t=100,
|
76 |
+
unconditional_guidance_scale=1.,
|
77 |
+
unconditional_conditioning=None,
|
78 |
+
# this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
|
79 |
+
**kwargs
|
80 |
+
):
|
81 |
+
if conditioning is not None:
|
82 |
+
if isinstance(conditioning, dict):
|
83 |
+
cbs = conditioning[list(conditioning.keys())[0]].shape[0]
|
84 |
+
if cbs != batch_size:
|
85 |
+
print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
|
86 |
+
else:
|
87 |
+
if conditioning.shape[0] != batch_size:
|
88 |
+
print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
|
89 |
+
|
90 |
+
self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose)
|
91 |
+
# sampling
|
92 |
+
C, H, W = shape
|
93 |
+
size = (batch_size, C, H, W)
|
94 |
+
print(f'Data shape for DDIM sampling is {size}, eta {eta}')
|
95 |
+
|
96 |
+
samples, intermediates = self.ddim_sampling(conditioning, size,
|
97 |
+
callback=callback,
|
98 |
+
img_callback=img_callback,
|
99 |
+
quantize_denoised=quantize_x0,
|
100 |
+
mask=mask, x0=x0,
|
101 |
+
ddim_use_original_steps=False,
|
102 |
+
noise_dropout=noise_dropout,
|
103 |
+
temperature=temperature,
|
104 |
+
score_corrector=score_corrector,
|
105 |
+
corrector_kwargs=corrector_kwargs,
|
106 |
+
x_T=x_T,
|
107 |
+
log_every_t=log_every_t,
|
108 |
+
unconditional_guidance_scale=unconditional_guidance_scale,
|
109 |
+
unconditional_conditioning=unconditional_conditioning,
|
110 |
+
)
|
111 |
+
return samples, intermediates
|
112 |
+
|
113 |
+
@torch.no_grad()
|
114 |
+
def ddim_sampling(self, cond, shape,
|
115 |
+
x_T=None, ddim_use_original_steps=False,
|
116 |
+
callback=None, timesteps=None, quantize_denoised=False,
|
117 |
+
mask=None, x0=None, img_callback=None, log_every_t=100,
|
118 |
+
temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
|
119 |
+
unconditional_guidance_scale=1., unconditional_conditioning=None,):
|
120 |
+
device = self.model.betas.device
|
121 |
+
b = shape[0]
|
122 |
+
if x_T is None:
|
123 |
+
img = torch.randn(shape, device=device)
|
124 |
+
else:
|
125 |
+
img = x_T
|
126 |
+
|
127 |
+
if timesteps is None:
|
128 |
+
timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
|
129 |
+
elif timesteps is not None and not ddim_use_original_steps:
|
130 |
+
subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
|
131 |
+
timesteps = self.ddim_timesteps[:subset_end]
|
132 |
+
|
133 |
+
intermediates = {'x_inter': [img], 'pred_x0': [img]}
|
134 |
+
time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps)
|
135 |
+
total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
|
136 |
+
print(f"Running DDIM Sampling with {total_steps} timesteps")
|
137 |
+
|
138 |
+
iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps)
|
139 |
+
|
140 |
+
for i, step in enumerate(iterator):
|
141 |
+
index = total_steps - i - 1
|
142 |
+
ts = torch.full((b,), step, device=device, dtype=torch.long)
|
143 |
+
|
144 |
+
if mask is not None:
|
145 |
+
assert x0 is not None
|
146 |
+
img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass?
|
147 |
+
img = img_orig * mask + (1. - mask) * img
|
148 |
+
|
149 |
+
outs = self.p_sample_ddim(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,
|
150 |
+
quantize_denoised=quantize_denoised, temperature=temperature,
|
151 |
+
noise_dropout=noise_dropout, score_corrector=score_corrector,
|
152 |
+
corrector_kwargs=corrector_kwargs,
|
153 |
+
unconditional_guidance_scale=unconditional_guidance_scale,
|
154 |
+
unconditional_conditioning=unconditional_conditioning)
|
155 |
+
img, pred_x0 = outs
|
156 |
+
if callback: callback(i)
|
157 |
+
if img_callback: img_callback(pred_x0, i)
|
158 |
+
|
159 |
+
if index % log_every_t == 0 or index == total_steps - 1:
|
160 |
+
intermediates['x_inter'].append(img)
|
161 |
+
intermediates['pred_x0'].append(pred_x0)
|
162 |
+
|
163 |
+
return img, intermediates
|
164 |
+
|
165 |
+
@torch.no_grad()
|
166 |
+
def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
|
167 |
+
temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
|
168 |
+
unconditional_guidance_scale=1., unconditional_conditioning=None):
|
169 |
+
b, *_, device = *x.shape, x.device
|
170 |
+
|
171 |
+
if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
|
172 |
+
e_t = self.model.apply_model(x, t, c)
|
173 |
+
else:
|
174 |
+
x_in = torch.cat([x] * 2)
|
175 |
+
t_in = torch.cat([t] * 2)
|
176 |
+
c_in = torch.cat([unconditional_conditioning, c])
|
177 |
+
e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2)
|
178 |
+
e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)
|
179 |
+
|
180 |
+
if score_corrector is not None:
|
181 |
+
assert self.model.parameterization == "eps"
|
182 |
+
e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
|
183 |
+
|
184 |
+
alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
|
185 |
+
alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
|
186 |
+
sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
|
187 |
+
sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
|
188 |
+
# select parameters corresponding to the currently considered timestep
|
189 |
+
a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
|
190 |
+
a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
|
191 |
+
sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
|
192 |
+
sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device)
|
193 |
+
|
194 |
+
# current prediction for x_0
|
195 |
+
pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
|
196 |
+
if quantize_denoised:
|
197 |
+
pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
|
198 |
+
# direction pointing to x_t
|
199 |
+
dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
|
200 |
+
noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
|
201 |
+
if noise_dropout > 0.:
|
202 |
+
noise = torch.nn.functional.dropout(noise, p=noise_dropout)
|
203 |
+
x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
|
204 |
+
return x_prev, pred_x0
|
205 |
+
|
206 |
+
@torch.no_grad()
|
207 |
+
def stochastic_encode(self, x0, t, use_original_steps=False, noise=None):
|
208 |
+
# fast, but does not allow for exact reconstruction
|
209 |
+
# t serves as an index to gather the correct alphas
|
210 |
+
if use_original_steps:
|
211 |
+
sqrt_alphas_cumprod = self.sqrt_alphas_cumprod
|
212 |
+
sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod
|
213 |
+
else:
|
214 |
+
sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas)
|
215 |
+
sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas
|
216 |
+
|
217 |
+
if noise is None:
|
218 |
+
noise = torch.randn_like(x0)
|
219 |
+
return (extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0 +
|
220 |
+
extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise)
|
221 |
+
|
222 |
+
@torch.no_grad()
|
223 |
+
def decode(self, x_latent, cond, t_start, unconditional_guidance_scale=1.0, unconditional_conditioning=None,
|
224 |
+
use_original_steps=False):
|
225 |
+
|
226 |
+
timesteps = np.arange(self.ddpm_num_timesteps) if use_original_steps else self.ddim_timesteps
|
227 |
+
timesteps = timesteps[:t_start]
|
228 |
+
|
229 |
+
time_range = np.flip(timesteps)
|
230 |
+
total_steps = timesteps.shape[0]
|
231 |
+
print(f"Running DDIM Sampling with {total_steps} timesteps")
|
232 |
+
|
233 |
+
iterator = tqdm(time_range, desc='Decoding image', total=total_steps)
|
234 |
+
x_dec = x_latent
|
235 |
+
for i, step in enumerate(iterator):
|
236 |
+
index = total_steps - i - 1
|
237 |
+
ts = torch.full((x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long)
|
238 |
+
x_dec, _ = self.p_sample_ddim(x_dec, cond, ts, index=index, use_original_steps=use_original_steps,
|
239 |
+
unconditional_guidance_scale=unconditional_guidance_scale,
|
240 |
+
unconditional_conditioning=unconditional_conditioning)
|
241 |
+
return x_dec
|
stable-diffusion/ldm/models/diffusion/ddpm.py
ADDED
@@ -0,0 +1,1445 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
wild mixture of
|
3 |
+
https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
|
4 |
+
https://github.com/openai/improved-diffusion/blob/e94489283bb876ac1477d5dd7709bbbd2d9902ce/improved_diffusion/gaussian_diffusion.py
|
5 |
+
https://github.com/CompVis/taming-transformers
|
6 |
+
-- merci
|
7 |
+
"""
|
8 |
+
|
9 |
+
import torch
|
10 |
+
import torch.nn as nn
|
11 |
+
import numpy as np
|
12 |
+
import pytorch_lightning as pl
|
13 |
+
from torch.optim.lr_scheduler import LambdaLR
|
14 |
+
from einops import rearrange, repeat
|
15 |
+
from contextlib import contextmanager
|
16 |
+
from functools import partial
|
17 |
+
from tqdm import tqdm
|
18 |
+
from torchvision.utils import make_grid
|
19 |
+
from pytorch_lightning.utilities.distributed import rank_zero_only
|
20 |
+
|
21 |
+
from ldm.util import log_txt_as_img, exists, default, ismap, isimage, mean_flat, count_params, instantiate_from_config
|
22 |
+
from ldm.modules.ema import LitEma
|
23 |
+
from ldm.modules.distributions.distributions import normal_kl, DiagonalGaussianDistribution
|
24 |
+
from ldm.models.autoencoder import VQModelInterface, IdentityFirstStage, AutoencoderKL
|
25 |
+
from ldm.modules.diffusionmodules.util import make_beta_schedule, extract_into_tensor, noise_like
|
26 |
+
from ldm.models.diffusion.ddim import DDIMSampler
|
27 |
+
|
28 |
+
|
29 |
+
__conditioning_keys__ = {'concat': 'c_concat',
|
30 |
+
'crossattn': 'c_crossattn',
|
31 |
+
'adm': 'y'}
|
32 |
+
|
33 |
+
|
34 |
+
def disabled_train(self, mode=True):
|
35 |
+
"""Overwrite model.train with this function to make sure train/eval mode
|
36 |
+
does not change anymore."""
|
37 |
+
return self
|
38 |
+
|
39 |
+
|
40 |
+
def uniform_on_device(r1, r2, shape, device):
|
41 |
+
return (r1 - r2) * torch.rand(*shape, device=device) + r2
|
42 |
+
|
43 |
+
|
44 |
+
class DDPM(pl.LightningModule):
|
45 |
+
# classic DDPM with Gaussian diffusion, in image space
|
46 |
+
def __init__(self,
|
47 |
+
unet_config,
|
48 |
+
timesteps=1000,
|
49 |
+
beta_schedule="linear",
|
50 |
+
loss_type="l2",
|
51 |
+
ckpt_path=None,
|
52 |
+
ignore_keys=[],
|
53 |
+
load_only_unet=False,
|
54 |
+
monitor="val/loss",
|
55 |
+
use_ema=True,
|
56 |
+
first_stage_key="image",
|
57 |
+
image_size=256,
|
58 |
+
channels=3,
|
59 |
+
log_every_t=100,
|
60 |
+
clip_denoised=True,
|
61 |
+
linear_start=1e-4,
|
62 |
+
linear_end=2e-2,
|
63 |
+
cosine_s=8e-3,
|
64 |
+
given_betas=None,
|
65 |
+
original_elbo_weight=0.,
|
66 |
+
v_posterior=0., # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta
|
67 |
+
l_simple_weight=1.,
|
68 |
+
conditioning_key=None,
|
69 |
+
parameterization="eps", # all assuming fixed variance schedules
|
70 |
+
scheduler_config=None,
|
71 |
+
use_positional_encodings=False,
|
72 |
+
learn_logvar=False,
|
73 |
+
logvar_init=0.,
|
74 |
+
):
|
75 |
+
super().__init__()
|
76 |
+
assert parameterization in ["eps", "x0"], 'currently only supporting "eps" and "x0"'
|
77 |
+
self.parameterization = parameterization
|
78 |
+
print(f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode")
|
79 |
+
self.cond_stage_model = None
|
80 |
+
self.clip_denoised = clip_denoised
|
81 |
+
self.log_every_t = log_every_t
|
82 |
+
self.first_stage_key = first_stage_key
|
83 |
+
self.image_size = image_size # try conv?
|
84 |
+
self.channels = channels
|
85 |
+
self.use_positional_encodings = use_positional_encodings
|
86 |
+
self.model = DiffusionWrapper(unet_config, conditioning_key)
|
87 |
+
count_params(self.model, verbose=True)
|
88 |
+
self.use_ema = use_ema
|
89 |
+
if self.use_ema:
|
90 |
+
self.model_ema = LitEma(self.model)
|
91 |
+
print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
|
92 |
+
|
93 |
+
self.use_scheduler = scheduler_config is not None
|
94 |
+
if self.use_scheduler:
|
95 |
+
self.scheduler_config = scheduler_config
|
96 |
+
|
97 |
+
self.v_posterior = v_posterior
|
98 |
+
self.original_elbo_weight = original_elbo_weight
|
99 |
+
self.l_simple_weight = l_simple_weight
|
100 |
+
|
101 |
+
if monitor is not None:
|
102 |
+
self.monitor = monitor
|
103 |
+
if ckpt_path is not None:
|
104 |
+
self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys, only_model=load_only_unet)
|
105 |
+
|
106 |
+
self.register_schedule(given_betas=given_betas, beta_schedule=beta_schedule, timesteps=timesteps,
|
107 |
+
linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s)
|
108 |
+
|
109 |
+
self.loss_type = loss_type
|
110 |
+
|
111 |
+
self.learn_logvar = learn_logvar
|
112 |
+
self.logvar = torch.full(fill_value=logvar_init, size=(self.num_timesteps,))
|
113 |
+
if self.learn_logvar:
|
114 |
+
self.logvar = nn.Parameter(self.logvar, requires_grad=True)
|
115 |
+
|
116 |
+
|
117 |
+
def register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000,
|
118 |
+
linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
|
119 |
+
if exists(given_betas):
|
120 |
+
betas = given_betas
|
121 |
+
else:
|
122 |
+
betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end,
|
123 |
+
cosine_s=cosine_s)
|
124 |
+
alphas = 1. - betas
|
125 |
+
alphas_cumprod = np.cumprod(alphas, axis=0)
|
126 |
+
alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
|
127 |
+
|
128 |
+
timesteps, = betas.shape
|
129 |
+
self.num_timesteps = int(timesteps)
|
130 |
+
self.linear_start = linear_start
|
131 |
+
self.linear_end = linear_end
|
132 |
+
assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep'
|
133 |
+
|
134 |
+
to_torch = partial(torch.tensor, dtype=torch.float32)
|
135 |
+
|
136 |
+
self.register_buffer('betas', to_torch(betas))
|
137 |
+
self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
|
138 |
+
self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev))
|
139 |
+
|
140 |
+
# calculations for diffusion q(x_t | x_{t-1}) and others
|
141 |
+
self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
|
142 |
+
self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
|
143 |
+
self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod)))
|
144 |
+
self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod)))
|
145 |
+
self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1)))
|
146 |
+
|
147 |
+
# calculations for posterior q(x_{t-1} | x_t, x_0)
|
148 |
+
posterior_variance = (1 - self.v_posterior) * betas * (1. - alphas_cumprod_prev) / (
|
149 |
+
1. - alphas_cumprod) + self.v_posterior * betas
|
150 |
+
# above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t)
|
151 |
+
self.register_buffer('posterior_variance', to_torch(posterior_variance))
|
152 |
+
# below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain
|
153 |
+
self.register_buffer('posterior_log_variance_clipped', to_torch(np.log(np.maximum(posterior_variance, 1e-20))))
|
154 |
+
self.register_buffer('posterior_mean_coef1', to_torch(
|
155 |
+
betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod)))
|
156 |
+
self.register_buffer('posterior_mean_coef2', to_torch(
|
157 |
+
(1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod)))
|
158 |
+
|
159 |
+
if self.parameterization == "eps":
|
160 |
+
lvlb_weights = self.betas ** 2 / (
|
161 |
+
2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod))
|
162 |
+
elif self.parameterization == "x0":
|
163 |
+
lvlb_weights = 0.5 * np.sqrt(torch.Tensor(alphas_cumprod)) / (2. * 1 - torch.Tensor(alphas_cumprod))
|
164 |
+
else:
|
165 |
+
raise NotImplementedError("mu not supported")
|
166 |
+
# TODO how to choose this term
|
167 |
+
lvlb_weights[0] = lvlb_weights[1]
|
168 |
+
self.register_buffer('lvlb_weights', lvlb_weights, persistent=False)
|
169 |
+
assert not torch.isnan(self.lvlb_weights).all()
|
170 |
+
|
171 |
+
@contextmanager
|
172 |
+
def ema_scope(self, context=None):
|
173 |
+
if self.use_ema:
|
174 |
+
self.model_ema.store(self.model.parameters())
|
175 |
+
self.model_ema.copy_to(self.model)
|
176 |
+
if context is not None:
|
177 |
+
print(f"{context}: Switched to EMA weights")
|
178 |
+
try:
|
179 |
+
yield None
|
180 |
+
finally:
|
181 |
+
if self.use_ema:
|
182 |
+
self.model_ema.restore(self.model.parameters())
|
183 |
+
if context is not None:
|
184 |
+
print(f"{context}: Restored training weights")
|
185 |
+
|
186 |
+
def init_from_ckpt(self, path, ignore_keys=list(), only_model=False):
|
187 |
+
sd = torch.load(path, map_location="cpu")
|
188 |
+
if "state_dict" in list(sd.keys()):
|
189 |
+
sd = sd["state_dict"]
|
190 |
+
keys = list(sd.keys())
|
191 |
+
for k in keys:
|
192 |
+
for ik in ignore_keys:
|
193 |
+
if k.startswith(ik):
|
194 |
+
print("Deleting key {} from state_dict.".format(k))
|
195 |
+
del sd[k]
|
196 |
+
missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(
|
197 |
+
sd, strict=False)
|
198 |
+
print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
|
199 |
+
if len(missing) > 0:
|
200 |
+
print(f"Missing Keys: {missing}")
|
201 |
+
if len(unexpected) > 0:
|
202 |
+
print(f"Unexpected Keys: {unexpected}")
|
203 |
+
|
204 |
+
def q_mean_variance(self, x_start, t):
|
205 |
+
"""
|
206 |
+
Get the distribution q(x_t | x_0).
|
207 |
+
:param x_start: the [N x C x ...] tensor of noiseless inputs.
|
208 |
+
:param t: the number of diffusion steps (minus 1). Here, 0 means one step.
|
209 |
+
:return: A tuple (mean, variance, log_variance), all of x_start's shape.
|
210 |
+
"""
|
211 |
+
mean = (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start)
|
212 |
+
variance = extract_into_tensor(1.0 - self.alphas_cumprod, t, x_start.shape)
|
213 |
+
log_variance = extract_into_tensor(self.log_one_minus_alphas_cumprod, t, x_start.shape)
|
214 |
+
return mean, variance, log_variance
|
215 |
+
|
216 |
+
def predict_start_from_noise(self, x_t, t, noise):
|
217 |
+
return (
|
218 |
+
extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t -
|
219 |
+
extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise
|
220 |
+
)
|
221 |
+
|
222 |
+
def q_posterior(self, x_start, x_t, t):
|
223 |
+
posterior_mean = (
|
224 |
+
extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * x_start +
|
225 |
+
extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t
|
226 |
+
)
|
227 |
+
posterior_variance = extract_into_tensor(self.posterior_variance, t, x_t.shape)
|
228 |
+
posterior_log_variance_clipped = extract_into_tensor(self.posterior_log_variance_clipped, t, x_t.shape)
|
229 |
+
return posterior_mean, posterior_variance, posterior_log_variance_clipped
|
230 |
+
|
231 |
+
def p_mean_variance(self, x, t, clip_denoised: bool):
|
232 |
+
model_out = self.model(x, t)
|
233 |
+
if self.parameterization == "eps":
|
234 |
+
x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
|
235 |
+
elif self.parameterization == "x0":
|
236 |
+
x_recon = model_out
|
237 |
+
if clip_denoised:
|
238 |
+
x_recon.clamp_(-1., 1.)
|
239 |
+
|
240 |
+
model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
|
241 |
+
return model_mean, posterior_variance, posterior_log_variance
|
242 |
+
|
243 |
+
# @torch.no_grad()
|
244 |
+
def p_sample(self, x, t, clip_denoised=True, repeat_noise=False):
|
245 |
+
b, *_, device = *x.shape, x.device
|
246 |
+
model_mean, _, model_log_variance = self.p_mean_variance(x=x, t=t, clip_denoised=clip_denoised)
|
247 |
+
noise = noise_like(x.shape, device, repeat_noise)
|
248 |
+
# no noise when t == 0
|
249 |
+
nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
|
250 |
+
return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
|
251 |
+
|
252 |
+
# @torch.no_grad()
|
253 |
+
def p_sample_loop(self, shape, return_intermediates=False):
|
254 |
+
device = self.betas.device
|
255 |
+
b = shape[0]
|
256 |
+
img = torch.randn(shape, device=device)
|
257 |
+
intermediates = [img]
|
258 |
+
for i in tqdm(reversed(range(0, self.num_timesteps)), desc='Sampling t', total=self.num_timesteps):
|
259 |
+
img = self.p_sample(img, torch.full((b,), i, device=device, dtype=torch.long),
|
260 |
+
clip_denoised=self.clip_denoised)
|
261 |
+
if i % self.log_every_t == 0 or i == self.num_timesteps - 1:
|
262 |
+
intermediates.append(img)
|
263 |
+
if return_intermediates:
|
264 |
+
return img, intermediates
|
265 |
+
return img
|
266 |
+
|
267 |
+
# @torch.no_grad()
|
268 |
+
def sample(self, batch_size=16, return_intermediates=False):
|
269 |
+
image_size = self.image_size
|
270 |
+
channels = self.channels
|
271 |
+
return self.p_sample_loop((batch_size, channels, image_size, image_size),
|
272 |
+
return_intermediates=return_intermediates)
|
273 |
+
|
274 |
+
def q_sample(self, x_start, t, noise=None):
|
275 |
+
noise = default(noise, lambda: torch.randn_like(x_start))
|
276 |
+
return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
|
277 |
+
extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise)
|
278 |
+
|
279 |
+
def get_loss(self, pred, target, mean=True):
|
280 |
+
if self.loss_type == 'l1':
|
281 |
+
loss = (target - pred).abs()
|
282 |
+
if mean:
|
283 |
+
loss = loss.mean()
|
284 |
+
elif self.loss_type == 'l2':
|
285 |
+
if mean:
|
286 |
+
loss = torch.nn.functional.mse_loss(target, pred)
|
287 |
+
else:
|
288 |
+
loss = torch.nn.functional.mse_loss(target, pred, reduction='none')
|
289 |
+
else:
|
290 |
+
raise NotImplementedError("unknown loss type '{loss_type}'")
|
291 |
+
|
292 |
+
return loss
|
293 |
+
|
294 |
+
def p_losses(self, x_start, t, noise=None):
|
295 |
+
noise = default(noise, lambda: torch.randn_like(x_start))
|
296 |
+
x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
|
297 |
+
model_out = self.model(x_noisy, t)
|
298 |
+
|
299 |
+
loss_dict = {}
|
300 |
+
if self.parameterization == "eps":
|
301 |
+
target = noise
|
302 |
+
elif self.parameterization == "x0":
|
303 |
+
target = x_start
|
304 |
+
else:
|
305 |
+
raise NotImplementedError(f"Paramterization {self.parameterization} not yet supported")
|
306 |
+
|
307 |
+
loss = self.get_loss(model_out, target, mean=False).mean(dim=[1, 2, 3])
|
308 |
+
|
309 |
+
log_prefix = 'train' if self.training else 'val'
|
310 |
+
|
311 |
+
loss_dict.update({f'{log_prefix}/loss_simple': loss.mean()})
|
312 |
+
loss_simple = loss.mean() * self.l_simple_weight
|
313 |
+
|
314 |
+
loss_vlb = (self.lvlb_weights[t] * loss).mean()
|
315 |
+
loss_dict.update({f'{log_prefix}/loss_vlb': loss_vlb})
|
316 |
+
|
317 |
+
loss = loss_simple + self.original_elbo_weight * loss_vlb
|
318 |
+
|
319 |
+
loss_dict.update({f'{log_prefix}/loss': loss})
|
320 |
+
|
321 |
+
return loss, loss_dict
|
322 |
+
|
323 |
+
def forward(self, x, *args, **kwargs):
|
324 |
+
# b, c, h, w, device, img_size, = *x.shape, x.device, self.image_size
|
325 |
+
# assert h == img_size and w == img_size, f'height and width of image must be {img_size}'
|
326 |
+
t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
|
327 |
+
return self.p_losses(x, t, *args, **kwargs)
|
328 |
+
|
329 |
+
def get_input(self, batch, k):
|
330 |
+
x = batch[k]
|
331 |
+
if len(x.shape) == 3:
|
332 |
+
x = x[..., None]
|
333 |
+
x = rearrange(x, 'b h w c -> b c h w')
|
334 |
+
x = x.to(memory_format=torch.contiguous_format).float()
|
335 |
+
return x
|
336 |
+
|
337 |
+
def shared_step(self, batch):
|
338 |
+
x = self.get_input(batch, self.first_stage_key)
|
339 |
+
loss, loss_dict = self(x)
|
340 |
+
return loss, loss_dict
|
341 |
+
|
342 |
+
def training_step(self, batch, batch_idx):
|
343 |
+
loss, loss_dict = self.shared_step(batch)
|
344 |
+
|
345 |
+
self.log_dict(loss_dict, prog_bar=True,
|
346 |
+
logger=True, on_step=True, on_epoch=True)
|
347 |
+
|
348 |
+
self.log("global_step", self.global_step,
|
349 |
+
prog_bar=True, logger=True, on_step=True, on_epoch=False)
|
350 |
+
|
351 |
+
if self.use_scheduler:
|
352 |
+
lr = self.optimizers().param_groups[0]['lr']
|
353 |
+
self.log('lr_abs', lr, prog_bar=True, logger=True, on_step=True, on_epoch=False)
|
354 |
+
|
355 |
+
return loss
|
356 |
+
|
357 |
+
# @torch.no_grad()
|
358 |
+
def validation_step(self, batch, batch_idx):
|
359 |
+
_, loss_dict_no_ema = self.shared_step(batch)
|
360 |
+
with self.ema_scope():
|
361 |
+
_, loss_dict_ema = self.shared_step(batch)
|
362 |
+
loss_dict_ema = {key + '_ema': loss_dict_ema[key] for key in loss_dict_ema}
|
363 |
+
self.log_dict(loss_dict_no_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True)
|
364 |
+
self.log_dict(loss_dict_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True)
|
365 |
+
|
366 |
+
def on_train_batch_end(self, *args, **kwargs):
|
367 |
+
if self.use_ema:
|
368 |
+
self.model_ema(self.model)
|
369 |
+
|
370 |
+
def _get_rows_from_list(self, samples):
|
371 |
+
n_imgs_per_row = len(samples)
|
372 |
+
denoise_grid = rearrange(samples, 'n b c h w -> b n c h w')
|
373 |
+
denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
|
374 |
+
denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
|
375 |
+
return denoise_grid
|
376 |
+
|
377 |
+
# @torch.no_grad()
|
378 |
+
def log_images(self, batch, N=8, n_row=2, sample=True, return_keys=None, **kwargs):
|
379 |
+
log = dict()
|
380 |
+
x = self.get_input(batch, self.first_stage_key)
|
381 |
+
N = min(x.shape[0], N)
|
382 |
+
n_row = min(x.shape[0], n_row)
|
383 |
+
x = x.to(self.device)[:N]
|
384 |
+
log["inputs"] = x
|
385 |
+
|
386 |
+
# get diffusion row
|
387 |
+
diffusion_row = list()
|
388 |
+
x_start = x[:n_row]
|
389 |
+
|
390 |
+
for t in range(self.num_timesteps):
|
391 |
+
if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
|
392 |
+
t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
|
393 |
+
t = t.to(self.device).long()
|
394 |
+
noise = torch.randn_like(x_start)
|
395 |
+
x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
|
396 |
+
diffusion_row.append(x_noisy)
|
397 |
+
|
398 |
+
log["diffusion_row"] = self._get_rows_from_list(diffusion_row)
|
399 |
+
|
400 |
+
if sample:
|
401 |
+
# get denoise row
|
402 |
+
with self.ema_scope("Plotting"):
|
403 |
+
samples, denoise_row = self.sample(batch_size=N, return_intermediates=True)
|
404 |
+
|
405 |
+
log["samples"] = samples
|
406 |
+
log["denoise_row"] = self._get_rows_from_list(denoise_row)
|
407 |
+
|
408 |
+
if return_keys:
|
409 |
+
if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
|
410 |
+
return log
|
411 |
+
else:
|
412 |
+
return {key: log[key] for key in return_keys}
|
413 |
+
return log
|
414 |
+
|
415 |
+
def configure_optimizers(self):
|
416 |
+
lr = self.learning_rate
|
417 |
+
params = list(self.model.parameters())
|
418 |
+
if self.learn_logvar:
|
419 |
+
params = params + [self.logvar]
|
420 |
+
opt = torch.optim.AdamW(params, lr=lr)
|
421 |
+
return opt
|
422 |
+
|
423 |
+
|
424 |
+
class LatentDiffusion(DDPM):
|
425 |
+
"""main class"""
|
426 |
+
def __init__(self,
|
427 |
+
first_stage_config,
|
428 |
+
cond_stage_config,
|
429 |
+
num_timesteps_cond=None,
|
430 |
+
cond_stage_key="image",
|
431 |
+
cond_stage_trainable=False,
|
432 |
+
concat_mode=True,
|
433 |
+
cond_stage_forward=None,
|
434 |
+
conditioning_key=None,
|
435 |
+
scale_factor=1.0,
|
436 |
+
scale_by_std=False,
|
437 |
+
*args, **kwargs):
|
438 |
+
self.num_timesteps_cond = default(num_timesteps_cond, 1)
|
439 |
+
self.scale_by_std = scale_by_std
|
440 |
+
assert self.num_timesteps_cond <= kwargs['timesteps']
|
441 |
+
# for backwards compatibility after implementation of DiffusionWrapper
|
442 |
+
if conditioning_key is None:
|
443 |
+
conditioning_key = 'concat' if concat_mode else 'crossattn'
|
444 |
+
if cond_stage_config == '__is_unconditional__':
|
445 |
+
conditioning_key = None
|
446 |
+
ckpt_path = kwargs.pop("ckpt_path", None)
|
447 |
+
ignore_keys = kwargs.pop("ignore_keys", [])
|
448 |
+
super().__init__(conditioning_key=conditioning_key, *args, **kwargs)
|
449 |
+
self.concat_mode = concat_mode
|
450 |
+
self.cond_stage_trainable = cond_stage_trainable
|
451 |
+
self.cond_stage_key = cond_stage_key
|
452 |
+
try:
|
453 |
+
self.num_downs = len(first_stage_config.params.ddconfig.ch_mult) - 1
|
454 |
+
except:
|
455 |
+
self.num_downs = 0
|
456 |
+
if not scale_by_std:
|
457 |
+
self.scale_factor = scale_factor
|
458 |
+
else:
|
459 |
+
self.register_buffer('scale_factor', torch.tensor(scale_factor))
|
460 |
+
self.instantiate_first_stage(first_stage_config)
|
461 |
+
self.instantiate_cond_stage(cond_stage_config)
|
462 |
+
self.cond_stage_forward = cond_stage_forward
|
463 |
+
self.clip_denoised = False
|
464 |
+
self.bbox_tokenizer = None
|
465 |
+
|
466 |
+
self.restarted_from_ckpt = False
|
467 |
+
if ckpt_path is not None:
|
468 |
+
self.init_from_ckpt(ckpt_path, ignore_keys)
|
469 |
+
self.restarted_from_ckpt = True
|
470 |
+
|
471 |
+
def make_cond_schedule(self, ):
|
472 |
+
self.cond_ids = torch.full(size=(self.num_timesteps,), fill_value=self.num_timesteps - 1, dtype=torch.long)
|
473 |
+
ids = torch.round(torch.linspace(0, self.num_timesteps - 1, self.num_timesteps_cond)).long()
|
474 |
+
self.cond_ids[:self.num_timesteps_cond] = ids
|
475 |
+
|
476 |
+
@rank_zero_only
|
477 |
+
# @torch.no_grad()
|
478 |
+
def on_train_batch_start(self, batch, batch_idx, dataloader_idx):
|
479 |
+
# only for very first batch
|
480 |
+
if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 and batch_idx == 0 and not self.restarted_from_ckpt:
|
481 |
+
assert self.scale_factor == 1., 'rather not use custom rescaling and std-rescaling simultaneously'
|
482 |
+
# set rescale weight to 1./std of encodings
|
483 |
+
print("### USING STD-RESCALING ###")
|
484 |
+
x = super().get_input(batch, self.first_stage_key)
|
485 |
+
x = x.to(self.device)
|
486 |
+
encoder_posterior = self.encode_first_stage(x)
|
487 |
+
z = self.get_first_stage_encoding(encoder_posterior).detach()
|
488 |
+
del self.scale_factor
|
489 |
+
self.register_buffer('scale_factor', 1. / z.flatten().std())
|
490 |
+
print(f"setting self.scale_factor to {self.scale_factor}")
|
491 |
+
print("### USING STD-RESCALING ###")
|
492 |
+
|
493 |
+
def register_schedule(self,
|
494 |
+
given_betas=None, beta_schedule="linear", timesteps=1000,
|
495 |
+
linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
|
496 |
+
super().register_schedule(given_betas, beta_schedule, timesteps, linear_start, linear_end, cosine_s)
|
497 |
+
|
498 |
+
self.shorten_cond_schedule = self.num_timesteps_cond > 1
|
499 |
+
if self.shorten_cond_schedule:
|
500 |
+
self.make_cond_schedule()
|
501 |
+
|
502 |
+
def instantiate_first_stage(self, config):
|
503 |
+
model = instantiate_from_config(config)
|
504 |
+
self.first_stage_model = model.eval()
|
505 |
+
self.first_stage_model.train = disabled_train
|
506 |
+
for param in self.first_stage_model.parameters():
|
507 |
+
param.requires_grad = False
|
508 |
+
|
509 |
+
def instantiate_cond_stage(self, config):
|
510 |
+
if not self.cond_stage_trainable:
|
511 |
+
if config == "__is_first_stage__":
|
512 |
+
print("Using first stage also as cond stage.")
|
513 |
+
self.cond_stage_model = self.first_stage_model
|
514 |
+
elif config == "__is_unconditional__":
|
515 |
+
print(f"Training {self.__class__.__name__} as an unconditional model.")
|
516 |
+
self.cond_stage_model = None
|
517 |
+
# self.be_unconditional = True
|
518 |
+
else:
|
519 |
+
model = instantiate_from_config(config)
|
520 |
+
self.cond_stage_model = model.eval()
|
521 |
+
self.cond_stage_model.train = disabled_train
|
522 |
+
for param in self.cond_stage_model.parameters():
|
523 |
+
param.requires_grad = False
|
524 |
+
else:
|
525 |
+
assert config != '__is_first_stage__'
|
526 |
+
assert config != '__is_unconditional__'
|
527 |
+
model = instantiate_from_config(config)
|
528 |
+
self.cond_stage_model = model
|
529 |
+
|
530 |
+
def _get_denoise_row_from_list(self, samples, desc='', force_no_decoder_quantization=False):
|
531 |
+
denoise_row = []
|
532 |
+
for zd in tqdm(samples, desc=desc):
|
533 |
+
denoise_row.append(self.decode_first_stage(zd.to(self.device),
|
534 |
+
force_not_quantize=force_no_decoder_quantization))
|
535 |
+
n_imgs_per_row = len(denoise_row)
|
536 |
+
denoise_row = torch.stack(denoise_row) # n_log_step, n_row, C, H, W
|
537 |
+
denoise_grid = rearrange(denoise_row, 'n b c h w -> b n c h w')
|
538 |
+
denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
|
539 |
+
denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
|
540 |
+
return denoise_grid
|
541 |
+
|
542 |
+
def get_first_stage_encoding(self, encoder_posterior):
|
543 |
+
if isinstance(encoder_posterior, DiagonalGaussianDistribution):
|
544 |
+
z = encoder_posterior.sample()
|
545 |
+
elif isinstance(encoder_posterior, torch.Tensor):
|
546 |
+
z = encoder_posterior
|
547 |
+
else:
|
548 |
+
raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented")
|
549 |
+
return self.scale_factor * z
|
550 |
+
|
551 |
+
def get_learned_conditioning(self, c):
|
552 |
+
if self.cond_stage_forward is None:
|
553 |
+
if hasattr(self.cond_stage_model, 'encode') and callable(self.cond_stage_model.encode):
|
554 |
+
c = self.cond_stage_model.encode(c)
|
555 |
+
if isinstance(c, DiagonalGaussianDistribution):
|
556 |
+
c = c.mode()
|
557 |
+
else:
|
558 |
+
c = self.cond_stage_model(c)
|
559 |
+
else:
|
560 |
+
assert hasattr(self.cond_stage_model, self.cond_stage_forward)
|
561 |
+
c = getattr(self.cond_stage_model, self.cond_stage_forward)(c)
|
562 |
+
return c
|
563 |
+
|
564 |
+
def meshgrid(self, h, w):
|
565 |
+
y = torch.arange(0, h).view(h, 1, 1).repeat(1, w, 1)
|
566 |
+
x = torch.arange(0, w).view(1, w, 1).repeat(h, 1, 1)
|
567 |
+
|
568 |
+
arr = torch.cat([y, x], dim=-1)
|
569 |
+
return arr
|
570 |
+
|
571 |
+
def delta_border(self, h, w):
|
572 |
+
"""
|
573 |
+
:param h: height
|
574 |
+
:param w: width
|
575 |
+
:return: normalized distance to image border,
|
576 |
+
wtith min distance = 0 at border and max dist = 0.5 at image center
|
577 |
+
"""
|
578 |
+
lower_right_corner = torch.tensor([h - 1, w - 1]).view(1, 1, 2)
|
579 |
+
arr = self.meshgrid(h, w) / lower_right_corner
|
580 |
+
dist_left_up = torch.min(arr, dim=-1, keepdims=True)[0]
|
581 |
+
dist_right_down = torch.min(1 - arr, dim=-1, keepdims=True)[0]
|
582 |
+
edge_dist = torch.min(torch.cat([dist_left_up, dist_right_down], dim=-1), dim=-1)[0]
|
583 |
+
return edge_dist
|
584 |
+
|
585 |
+
def get_weighting(self, h, w, Ly, Lx, device):
|
586 |
+
weighting = self.delta_border(h, w)
|
587 |
+
weighting = torch.clip(weighting, self.split_input_params["clip_min_weight"],
|
588 |
+
self.split_input_params["clip_max_weight"], )
|
589 |
+
weighting = weighting.view(1, h * w, 1).repeat(1, 1, Ly * Lx).to(device)
|
590 |
+
|
591 |
+
if self.split_input_params["tie_braker"]:
|
592 |
+
L_weighting = self.delta_border(Ly, Lx)
|
593 |
+
L_weighting = torch.clip(L_weighting,
|
594 |
+
self.split_input_params["clip_min_tie_weight"],
|
595 |
+
self.split_input_params["clip_max_tie_weight"])
|
596 |
+
|
597 |
+
L_weighting = L_weighting.view(1, 1, Ly * Lx).to(device)
|
598 |
+
weighting = weighting * L_weighting
|
599 |
+
return weighting
|
600 |
+
|
601 |
+
def get_fold_unfold(self, x, kernel_size, stride, uf=1, df=1): # todo load once not every time, shorten code
|
602 |
+
"""
|
603 |
+
:param x: img of size (bs, c, h, w)
|
604 |
+
:return: n img crops of size (n, bs, c, kernel_size[0], kernel_size[1])
|
605 |
+
"""
|
606 |
+
bs, nc, h, w = x.shape
|
607 |
+
|
608 |
+
# number of crops in image
|
609 |
+
Ly = (h - kernel_size[0]) // stride[0] + 1
|
610 |
+
Lx = (w - kernel_size[1]) // stride[1] + 1
|
611 |
+
|
612 |
+
if uf == 1 and df == 1:
|
613 |
+
fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
|
614 |
+
unfold = torch.nn.Unfold(**fold_params)
|
615 |
+
|
616 |
+
fold = torch.nn.Fold(output_size=x.shape[2:], **fold_params)
|
617 |
+
|
618 |
+
weighting = self.get_weighting(kernel_size[0], kernel_size[1], Ly, Lx, x.device).to(x.dtype)
|
619 |
+
normalization = fold(weighting).view(1, 1, h, w) # normalizes the overlap
|
620 |
+
weighting = weighting.view((1, 1, kernel_size[0], kernel_size[1], Ly * Lx))
|
621 |
+
|
622 |
+
elif uf > 1 and df == 1:
|
623 |
+
fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
|
624 |
+
unfold = torch.nn.Unfold(**fold_params)
|
625 |
+
|
626 |
+
fold_params2 = dict(kernel_size=(kernel_size[0] * uf, kernel_size[0] * uf),
|
627 |
+
dilation=1, padding=0,
|
628 |
+
stride=(stride[0] * uf, stride[1] * uf))
|
629 |
+
fold = torch.nn.Fold(output_size=(x.shape[2] * uf, x.shape[3] * uf), **fold_params2)
|
630 |
+
|
631 |
+
weighting = self.get_weighting(kernel_size[0] * uf, kernel_size[1] * uf, Ly, Lx, x.device).to(x.dtype)
|
632 |
+
normalization = fold(weighting).view(1, 1, h * uf, w * uf) # normalizes the overlap
|
633 |
+
weighting = weighting.view((1, 1, kernel_size[0] * uf, kernel_size[1] * uf, Ly * Lx))
|
634 |
+
|
635 |
+
elif df > 1 and uf == 1:
|
636 |
+
fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
|
637 |
+
unfold = torch.nn.Unfold(**fold_params)
|
638 |
+
|
639 |
+
fold_params2 = dict(kernel_size=(kernel_size[0] // df, kernel_size[0] // df),
|
640 |
+
dilation=1, padding=0,
|
641 |
+
stride=(stride[0] // df, stride[1] // df))
|
642 |
+
fold = torch.nn.Fold(output_size=(x.shape[2] // df, x.shape[3] // df), **fold_params2)
|
643 |
+
|
644 |
+
weighting = self.get_weighting(kernel_size[0] // df, kernel_size[1] // df, Ly, Lx, x.device).to(x.dtype)
|
645 |
+
normalization = fold(weighting).view(1, 1, h // df, w // df) # normalizes the overlap
|
646 |
+
weighting = weighting.view((1, 1, kernel_size[0] // df, kernel_size[1] // df, Ly * Lx))
|
647 |
+
|
648 |
+
else:
|
649 |
+
raise NotImplementedError
|
650 |
+
|
651 |
+
return fold, unfold, normalization, weighting
|
652 |
+
|
653 |
+
# @torch.no_grad()
|
654 |
+
def get_input(self, batch, k, return_first_stage_outputs=False, force_c_encode=False,
|
655 |
+
cond_key=None, return_original_cond=False, bs=None):
|
656 |
+
x = super().get_input(batch, k)
|
657 |
+
if bs is not None:
|
658 |
+
x = x[:bs]
|
659 |
+
x = x.to(self.device)
|
660 |
+
encoder_posterior = self.encode_first_stage(x)
|
661 |
+
z = self.get_first_stage_encoding(encoder_posterior).detach()
|
662 |
+
|
663 |
+
if self.model.conditioning_key is not None:
|
664 |
+
if cond_key is None:
|
665 |
+
cond_key = self.cond_stage_key
|
666 |
+
if cond_key != self.first_stage_key:
|
667 |
+
if cond_key in ['caption', 'coordinates_bbox']:
|
668 |
+
xc = batch[cond_key]
|
669 |
+
elif cond_key == 'class_label':
|
670 |
+
xc = batch
|
671 |
+
else:
|
672 |
+
xc = super().get_input(batch, cond_key).to(self.device)
|
673 |
+
else:
|
674 |
+
xc = x
|
675 |
+
if not self.cond_stage_trainable or force_c_encode:
|
676 |
+
if isinstance(xc, dict) or isinstance(xc, list):
|
677 |
+
# import pudb; pudb.set_trace()
|
678 |
+
c = self.get_learned_conditioning(xc)
|
679 |
+
else:
|
680 |
+
c = self.get_learned_conditioning(xc.to(self.device))
|
681 |
+
else:
|
682 |
+
c = xc
|
683 |
+
if bs is not None:
|
684 |
+
c = c[:bs]
|
685 |
+
|
686 |
+
if self.use_positional_encodings:
|
687 |
+
pos_x, pos_y = self.compute_latent_shifts(batch)
|
688 |
+
ckey = __conditioning_keys__[self.model.conditioning_key]
|
689 |
+
c = {ckey: c, 'pos_x': pos_x, 'pos_y': pos_y}
|
690 |
+
|
691 |
+
else:
|
692 |
+
c = None
|
693 |
+
xc = None
|
694 |
+
if self.use_positional_encodings:
|
695 |
+
pos_x, pos_y = self.compute_latent_shifts(batch)
|
696 |
+
c = {'pos_x': pos_x, 'pos_y': pos_y}
|
697 |
+
out = [z, c]
|
698 |
+
if return_first_stage_outputs:
|
699 |
+
xrec = self.decode_first_stage(z)
|
700 |
+
out.extend([x, xrec])
|
701 |
+
if return_original_cond:
|
702 |
+
out.append(xc)
|
703 |
+
return out
|
704 |
+
|
705 |
+
# @torch.no_grad()
|
706 |
+
def decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
|
707 |
+
if predict_cids:
|
708 |
+
if z.dim() == 4:
|
709 |
+
z = torch.argmax(z.exp(), dim=1).long()
|
710 |
+
z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
|
711 |
+
z = rearrange(z, 'b h w c -> b c h w').contiguous()
|
712 |
+
|
713 |
+
z = 1. / self.scale_factor * z
|
714 |
+
|
715 |
+
if hasattr(self, "split_input_params"):
|
716 |
+
if self.split_input_params["patch_distributed_vq"]:
|
717 |
+
ks = self.split_input_params["ks"] # eg. (128, 128)
|
718 |
+
stride = self.split_input_params["stride"] # eg. (64, 64)
|
719 |
+
uf = self.split_input_params["vqf"]
|
720 |
+
bs, nc, h, w = z.shape
|
721 |
+
if ks[0] > h or ks[1] > w:
|
722 |
+
ks = (min(ks[0], h), min(ks[1], w))
|
723 |
+
print("reducing Kernel")
|
724 |
+
|
725 |
+
if stride[0] > h or stride[1] > w:
|
726 |
+
stride = (min(stride[0], h), min(stride[1], w))
|
727 |
+
print("reducing stride")
|
728 |
+
|
729 |
+
fold, unfold, normalization, weighting = self.get_fold_unfold(z, ks, stride, uf=uf)
|
730 |
+
|
731 |
+
z = unfold(z) # (bn, nc * prod(**ks), L)
|
732 |
+
# 1. Reshape to img shape
|
733 |
+
z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1])) # (bn, nc, ks[0], ks[1], L )
|
734 |
+
|
735 |
+
# 2. apply model loop over last dim
|
736 |
+
if isinstance(self.first_stage_model, VQModelInterface):
|
737 |
+
output_list = [self.first_stage_model.decode(z[:, :, :, :, i],
|
738 |
+
force_not_quantize=predict_cids or force_not_quantize)
|
739 |
+
for i in range(z.shape[-1])]
|
740 |
+
else:
|
741 |
+
|
742 |
+
output_list = [self.first_stage_model.decode(z[:, :, :, :, i])
|
743 |
+
for i in range(z.shape[-1])]
|
744 |
+
|
745 |
+
o = torch.stack(output_list, axis=-1) # # (bn, nc, ks[0], ks[1], L)
|
746 |
+
o = o * weighting
|
747 |
+
# Reverse 1. reshape to img shape
|
748 |
+
o = o.view((o.shape[0], -1, o.shape[-1])) # (bn, nc * ks[0] * ks[1], L)
|
749 |
+
# stitch crops together
|
750 |
+
decoded = fold(o)
|
751 |
+
decoded = decoded / normalization # norm is shape (1, 1, h, w)
|
752 |
+
return decoded
|
753 |
+
else:
|
754 |
+
if isinstance(self.first_stage_model, VQModelInterface):
|
755 |
+
return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
|
756 |
+
else:
|
757 |
+
return self.first_stage_model.decode(z)
|
758 |
+
|
759 |
+
else:
|
760 |
+
if isinstance(self.first_stage_model, VQModelInterface):
|
761 |
+
return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
|
762 |
+
else:
|
763 |
+
return self.first_stage_model.decode(z)
|
764 |
+
|
765 |
+
# same as above but without decorator
|
766 |
+
def differentiable_decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
|
767 |
+
if predict_cids:
|
768 |
+
if z.dim() == 4:
|
769 |
+
z = torch.argmax(z.exp(), dim=1).long()
|
770 |
+
z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
|
771 |
+
z = rearrange(z, 'b h w c -> b c h w').contiguous()
|
772 |
+
|
773 |
+
z = 1. / self.scale_factor * z
|
774 |
+
|
775 |
+
if hasattr(self, "split_input_params"):
|
776 |
+
if self.split_input_params["patch_distributed_vq"]:
|
777 |
+
ks = self.split_input_params["ks"] # eg. (128, 128)
|
778 |
+
stride = self.split_input_params["stride"] # eg. (64, 64)
|
779 |
+
uf = self.split_input_params["vqf"]
|
780 |
+
bs, nc, h, w = z.shape
|
781 |
+
if ks[0] > h or ks[1] > w:
|
782 |
+
ks = (min(ks[0], h), min(ks[1], w))
|
783 |
+
print("reducing Kernel")
|
784 |
+
|
785 |
+
if stride[0] > h or stride[1] > w:
|
786 |
+
stride = (min(stride[0], h), min(stride[1], w))
|
787 |
+
print("reducing stride")
|
788 |
+
|
789 |
+
fold, unfold, normalization, weighting = self.get_fold_unfold(z, ks, stride, uf=uf)
|
790 |
+
|
791 |
+
z = unfold(z) # (bn, nc * prod(**ks), L)
|
792 |
+
# 1. Reshape to img shape
|
793 |
+
z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1])) # (bn, nc, ks[0], ks[1], L )
|
794 |
+
|
795 |
+
# 2. apply model loop over last dim
|
796 |
+
if isinstance(self.first_stage_model, VQModelInterface):
|
797 |
+
output_list = [self.first_stage_model.decode(z[:, :, :, :, i],
|
798 |
+
force_not_quantize=predict_cids or force_not_quantize)
|
799 |
+
for i in range(z.shape[-1])]
|
800 |
+
else:
|
801 |
+
|
802 |
+
output_list = [self.first_stage_model.decode(z[:, :, :, :, i])
|
803 |
+
for i in range(z.shape[-1])]
|
804 |
+
|
805 |
+
o = torch.stack(output_list, axis=-1) # # (bn, nc, ks[0], ks[1], L)
|
806 |
+
o = o * weighting
|
807 |
+
# Reverse 1. reshape to img shape
|
808 |
+
o = o.view((o.shape[0], -1, o.shape[-1])) # (bn, nc * ks[0] * ks[1], L)
|
809 |
+
# stitch crops together
|
810 |
+
decoded = fold(o)
|
811 |
+
decoded = decoded / normalization # norm is shape (1, 1, h, w)
|
812 |
+
return decoded
|
813 |
+
else:
|
814 |
+
if isinstance(self.first_stage_model, VQModelInterface):
|
815 |
+
return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
|
816 |
+
else:
|
817 |
+
return self.first_stage_model.decode(z)
|
818 |
+
|
819 |
+
else:
|
820 |
+
if isinstance(self.first_stage_model, VQModelInterface):
|
821 |
+
return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
|
822 |
+
else:
|
823 |
+
return self.first_stage_model.decode(z)
|
824 |
+
|
825 |
+
# @torch.no_grad()
|
826 |
+
def encode_first_stage(self, x):
|
827 |
+
if hasattr(self, "split_input_params"):
|
828 |
+
if self.split_input_params["patch_distributed_vq"]:
|
829 |
+
ks = self.split_input_params["ks"] # eg. (128, 128)
|
830 |
+
stride = self.split_input_params["stride"] # eg. (64, 64)
|
831 |
+
df = self.split_input_params["vqf"]
|
832 |
+
self.split_input_params['original_image_size'] = x.shape[-2:]
|
833 |
+
bs, nc, h, w = x.shape
|
834 |
+
if ks[0] > h or ks[1] > w:
|
835 |
+
ks = (min(ks[0], h), min(ks[1], w))
|
836 |
+
print("reducing Kernel")
|
837 |
+
|
838 |
+
if stride[0] > h or stride[1] > w:
|
839 |
+
stride = (min(stride[0], h), min(stride[1], w))
|
840 |
+
print("reducing stride")
|
841 |
+
|
842 |
+
fold, unfold, normalization, weighting = self.get_fold_unfold(x, ks, stride, df=df)
|
843 |
+
z = unfold(x) # (bn, nc * prod(**ks), L)
|
844 |
+
# Reshape to img shape
|
845 |
+
z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1])) # (bn, nc, ks[0], ks[1], L )
|
846 |
+
|
847 |
+
output_list = [self.first_stage_model.encode(z[:, :, :, :, i])
|
848 |
+
for i in range(z.shape[-1])]
|
849 |
+
|
850 |
+
o = torch.stack(output_list, axis=-1)
|
851 |
+
o = o * weighting
|
852 |
+
|
853 |
+
# Reverse reshape to img shape
|
854 |
+
o = o.view((o.shape[0], -1, o.shape[-1])) # (bn, nc * ks[0] * ks[1], L)
|
855 |
+
# stitch crops together
|
856 |
+
decoded = fold(o)
|
857 |
+
decoded = decoded / normalization
|
858 |
+
return decoded
|
859 |
+
|
860 |
+
else:
|
861 |
+
return self.first_stage_model.encode(x)
|
862 |
+
else:
|
863 |
+
return self.first_stage_model.encode(x)
|
864 |
+
|
865 |
+
def shared_step(self, batch, **kwargs):
|
866 |
+
x, c = self.get_input(batch, self.first_stage_key)
|
867 |
+
loss = self(x, c)
|
868 |
+
return loss
|
869 |
+
|
870 |
+
def forward(self, x, c, *args, **kwargs):
|
871 |
+
t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
|
872 |
+
if self.model.conditioning_key is not None:
|
873 |
+
assert c is not None
|
874 |
+
if self.cond_stage_trainable:
|
875 |
+
c = self.get_learned_conditioning(c)
|
876 |
+
if self.shorten_cond_schedule: # TODO: drop this option
|
877 |
+
tc = self.cond_ids[t].to(self.device)
|
878 |
+
c = self.q_sample(x_start=c, t=tc, noise=torch.randn_like(c.float()))
|
879 |
+
return self.p_losses(x, c, t, *args, **kwargs)
|
880 |
+
|
881 |
+
def _rescale_annotations(self, bboxes, crop_coordinates): # TODO: move to dataset
|
882 |
+
def rescale_bbox(bbox):
|
883 |
+
x0 = clamp((bbox[0] - crop_coordinates[0]) / crop_coordinates[2])
|
884 |
+
y0 = clamp((bbox[1] - crop_coordinates[1]) / crop_coordinates[3])
|
885 |
+
w = min(bbox[2] / crop_coordinates[2], 1 - x0)
|
886 |
+
h = min(bbox[3] / crop_coordinates[3], 1 - y0)
|
887 |
+
return x0, y0, w, h
|
888 |
+
|
889 |
+
return [rescale_bbox(b) for b in bboxes]
|
890 |
+
|
891 |
+
def apply_model(self, x_noisy, t, cond, return_ids=False):
|
892 |
+
|
893 |
+
if isinstance(cond, dict):
|
894 |
+
# hybrid case, cond is exptected to be a dict
|
895 |
+
pass
|
896 |
+
else:
|
897 |
+
if not isinstance(cond, list):
|
898 |
+
cond = [cond]
|
899 |
+
key = 'c_concat' if self.model.conditioning_key == 'concat' else 'c_crossattn'
|
900 |
+
cond = {key: cond}
|
901 |
+
|
902 |
+
if hasattr(self, "split_input_params"):
|
903 |
+
assert len(cond) == 1 # todo can only deal with one conditioning atm
|
904 |
+
assert not return_ids
|
905 |
+
ks = self.split_input_params["ks"] # eg. (128, 128)
|
906 |
+
stride = self.split_input_params["stride"] # eg. (64, 64)
|
907 |
+
|
908 |
+
h, w = x_noisy.shape[-2:]
|
909 |
+
|
910 |
+
fold, unfold, normalization, weighting = self.get_fold_unfold(x_noisy, ks, stride)
|
911 |
+
|
912 |
+
z = unfold(x_noisy) # (bn, nc * prod(**ks), L)
|
913 |
+
# Reshape to img shape
|
914 |
+
z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1])) # (bn, nc, ks[0], ks[1], L )
|
915 |
+
z_list = [z[:, :, :, :, i] for i in range(z.shape[-1])]
|
916 |
+
|
917 |
+
if self.cond_stage_key in ["image", "LR_image", "segmentation",
|
918 |
+
'bbox_img'] and self.model.conditioning_key: # todo check for completeness
|
919 |
+
c_key = next(iter(cond.keys())) # get key
|
920 |
+
c = next(iter(cond.values())) # get value
|
921 |
+
assert (len(c) == 1) # todo extend to list with more than one elem
|
922 |
+
c = c[0] # get element
|
923 |
+
|
924 |
+
c = unfold(c)
|
925 |
+
c = c.view((c.shape[0], -1, ks[0], ks[1], c.shape[-1])) # (bn, nc, ks[0], ks[1], L )
|
926 |
+
|
927 |
+
cond_list = [{c_key: [c[:, :, :, :, i]]} for i in range(c.shape[-1])]
|
928 |
+
|
929 |
+
elif self.cond_stage_key == 'coordinates_bbox':
|
930 |
+
assert 'original_image_size' in self.split_input_params, 'BoudingBoxRescaling is missing original_image_size'
|
931 |
+
|
932 |
+
# assuming padding of unfold is always 0 and its dilation is always 1
|
933 |
+
n_patches_per_row = int((w - ks[0]) / stride[0] + 1)
|
934 |
+
full_img_h, full_img_w = self.split_input_params['original_image_size']
|
935 |
+
# as we are operating on latents, we need the factor from the original image size to the
|
936 |
+
# spatial latent size to properly rescale the crops for regenerating the bbox annotations
|
937 |
+
num_downs = self.first_stage_model.encoder.num_resolutions - 1
|
938 |
+
rescale_latent = 2 ** (num_downs)
|
939 |
+
|
940 |
+
# get top left postions of patches as conforming for the bbbox tokenizer, therefore we
|
941 |
+
# need to rescale the tl patch coordinates to be in between (0,1)
|
942 |
+
tl_patch_coordinates = [(rescale_latent * stride[0] * (patch_nr % n_patches_per_row) / full_img_w,
|
943 |
+
rescale_latent * stride[1] * (patch_nr // n_patches_per_row) / full_img_h)
|
944 |
+
for patch_nr in range(z.shape[-1])]
|
945 |
+
|
946 |
+
# patch_limits are tl_coord, width and height coordinates as (x_tl, y_tl, h, w)
|
947 |
+
patch_limits = [(x_tl, y_tl,
|
948 |
+
rescale_latent * ks[0] / full_img_w,
|
949 |
+
rescale_latent * ks[1] / full_img_h) for x_tl, y_tl in tl_patch_coordinates]
|
950 |
+
# patch_values = [(np.arange(x_tl,min(x_tl+ks, 1.)),np.arange(y_tl,min(y_tl+ks, 1.))) for x_tl, y_tl in tl_patch_coordinates]
|
951 |
+
|
952 |
+
# tokenize crop coordinates for the bounding boxes of the respective patches
|
953 |
+
patch_limits_tknzd = [torch.LongTensor(self.bbox_tokenizer._crop_encoder(bbox))[None].to(self.device)
|
954 |
+
for bbox in patch_limits] # list of length l with tensors of shape (1, 2)
|
955 |
+
print(patch_limits_tknzd[0].shape)
|
956 |
+
# cut tknzd crop position from conditioning
|
957 |
+
assert isinstance(cond, dict), 'cond must be dict to be fed into model'
|
958 |
+
cut_cond = cond['c_crossattn'][0][..., :-2].to(self.device)
|
959 |
+
print(cut_cond.shape)
|
960 |
+
|
961 |
+
adapted_cond = torch.stack([torch.cat([cut_cond, p], dim=1) for p in patch_limits_tknzd])
|
962 |
+
adapted_cond = rearrange(adapted_cond, 'l b n -> (l b) n')
|
963 |
+
print(adapted_cond.shape)
|
964 |
+
adapted_cond = self.get_learned_conditioning(adapted_cond)
|
965 |
+
print(adapted_cond.shape)
|
966 |
+
adapted_cond = rearrange(adapted_cond, '(l b) n d -> l b n d', l=z.shape[-1])
|
967 |
+
print(adapted_cond.shape)
|
968 |
+
|
969 |
+
cond_list = [{'c_crossattn': [e]} for e in adapted_cond]
|
970 |
+
|
971 |
+
else:
|
972 |
+
cond_list = [cond for i in range(z.shape[-1])] # Todo make this more efficient
|
973 |
+
|
974 |
+
# apply model by loop over crops
|
975 |
+
output_list = [self.model(z_list[i], t, **cond_list[i]) for i in range(z.shape[-1])]
|
976 |
+
assert not isinstance(output_list[0],
|
977 |
+
tuple) # todo cant deal with multiple model outputs check this never happens
|
978 |
+
|
979 |
+
o = torch.stack(output_list, axis=-1)
|
980 |
+
o = o * weighting
|
981 |
+
# Reverse reshape to img shape
|
982 |
+
o = o.view((o.shape[0], -1, o.shape[-1])) # (bn, nc * ks[0] * ks[1], L)
|
983 |
+
# stitch crops together
|
984 |
+
x_recon = fold(o) / normalization
|
985 |
+
|
986 |
+
else:
|
987 |
+
x_recon = self.model(x_noisy, t, **cond)
|
988 |
+
|
989 |
+
if isinstance(x_recon, tuple) and not return_ids:
|
990 |
+
return x_recon[0]
|
991 |
+
else:
|
992 |
+
return x_recon
|
993 |
+
|
994 |
+
def _predict_eps_from_xstart(self, x_t, t, pred_xstart):
|
995 |
+
return (extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - pred_xstart) / \
|
996 |
+
extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape)
|
997 |
+
|
998 |
+
def _prior_bpd(self, x_start):
|
999 |
+
"""
|
1000 |
+
Get the prior KL term for the variational lower-bound, measured in
|
1001 |
+
bits-per-dim.
|
1002 |
+
This term can't be optimized, as it only depends on the encoder.
|
1003 |
+
:param x_start: the [N x C x ...] tensor of inputs.
|
1004 |
+
:return: a batch of [N] KL values (in bits), one per batch element.
|
1005 |
+
"""
|
1006 |
+
batch_size = x_start.shape[0]
|
1007 |
+
t = torch.tensor([self.num_timesteps - 1] * batch_size, device=x_start.device)
|
1008 |
+
qt_mean, _, qt_log_variance = self.q_mean_variance(x_start, t)
|
1009 |
+
kl_prior = normal_kl(mean1=qt_mean, logvar1=qt_log_variance, mean2=0.0, logvar2=0.0)
|
1010 |
+
return mean_flat(kl_prior) / np.log(2.0)
|
1011 |
+
|
1012 |
+
def p_losses(self, x_start, cond, t, noise=None):
|
1013 |
+
noise = default(noise, lambda: torch.randn_like(x_start))
|
1014 |
+
x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
|
1015 |
+
model_output = self.apply_model(x_noisy, t, cond)
|
1016 |
+
|
1017 |
+
loss_dict = {}
|
1018 |
+
prefix = 'train' if self.training else 'val'
|
1019 |
+
|
1020 |
+
if self.parameterization == "x0":
|
1021 |
+
target = x_start
|
1022 |
+
elif self.parameterization == "eps":
|
1023 |
+
target = noise
|
1024 |
+
else:
|
1025 |
+
raise NotImplementedError()
|
1026 |
+
|
1027 |
+
loss_simple = self.get_loss(model_output, target, mean=False).mean([1, 2, 3])
|
1028 |
+
loss_dict.update({f'{prefix}/loss_simple': loss_simple.mean()})
|
1029 |
+
|
1030 |
+
logvar_t = self.logvar[t].to(self.device)
|
1031 |
+
loss = loss_simple / torch.exp(logvar_t) + logvar_t
|
1032 |
+
# loss = loss_simple / torch.exp(self.logvar) + self.logvar
|
1033 |
+
if self.learn_logvar:
|
1034 |
+
loss_dict.update({f'{prefix}/loss_gamma': loss.mean()})
|
1035 |
+
loss_dict.update({'logvar': self.logvar.data.mean()})
|
1036 |
+
|
1037 |
+
loss = self.l_simple_weight * loss.mean()
|
1038 |
+
|
1039 |
+
loss_vlb = self.get_loss(model_output, target, mean=False).mean(dim=(1, 2, 3))
|
1040 |
+
loss_vlb = (self.lvlb_weights[t] * loss_vlb).mean()
|
1041 |
+
loss_dict.update({f'{prefix}/loss_vlb': loss_vlb})
|
1042 |
+
loss += (self.original_elbo_weight * loss_vlb)
|
1043 |
+
loss_dict.update({f'{prefix}/loss': loss})
|
1044 |
+
|
1045 |
+
return loss, loss_dict
|
1046 |
+
|
1047 |
+
def p_mean_variance(self, x, c, t, clip_denoised: bool, return_codebook_ids=False, quantize_denoised=False,
|
1048 |
+
return_x0=False, score_corrector=None, corrector_kwargs=None):
|
1049 |
+
t_in = t
|
1050 |
+
model_out = self.apply_model(x, t_in, c, return_ids=return_codebook_ids)
|
1051 |
+
|
1052 |
+
if score_corrector is not None:
|
1053 |
+
assert self.parameterization == "eps"
|
1054 |
+
model_out = score_corrector.modify_score(self, model_out, x, t, c, **corrector_kwargs)
|
1055 |
+
|
1056 |
+
if return_codebook_ids:
|
1057 |
+
model_out, logits = model_out
|
1058 |
+
|
1059 |
+
if self.parameterization == "eps":
|
1060 |
+
x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
|
1061 |
+
elif self.parameterization == "x0":
|
1062 |
+
x_recon = model_out
|
1063 |
+
else:
|
1064 |
+
raise NotImplementedError()
|
1065 |
+
|
1066 |
+
if clip_denoised:
|
1067 |
+
x_recon.clamp_(-1., 1.)
|
1068 |
+
if quantize_denoised:
|
1069 |
+
x_recon, _, [_, _, indices] = self.first_stage_model.quantize(x_recon)
|
1070 |
+
model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
|
1071 |
+
if return_codebook_ids:
|
1072 |
+
return model_mean, posterior_variance, posterior_log_variance, logits
|
1073 |
+
elif return_x0:
|
1074 |
+
return model_mean, posterior_variance, posterior_log_variance, x_recon
|
1075 |
+
else:
|
1076 |
+
return model_mean, posterior_variance, posterior_log_variance
|
1077 |
+
|
1078 |
+
# @torch.no_grad()
|
1079 |
+
def p_sample(self, x, c, t, clip_denoised=False, repeat_noise=False,
|
1080 |
+
return_codebook_ids=False, quantize_denoised=False, return_x0=False,
|
1081 |
+
temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None):
|
1082 |
+
b, *_, device = *x.shape, x.device
|
1083 |
+
outputs = self.p_mean_variance(x=x, c=c, t=t, clip_denoised=clip_denoised,
|
1084 |
+
return_codebook_ids=return_codebook_ids,
|
1085 |
+
quantize_denoised=quantize_denoised,
|
1086 |
+
return_x0=return_x0,
|
1087 |
+
score_corrector=score_corrector, corrector_kwargs=corrector_kwargs)
|
1088 |
+
if return_codebook_ids:
|
1089 |
+
raise DeprecationWarning("Support dropped.")
|
1090 |
+
model_mean, _, model_log_variance, logits = outputs
|
1091 |
+
elif return_x0:
|
1092 |
+
model_mean, _, model_log_variance, x0 = outputs
|
1093 |
+
else:
|
1094 |
+
model_mean, _, model_log_variance = outputs
|
1095 |
+
|
1096 |
+
noise = noise_like(x.shape, device, repeat_noise) * temperature
|
1097 |
+
if noise_dropout > 0.:
|
1098 |
+
noise = torch.nn.functional.dropout(noise, p=noise_dropout)
|
1099 |
+
# no noise when t == 0
|
1100 |
+
nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
|
1101 |
+
|
1102 |
+
if return_codebook_ids:
|
1103 |
+
return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, logits.argmax(dim=1)
|
1104 |
+
if return_x0:
|
1105 |
+
return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, x0
|
1106 |
+
else:
|
1107 |
+
return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
|
1108 |
+
|
1109 |
+
# @torch.no_grad()
|
1110 |
+
def progressive_denoising(self, cond, shape, verbose=True, callback=None, quantize_denoised=False,
|
1111 |
+
img_callback=None, mask=None, x0=None, temperature=1., noise_dropout=0.,
|
1112 |
+
score_corrector=None, corrector_kwargs=None, batch_size=None, x_T=None, start_T=None,
|
1113 |
+
log_every_t=None):
|
1114 |
+
if not log_every_t:
|
1115 |
+
log_every_t = self.log_every_t
|
1116 |
+
timesteps = self.num_timesteps
|
1117 |
+
if batch_size is not None:
|
1118 |
+
b = batch_size if batch_size is not None else shape[0]
|
1119 |
+
shape = [batch_size] + list(shape)
|
1120 |
+
else:
|
1121 |
+
b = batch_size = shape[0]
|
1122 |
+
if x_T is None:
|
1123 |
+
img = torch.randn(shape, device=self.device)
|
1124 |
+
else:
|
1125 |
+
img = x_T
|
1126 |
+
intermediates = []
|
1127 |
+
if cond is not None:
|
1128 |
+
if isinstance(cond, dict):
|
1129 |
+
cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
|
1130 |
+
list(map(lambda x: x[:batch_size], cond[key])) for key in cond}
|
1131 |
+
else:
|
1132 |
+
cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
|
1133 |
+
|
1134 |
+
if start_T is not None:
|
1135 |
+
timesteps = min(timesteps, start_T)
|
1136 |
+
iterator = tqdm(reversed(range(0, timesteps)), desc='Progressive Generation',
|
1137 |
+
total=timesteps) if verbose else reversed(
|
1138 |
+
range(0, timesteps))
|
1139 |
+
if type(temperature) == float:
|
1140 |
+
temperature = [temperature] * timesteps
|
1141 |
+
|
1142 |
+
for i in iterator:
|
1143 |
+
ts = torch.full((b,), i, device=self.device, dtype=torch.long)
|
1144 |
+
if self.shorten_cond_schedule:
|
1145 |
+
assert self.model.conditioning_key != 'hybrid'
|
1146 |
+
tc = self.cond_ids[ts].to(cond.device)
|
1147 |
+
cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
|
1148 |
+
|
1149 |
+
img, x0_partial = self.p_sample(img, cond, ts,
|
1150 |
+
clip_denoised=self.clip_denoised,
|
1151 |
+
quantize_denoised=quantize_denoised, return_x0=True,
|
1152 |
+
temperature=temperature[i], noise_dropout=noise_dropout,
|
1153 |
+
score_corrector=score_corrector, corrector_kwargs=corrector_kwargs)
|
1154 |
+
if mask is not None:
|
1155 |
+
assert x0 is not None
|
1156 |
+
img_orig = self.q_sample(x0, ts)
|
1157 |
+
img = img_orig * mask + (1. - mask) * img
|
1158 |
+
|
1159 |
+
if i % log_every_t == 0 or i == timesteps - 1:
|
1160 |
+
intermediates.append(x0_partial)
|
1161 |
+
if callback: callback(i)
|
1162 |
+
if img_callback: img_callback(img, i)
|
1163 |
+
return img, intermediates
|
1164 |
+
|
1165 |
+
# @torch.no_grad()
|
1166 |
+
def p_sample_loop(self, cond, shape, return_intermediates=False,
|
1167 |
+
x_T=None, verbose=True, callback=None, timesteps=None, quantize_denoised=False,
|
1168 |
+
mask=None, x0=None, img_callback=None, start_T=None,
|
1169 |
+
log_every_t=None):
|
1170 |
+
|
1171 |
+
if not log_every_t:
|
1172 |
+
log_every_t = self.log_every_t
|
1173 |
+
device = self.betas.device
|
1174 |
+
b = shape[0]
|
1175 |
+
if x_T is None:
|
1176 |
+
img = torch.randn(shape, device=device)
|
1177 |
+
else:
|
1178 |
+
img = x_T
|
1179 |
+
|
1180 |
+
intermediates = [img]
|
1181 |
+
if timesteps is None:
|
1182 |
+
timesteps = self.num_timesteps
|
1183 |
+
|
1184 |
+
if start_T is not None:
|
1185 |
+
timesteps = min(timesteps, start_T)
|
1186 |
+
iterator = tqdm(reversed(range(0, timesteps)), desc='Sampling t', total=timesteps) if verbose else reversed(
|
1187 |
+
range(0, timesteps))
|
1188 |
+
|
1189 |
+
if mask is not None:
|
1190 |
+
assert x0 is not None
|
1191 |
+
assert x0.shape[2:3] == mask.shape[2:3] # spatial size has to match
|
1192 |
+
|
1193 |
+
for i in iterator:
|
1194 |
+
ts = torch.full((b,), i, device=device, dtype=torch.long)
|
1195 |
+
if self.shorten_cond_schedule:
|
1196 |
+
assert self.model.conditioning_key != 'hybrid'
|
1197 |
+
tc = self.cond_ids[ts].to(cond.device)
|
1198 |
+
cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
|
1199 |
+
|
1200 |
+
img = self.p_sample(img, cond, ts,
|
1201 |
+
clip_denoised=self.clip_denoised,
|
1202 |
+
quantize_denoised=quantize_denoised)
|
1203 |
+
if mask is not None:
|
1204 |
+
img_orig = self.q_sample(x0, ts)
|
1205 |
+
img = img_orig * mask + (1. - mask) * img
|
1206 |
+
|
1207 |
+
if i % log_every_t == 0 or i == timesteps - 1:
|
1208 |
+
intermediates.append(img)
|
1209 |
+
if callback: callback(i)
|
1210 |
+
if img_callback: img_callback(img, i)
|
1211 |
+
|
1212 |
+
if return_intermediates:
|
1213 |
+
return img, intermediates
|
1214 |
+
return img
|
1215 |
+
|
1216 |
+
# @torch.no_grad()
|
1217 |
+
def sample(self, cond, batch_size=16, return_intermediates=False, x_T=None,
|
1218 |
+
verbose=True, timesteps=None, quantize_denoised=False,
|
1219 |
+
mask=None, x0=None, shape=None,**kwargs):
|
1220 |
+
if shape is None:
|
1221 |
+
shape = (batch_size, self.channels, self.image_size, self.image_size)
|
1222 |
+
if cond is not None:
|
1223 |
+
if isinstance(cond, dict):
|
1224 |
+
cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
|
1225 |
+
list(map(lambda x: x[:batch_size], cond[key])) for key in cond}
|
1226 |
+
else:
|
1227 |
+
cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
|
1228 |
+
return self.p_sample_loop(cond,
|
1229 |
+
shape,
|
1230 |
+
return_intermediates=return_intermediates, x_T=x_T,
|
1231 |
+
verbose=verbose, timesteps=timesteps, quantize_denoised=quantize_denoised,
|
1232 |
+
mask=mask, x0=x0)
|
1233 |
+
|
1234 |
+
# @torch.no_grad()
|
1235 |
+
def sample_log(self,cond,batch_size,ddim, ddim_steps,**kwargs):
|
1236 |
+
|
1237 |
+
if ddim:
|
1238 |
+
ddim_sampler = DDIMSampler(self)
|
1239 |
+
shape = (self.channels, self.image_size, self.image_size)
|
1240 |
+
samples, intermediates =ddim_sampler.sample(ddim_steps,batch_size,
|
1241 |
+
shape,cond,verbose=False,**kwargs)
|
1242 |
+
|
1243 |
+
else:
|
1244 |
+
samples, intermediates = self.sample(cond=cond, batch_size=batch_size,
|
1245 |
+
return_intermediates=True,**kwargs)
|
1246 |
+
|
1247 |
+
return samples, intermediates
|
1248 |
+
|
1249 |
+
|
1250 |
+
# @torch.no_grad()
|
1251 |
+
def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None,
|
1252 |
+
quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True,
|
1253 |
+
plot_diffusion_rows=True, **kwargs):
|
1254 |
+
|
1255 |
+
use_ddim = ddim_steps is not None
|
1256 |
+
|
1257 |
+
log = dict()
|
1258 |
+
z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key,
|
1259 |
+
return_first_stage_outputs=True,
|
1260 |
+
force_c_encode=True,
|
1261 |
+
return_original_cond=True,
|
1262 |
+
bs=N)
|
1263 |
+
N = min(x.shape[0], N)
|
1264 |
+
n_row = min(x.shape[0], n_row)
|
1265 |
+
log["inputs"] = x
|
1266 |
+
log["reconstruction"] = xrec
|
1267 |
+
if self.model.conditioning_key is not None:
|
1268 |
+
if hasattr(self.cond_stage_model, "decode"):
|
1269 |
+
xc = self.cond_stage_model.decode(c)
|
1270 |
+
log["conditioning"] = xc
|
1271 |
+
elif self.cond_stage_key in ["caption"]:
|
1272 |
+
xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["caption"])
|
1273 |
+
log["conditioning"] = xc
|
1274 |
+
elif self.cond_stage_key == 'class_label':
|
1275 |
+
xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"])
|
1276 |
+
log['conditioning'] = xc
|
1277 |
+
elif isimage(xc):
|
1278 |
+
log["conditioning"] = xc
|
1279 |
+
if ismap(xc):
|
1280 |
+
log["original_conditioning"] = self.to_rgb(xc)
|
1281 |
+
|
1282 |
+
if plot_diffusion_rows:
|
1283 |
+
# get diffusion row
|
1284 |
+
diffusion_row = list()
|
1285 |
+
z_start = z[:n_row]
|
1286 |
+
for t in range(self.num_timesteps):
|
1287 |
+
if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
|
1288 |
+
t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
|
1289 |
+
t = t.to(self.device).long()
|
1290 |
+
noise = torch.randn_like(z_start)
|
1291 |
+
z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise)
|
1292 |
+
diffusion_row.append(self.decode_first_stage(z_noisy))
|
1293 |
+
|
1294 |
+
diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W
|
1295 |
+
diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w')
|
1296 |
+
diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w')
|
1297 |
+
diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0])
|
1298 |
+
log["diffusion_row"] = diffusion_grid
|
1299 |
+
|
1300 |
+
if sample:
|
1301 |
+
# get denoise row
|
1302 |
+
with self.ema_scope("Plotting"):
|
1303 |
+
samples, z_denoise_row = self.sample_log(cond=c,batch_size=N,ddim=use_ddim,
|
1304 |
+
ddim_steps=ddim_steps,eta=ddim_eta)
|
1305 |
+
# samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True)
|
1306 |
+
x_samples = self.decode_first_stage(samples)
|
1307 |
+
log["samples"] = x_samples
|
1308 |
+
if plot_denoise_rows:
|
1309 |
+
denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
|
1310 |
+
log["denoise_row"] = denoise_grid
|
1311 |
+
|
1312 |
+
if quantize_denoised and not isinstance(self.first_stage_model, AutoencoderKL) and not isinstance(
|
1313 |
+
self.first_stage_model, IdentityFirstStage):
|
1314 |
+
# also display when quantizing x0 while sampling
|
1315 |
+
with self.ema_scope("Plotting Quantized Denoised"):
|
1316 |
+
samples, z_denoise_row = self.sample_log(cond=c,batch_size=N,ddim=use_ddim,
|
1317 |
+
ddim_steps=ddim_steps,eta=ddim_eta,
|
1318 |
+
quantize_denoised=True)
|
1319 |
+
# samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True,
|
1320 |
+
# quantize_denoised=True)
|
1321 |
+
x_samples = self.decode_first_stage(samples.to(self.device))
|
1322 |
+
log["samples_x0_quantized"] = x_samples
|
1323 |
+
|
1324 |
+
if inpaint:
|
1325 |
+
# make a simple center square
|
1326 |
+
b, h, w = z.shape[0], z.shape[2], z.shape[3]
|
1327 |
+
mask = torch.ones(N, h, w).to(self.device)
|
1328 |
+
# zeros will be filled in
|
1329 |
+
mask[:, h // 4:3 * h // 4, w // 4:3 * w // 4] = 0.
|
1330 |
+
mask = mask[:, None, ...]
|
1331 |
+
with self.ema_scope("Plotting Inpaint"):
|
1332 |
+
|
1333 |
+
samples, _ = self.sample_log(cond=c,batch_size=N,ddim=use_ddim, eta=ddim_eta,
|
1334 |
+
ddim_steps=ddim_steps, x0=z[:N], mask=mask)
|
1335 |
+
x_samples = self.decode_first_stage(samples.to(self.device))
|
1336 |
+
log["samples_inpainting"] = x_samples
|
1337 |
+
log["mask"] = mask
|
1338 |
+
|
1339 |
+
# outpaint
|
1340 |
+
with self.ema_scope("Plotting Outpaint"):
|
1341 |
+
samples, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,eta=ddim_eta,
|
1342 |
+
ddim_steps=ddim_steps, x0=z[:N], mask=mask)
|
1343 |
+
x_samples = self.decode_first_stage(samples.to(self.device))
|
1344 |
+
log["samples_outpainting"] = x_samples
|
1345 |
+
|
1346 |
+
if plot_progressive_rows:
|
1347 |
+
with self.ema_scope("Plotting Progressives"):
|
1348 |
+
img, progressives = self.progressive_denoising(c,
|
1349 |
+
shape=(self.channels, self.image_size, self.image_size),
|
1350 |
+
batch_size=N)
|
1351 |
+
prog_row = self._get_denoise_row_from_list(progressives, desc="Progressive Generation")
|
1352 |
+
log["progressive_row"] = prog_row
|
1353 |
+
|
1354 |
+
if return_keys:
|
1355 |
+
if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
|
1356 |
+
return log
|
1357 |
+
else:
|
1358 |
+
return {key: log[key] for key in return_keys}
|
1359 |
+
return log
|
1360 |
+
|
1361 |
+
def configure_optimizers(self):
|
1362 |
+
lr = self.learning_rate
|
1363 |
+
params = list(self.model.parameters())
|
1364 |
+
if self.cond_stage_trainable:
|
1365 |
+
print(f"{self.__class__.__name__}: Also optimizing conditioner params!")
|
1366 |
+
params = params + list(self.cond_stage_model.parameters())
|
1367 |
+
if self.learn_logvar:
|
1368 |
+
print('Diffusion model optimizing logvar')
|
1369 |
+
params.append(self.logvar)
|
1370 |
+
opt = torch.optim.AdamW(params, lr=lr)
|
1371 |
+
if self.use_scheduler:
|
1372 |
+
assert 'target' in self.scheduler_config
|
1373 |
+
scheduler = instantiate_from_config(self.scheduler_config)
|
1374 |
+
|
1375 |
+
print("Setting up LambdaLR scheduler...")
|
1376 |
+
scheduler = [
|
1377 |
+
{
|
1378 |
+
'scheduler': LambdaLR(opt, lr_lambda=scheduler.schedule),
|
1379 |
+
'interval': 'step',
|
1380 |
+
'frequency': 1
|
1381 |
+
}]
|
1382 |
+
return [opt], scheduler
|
1383 |
+
return opt
|
1384 |
+
|
1385 |
+
# @torch.no_grad()
|
1386 |
+
def to_rgb(self, x):
|
1387 |
+
x = x.float()
|
1388 |
+
if not hasattr(self, "colorize"):
|
1389 |
+
self.colorize = torch.randn(3, x.shape[1], 1, 1).to(x)
|
1390 |
+
x = nn.functional.conv2d(x, weight=self.colorize)
|
1391 |
+
x = 2. * (x - x.min()) / (x.max() - x.min()) - 1.
|
1392 |
+
return x
|
1393 |
+
|
1394 |
+
|
1395 |
+
class DiffusionWrapper(pl.LightningModule):
|
1396 |
+
def __init__(self, diff_model_config, conditioning_key):
|
1397 |
+
super().__init__()
|
1398 |
+
self.diffusion_model = instantiate_from_config(diff_model_config)
|
1399 |
+
self.conditioning_key = conditioning_key
|
1400 |
+
assert self.conditioning_key in [None, 'concat', 'crossattn', 'hybrid', 'adm']
|
1401 |
+
|
1402 |
+
def forward(self, x, t, c_concat: list = None, c_crossattn: list = None):
|
1403 |
+
if self.conditioning_key is None:
|
1404 |
+
out = self.diffusion_model(x, t)
|
1405 |
+
elif self.conditioning_key == 'concat':
|
1406 |
+
xc = torch.cat([x] + c_concat, dim=1)
|
1407 |
+
out = self.diffusion_model(xc, t)
|
1408 |
+
elif self.conditioning_key == 'crossattn':
|
1409 |
+
cc = torch.cat(c_crossattn, 1)
|
1410 |
+
out = self.diffusion_model(x, t, context=cc)
|
1411 |
+
elif self.conditioning_key == 'hybrid':
|
1412 |
+
xc = torch.cat([x] + c_concat, dim=1)
|
1413 |
+
cc = torch.cat(c_crossattn, 1)
|
1414 |
+
out = self.diffusion_model(xc, t, context=cc)
|
1415 |
+
elif self.conditioning_key == 'adm':
|
1416 |
+
cc = c_crossattn[0]
|
1417 |
+
out = self.diffusion_model(x, t, y=cc)
|
1418 |
+
else:
|
1419 |
+
raise NotImplementedError()
|
1420 |
+
|
1421 |
+
return out
|
1422 |
+
|
1423 |
+
|
1424 |
+
class Layout2ImgDiffusion(LatentDiffusion):
|
1425 |
+
# TODO: move all layout-specific hacks to this class
|
1426 |
+
def __init__(self, cond_stage_key, *args, **kwargs):
|
1427 |
+
assert cond_stage_key == 'coordinates_bbox', 'Layout2ImgDiffusion only for cond_stage_key="coordinates_bbox"'
|
1428 |
+
super().__init__(cond_stage_key=cond_stage_key, *args, **kwargs)
|
1429 |
+
|
1430 |
+
def log_images(self, batch, N=8, *args, **kwargs):
|
1431 |
+
logs = super().log_images(batch=batch, N=N, *args, **kwargs)
|
1432 |
+
|
1433 |
+
key = 'train' if self.training else 'validation'
|
1434 |
+
dset = self.trainer.datamodule.datasets[key]
|
1435 |
+
mapper = dset.conditional_builders[self.cond_stage_key]
|
1436 |
+
|
1437 |
+
bbox_imgs = []
|
1438 |
+
map_fn = lambda catno: dset.get_textual_label(dset.get_category_id(catno))
|
1439 |
+
for tknzd_bbox in batch[self.cond_stage_key][:N]:
|
1440 |
+
bboximg = mapper.plot(tknzd_bbox.detach().cpu(), map_fn, (256, 256))
|
1441 |
+
bbox_imgs.append(bboximg)
|
1442 |
+
|
1443 |
+
cond_img = torch.stack(bbox_imgs, dim=0)
|
1444 |
+
logs['bbox_image'] = cond_img
|
1445 |
+
return logs
|
stable-diffusion/ldm/models/diffusion/dpm_solver/__init__.py
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
from .sampler import DPMSolverSampler
|
stable-diffusion/ldm/models/diffusion/dpm_solver/__pycache__/__init__.cpython-38.pyc
ADDED
Binary file (228 Bytes). View file
|
|
stable-diffusion/ldm/models/diffusion/dpm_solver/__pycache__/__init__.cpython-39.pyc
ADDED
Binary file (228 Bytes). View file
|
|
stable-diffusion/ldm/models/diffusion/dpm_solver/__pycache__/dpm_solver.cpython-38.pyc
ADDED
Binary file (51.4 kB). View file
|
|
stable-diffusion/ldm/models/diffusion/dpm_solver/__pycache__/dpm_solver.cpython-39.pyc
ADDED
Binary file (51.4 kB). View file
|
|
stable-diffusion/ldm/models/diffusion/dpm_solver/__pycache__/sampler.cpython-38.pyc
ADDED
Binary file (2.65 kB). View file
|
|
stable-diffusion/ldm/models/diffusion/dpm_solver/__pycache__/sampler.cpython-39.pyc
ADDED
Binary file (2.65 kB). View file
|
|
stable-diffusion/ldm/models/diffusion/dpm_solver/dpm_solver.py
ADDED
@@ -0,0 +1,1184 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn.functional as F
|
3 |
+
import math
|
4 |
+
|
5 |
+
|
6 |
+
class NoiseScheduleVP:
|
7 |
+
def __init__(
|
8 |
+
self,
|
9 |
+
schedule='discrete',
|
10 |
+
betas=None,
|
11 |
+
alphas_cumprod=None,
|
12 |
+
continuous_beta_0=0.1,
|
13 |
+
continuous_beta_1=20.,
|
14 |
+
):
|
15 |
+
"""Create a wrapper class for the forward SDE (VP type).
|
16 |
+
|
17 |
+
***
|
18 |
+
Update: We support discrete-time diffusion models by implementing a picewise linear interpolation for log_alpha_t.
|
19 |
+
We recommend to use schedule='discrete' for the discrete-time diffusion models, especially for high-resolution images.
|
20 |
+
***
|
21 |
+
|
22 |
+
The forward SDE ensures that the condition distribution q_{t|0}(x_t | x_0) = N ( alpha_t * x_0, sigma_t^2 * I ).
|
23 |
+
We further define lambda_t = log(alpha_t) - log(sigma_t), which is the half-logSNR (described in the DPM-Solver paper).
|
24 |
+
Therefore, we implement the functions for computing alpha_t, sigma_t and lambda_t. For t in [0, T], we have:
|
25 |
+
|
26 |
+
log_alpha_t = self.marginal_log_mean_coeff(t)
|
27 |
+
sigma_t = self.marginal_std(t)
|
28 |
+
lambda_t = self.marginal_lambda(t)
|
29 |
+
|
30 |
+
Moreover, as lambda(t) is an invertible function, we also support its inverse function:
|
31 |
+
|
32 |
+
t = self.inverse_lambda(lambda_t)
|
33 |
+
|
34 |
+
===============================================================
|
35 |
+
|
36 |
+
We support both discrete-time DPMs (trained on n = 0, 1, ..., N-1) and continuous-time DPMs (trained on t in [t_0, T]).
|
37 |
+
|
38 |
+
1. For discrete-time DPMs:
|
39 |
+
|
40 |
+
For discrete-time DPMs trained on n = 0, 1, ..., N-1, we convert the discrete steps to continuous time steps by:
|
41 |
+
t_i = (i + 1) / N
|
42 |
+
e.g. for N = 1000, we have t_0 = 1e-3 and T = t_{N-1} = 1.
|
43 |
+
We solve the corresponding diffusion ODE from time T = 1 to time t_0 = 1e-3.
|
44 |
+
|
45 |
+
Args:
|
46 |
+
betas: A `torch.Tensor`. The beta array for the discrete-time DPM. (See the original DDPM paper for details)
|
47 |
+
alphas_cumprod: A `torch.Tensor`. The cumprod alphas for the discrete-time DPM. (See the original DDPM paper for details)
|
48 |
+
|
49 |
+
Note that we always have alphas_cumprod = cumprod(betas). Therefore, we only need to set one of `betas` and `alphas_cumprod`.
|
50 |
+
|
51 |
+
**Important**: Please pay special attention for the args for `alphas_cumprod`:
|
52 |
+
The `alphas_cumprod` is the \hat{alpha_n} arrays in the notations of DDPM. Specifically, DDPMs assume that
|
53 |
+
q_{t_n | 0}(x_{t_n} | x_0) = N ( \sqrt{\hat{alpha_n}} * x_0, (1 - \hat{alpha_n}) * I ).
|
54 |
+
Therefore, the notation \hat{alpha_n} is different from the notation alpha_t in DPM-Solver. In fact, we have
|
55 |
+
alpha_{t_n} = \sqrt{\hat{alpha_n}},
|
56 |
+
and
|
57 |
+
log(alpha_{t_n}) = 0.5 * log(\hat{alpha_n}).
|
58 |
+
|
59 |
+
|
60 |
+
2. For continuous-time DPMs:
|
61 |
+
|
62 |
+
We support two types of VPSDEs: linear (DDPM) and cosine (improved-DDPM). The hyperparameters for the noise
|
63 |
+
schedule are the default settings in DDPM and improved-DDPM:
|
64 |
+
|
65 |
+
Args:
|
66 |
+
beta_min: A `float` number. The smallest beta for the linear schedule.
|
67 |
+
beta_max: A `float` number. The largest beta for the linear schedule.
|
68 |
+
cosine_s: A `float` number. The hyperparameter in the cosine schedule.
|
69 |
+
cosine_beta_max: A `float` number. The hyperparameter in the cosine schedule.
|
70 |
+
T: A `float` number. The ending time of the forward process.
|
71 |
+
|
72 |
+
===============================================================
|
73 |
+
|
74 |
+
Args:
|
75 |
+
schedule: A `str`. The noise schedule of the forward SDE. 'discrete' for discrete-time DPMs,
|
76 |
+
'linear' or 'cosine' for continuous-time DPMs.
|
77 |
+
Returns:
|
78 |
+
A wrapper object of the forward SDE (VP type).
|
79 |
+
|
80 |
+
===============================================================
|
81 |
+
|
82 |
+
Example:
|
83 |
+
|
84 |
+
# For discrete-time DPMs, given betas (the beta array for n = 0, 1, ..., N - 1):
|
85 |
+
>>> ns = NoiseScheduleVP('discrete', betas=betas)
|
86 |
+
|
87 |
+
# For discrete-time DPMs, given alphas_cumprod (the \hat{alpha_n} array for n = 0, 1, ..., N - 1):
|
88 |
+
>>> ns = NoiseScheduleVP('discrete', alphas_cumprod=alphas_cumprod)
|
89 |
+
|
90 |
+
# For continuous-time DPMs (VPSDE), linear schedule:
|
91 |
+
>>> ns = NoiseScheduleVP('linear', continuous_beta_0=0.1, continuous_beta_1=20.)
|
92 |
+
|
93 |
+
"""
|
94 |
+
|
95 |
+
if schedule not in ['discrete', 'linear', 'cosine']:
|
96 |
+
raise ValueError("Unsupported noise schedule {}. The schedule needs to be 'discrete' or 'linear' or 'cosine'".format(schedule))
|
97 |
+
|
98 |
+
self.schedule = schedule
|
99 |
+
if schedule == 'discrete':
|
100 |
+
if betas is not None:
|
101 |
+
log_alphas = 0.5 * torch.log(1 - betas).cumsum(dim=0)
|
102 |
+
else:
|
103 |
+
assert alphas_cumprod is not None
|
104 |
+
log_alphas = 0.5 * torch.log(alphas_cumprod)
|
105 |
+
self.total_N = len(log_alphas)
|
106 |
+
self.T = 1.
|
107 |
+
self.t_array = torch.linspace(0., 1., self.total_N + 1)[1:].reshape((1, -1))
|
108 |
+
self.log_alpha_array = log_alphas.reshape((1, -1,))
|
109 |
+
else:
|
110 |
+
self.total_N = 1000
|
111 |
+
self.beta_0 = continuous_beta_0
|
112 |
+
self.beta_1 = continuous_beta_1
|
113 |
+
self.cosine_s = 0.008
|
114 |
+
self.cosine_beta_max = 999.
|
115 |
+
self.cosine_t_max = math.atan(self.cosine_beta_max * (1. + self.cosine_s) / math.pi) * 2. * (1. + self.cosine_s) / math.pi - self.cosine_s
|
116 |
+
self.cosine_log_alpha_0 = math.log(math.cos(self.cosine_s / (1. + self.cosine_s) * math.pi / 2.))
|
117 |
+
self.schedule = schedule
|
118 |
+
if schedule == 'cosine':
|
119 |
+
# For the cosine schedule, T = 1 will have numerical issues. So we manually set the ending time T.
|
120 |
+
# Note that T = 0.9946 may be not the optimal setting. However, we find it works well.
|
121 |
+
self.T = 0.9946
|
122 |
+
else:
|
123 |
+
self.T = 1.
|
124 |
+
|
125 |
+
def marginal_log_mean_coeff(self, t):
|
126 |
+
"""
|
127 |
+
Compute log(alpha_t) of a given continuous-time label t in [0, T].
|
128 |
+
"""
|
129 |
+
if self.schedule == 'discrete':
|
130 |
+
return interpolate_fn(t.reshape((-1, 1)), self.t_array.to(t.device), self.log_alpha_array.to(t.device)).reshape((-1))
|
131 |
+
elif self.schedule == 'linear':
|
132 |
+
return -0.25 * t ** 2 * (self.beta_1 - self.beta_0) - 0.5 * t * self.beta_0
|
133 |
+
elif self.schedule == 'cosine':
|
134 |
+
log_alpha_fn = lambda s: torch.log(torch.cos((s + self.cosine_s) / (1. + self.cosine_s) * math.pi / 2.))
|
135 |
+
log_alpha_t = log_alpha_fn(t) - self.cosine_log_alpha_0
|
136 |
+
return log_alpha_t
|
137 |
+
|
138 |
+
def marginal_alpha(self, t):
|
139 |
+
"""
|
140 |
+
Compute alpha_t of a given continuous-time label t in [0, T].
|
141 |
+
"""
|
142 |
+
return torch.exp(self.marginal_log_mean_coeff(t))
|
143 |
+
|
144 |
+
def marginal_std(self, t):
|
145 |
+
"""
|
146 |
+
Compute sigma_t of a given continuous-time label t in [0, T].
|
147 |
+
"""
|
148 |
+
return torch.sqrt(1. - torch.exp(2. * self.marginal_log_mean_coeff(t)))
|
149 |
+
|
150 |
+
def marginal_lambda(self, t):
|
151 |
+
"""
|
152 |
+
Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T].
|
153 |
+
"""
|
154 |
+
log_mean_coeff = self.marginal_log_mean_coeff(t)
|
155 |
+
log_std = 0.5 * torch.log(1. - torch.exp(2. * log_mean_coeff))
|
156 |
+
return log_mean_coeff - log_std
|
157 |
+
|
158 |
+
def inverse_lambda(self, lamb):
|
159 |
+
"""
|
160 |
+
Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t.
|
161 |
+
"""
|
162 |
+
if self.schedule == 'linear':
|
163 |
+
tmp = 2. * (self.beta_1 - self.beta_0) * torch.logaddexp(-2. * lamb, torch.zeros((1,)).to(lamb))
|
164 |
+
Delta = self.beta_0**2 + tmp
|
165 |
+
return tmp / (torch.sqrt(Delta) + self.beta_0) / (self.beta_1 - self.beta_0)
|
166 |
+
elif self.schedule == 'discrete':
|
167 |
+
log_alpha = -0.5 * torch.logaddexp(torch.zeros((1,)).to(lamb.device), -2. * lamb)
|
168 |
+
t = interpolate_fn(log_alpha.reshape((-1, 1)), torch.flip(self.log_alpha_array.to(lamb.device), [1]), torch.flip(self.t_array.to(lamb.device), [1]))
|
169 |
+
return t.reshape((-1,))
|
170 |
+
else:
|
171 |
+
log_alpha = -0.5 * torch.logaddexp(-2. * lamb, torch.zeros((1,)).to(lamb))
|
172 |
+
t_fn = lambda log_alpha_t: torch.arccos(torch.exp(log_alpha_t + self.cosine_log_alpha_0)) * 2. * (1. + self.cosine_s) / math.pi - self.cosine_s
|
173 |
+
t = t_fn(log_alpha)
|
174 |
+
return t
|
175 |
+
|
176 |
+
|
177 |
+
def model_wrapper(
|
178 |
+
model,
|
179 |
+
noise_schedule,
|
180 |
+
model_type="noise",
|
181 |
+
model_kwargs={},
|
182 |
+
guidance_type="uncond",
|
183 |
+
condition=None,
|
184 |
+
unconditional_condition=None,
|
185 |
+
guidance_scale=1.,
|
186 |
+
classifier_fn=None,
|
187 |
+
classifier_kwargs={},
|
188 |
+
):
|
189 |
+
"""Create a wrapper function for the noise prediction model.
|
190 |
+
|
191 |
+
DPM-Solver needs to solve the continuous-time diffusion ODEs. For DPMs trained on discrete-time labels, we need to
|
192 |
+
firstly wrap the model function to a noise prediction model that accepts the continuous time as the input.
|
193 |
+
|
194 |
+
We support four types of the diffusion model by setting `model_type`:
|
195 |
+
|
196 |
+
1. "noise": noise prediction model. (Trained by predicting noise).
|
197 |
+
|
198 |
+
2. "x_start": data prediction model. (Trained by predicting the data x_0 at time 0).
|
199 |
+
|
200 |
+
3. "v": velocity prediction model. (Trained by predicting the velocity).
|
201 |
+
The "v" prediction is derivation detailed in Appendix D of [1], and is used in Imagen-Video [2].
|
202 |
+
|
203 |
+
[1] Salimans, Tim, and Jonathan Ho. "Progressive distillation for fast sampling of diffusion models."
|
204 |
+
arXiv preprint arXiv:2202.00512 (2022).
|
205 |
+
[2] Ho, Jonathan, et al. "Imagen Video: High Definition Video Generation with Diffusion Models."
|
206 |
+
arXiv preprint arXiv:2210.02303 (2022).
|
207 |
+
|
208 |
+
4. "score": marginal score function. (Trained by denoising score matching).
|
209 |
+
Note that the score function and the noise prediction model follows a simple relationship:
|
210 |
+
```
|
211 |
+
noise(x_t, t) = -sigma_t * score(x_t, t)
|
212 |
+
```
|
213 |
+
|
214 |
+
We support three types of guided sampling by DPMs by setting `guidance_type`:
|
215 |
+
1. "uncond": unconditional sampling by DPMs.
|
216 |
+
The input `model` has the following format:
|
217 |
+
``
|
218 |
+
model(x, t_input, **model_kwargs) -> noise | x_start | v | score
|
219 |
+
``
|
220 |
+
|
221 |
+
2. "classifier": classifier guidance sampling [3] by DPMs and another classifier.
|
222 |
+
The input `model` has the following format:
|
223 |
+
``
|
224 |
+
model(x, t_input, **model_kwargs) -> noise | x_start | v | score
|
225 |
+
``
|
226 |
+
|
227 |
+
The input `classifier_fn` has the following format:
|
228 |
+
``
|
229 |
+
classifier_fn(x, t_input, cond, **classifier_kwargs) -> logits(x, t_input, cond)
|
230 |
+
``
|
231 |
+
|
232 |
+
[3] P. Dhariwal and A. Q. Nichol, "Diffusion models beat GANs on image synthesis,"
|
233 |
+
in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 8780-8794.
|
234 |
+
|
235 |
+
3. "classifier-free": classifier-free guidance sampling by conditional DPMs.
|
236 |
+
The input `model` has the following format:
|
237 |
+
``
|
238 |
+
model(x, t_input, cond, **model_kwargs) -> noise | x_start | v | score
|
239 |
+
``
|
240 |
+
And if cond == `unconditional_condition`, the model output is the unconditional DPM output.
|
241 |
+
|
242 |
+
[4] Ho, Jonathan, and Tim Salimans. "Classifier-free diffusion guidance."
|
243 |
+
arXiv preprint arXiv:2207.12598 (2022).
|
244 |
+
|
245 |
+
|
246 |
+
The `t_input` is the time label of the model, which may be discrete-time labels (i.e. 0 to 999)
|
247 |
+
or continuous-time labels (i.e. epsilon to T).
|
248 |
+
|
249 |
+
We wrap the model function to accept only `x` and `t_continuous` as inputs, and outputs the predicted noise:
|
250 |
+
``
|
251 |
+
def model_fn(x, t_continuous) -> noise:
|
252 |
+
t_input = get_model_input_time(t_continuous)
|
253 |
+
return noise_pred(model, x, t_input, **model_kwargs)
|
254 |
+
``
|
255 |
+
where `t_continuous` is the continuous time labels (i.e. epsilon to T). And we use `model_fn` for DPM-Solver.
|
256 |
+
|
257 |
+
===============================================================
|
258 |
+
|
259 |
+
Args:
|
260 |
+
model: A diffusion model with the corresponding format described above.
|
261 |
+
noise_schedule: A noise schedule object, such as NoiseScheduleVP.
|
262 |
+
model_type: A `str`. The parameterization type of the diffusion model.
|
263 |
+
"noise" or "x_start" or "v" or "score".
|
264 |
+
model_kwargs: A `dict`. A dict for the other inputs of the model function.
|
265 |
+
guidance_type: A `str`. The type of the guidance for sampling.
|
266 |
+
"uncond" or "classifier" or "classifier-free".
|
267 |
+
condition: A pytorch tensor. The condition for the guided sampling.
|
268 |
+
Only used for "classifier" or "classifier-free" guidance type.
|
269 |
+
unconditional_condition: A pytorch tensor. The condition for the unconditional sampling.
|
270 |
+
Only used for "classifier-free" guidance type.
|
271 |
+
guidance_scale: A `float`. The scale for the guided sampling.
|
272 |
+
classifier_fn: A classifier function. Only used for the classifier guidance.
|
273 |
+
classifier_kwargs: A `dict`. A dict for the other inputs of the classifier function.
|
274 |
+
Returns:
|
275 |
+
A noise prediction model that accepts the noised data and the continuous time as the inputs.
|
276 |
+
"""
|
277 |
+
|
278 |
+
def get_model_input_time(t_continuous):
|
279 |
+
"""
|
280 |
+
Convert the continuous-time `t_continuous` (in [epsilon, T]) to the model input time.
|
281 |
+
For discrete-time DPMs, we convert `t_continuous` in [1 / N, 1] to `t_input` in [0, 1000 * (N - 1) / N].
|
282 |
+
For continuous-time DPMs, we just use `t_continuous`.
|
283 |
+
"""
|
284 |
+
if noise_schedule.schedule == 'discrete':
|
285 |
+
return (t_continuous - 1. / noise_schedule.total_N) * 1000.
|
286 |
+
else:
|
287 |
+
return t_continuous
|
288 |
+
|
289 |
+
def noise_pred_fn(x, t_continuous, cond=None):
|
290 |
+
if t_continuous.reshape((-1,)).shape[0] == 1:
|
291 |
+
t_continuous = t_continuous.expand((x.shape[0]))
|
292 |
+
t_input = get_model_input_time(t_continuous)
|
293 |
+
if cond is None:
|
294 |
+
output = model(x, t_input, **model_kwargs)
|
295 |
+
else:
|
296 |
+
output = model(x, t_input, cond, **model_kwargs)
|
297 |
+
if model_type == "noise":
|
298 |
+
return output
|
299 |
+
elif model_type == "x_start":
|
300 |
+
alpha_t, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous)
|
301 |
+
dims = x.dim()
|
302 |
+
return (x - expand_dims(alpha_t, dims) * output) / expand_dims(sigma_t, dims)
|
303 |
+
elif model_type == "v":
|
304 |
+
alpha_t, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous)
|
305 |
+
dims = x.dim()
|
306 |
+
return expand_dims(alpha_t, dims) * output + expand_dims(sigma_t, dims) * x
|
307 |
+
elif model_type == "score":
|
308 |
+
sigma_t = noise_schedule.marginal_std(t_continuous)
|
309 |
+
dims = x.dim()
|
310 |
+
return -expand_dims(sigma_t, dims) * output
|
311 |
+
|
312 |
+
def cond_grad_fn(x, t_input):
|
313 |
+
"""
|
314 |
+
Compute the gradient of the classifier, i.e. nabla_{x} log p_t(cond | x_t).
|
315 |
+
"""
|
316 |
+
with torch.enable_grad():
|
317 |
+
x_in = x.detach().requires_grad_(True)
|
318 |
+
log_prob = classifier_fn(x_in, t_input, condition, **classifier_kwargs)
|
319 |
+
return torch.autograd.grad(log_prob.sum(), x_in)[0]
|
320 |
+
|
321 |
+
def model_fn(x, t_continuous):
|
322 |
+
"""
|
323 |
+
The noise predicition model function that is used for DPM-Solver.
|
324 |
+
"""
|
325 |
+
if t_continuous.reshape((-1,)).shape[0] == 1:
|
326 |
+
t_continuous = t_continuous.expand((x.shape[0]))
|
327 |
+
if guidance_type == "uncond":
|
328 |
+
return noise_pred_fn(x, t_continuous)
|
329 |
+
elif guidance_type == "classifier":
|
330 |
+
assert classifier_fn is not None
|
331 |
+
t_input = get_model_input_time(t_continuous)
|
332 |
+
cond_grad = cond_grad_fn(x, t_input)
|
333 |
+
sigma_t = noise_schedule.marginal_std(t_continuous)
|
334 |
+
noise = noise_pred_fn(x, t_continuous)
|
335 |
+
return noise - guidance_scale * expand_dims(sigma_t, dims=cond_grad.dim()) * cond_grad
|
336 |
+
elif guidance_type == "classifier-free":
|
337 |
+
if guidance_scale == 1. or unconditional_condition is None:
|
338 |
+
return noise_pred_fn(x, t_continuous, cond=condition)
|
339 |
+
else:
|
340 |
+
x_in = torch.cat([x] * 2)
|
341 |
+
t_in = torch.cat([t_continuous] * 2)
|
342 |
+
c_in = torch.cat([unconditional_condition, condition])
|
343 |
+
noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in).chunk(2)
|
344 |
+
return noise_uncond + guidance_scale * (noise - noise_uncond)
|
345 |
+
|
346 |
+
assert model_type in ["noise", "x_start", "v"]
|
347 |
+
assert guidance_type in ["uncond", "classifier", "classifier-free"]
|
348 |
+
return model_fn
|
349 |
+
|
350 |
+
|
351 |
+
class DPM_Solver:
|
352 |
+
def __init__(self, model_fn, noise_schedule, predict_x0=False, thresholding=False, max_val=1.):
|
353 |
+
"""Construct a DPM-Solver.
|
354 |
+
|
355 |
+
We support both the noise prediction model ("predicting epsilon") and the data prediction model ("predicting x0").
|
356 |
+
If `predict_x0` is False, we use the solver for the noise prediction model (DPM-Solver).
|
357 |
+
If `predict_x0` is True, we use the solver for the data prediction model (DPM-Solver++).
|
358 |
+
In such case, we further support the "dynamic thresholding" in [1] when `thresholding` is True.
|
359 |
+
The "dynamic thresholding" can greatly improve the sample quality for pixel-space DPMs with large guidance scales.
|
360 |
+
|
361 |
+
Args:
|
362 |
+
model_fn: A noise prediction model function which accepts the continuous-time input (t in [epsilon, T]):
|
363 |
+
``
|
364 |
+
def model_fn(x, t_continuous):
|
365 |
+
return noise
|
366 |
+
``
|
367 |
+
noise_schedule: A noise schedule object, such as NoiseScheduleVP.
|
368 |
+
predict_x0: A `bool`. If true, use the data prediction model; else, use the noise prediction model.
|
369 |
+
thresholding: A `bool`. Valid when `predict_x0` is True. Whether to use the "dynamic thresholding" in [1].
|
370 |
+
max_val: A `float`. Valid when both `predict_x0` and `thresholding` are True. The max value for thresholding.
|
371 |
+
|
372 |
+
[1] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022b.
|
373 |
+
"""
|
374 |
+
self.model = model_fn
|
375 |
+
self.noise_schedule = noise_schedule
|
376 |
+
self.predict_x0 = predict_x0
|
377 |
+
self.thresholding = thresholding
|
378 |
+
self.max_val = max_val
|
379 |
+
|
380 |
+
def noise_prediction_fn(self, x, t):
|
381 |
+
"""
|
382 |
+
Return the noise prediction model.
|
383 |
+
"""
|
384 |
+
return self.model(x, t)
|
385 |
+
|
386 |
+
def data_prediction_fn(self, x, t):
|
387 |
+
"""
|
388 |
+
Return the data prediction model (with thresholding).
|
389 |
+
"""
|
390 |
+
noise = self.noise_prediction_fn(x, t)
|
391 |
+
dims = x.dim()
|
392 |
+
alpha_t, sigma_t = self.noise_schedule.marginal_alpha(t), self.noise_schedule.marginal_std(t)
|
393 |
+
x0 = (x - expand_dims(sigma_t, dims) * noise) / expand_dims(alpha_t, dims)
|
394 |
+
if self.thresholding:
|
395 |
+
p = 0.995 # A hyperparameter in the paper of "Imagen" [1].
|
396 |
+
s = torch.quantile(torch.abs(x0).reshape((x0.shape[0], -1)), p, dim=1)
|
397 |
+
s = expand_dims(torch.maximum(s, self.max_val * torch.ones_like(s).to(s.device)), dims)
|
398 |
+
x0 = torch.clamp(x0, -s, s) / s
|
399 |
+
return x0
|
400 |
+
|
401 |
+
def model_fn(self, x, t):
|
402 |
+
"""
|
403 |
+
Convert the model to the noise prediction model or the data prediction model.
|
404 |
+
"""
|
405 |
+
if self.predict_x0:
|
406 |
+
return self.data_prediction_fn(x, t)
|
407 |
+
else:
|
408 |
+
return self.noise_prediction_fn(x, t)
|
409 |
+
|
410 |
+
def get_time_steps(self, skip_type, t_T, t_0, N, device):
|
411 |
+
"""Compute the intermediate time steps for sampling.
|
412 |
+
|
413 |
+
Args:
|
414 |
+
skip_type: A `str`. The type for the spacing of the time steps. We support three types:
|
415 |
+
- 'logSNR': uniform logSNR for the time steps.
|
416 |
+
- 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.)
|
417 |
+
- 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.)
|
418 |
+
t_T: A `float`. The starting time of the sampling (default is T).
|
419 |
+
t_0: A `float`. The ending time of the sampling (default is epsilon).
|
420 |
+
N: A `int`. The total number of the spacing of the time steps.
|
421 |
+
device: A torch device.
|
422 |
+
Returns:
|
423 |
+
A pytorch tensor of the time steps, with the shape (N + 1,).
|
424 |
+
"""
|
425 |
+
if skip_type == 'logSNR':
|
426 |
+
lambda_T = self.noise_schedule.marginal_lambda(torch.tensor(t_T).to(device))
|
427 |
+
lambda_0 = self.noise_schedule.marginal_lambda(torch.tensor(t_0).to(device))
|
428 |
+
logSNR_steps = torch.linspace(lambda_T.cpu().item(), lambda_0.cpu().item(), N + 1).to(device)
|
429 |
+
return self.noise_schedule.inverse_lambda(logSNR_steps)
|
430 |
+
elif skip_type == 'time_uniform':
|
431 |
+
return torch.linspace(t_T, t_0, N + 1).to(device)
|
432 |
+
elif skip_type == 'time_quadratic':
|
433 |
+
t_order = 2
|
434 |
+
t = torch.linspace(t_T**(1. / t_order), t_0**(1. / t_order), N + 1).pow(t_order).to(device)
|
435 |
+
return t
|
436 |
+
else:
|
437 |
+
raise ValueError("Unsupported skip_type {}, need to be 'logSNR' or 'time_uniform' or 'time_quadratic'".format(skip_type))
|
438 |
+
|
439 |
+
def get_orders_and_timesteps_for_singlestep_solver(self, steps, order, skip_type, t_T, t_0, device):
|
440 |
+
"""
|
441 |
+
Get the order of each step for sampling by the singlestep DPM-Solver.
|
442 |
+
|
443 |
+
We combine both DPM-Solver-1,2,3 to use all the function evaluations, which is named as "DPM-Solver-fast".
|
444 |
+
Given a fixed number of function evaluations by `steps`, the sampling procedure by DPM-Solver-fast is:
|
445 |
+
- If order == 1:
|
446 |
+
We take `steps` of DPM-Solver-1 (i.e. DDIM).
|
447 |
+
- If order == 2:
|
448 |
+
- Denote K = (steps // 2). We take K or (K + 1) intermediate time steps for sampling.
|
449 |
+
- If steps % 2 == 0, we use K steps of DPM-Solver-2.
|
450 |
+
- If steps % 2 == 1, we use K steps of DPM-Solver-2 and 1 step of DPM-Solver-1.
|
451 |
+
- If order == 3:
|
452 |
+
- Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling.
|
453 |
+
- If steps % 3 == 0, we use (K - 2) steps of DPM-Solver-3, and 1 step of DPM-Solver-2 and 1 step of DPM-Solver-1.
|
454 |
+
- If steps % 3 == 1, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-1.
|
455 |
+
- If steps % 3 == 2, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-2.
|
456 |
+
|
457 |
+
============================================
|
458 |
+
Args:
|
459 |
+
order: A `int`. The max order for the solver (2 or 3).
|
460 |
+
steps: A `int`. The total number of function evaluations (NFE).
|
461 |
+
skip_type: A `str`. The type for the spacing of the time steps. We support three types:
|
462 |
+
- 'logSNR': uniform logSNR for the time steps.
|
463 |
+
- 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.)
|
464 |
+
- 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.)
|
465 |
+
t_T: A `float`. The starting time of the sampling (default is T).
|
466 |
+
t_0: A `float`. The ending time of the sampling (default is epsilon).
|
467 |
+
device: A torch device.
|
468 |
+
Returns:
|
469 |
+
orders: A list of the solver order of each step.
|
470 |
+
"""
|
471 |
+
if order == 3:
|
472 |
+
K = steps // 3 + 1
|
473 |
+
if steps % 3 == 0:
|
474 |
+
orders = [3,] * (K - 2) + [2, 1]
|
475 |
+
elif steps % 3 == 1:
|
476 |
+
orders = [3,] * (K - 1) + [1]
|
477 |
+
else:
|
478 |
+
orders = [3,] * (K - 1) + [2]
|
479 |
+
elif order == 2:
|
480 |
+
if steps % 2 == 0:
|
481 |
+
K = steps // 2
|
482 |
+
orders = [2,] * K
|
483 |
+
else:
|
484 |
+
K = steps // 2 + 1
|
485 |
+
orders = [2,] * (K - 1) + [1]
|
486 |
+
elif order == 1:
|
487 |
+
K = 1
|
488 |
+
orders = [1,] * steps
|
489 |
+
else:
|
490 |
+
raise ValueError("'order' must be '1' or '2' or '3'.")
|
491 |
+
if skip_type == 'logSNR':
|
492 |
+
# To reproduce the results in DPM-Solver paper
|
493 |
+
timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, K, device)
|
494 |
+
else:
|
495 |
+
timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, steps, device)[torch.cumsum(torch.tensor([0,] + orders)).to(device)]
|
496 |
+
return timesteps_outer, orders
|
497 |
+
|
498 |
+
def denoise_to_zero_fn(self, x, s):
|
499 |
+
"""
|
500 |
+
Denoise at the final step, which is equivalent to solve the ODE from lambda_s to infty by first-order discretization.
|
501 |
+
"""
|
502 |
+
return self.data_prediction_fn(x, s)
|
503 |
+
|
504 |
+
def dpm_solver_first_update(self, x, s, t, model_s=None, return_intermediate=False):
|
505 |
+
"""
|
506 |
+
DPM-Solver-1 (equivalent to DDIM) from time `s` to time `t`.
|
507 |
+
|
508 |
+
Args:
|
509 |
+
x: A pytorch tensor. The initial value at time `s`.
|
510 |
+
s: A pytorch tensor. The starting time, with the shape (x.shape[0],).
|
511 |
+
t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
|
512 |
+
model_s: A pytorch tensor. The model function evaluated at time `s`.
|
513 |
+
If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
|
514 |
+
return_intermediate: A `bool`. If true, also return the model value at time `s`.
|
515 |
+
Returns:
|
516 |
+
x_t: A pytorch tensor. The approximated solution at time `t`.
|
517 |
+
"""
|
518 |
+
ns = self.noise_schedule
|
519 |
+
dims = x.dim()
|
520 |
+
lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
|
521 |
+
h = lambda_t - lambda_s
|
522 |
+
log_alpha_s, log_alpha_t = ns.marginal_log_mean_coeff(s), ns.marginal_log_mean_coeff(t)
|
523 |
+
sigma_s, sigma_t = ns.marginal_std(s), ns.marginal_std(t)
|
524 |
+
alpha_t = torch.exp(log_alpha_t)
|
525 |
+
|
526 |
+
if self.predict_x0:
|
527 |
+
phi_1 = torch.expm1(-h)
|
528 |
+
if model_s is None:
|
529 |
+
model_s = self.model_fn(x, s)
|
530 |
+
x_t = (
|
531 |
+
expand_dims(sigma_t / sigma_s, dims) * x
|
532 |
+
- expand_dims(alpha_t * phi_1, dims) * model_s
|
533 |
+
)
|
534 |
+
if return_intermediate:
|
535 |
+
return x_t, {'model_s': model_s}
|
536 |
+
else:
|
537 |
+
return x_t
|
538 |
+
else:
|
539 |
+
phi_1 = torch.expm1(h)
|
540 |
+
if model_s is None:
|
541 |
+
model_s = self.model_fn(x, s)
|
542 |
+
x_t = (
|
543 |
+
expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
|
544 |
+
- expand_dims(sigma_t * phi_1, dims) * model_s
|
545 |
+
)
|
546 |
+
if return_intermediate:
|
547 |
+
return x_t, {'model_s': model_s}
|
548 |
+
else:
|
549 |
+
return x_t
|
550 |
+
|
551 |
+
def singlestep_dpm_solver_second_update(self, x, s, t, r1=0.5, model_s=None, return_intermediate=False, solver_type='dpm_solver'):
|
552 |
+
"""
|
553 |
+
Singlestep solver DPM-Solver-2 from time `s` to time `t`.
|
554 |
+
|
555 |
+
Args:
|
556 |
+
x: A pytorch tensor. The initial value at time `s`.
|
557 |
+
s: A pytorch tensor. The starting time, with the shape (x.shape[0],).
|
558 |
+
t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
|
559 |
+
r1: A `float`. The hyperparameter of the second-order solver.
|
560 |
+
model_s: A pytorch tensor. The model function evaluated at time `s`.
|
561 |
+
If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
|
562 |
+
return_intermediate: A `bool`. If true, also return the model value at time `s` and `s1` (the intermediate time).
|
563 |
+
solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
|
564 |
+
The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
|
565 |
+
Returns:
|
566 |
+
x_t: A pytorch tensor. The approximated solution at time `t`.
|
567 |
+
"""
|
568 |
+
if solver_type not in ['dpm_solver', 'taylor']:
|
569 |
+
raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type))
|
570 |
+
if r1 is None:
|
571 |
+
r1 = 0.5
|
572 |
+
ns = self.noise_schedule
|
573 |
+
dims = x.dim()
|
574 |
+
lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
|
575 |
+
h = lambda_t - lambda_s
|
576 |
+
lambda_s1 = lambda_s + r1 * h
|
577 |
+
s1 = ns.inverse_lambda(lambda_s1)
|
578 |
+
log_alpha_s, log_alpha_s1, log_alpha_t = ns.marginal_log_mean_coeff(s), ns.marginal_log_mean_coeff(s1), ns.marginal_log_mean_coeff(t)
|
579 |
+
sigma_s, sigma_s1, sigma_t = ns.marginal_std(s), ns.marginal_std(s1), ns.marginal_std(t)
|
580 |
+
alpha_s1, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_t)
|
581 |
+
|
582 |
+
if self.predict_x0:
|
583 |
+
phi_11 = torch.expm1(-r1 * h)
|
584 |
+
phi_1 = torch.expm1(-h)
|
585 |
+
|
586 |
+
if model_s is None:
|
587 |
+
model_s = self.model_fn(x, s)
|
588 |
+
x_s1 = (
|
589 |
+
expand_dims(sigma_s1 / sigma_s, dims) * x
|
590 |
+
- expand_dims(alpha_s1 * phi_11, dims) * model_s
|
591 |
+
)
|
592 |
+
model_s1 = self.model_fn(x_s1, s1)
|
593 |
+
if solver_type == 'dpm_solver':
|
594 |
+
x_t = (
|
595 |
+
expand_dims(sigma_t / sigma_s, dims) * x
|
596 |
+
- expand_dims(alpha_t * phi_1, dims) * model_s
|
597 |
+
- (0.5 / r1) * expand_dims(alpha_t * phi_1, dims) * (model_s1 - model_s)
|
598 |
+
)
|
599 |
+
elif solver_type == 'taylor':
|
600 |
+
x_t = (
|
601 |
+
expand_dims(sigma_t / sigma_s, dims) * x
|
602 |
+
- expand_dims(alpha_t * phi_1, dims) * model_s
|
603 |
+
+ (1. / r1) * expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * (model_s1 - model_s)
|
604 |
+
)
|
605 |
+
else:
|
606 |
+
phi_11 = torch.expm1(r1 * h)
|
607 |
+
phi_1 = torch.expm1(h)
|
608 |
+
|
609 |
+
if model_s is None:
|
610 |
+
model_s = self.model_fn(x, s)
|
611 |
+
x_s1 = (
|
612 |
+
expand_dims(torch.exp(log_alpha_s1 - log_alpha_s), dims) * x
|
613 |
+
- expand_dims(sigma_s1 * phi_11, dims) * model_s
|
614 |
+
)
|
615 |
+
model_s1 = self.model_fn(x_s1, s1)
|
616 |
+
if solver_type == 'dpm_solver':
|
617 |
+
x_t = (
|
618 |
+
expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
|
619 |
+
- expand_dims(sigma_t * phi_1, dims) * model_s
|
620 |
+
- (0.5 / r1) * expand_dims(sigma_t * phi_1, dims) * (model_s1 - model_s)
|
621 |
+
)
|
622 |
+
elif solver_type == 'taylor':
|
623 |
+
x_t = (
|
624 |
+
expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
|
625 |
+
- expand_dims(sigma_t * phi_1, dims) * model_s
|
626 |
+
- (1. / r1) * expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * (model_s1 - model_s)
|
627 |
+
)
|
628 |
+
if return_intermediate:
|
629 |
+
return x_t, {'model_s': model_s, 'model_s1': model_s1}
|
630 |
+
else:
|
631 |
+
return x_t
|
632 |
+
|
633 |
+
def singlestep_dpm_solver_third_update(self, x, s, t, r1=1./3., r2=2./3., model_s=None, model_s1=None, return_intermediate=False, solver_type='dpm_solver'):
|
634 |
+
"""
|
635 |
+
Singlestep solver DPM-Solver-3 from time `s` to time `t`.
|
636 |
+
|
637 |
+
Args:
|
638 |
+
x: A pytorch tensor. The initial value at time `s`.
|
639 |
+
s: A pytorch tensor. The starting time, with the shape (x.shape[0],).
|
640 |
+
t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
|
641 |
+
r1: A `float`. The hyperparameter of the third-order solver.
|
642 |
+
r2: A `float`. The hyperparameter of the third-order solver.
|
643 |
+
model_s: A pytorch tensor. The model function evaluated at time `s`.
|
644 |
+
If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
|
645 |
+
model_s1: A pytorch tensor. The model function evaluated at time `s1` (the intermediate time given by `r1`).
|
646 |
+
If `model_s1` is None, we evaluate the model at `s1`; otherwise we directly use it.
|
647 |
+
return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times).
|
648 |
+
solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
|
649 |
+
The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
|
650 |
+
Returns:
|
651 |
+
x_t: A pytorch tensor. The approximated solution at time `t`.
|
652 |
+
"""
|
653 |
+
if solver_type not in ['dpm_solver', 'taylor']:
|
654 |
+
raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type))
|
655 |
+
if r1 is None:
|
656 |
+
r1 = 1. / 3.
|
657 |
+
if r2 is None:
|
658 |
+
r2 = 2. / 3.
|
659 |
+
ns = self.noise_schedule
|
660 |
+
dims = x.dim()
|
661 |
+
lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
|
662 |
+
h = lambda_t - lambda_s
|
663 |
+
lambda_s1 = lambda_s + r1 * h
|
664 |
+
lambda_s2 = lambda_s + r2 * h
|
665 |
+
s1 = ns.inverse_lambda(lambda_s1)
|
666 |
+
s2 = ns.inverse_lambda(lambda_s2)
|
667 |
+
log_alpha_s, log_alpha_s1, log_alpha_s2, log_alpha_t = ns.marginal_log_mean_coeff(s), ns.marginal_log_mean_coeff(s1), ns.marginal_log_mean_coeff(s2), ns.marginal_log_mean_coeff(t)
|
668 |
+
sigma_s, sigma_s1, sigma_s2, sigma_t = ns.marginal_std(s), ns.marginal_std(s1), ns.marginal_std(s2), ns.marginal_std(t)
|
669 |
+
alpha_s1, alpha_s2, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_s2), torch.exp(log_alpha_t)
|
670 |
+
|
671 |
+
if self.predict_x0:
|
672 |
+
phi_11 = torch.expm1(-r1 * h)
|
673 |
+
phi_12 = torch.expm1(-r2 * h)
|
674 |
+
phi_1 = torch.expm1(-h)
|
675 |
+
phi_22 = torch.expm1(-r2 * h) / (r2 * h) + 1.
|
676 |
+
phi_2 = phi_1 / h + 1.
|
677 |
+
phi_3 = phi_2 / h - 0.5
|
678 |
+
|
679 |
+
if model_s is None:
|
680 |
+
model_s = self.model_fn(x, s)
|
681 |
+
if model_s1 is None:
|
682 |
+
x_s1 = (
|
683 |
+
expand_dims(sigma_s1 / sigma_s, dims) * x
|
684 |
+
- expand_dims(alpha_s1 * phi_11, dims) * model_s
|
685 |
+
)
|
686 |
+
model_s1 = self.model_fn(x_s1, s1)
|
687 |
+
x_s2 = (
|
688 |
+
expand_dims(sigma_s2 / sigma_s, dims) * x
|
689 |
+
- expand_dims(alpha_s2 * phi_12, dims) * model_s
|
690 |
+
+ r2 / r1 * expand_dims(alpha_s2 * phi_22, dims) * (model_s1 - model_s)
|
691 |
+
)
|
692 |
+
model_s2 = self.model_fn(x_s2, s2)
|
693 |
+
if solver_type == 'dpm_solver':
|
694 |
+
x_t = (
|
695 |
+
expand_dims(sigma_t / sigma_s, dims) * x
|
696 |
+
- expand_dims(alpha_t * phi_1, dims) * model_s
|
697 |
+
+ (1. / r2) * expand_dims(alpha_t * phi_2, dims) * (model_s2 - model_s)
|
698 |
+
)
|
699 |
+
elif solver_type == 'taylor':
|
700 |
+
D1_0 = (1. / r1) * (model_s1 - model_s)
|
701 |
+
D1_1 = (1. / r2) * (model_s2 - model_s)
|
702 |
+
D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1)
|
703 |
+
D2 = 2. * (D1_1 - D1_0) / (r2 - r1)
|
704 |
+
x_t = (
|
705 |
+
expand_dims(sigma_t / sigma_s, dims) * x
|
706 |
+
- expand_dims(alpha_t * phi_1, dims) * model_s
|
707 |
+
+ expand_dims(alpha_t * phi_2, dims) * D1
|
708 |
+
- expand_dims(alpha_t * phi_3, dims) * D2
|
709 |
+
)
|
710 |
+
else:
|
711 |
+
phi_11 = torch.expm1(r1 * h)
|
712 |
+
phi_12 = torch.expm1(r2 * h)
|
713 |
+
phi_1 = torch.expm1(h)
|
714 |
+
phi_22 = torch.expm1(r2 * h) / (r2 * h) - 1.
|
715 |
+
phi_2 = phi_1 / h - 1.
|
716 |
+
phi_3 = phi_2 / h - 0.5
|
717 |
+
|
718 |
+
if model_s is None:
|
719 |
+
model_s = self.model_fn(x, s)
|
720 |
+
if model_s1 is None:
|
721 |
+
x_s1 = (
|
722 |
+
expand_dims(torch.exp(log_alpha_s1 - log_alpha_s), dims) * x
|
723 |
+
- expand_dims(sigma_s1 * phi_11, dims) * model_s
|
724 |
+
)
|
725 |
+
model_s1 = self.model_fn(x_s1, s1)
|
726 |
+
x_s2 = (
|
727 |
+
expand_dims(torch.exp(log_alpha_s2 - log_alpha_s), dims) * x
|
728 |
+
- expand_dims(sigma_s2 * phi_12, dims) * model_s
|
729 |
+
- r2 / r1 * expand_dims(sigma_s2 * phi_22, dims) * (model_s1 - model_s)
|
730 |
+
)
|
731 |
+
model_s2 = self.model_fn(x_s2, s2)
|
732 |
+
if solver_type == 'dpm_solver':
|
733 |
+
x_t = (
|
734 |
+
expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
|
735 |
+
- expand_dims(sigma_t * phi_1, dims) * model_s
|
736 |
+
- (1. / r2) * expand_dims(sigma_t * phi_2, dims) * (model_s2 - model_s)
|
737 |
+
)
|
738 |
+
elif solver_type == 'taylor':
|
739 |
+
D1_0 = (1. / r1) * (model_s1 - model_s)
|
740 |
+
D1_1 = (1. / r2) * (model_s2 - model_s)
|
741 |
+
D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1)
|
742 |
+
D2 = 2. * (D1_1 - D1_0) / (r2 - r1)
|
743 |
+
x_t = (
|
744 |
+
expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
|
745 |
+
- expand_dims(sigma_t * phi_1, dims) * model_s
|
746 |
+
- expand_dims(sigma_t * phi_2, dims) * D1
|
747 |
+
- expand_dims(sigma_t * phi_3, dims) * D2
|
748 |
+
)
|
749 |
+
|
750 |
+
if return_intermediate:
|
751 |
+
return x_t, {'model_s': model_s, 'model_s1': model_s1, 'model_s2': model_s2}
|
752 |
+
else:
|
753 |
+
return x_t
|
754 |
+
|
755 |
+
def multistep_dpm_solver_second_update(self, x, model_prev_list, t_prev_list, t, solver_type="dpm_solver"):
|
756 |
+
"""
|
757 |
+
Multistep solver DPM-Solver-2 from time `t_prev_list[-1]` to time `t`.
|
758 |
+
|
759 |
+
Args:
|
760 |
+
x: A pytorch tensor. The initial value at time `s`.
|
761 |
+
model_prev_list: A list of pytorch tensor. The previous computed model values.
|
762 |
+
t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],)
|
763 |
+
t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
|
764 |
+
solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
|
765 |
+
The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
|
766 |
+
Returns:
|
767 |
+
x_t: A pytorch tensor. The approximated solution at time `t`.
|
768 |
+
"""
|
769 |
+
if solver_type not in ['dpm_solver', 'taylor']:
|
770 |
+
raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type))
|
771 |
+
ns = self.noise_schedule
|
772 |
+
dims = x.dim()
|
773 |
+
model_prev_1, model_prev_0 = model_prev_list
|
774 |
+
t_prev_1, t_prev_0 = t_prev_list
|
775 |
+
lambda_prev_1, lambda_prev_0, lambda_t = ns.marginal_lambda(t_prev_1), ns.marginal_lambda(t_prev_0), ns.marginal_lambda(t)
|
776 |
+
log_alpha_prev_0, log_alpha_t = ns.marginal_log_mean_coeff(t_prev_0), ns.marginal_log_mean_coeff(t)
|
777 |
+
sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t)
|
778 |
+
alpha_t = torch.exp(log_alpha_t)
|
779 |
+
|
780 |
+
h_0 = lambda_prev_0 - lambda_prev_1
|
781 |
+
h = lambda_t - lambda_prev_0
|
782 |
+
r0 = h_0 / h
|
783 |
+
D1_0 = expand_dims(1. / r0, dims) * (model_prev_0 - model_prev_1)
|
784 |
+
if self.predict_x0:
|
785 |
+
if solver_type == 'dpm_solver':
|
786 |
+
x_t = (
|
787 |
+
expand_dims(sigma_t / sigma_prev_0, dims) * x
|
788 |
+
- expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0
|
789 |
+
- 0.5 * expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * D1_0
|
790 |
+
)
|
791 |
+
elif solver_type == 'taylor':
|
792 |
+
x_t = (
|
793 |
+
expand_dims(sigma_t / sigma_prev_0, dims) * x
|
794 |
+
- expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0
|
795 |
+
+ expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * D1_0
|
796 |
+
)
|
797 |
+
else:
|
798 |
+
if solver_type == 'dpm_solver':
|
799 |
+
x_t = (
|
800 |
+
expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x
|
801 |
+
- expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0
|
802 |
+
- 0.5 * expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * D1_0
|
803 |
+
)
|
804 |
+
elif solver_type == 'taylor':
|
805 |
+
x_t = (
|
806 |
+
expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x
|
807 |
+
- expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0
|
808 |
+
- expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * D1_0
|
809 |
+
)
|
810 |
+
return x_t
|
811 |
+
|
812 |
+
def multistep_dpm_solver_third_update(self, x, model_prev_list, t_prev_list, t, solver_type='dpm_solver'):
|
813 |
+
"""
|
814 |
+
Multistep solver DPM-Solver-3 from time `t_prev_list[-1]` to time `t`.
|
815 |
+
|
816 |
+
Args:
|
817 |
+
x: A pytorch tensor. The initial value at time `s`.
|
818 |
+
model_prev_list: A list of pytorch tensor. The previous computed model values.
|
819 |
+
t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],)
|
820 |
+
t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
|
821 |
+
solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
|
822 |
+
The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
|
823 |
+
Returns:
|
824 |
+
x_t: A pytorch tensor. The approximated solution at time `t`.
|
825 |
+
"""
|
826 |
+
ns = self.noise_schedule
|
827 |
+
dims = x.dim()
|
828 |
+
model_prev_2, model_prev_1, model_prev_0 = model_prev_list
|
829 |
+
t_prev_2, t_prev_1, t_prev_0 = t_prev_list
|
830 |
+
lambda_prev_2, lambda_prev_1, lambda_prev_0, lambda_t = ns.marginal_lambda(t_prev_2), ns.marginal_lambda(t_prev_1), ns.marginal_lambda(t_prev_0), ns.marginal_lambda(t)
|
831 |
+
log_alpha_prev_0, log_alpha_t = ns.marginal_log_mean_coeff(t_prev_0), ns.marginal_log_mean_coeff(t)
|
832 |
+
sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t)
|
833 |
+
alpha_t = torch.exp(log_alpha_t)
|
834 |
+
|
835 |
+
h_1 = lambda_prev_1 - lambda_prev_2
|
836 |
+
h_0 = lambda_prev_0 - lambda_prev_1
|
837 |
+
h = lambda_t - lambda_prev_0
|
838 |
+
r0, r1 = h_0 / h, h_1 / h
|
839 |
+
D1_0 = expand_dims(1. / r0, dims) * (model_prev_0 - model_prev_1)
|
840 |
+
D1_1 = expand_dims(1. / r1, dims) * (model_prev_1 - model_prev_2)
|
841 |
+
D1 = D1_0 + expand_dims(r0 / (r0 + r1), dims) * (D1_0 - D1_1)
|
842 |
+
D2 = expand_dims(1. / (r0 + r1), dims) * (D1_0 - D1_1)
|
843 |
+
if self.predict_x0:
|
844 |
+
x_t = (
|
845 |
+
expand_dims(sigma_t / sigma_prev_0, dims) * x
|
846 |
+
- expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0
|
847 |
+
+ expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * D1
|
848 |
+
- expand_dims(alpha_t * ((torch.exp(-h) - 1. + h) / h**2 - 0.5), dims) * D2
|
849 |
+
)
|
850 |
+
else:
|
851 |
+
x_t = (
|
852 |
+
expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x
|
853 |
+
- expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0
|
854 |
+
- expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * D1
|
855 |
+
- expand_dims(sigma_t * ((torch.exp(h) - 1. - h) / h**2 - 0.5), dims) * D2
|
856 |
+
)
|
857 |
+
return x_t
|
858 |
+
|
859 |
+
def singlestep_dpm_solver_update(self, x, s, t, order, return_intermediate=False, solver_type='dpm_solver', r1=None, r2=None):
|
860 |
+
"""
|
861 |
+
Singlestep DPM-Solver with the order `order` from time `s` to time `t`.
|
862 |
+
|
863 |
+
Args:
|
864 |
+
x: A pytorch tensor. The initial value at time `s`.
|
865 |
+
s: A pytorch tensor. The starting time, with the shape (x.shape[0],).
|
866 |
+
t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
|
867 |
+
order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3.
|
868 |
+
return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times).
|
869 |
+
solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
|
870 |
+
The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
|
871 |
+
r1: A `float`. The hyperparameter of the second-order or third-order solver.
|
872 |
+
r2: A `float`. The hyperparameter of the third-order solver.
|
873 |
+
Returns:
|
874 |
+
x_t: A pytorch tensor. The approximated solution at time `t`.
|
875 |
+
"""
|
876 |
+
if order == 1:
|
877 |
+
return self.dpm_solver_first_update(x, s, t, return_intermediate=return_intermediate)
|
878 |
+
elif order == 2:
|
879 |
+
return self.singlestep_dpm_solver_second_update(x, s, t, return_intermediate=return_intermediate, solver_type=solver_type, r1=r1)
|
880 |
+
elif order == 3:
|
881 |
+
return self.singlestep_dpm_solver_third_update(x, s, t, return_intermediate=return_intermediate, solver_type=solver_type, r1=r1, r2=r2)
|
882 |
+
else:
|
883 |
+
raise ValueError("Solver order must be 1 or 2 or 3, got {}".format(order))
|
884 |
+
|
885 |
+
def multistep_dpm_solver_update(self, x, model_prev_list, t_prev_list, t, order, solver_type='dpm_solver'):
|
886 |
+
"""
|
887 |
+
Multistep DPM-Solver with the order `order` from time `t_prev_list[-1]` to time `t`.
|
888 |
+
|
889 |
+
Args:
|
890 |
+
x: A pytorch tensor. The initial value at time `s`.
|
891 |
+
model_prev_list: A list of pytorch tensor. The previous computed model values.
|
892 |
+
t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],)
|
893 |
+
t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
|
894 |
+
order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3.
|
895 |
+
solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
|
896 |
+
The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
|
897 |
+
Returns:
|
898 |
+
x_t: A pytorch tensor. The approximated solution at time `t`.
|
899 |
+
"""
|
900 |
+
if order == 1:
|
901 |
+
return self.dpm_solver_first_update(x, t_prev_list[-1], t, model_s=model_prev_list[-1])
|
902 |
+
elif order == 2:
|
903 |
+
return self.multistep_dpm_solver_second_update(x, model_prev_list, t_prev_list, t, solver_type=solver_type)
|
904 |
+
elif order == 3:
|
905 |
+
return self.multistep_dpm_solver_third_update(x, model_prev_list, t_prev_list, t, solver_type=solver_type)
|
906 |
+
else:
|
907 |
+
raise ValueError("Solver order must be 1 or 2 or 3, got {}".format(order))
|
908 |
+
|
909 |
+
def dpm_solver_adaptive(self, x, order, t_T, t_0, h_init=0.05, atol=0.0078, rtol=0.05, theta=0.9, t_err=1e-5, solver_type='dpm_solver'):
|
910 |
+
"""
|
911 |
+
The adaptive step size solver based on singlestep DPM-Solver.
|
912 |
+
|
913 |
+
Args:
|
914 |
+
x: A pytorch tensor. The initial value at time `t_T`.
|
915 |
+
order: A `int`. The (higher) order of the solver. We only support order == 2 or 3.
|
916 |
+
t_T: A `float`. The starting time of the sampling (default is T).
|
917 |
+
t_0: A `float`. The ending time of the sampling (default is epsilon).
|
918 |
+
h_init: A `float`. The initial step size (for logSNR).
|
919 |
+
atol: A `float`. The absolute tolerance of the solver. For image data, the default setting is 0.0078, followed [1].
|
920 |
+
rtol: A `float`. The relative tolerance of the solver. The default setting is 0.05.
|
921 |
+
theta: A `float`. The safety hyperparameter for adapting the step size. The default setting is 0.9, followed [1].
|
922 |
+
t_err: A `float`. The tolerance for the time. We solve the diffusion ODE until the absolute error between the
|
923 |
+
current time and `t_0` is less than `t_err`. The default setting is 1e-5.
|
924 |
+
solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
|
925 |
+
The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
|
926 |
+
Returns:
|
927 |
+
x_0: A pytorch tensor. The approximated solution at time `t_0`.
|
928 |
+
|
929 |
+
[1] A. Jolicoeur-Martineau, K. Li, R. Piché-Taillefer, T. Kachman, and I. Mitliagkas, "Gotta go fast when generating data with score-based models," arXiv preprint arXiv:2105.14080, 2021.
|
930 |
+
"""
|
931 |
+
ns = self.noise_schedule
|
932 |
+
s = t_T * torch.ones((x.shape[0],)).to(x)
|
933 |
+
lambda_s = ns.marginal_lambda(s)
|
934 |
+
lambda_0 = ns.marginal_lambda(t_0 * torch.ones_like(s).to(x))
|
935 |
+
h = h_init * torch.ones_like(s).to(x)
|
936 |
+
x_prev = x
|
937 |
+
nfe = 0
|
938 |
+
if order == 2:
|
939 |
+
r1 = 0.5
|
940 |
+
lower_update = lambda x, s, t: self.dpm_solver_first_update(x, s, t, return_intermediate=True)
|
941 |
+
higher_update = lambda x, s, t, **kwargs: self.singlestep_dpm_solver_second_update(x, s, t, r1=r1, solver_type=solver_type, **kwargs)
|
942 |
+
elif order == 3:
|
943 |
+
r1, r2 = 1. / 3., 2. / 3.
|
944 |
+
lower_update = lambda x, s, t: self.singlestep_dpm_solver_second_update(x, s, t, r1=r1, return_intermediate=True, solver_type=solver_type)
|
945 |
+
higher_update = lambda x, s, t, **kwargs: self.singlestep_dpm_solver_third_update(x, s, t, r1=r1, r2=r2, solver_type=solver_type, **kwargs)
|
946 |
+
else:
|
947 |
+
raise ValueError("For adaptive step size solver, order must be 2 or 3, got {}".format(order))
|
948 |
+
while torch.abs((s - t_0)).mean() > t_err:
|
949 |
+
t = ns.inverse_lambda(lambda_s + h)
|
950 |
+
x_lower, lower_noise_kwargs = lower_update(x, s, t)
|
951 |
+
x_higher = higher_update(x, s, t, **lower_noise_kwargs)
|
952 |
+
delta = torch.max(torch.ones_like(x).to(x) * atol, rtol * torch.max(torch.abs(x_lower), torch.abs(x_prev)))
|
953 |
+
norm_fn = lambda v: torch.sqrt(torch.square(v.reshape((v.shape[0], -1))).mean(dim=-1, keepdim=True))
|
954 |
+
E = norm_fn((x_higher - x_lower) / delta).max()
|
955 |
+
if torch.all(E <= 1.):
|
956 |
+
x = x_higher
|
957 |
+
s = t
|
958 |
+
x_prev = x_lower
|
959 |
+
lambda_s = ns.marginal_lambda(s)
|
960 |
+
h = torch.min(theta * h * torch.float_power(E, -1. / order).float(), lambda_0 - lambda_s)
|
961 |
+
nfe += order
|
962 |
+
print('adaptive solver nfe', nfe)
|
963 |
+
return x
|
964 |
+
|
965 |
+
def sample(self, x, steps=20, t_start=None, t_end=None, order=3, skip_type='time_uniform',
|
966 |
+
method='singlestep', lower_order_final=True, denoise_to_zero=False, solver_type='dpm_solver',
|
967 |
+
atol=0.0078, rtol=0.05,
|
968 |
+
):
|
969 |
+
"""
|
970 |
+
Compute the sample at time `t_end` by DPM-Solver, given the initial `x` at time `t_start`.
|
971 |
+
|
972 |
+
=====================================================
|
973 |
+
|
974 |
+
We support the following algorithms for both noise prediction model and data prediction model:
|
975 |
+
- 'singlestep':
|
976 |
+
Singlestep DPM-Solver (i.e. "DPM-Solver-fast" in the paper), which combines different orders of singlestep DPM-Solver.
|
977 |
+
We combine all the singlestep solvers with order <= `order` to use up all the function evaluations (steps).
|
978 |
+
The total number of function evaluations (NFE) == `steps`.
|
979 |
+
Given a fixed NFE == `steps`, the sampling procedure is:
|
980 |
+
- If `order` == 1:
|
981 |
+
- Denote K = steps. We use K steps of DPM-Solver-1 (i.e. DDIM).
|
982 |
+
- If `order` == 2:
|
983 |
+
- Denote K = (steps // 2) + (steps % 2). We take K intermediate time steps for sampling.
|
984 |
+
- If steps % 2 == 0, we use K steps of singlestep DPM-Solver-2.
|
985 |
+
- If steps % 2 == 1, we use (K - 1) steps of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1.
|
986 |
+
- If `order` == 3:
|
987 |
+
- Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling.
|
988 |
+
- If steps % 3 == 0, we use (K - 2) steps of singlestep DPM-Solver-3, and 1 step of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1.
|
989 |
+
- If steps % 3 == 1, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of DPM-Solver-1.
|
990 |
+
- If steps % 3 == 2, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of singlestep DPM-Solver-2.
|
991 |
+
- 'multistep':
|
992 |
+
Multistep DPM-Solver with the order of `order`. The total number of function evaluations (NFE) == `steps`.
|
993 |
+
We initialize the first `order` values by lower order multistep solvers.
|
994 |
+
Given a fixed NFE == `steps`, the sampling procedure is:
|
995 |
+
Denote K = steps.
|
996 |
+
- If `order` == 1:
|
997 |
+
- We use K steps of DPM-Solver-1 (i.e. DDIM).
|
998 |
+
- If `order` == 2:
|
999 |
+
- We firstly use 1 step of DPM-Solver-1, then use (K - 1) step of multistep DPM-Solver-2.
|
1000 |
+
- If `order` == 3:
|
1001 |
+
- We firstly use 1 step of DPM-Solver-1, then 1 step of multistep DPM-Solver-2, then (K - 2) step of multistep DPM-Solver-3.
|
1002 |
+
- 'singlestep_fixed':
|
1003 |
+
Fixed order singlestep DPM-Solver (i.e. DPM-Solver-1 or singlestep DPM-Solver-2 or singlestep DPM-Solver-3).
|
1004 |
+
We use singlestep DPM-Solver-`order` for `order`=1 or 2 or 3, with total [`steps` // `order`] * `order` NFE.
|
1005 |
+
- 'adaptive':
|
1006 |
+
Adaptive step size DPM-Solver (i.e. "DPM-Solver-12" and "DPM-Solver-23" in the paper).
|
1007 |
+
We ignore `steps` and use adaptive step size DPM-Solver with a higher order of `order`.
|
1008 |
+
You can adjust the absolute tolerance `atol` and the relative tolerance `rtol` to balance the computatation costs
|
1009 |
+
(NFE) and the sample quality.
|
1010 |
+
- If `order` == 2, we use DPM-Solver-12 which combines DPM-Solver-1 and singlestep DPM-Solver-2.
|
1011 |
+
- If `order` == 3, we use DPM-Solver-23 which combines singlestep DPM-Solver-2 and singlestep DPM-Solver-3.
|
1012 |
+
|
1013 |
+
=====================================================
|
1014 |
+
|
1015 |
+
Some advices for choosing the algorithm:
|
1016 |
+
- For **unconditional sampling** or **guided sampling with small guidance scale** by DPMs:
|
1017 |
+
Use singlestep DPM-Solver ("DPM-Solver-fast" in the paper) with `order = 3`.
|
1018 |
+
e.g.
|
1019 |
+
>>> dpm_solver = DPM_Solver(model_fn, noise_schedule, predict_x0=False)
|
1020 |
+
>>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=3,
|
1021 |
+
skip_type='time_uniform', method='singlestep')
|
1022 |
+
- For **guided sampling with large guidance scale** by DPMs:
|
1023 |
+
Use multistep DPM-Solver with `predict_x0 = True` and `order = 2`.
|
1024 |
+
e.g.
|
1025 |
+
>>> dpm_solver = DPM_Solver(model_fn, noise_schedule, predict_x0=True)
|
1026 |
+
>>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=2,
|
1027 |
+
skip_type='time_uniform', method='multistep')
|
1028 |
+
|
1029 |
+
We support three types of `skip_type`:
|
1030 |
+
- 'logSNR': uniform logSNR for the time steps. **Recommended for low-resolutional images**
|
1031 |
+
- 'time_uniform': uniform time for the time steps. **Recommended for high-resolutional images**.
|
1032 |
+
- 'time_quadratic': quadratic time for the time steps.
|
1033 |
+
|
1034 |
+
=====================================================
|
1035 |
+
Args:
|
1036 |
+
x: A pytorch tensor. The initial value at time `t_start`
|
1037 |
+
e.g. if `t_start` == T, then `x` is a sample from the standard normal distribution.
|
1038 |
+
steps: A `int`. The total number of function evaluations (NFE).
|
1039 |
+
t_start: A `float`. The starting time of the sampling.
|
1040 |
+
If `T` is None, we use self.noise_schedule.T (default is 1.0).
|
1041 |
+
t_end: A `float`. The ending time of the sampling.
|
1042 |
+
If `t_end` is None, we use 1. / self.noise_schedule.total_N.
|
1043 |
+
e.g. if total_N == 1000, we have `t_end` == 1e-3.
|
1044 |
+
For discrete-time DPMs:
|
1045 |
+
- We recommend `t_end` == 1. / self.noise_schedule.total_N.
|
1046 |
+
For continuous-time DPMs:
|
1047 |
+
- We recommend `t_end` == 1e-3 when `steps` <= 15; and `t_end` == 1e-4 when `steps` > 15.
|
1048 |
+
order: A `int`. The order of DPM-Solver.
|
1049 |
+
skip_type: A `str`. The type for the spacing of the time steps. 'time_uniform' or 'logSNR' or 'time_quadratic'.
|
1050 |
+
method: A `str`. The method for sampling. 'singlestep' or 'multistep' or 'singlestep_fixed' or 'adaptive'.
|
1051 |
+
denoise_to_zero: A `bool`. Whether to denoise to time 0 at the final step.
|
1052 |
+
Default is `False`. If `denoise_to_zero` is `True`, the total NFE is (`steps` + 1).
|
1053 |
+
|
1054 |
+
This trick is firstly proposed by DDPM (https://arxiv.org/abs/2006.11239) and
|
1055 |
+
score_sde (https://arxiv.org/abs/2011.13456). Such trick can improve the FID
|
1056 |
+
for diffusion models sampling by diffusion SDEs for low-resolutional images
|
1057 |
+
(such as CIFAR-10). However, we observed that such trick does not matter for
|
1058 |
+
high-resolutional images. As it needs an additional NFE, we do not recommend
|
1059 |
+
it for high-resolutional images.
|
1060 |
+
lower_order_final: A `bool`. Whether to use lower order solvers at the final steps.
|
1061 |
+
Only valid for `method=multistep` and `steps < 15`. We empirically find that
|
1062 |
+
this trick is a key to stabilizing the sampling by DPM-Solver with very few steps
|
1063 |
+
(especially for steps <= 10). So we recommend to set it to be `True`.
|
1064 |
+
solver_type: A `str`. The taylor expansion type for the solver. `dpm_solver` or `taylor`. We recommend `dpm_solver`.
|
1065 |
+
atol: A `float`. The absolute tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'.
|
1066 |
+
rtol: A `float`. The relative tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'.
|
1067 |
+
Returns:
|
1068 |
+
x_end: A pytorch tensor. The approximated solution at time `t_end`.
|
1069 |
+
|
1070 |
+
"""
|
1071 |
+
t_0 = 1. / self.noise_schedule.total_N if t_end is None else t_end
|
1072 |
+
t_T = self.noise_schedule.T if t_start is None else t_start
|
1073 |
+
device = x.device
|
1074 |
+
if method == 'adaptive':
|
1075 |
+
with torch.no_grad():
|
1076 |
+
x = self.dpm_solver_adaptive(x, order=order, t_T=t_T, t_0=t_0, atol=atol, rtol=rtol, solver_type=solver_type)
|
1077 |
+
elif method == 'multistep':
|
1078 |
+
assert steps >= order
|
1079 |
+
timesteps = self.get_time_steps(skip_type=skip_type, t_T=t_T, t_0=t_0, N=steps, device=device)
|
1080 |
+
assert timesteps.shape[0] - 1 == steps
|
1081 |
+
with torch.no_grad():
|
1082 |
+
vec_t = timesteps[0].expand((x.shape[0]))
|
1083 |
+
model_prev_list = [self.model_fn(x, vec_t)]
|
1084 |
+
t_prev_list = [vec_t]
|
1085 |
+
# Init the first `order` values by lower order multistep DPM-Solver.
|
1086 |
+
for init_order in range(1, order):
|
1087 |
+
vec_t = timesteps[init_order].expand(x.shape[0])
|
1088 |
+
x = self.multistep_dpm_solver_update(x, model_prev_list, t_prev_list, vec_t, init_order, solver_type=solver_type)
|
1089 |
+
model_prev_list.append(self.model_fn(x, vec_t))
|
1090 |
+
t_prev_list.append(vec_t)
|
1091 |
+
# Compute the remaining values by `order`-th order multistep DPM-Solver.
|
1092 |
+
for step in range(order, steps + 1):
|
1093 |
+
vec_t = timesteps[step].expand(x.shape[0])
|
1094 |
+
if lower_order_final and steps < 15:
|
1095 |
+
step_order = min(order, steps + 1 - step)
|
1096 |
+
else:
|
1097 |
+
step_order = order
|
1098 |
+
x = self.multistep_dpm_solver_update(x, model_prev_list, t_prev_list, vec_t, step_order, solver_type=solver_type)
|
1099 |
+
for i in range(order - 1):
|
1100 |
+
t_prev_list[i] = t_prev_list[i + 1]
|
1101 |
+
model_prev_list[i] = model_prev_list[i + 1]
|
1102 |
+
t_prev_list[-1] = vec_t
|
1103 |
+
# We do not need to evaluate the final model value.
|
1104 |
+
if step < steps:
|
1105 |
+
model_prev_list[-1] = self.model_fn(x, vec_t)
|
1106 |
+
elif method in ['singlestep', 'singlestep_fixed']:
|
1107 |
+
if method == 'singlestep':
|
1108 |
+
timesteps_outer, orders = self.get_orders_and_timesteps_for_singlestep_solver(steps=steps, order=order, skip_type=skip_type, t_T=t_T, t_0=t_0, device=device)
|
1109 |
+
elif method == 'singlestep_fixed':
|
1110 |
+
K = steps // order
|
1111 |
+
orders = [order,] * K
|
1112 |
+
timesteps_outer = self.get_time_steps(skip_type=skip_type, t_T=t_T, t_0=t_0, N=K, device=device)
|
1113 |
+
for i, order in enumerate(orders):
|
1114 |
+
t_T_inner, t_0_inner = timesteps_outer[i], timesteps_outer[i + 1]
|
1115 |
+
timesteps_inner = self.get_time_steps(skip_type=skip_type, t_T=t_T_inner.item(), t_0=t_0_inner.item(), N=order, device=device)
|
1116 |
+
lambda_inner = self.noise_schedule.marginal_lambda(timesteps_inner)
|
1117 |
+
vec_s, vec_t = t_T_inner.tile(x.shape[0]), t_0_inner.tile(x.shape[0])
|
1118 |
+
h = lambda_inner[-1] - lambda_inner[0]
|
1119 |
+
r1 = None if order <= 1 else (lambda_inner[1] - lambda_inner[0]) / h
|
1120 |
+
r2 = None if order <= 2 else (lambda_inner[2] - lambda_inner[0]) / h
|
1121 |
+
x = self.singlestep_dpm_solver_update(x, vec_s, vec_t, order, solver_type=solver_type, r1=r1, r2=r2)
|
1122 |
+
if denoise_to_zero:
|
1123 |
+
x = self.denoise_to_zero_fn(x, torch.ones((x.shape[0],)).to(device) * t_0)
|
1124 |
+
return x
|
1125 |
+
|
1126 |
+
|
1127 |
+
|
1128 |
+
#############################################################
|
1129 |
+
# other utility functions
|
1130 |
+
#############################################################
|
1131 |
+
|
1132 |
+
def interpolate_fn(x, xp, yp):
|
1133 |
+
"""
|
1134 |
+
A piecewise linear function y = f(x), using xp and yp as keypoints.
|
1135 |
+
We implement f(x) in a differentiable way (i.e. applicable for autograd).
|
1136 |
+
The function f(x) is well-defined for all x-axis. (For x beyond the bounds of xp, we use the outmost points of xp to define the linear function.)
|
1137 |
+
|
1138 |
+
Args:
|
1139 |
+
x: PyTorch tensor with shape [N, C], where N is the batch size, C is the number of channels (we use C = 1 for DPM-Solver).
|
1140 |
+
xp: PyTorch tensor with shape [C, K], where K is the number of keypoints.
|
1141 |
+
yp: PyTorch tensor with shape [C, K].
|
1142 |
+
Returns:
|
1143 |
+
The function values f(x), with shape [N, C].
|
1144 |
+
"""
|
1145 |
+
N, K = x.shape[0], xp.shape[1]
|
1146 |
+
all_x = torch.cat([x.unsqueeze(2), xp.unsqueeze(0).repeat((N, 1, 1))], dim=2)
|
1147 |
+
sorted_all_x, x_indices = torch.sort(all_x, dim=2)
|
1148 |
+
x_idx = torch.argmin(x_indices, dim=2)
|
1149 |
+
cand_start_idx = x_idx - 1
|
1150 |
+
start_idx = torch.where(
|
1151 |
+
torch.eq(x_idx, 0),
|
1152 |
+
torch.tensor(1, device=x.device),
|
1153 |
+
torch.where(
|
1154 |
+
torch.eq(x_idx, K), torch.tensor(K - 2, device=x.device), cand_start_idx,
|
1155 |
+
),
|
1156 |
+
)
|
1157 |
+
end_idx = torch.where(torch.eq(start_idx, cand_start_idx), start_idx + 2, start_idx + 1)
|
1158 |
+
start_x = torch.gather(sorted_all_x, dim=2, index=start_idx.unsqueeze(2)).squeeze(2)
|
1159 |
+
end_x = torch.gather(sorted_all_x, dim=2, index=end_idx.unsqueeze(2)).squeeze(2)
|
1160 |
+
start_idx2 = torch.where(
|
1161 |
+
torch.eq(x_idx, 0),
|
1162 |
+
torch.tensor(0, device=x.device),
|
1163 |
+
torch.where(
|
1164 |
+
torch.eq(x_idx, K), torch.tensor(K - 2, device=x.device), cand_start_idx,
|
1165 |
+
),
|
1166 |
+
)
|
1167 |
+
y_positions_expanded = yp.unsqueeze(0).expand(N, -1, -1)
|
1168 |
+
start_y = torch.gather(y_positions_expanded, dim=2, index=start_idx2.unsqueeze(2)).squeeze(2)
|
1169 |
+
end_y = torch.gather(y_positions_expanded, dim=2, index=(start_idx2 + 1).unsqueeze(2)).squeeze(2)
|
1170 |
+
cand = start_y + (x - start_x) * (end_y - start_y) / (end_x - start_x)
|
1171 |
+
return cand
|
1172 |
+
|
1173 |
+
|
1174 |
+
def expand_dims(v, dims):
|
1175 |
+
"""
|
1176 |
+
Expand the tensor `v` to the dim `dims`.
|
1177 |
+
|
1178 |
+
Args:
|
1179 |
+
`v`: a PyTorch tensor with shape [N].
|
1180 |
+
`dim`: a `int`.
|
1181 |
+
Returns:
|
1182 |
+
a PyTorch tensor with shape [N, 1, 1, ..., 1] and the total dimension is `dims`.
|
1183 |
+
"""
|
1184 |
+
return v[(...,) + (None,)*(dims - 1)]
|
stable-diffusion/ldm/models/diffusion/dpm_solver/sampler.py
ADDED
@@ -0,0 +1,82 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""SAMPLING ONLY."""
|
2 |
+
|
3 |
+
import torch
|
4 |
+
|
5 |
+
from .dpm_solver import NoiseScheduleVP, model_wrapper, DPM_Solver
|
6 |
+
|
7 |
+
|
8 |
+
class DPMSolverSampler(object):
|
9 |
+
def __init__(self, model, **kwargs):
|
10 |
+
super().__init__()
|
11 |
+
self.model = model
|
12 |
+
to_torch = lambda x: x.clone().detach().to(torch.float32).to(model.device)
|
13 |
+
self.register_buffer('alphas_cumprod', to_torch(model.alphas_cumprod))
|
14 |
+
|
15 |
+
def register_buffer(self, name, attr):
|
16 |
+
if type(attr) == torch.Tensor:
|
17 |
+
if attr.device != torch.device("cuda"):
|
18 |
+
attr = attr.to(torch.device("cuda"))
|
19 |
+
setattr(self, name, attr)
|
20 |
+
|
21 |
+
@torch.no_grad()
|
22 |
+
def sample(self,
|
23 |
+
S,
|
24 |
+
batch_size,
|
25 |
+
shape,
|
26 |
+
conditioning=None,
|
27 |
+
callback=None,
|
28 |
+
normals_sequence=None,
|
29 |
+
img_callback=None,
|
30 |
+
quantize_x0=False,
|
31 |
+
eta=0.,
|
32 |
+
mask=None,
|
33 |
+
x0=None,
|
34 |
+
temperature=1.,
|
35 |
+
noise_dropout=0.,
|
36 |
+
score_corrector=None,
|
37 |
+
corrector_kwargs=None,
|
38 |
+
verbose=True,
|
39 |
+
x_T=None,
|
40 |
+
log_every_t=100,
|
41 |
+
unconditional_guidance_scale=1.,
|
42 |
+
unconditional_conditioning=None,
|
43 |
+
# this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
|
44 |
+
**kwargs
|
45 |
+
):
|
46 |
+
if conditioning is not None:
|
47 |
+
if isinstance(conditioning, dict):
|
48 |
+
cbs = conditioning[list(conditioning.keys())[0]].shape[0]
|
49 |
+
if cbs != batch_size:
|
50 |
+
print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
|
51 |
+
else:
|
52 |
+
if conditioning.shape[0] != batch_size:
|
53 |
+
print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
|
54 |
+
|
55 |
+
# sampling
|
56 |
+
C, H, W = shape
|
57 |
+
size = (batch_size, C, H, W)
|
58 |
+
|
59 |
+
# print(f'Data shape for DPM-Solver sampling is {size}, sampling steps {S}')
|
60 |
+
|
61 |
+
device = self.model.betas.device
|
62 |
+
if x_T is None:
|
63 |
+
img = torch.randn(size, device=device)
|
64 |
+
else:
|
65 |
+
img = x_T
|
66 |
+
|
67 |
+
ns = NoiseScheduleVP('discrete', alphas_cumprod=self.alphas_cumprod)
|
68 |
+
|
69 |
+
model_fn = model_wrapper(
|
70 |
+
lambda x, t, c: self.model.apply_model(x, t, c),
|
71 |
+
ns,
|
72 |
+
model_type="noise",
|
73 |
+
guidance_type="classifier-free",
|
74 |
+
condition=conditioning,
|
75 |
+
unconditional_condition=unconditional_conditioning,
|
76 |
+
guidance_scale=unconditional_guidance_scale,
|
77 |
+
)
|
78 |
+
|
79 |
+
dpm_solver = DPM_Solver(model_fn, ns, predict_x0=True, thresholding=False)
|
80 |
+
x = dpm_solver.sample(img, steps=S, skip_type="time_uniform", method="multistep", order=2, lower_order_final=True)
|
81 |
+
|
82 |
+
return x.to(device), None
|
stable-diffusion/ldm/models/diffusion/plms.py
ADDED
@@ -0,0 +1,236 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""SAMPLING ONLY."""
|
2 |
+
|
3 |
+
import torch
|
4 |
+
import numpy as np
|
5 |
+
from tqdm import tqdm
|
6 |
+
from functools import partial
|
7 |
+
|
8 |
+
from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like
|
9 |
+
|
10 |
+
|
11 |
+
class PLMSSampler(object):
|
12 |
+
def __init__(self, model, schedule="linear", **kwargs):
|
13 |
+
super().__init__()
|
14 |
+
self.model = model
|
15 |
+
self.ddpm_num_timesteps = model.num_timesteps
|
16 |
+
self.schedule = schedule
|
17 |
+
|
18 |
+
def register_buffer(self, name, attr):
|
19 |
+
if type(attr) == torch.Tensor:
|
20 |
+
if attr.device != torch.device("cuda"):
|
21 |
+
attr = attr.to(torch.device("cuda"))
|
22 |
+
setattr(self, name, attr)
|
23 |
+
|
24 |
+
def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
|
25 |
+
if ddim_eta != 0:
|
26 |
+
raise ValueError('ddim_eta must be 0 for PLMS')
|
27 |
+
self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
|
28 |
+
num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)
|
29 |
+
alphas_cumprod = self.model.alphas_cumprod
|
30 |
+
assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
|
31 |
+
to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
|
32 |
+
|
33 |
+
self.register_buffer('betas', to_torch(self.model.betas))
|
34 |
+
self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
|
35 |
+
self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev))
|
36 |
+
|
37 |
+
# calculations for diffusion q(x_t | x_{t-1}) and others
|
38 |
+
self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))
|
39 |
+
self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
|
40 |
+
self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))
|
41 |
+
self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))
|
42 |
+
self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))
|
43 |
+
|
44 |
+
# ddim sampling parameters
|
45 |
+
ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),
|
46 |
+
ddim_timesteps=self.ddim_timesteps,
|
47 |
+
eta=ddim_eta,verbose=verbose)
|
48 |
+
self.register_buffer('ddim_sigmas', ddim_sigmas)
|
49 |
+
self.register_buffer('ddim_alphas', ddim_alphas)
|
50 |
+
self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
|
51 |
+
self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
|
52 |
+
sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
|
53 |
+
(1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
|
54 |
+
1 - self.alphas_cumprod / self.alphas_cumprod_prev))
|
55 |
+
self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
|
56 |
+
|
57 |
+
@torch.no_grad()
|
58 |
+
def sample(self,
|
59 |
+
S,
|
60 |
+
batch_size,
|
61 |
+
shape,
|
62 |
+
conditioning=None,
|
63 |
+
callback=None,
|
64 |
+
normals_sequence=None,
|
65 |
+
img_callback=None,
|
66 |
+
quantize_x0=False,
|
67 |
+
eta=0.,
|
68 |
+
mask=None,
|
69 |
+
x0=None,
|
70 |
+
temperature=1.,
|
71 |
+
noise_dropout=0.,
|
72 |
+
score_corrector=None,
|
73 |
+
corrector_kwargs=None,
|
74 |
+
verbose=True,
|
75 |
+
x_T=None,
|
76 |
+
log_every_t=100,
|
77 |
+
unconditional_guidance_scale=1.,
|
78 |
+
unconditional_conditioning=None,
|
79 |
+
# this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
|
80 |
+
**kwargs
|
81 |
+
):
|
82 |
+
if conditioning is not None:
|
83 |
+
if isinstance(conditioning, dict):
|
84 |
+
cbs = conditioning[list(conditioning.keys())[0]].shape[0]
|
85 |
+
if cbs != batch_size:
|
86 |
+
print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
|
87 |
+
else:
|
88 |
+
if conditioning.shape[0] != batch_size:
|
89 |
+
print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
|
90 |
+
|
91 |
+
self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose)
|
92 |
+
# sampling
|
93 |
+
C, H, W = shape
|
94 |
+
size = (batch_size, C, H, W)
|
95 |
+
print(f'Data shape for PLMS sampling is {size}')
|
96 |
+
|
97 |
+
samples, intermediates = self.plms_sampling(conditioning, size,
|
98 |
+
callback=callback,
|
99 |
+
img_callback=img_callback,
|
100 |
+
quantize_denoised=quantize_x0,
|
101 |
+
mask=mask, x0=x0,
|
102 |
+
ddim_use_original_steps=False,
|
103 |
+
noise_dropout=noise_dropout,
|
104 |
+
temperature=temperature,
|
105 |
+
score_corrector=score_corrector,
|
106 |
+
corrector_kwargs=corrector_kwargs,
|
107 |
+
x_T=x_T,
|
108 |
+
log_every_t=log_every_t,
|
109 |
+
unconditional_guidance_scale=unconditional_guidance_scale,
|
110 |
+
unconditional_conditioning=unconditional_conditioning,
|
111 |
+
)
|
112 |
+
return samples, intermediates
|
113 |
+
|
114 |
+
@torch.no_grad()
|
115 |
+
def plms_sampling(self, cond, shape,
|
116 |
+
x_T=None, ddim_use_original_steps=False,
|
117 |
+
callback=None, timesteps=None, quantize_denoised=False,
|
118 |
+
mask=None, x0=None, img_callback=None, log_every_t=100,
|
119 |
+
temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
|
120 |
+
unconditional_guidance_scale=1., unconditional_conditioning=None,):
|
121 |
+
device = self.model.betas.device
|
122 |
+
b = shape[0]
|
123 |
+
if x_T is None:
|
124 |
+
img = torch.randn(shape, device=device)
|
125 |
+
else:
|
126 |
+
img = x_T
|
127 |
+
|
128 |
+
if timesteps is None:
|
129 |
+
timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
|
130 |
+
elif timesteps is not None and not ddim_use_original_steps:
|
131 |
+
subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
|
132 |
+
timesteps = self.ddim_timesteps[:subset_end]
|
133 |
+
|
134 |
+
intermediates = {'x_inter': [img], 'pred_x0': [img]}
|
135 |
+
time_range = list(reversed(range(0,timesteps))) if ddim_use_original_steps else np.flip(timesteps)
|
136 |
+
total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
|
137 |
+
print(f"Running PLMS Sampling with {total_steps} timesteps")
|
138 |
+
|
139 |
+
iterator = tqdm(time_range, desc='PLMS Sampler', total=total_steps)
|
140 |
+
old_eps = []
|
141 |
+
|
142 |
+
for i, step in enumerate(iterator):
|
143 |
+
index = total_steps - i - 1
|
144 |
+
ts = torch.full((b,), step, device=device, dtype=torch.long)
|
145 |
+
ts_next = torch.full((b,), time_range[min(i + 1, len(time_range) - 1)], device=device, dtype=torch.long)
|
146 |
+
|
147 |
+
if mask is not None:
|
148 |
+
assert x0 is not None
|
149 |
+
img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass?
|
150 |
+
img = img_orig * mask + (1. - mask) * img
|
151 |
+
|
152 |
+
outs = self.p_sample_plms(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,
|
153 |
+
quantize_denoised=quantize_denoised, temperature=temperature,
|
154 |
+
noise_dropout=noise_dropout, score_corrector=score_corrector,
|
155 |
+
corrector_kwargs=corrector_kwargs,
|
156 |
+
unconditional_guidance_scale=unconditional_guidance_scale,
|
157 |
+
unconditional_conditioning=unconditional_conditioning,
|
158 |
+
old_eps=old_eps, t_next=ts_next)
|
159 |
+
img, pred_x0, e_t = outs
|
160 |
+
old_eps.append(e_t)
|
161 |
+
if len(old_eps) >= 4:
|
162 |
+
old_eps.pop(0)
|
163 |
+
if callback: callback(i)
|
164 |
+
if img_callback: img_callback(pred_x0, i)
|
165 |
+
|
166 |
+
if index % log_every_t == 0 or index == total_steps - 1:
|
167 |
+
intermediates['x_inter'].append(img)
|
168 |
+
intermediates['pred_x0'].append(pred_x0)
|
169 |
+
|
170 |
+
return img, intermediates
|
171 |
+
|
172 |
+
@torch.no_grad()
|
173 |
+
def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
|
174 |
+
temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
|
175 |
+
unconditional_guidance_scale=1., unconditional_conditioning=None, old_eps=None, t_next=None):
|
176 |
+
b, *_, device = *x.shape, x.device
|
177 |
+
|
178 |
+
def get_model_output(x, t):
|
179 |
+
if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
|
180 |
+
e_t = self.model.apply_model(x, t, c)
|
181 |
+
else:
|
182 |
+
x_in = torch.cat([x] * 2)
|
183 |
+
t_in = torch.cat([t] * 2)
|
184 |
+
c_in = torch.cat([unconditional_conditioning, c])
|
185 |
+
e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2)
|
186 |
+
e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)
|
187 |
+
|
188 |
+
if score_corrector is not None:
|
189 |
+
assert self.model.parameterization == "eps"
|
190 |
+
e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
|
191 |
+
|
192 |
+
return e_t
|
193 |
+
|
194 |
+
alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
|
195 |
+
alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
|
196 |
+
sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
|
197 |
+
sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
|
198 |
+
|
199 |
+
def get_x_prev_and_pred_x0(e_t, index):
|
200 |
+
# select parameters corresponding to the currently considered timestep
|
201 |
+
a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
|
202 |
+
a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
|
203 |
+
sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
|
204 |
+
sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device)
|
205 |
+
|
206 |
+
# current prediction for x_0
|
207 |
+
pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
|
208 |
+
if quantize_denoised:
|
209 |
+
pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
|
210 |
+
# direction pointing to x_t
|
211 |
+
dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
|
212 |
+
noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
|
213 |
+
if noise_dropout > 0.:
|
214 |
+
noise = torch.nn.functional.dropout(noise, p=noise_dropout)
|
215 |
+
x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
|
216 |
+
return x_prev, pred_x0
|
217 |
+
|
218 |
+
e_t = get_model_output(x, t)
|
219 |
+
if len(old_eps) == 0:
|
220 |
+
# Pseudo Improved Euler (2nd order)
|
221 |
+
x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t, index)
|
222 |
+
e_t_next = get_model_output(x_prev, t_next)
|
223 |
+
e_t_prime = (e_t + e_t_next) / 2
|
224 |
+
elif len(old_eps) == 1:
|
225 |
+
# 2nd order Pseudo Linear Multistep (Adams-Bashforth)
|
226 |
+
e_t_prime = (3 * e_t - old_eps[-1]) / 2
|
227 |
+
elif len(old_eps) == 2:
|
228 |
+
# 3nd order Pseudo Linear Multistep (Adams-Bashforth)
|
229 |
+
e_t_prime = (23 * e_t - 16 * old_eps[-1] + 5 * old_eps[-2]) / 12
|
230 |
+
elif len(old_eps) >= 3:
|
231 |
+
# 4nd order Pseudo Linear Multistep (Adams-Bashforth)
|
232 |
+
e_t_prime = (55 * e_t - 59 * old_eps[-1] + 37 * old_eps[-2] - 9 * old_eps[-3]) / 24
|
233 |
+
|
234 |
+
x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t_prime, index)
|
235 |
+
|
236 |
+
return x_prev, pred_x0, e_t
|