diffusion-model-framework 1.0.0rc1__tar.gz → 1.0.0rc2__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- {diffusion_model_framework-1.0.0rc1 → diffusion_model_framework-1.0.0rc2}/PKG-INFO +1 -1
- diffusion_model_framework-1.0.0rc2/diffusion_models/__init__.py +1 -0
- {diffusion_model_framework-1.0.0rc1 → diffusion_model_framework-1.0.0rc2}/diffusion_models/diffusion_inference.py +3 -5
- {diffusion_model_framework-1.0.0rc1 → diffusion_model_framework-1.0.0rc2}/diffusion_models/diffusion_trainer.py +6 -2
- {diffusion_model_framework-1.0.0rc1 → diffusion_model_framework-1.0.0rc2}/diffusion_models/gaussian_diffusion/base_diffuser.py +12 -0
- {diffusion_model_framework-1.0.0rc1 → diffusion_model_framework-1.0.0rc2}/diffusion_models/gaussian_diffusion/ddimm_diffuser.py +63 -35
- {diffusion_model_framework-1.0.0rc1 → diffusion_model_framework-1.0.0rc2}/diffusion_models/gaussian_diffusion/gaussian_diffuser.py +22 -13
- {diffusion_model_framework-1.0.0rc1 → diffusion_model_framework-1.0.0rc2}/diffusion_models/utils/schemas.py +5 -0
- {diffusion_model_framework-1.0.0rc1 → diffusion_model_framework-1.0.0rc2}/pyproject.toml +1 -1
- diffusion_model_framework-1.0.0rc1/diffusion_models/__init__.py +0 -1
- {diffusion_model_framework-1.0.0rc1 → diffusion_model_framework-1.0.0rc2}/README.md +0 -0
- {diffusion_model_framework-1.0.0rc1 → diffusion_model_framework-1.0.0rc2}/diffusion_models/gaussian_diffusion/__init__.py +0 -0
- {diffusion_model_framework-1.0.0rc1 → diffusion_model_framework-1.0.0rc2}/diffusion_models/gaussian_diffusion/beta_schedulers.py +0 -0
- {diffusion_model_framework-1.0.0rc1 → diffusion_model_framework-1.0.0rc2}/diffusion_models/models/SimpleUnet.py +0 -0
- {diffusion_model_framework-1.0.0rc1 → diffusion_model_framework-1.0.0rc2}/diffusion_models/models/__init__.py +0 -0
- {diffusion_model_framework-1.0.0rc1 → diffusion_model_framework-1.0.0rc2}/diffusion_models/models/base_diffusion_model.py +0 -0
- {diffusion_model_framework-1.0.0rc1 → diffusion_model_framework-1.0.0rc2}/diffusion_models/utils/__init__.py +0 -0
- {diffusion_model_framework-1.0.0rc1 → diffusion_model_framework-1.0.0rc2}/diffusion_models/utils/tensorboard.py +0 -0
|
@@ -0,0 +1 @@
|
|
|
1
|
+
__version__ = "1.0.0-rc.2"
|
|
@@ -49,9 +49,7 @@ class DiffusionInference:
|
|
|
49
49
|
pil_images = transforms.ToPILImage()(image_grid)
|
|
50
50
|
return pil_images
|
|
51
51
|
|
|
52
|
-
def generate(
|
|
53
|
-
self, number_of_images: int, save_gif: bool = False
|
|
54
|
-
) -> Image:
|
|
52
|
+
def generate(self, number_of_images: int, save_gif: bool = False) -> Image:
|
|
55
53
|
"""Generate a batch of images.
|
|
56
54
|
|
|
57
55
|
Args:
|
|
@@ -117,8 +115,8 @@ class DiffusionInference:
|
|
|
117
115
|
device=self.device,
|
|
118
116
|
)
|
|
119
117
|
|
|
120
|
-
for i in
|
|
121
|
-
timestep =
|
|
118
|
+
for i in self.model.diffuser.steps:
|
|
119
|
+
timestep = self.model.diffuser.get_timestep(images.shape[0], i)
|
|
122
120
|
images = self.model.diffuser._denoise_step(
|
|
123
121
|
images, model=self.model, timestep=timestep
|
|
124
122
|
)
|
|
@@ -7,8 +7,12 @@ from torch.utils.data import DataLoader, Dataset
|
|
|
7
7
|
from tqdm import tqdm
|
|
8
8
|
|
|
9
9
|
from diffusion_models.models.base_diffusion_model import BaseDiffusionModel
|
|
10
|
-
from diffusion_models.utils.schemas import
|
|
11
|
-
|
|
10
|
+
from diffusion_models.utils.schemas import (
|
|
11
|
+
BetaSchedulerConfiguration,
|
|
12
|
+
Checkpoint,
|
|
13
|
+
LogConfiguration,
|
|
14
|
+
TrainingConfiguration,
|
|
15
|
+
)
|
|
12
16
|
from diffusion_models.utils.tensorboard import TensorboardManager
|
|
13
17
|
|
|
14
18
|
|
|
@@ -7,6 +7,7 @@ import torch
|
|
|
7
7
|
from diffusion_models.gaussian_diffusion.beta_schedulers import (
|
|
8
8
|
BaseBetaScheduler,
|
|
9
9
|
)
|
|
10
|
+
from diffusion_models.utils.schemas import Timestep
|
|
10
11
|
|
|
11
12
|
if TYPE_CHECKING:
|
|
12
13
|
from diffusion_models.models.base_diffusion_model import BaseDiffusionModel
|
|
@@ -32,6 +33,17 @@ class BaseDiffuser(abc.ABC):
|
|
|
32
33
|
self.beta_scheduler: BaseBetaScheduler = beta_scheduler
|
|
33
34
|
"""The beta scheduler used by the diffuser."""
|
|
34
35
|
|
|
36
|
+
@property
|
|
37
|
+
@abc.abstractmethod
|
|
38
|
+
def steps(self) -> List[int]:
|
|
39
|
+
"""Returns the list of steps used in the denoising process."""
|
|
40
|
+
raise NotImplementedError
|
|
41
|
+
|
|
42
|
+
@abc.abstractmethod
|
|
43
|
+
def get_timestep(self, number_of_images: int, idx: int) -> Timestep:
|
|
44
|
+
"""Get timestep information used for denoising."""
|
|
45
|
+
raise NotImplementedError
|
|
46
|
+
|
|
35
47
|
@abc.abstractmethod
|
|
36
48
|
def diffuse_batch(
|
|
37
49
|
self, images: torch.Tensor
|
|
@@ -1,22 +1,31 @@
|
|
|
1
1
|
from enum import Enum
|
|
2
|
-
from typing import List
|
|
2
|
+
from typing import List
|
|
3
|
+
from typing import Tuple
|
|
3
4
|
|
|
4
5
|
import numpy as np
|
|
5
6
|
import torch
|
|
6
7
|
from tqdm import tqdm
|
|
7
8
|
|
|
8
9
|
from diffusion_models.gaussian_diffusion.base_diffuser import BaseDiffuser
|
|
9
|
-
from diffusion_models.gaussian_diffusion.beta_schedulers import
|
|
10
|
-
BaseBetaScheduler
|
|
10
|
+
from diffusion_models.gaussian_diffusion.beta_schedulers import (
|
|
11
|
+
BaseBetaScheduler,
|
|
12
|
+
)
|
|
11
13
|
from diffusion_models.models.base_diffusion_model import BaseDiffusionModel
|
|
12
|
-
from diffusion_models.utils.schemas import Checkpoint
|
|
14
|
+
from diffusion_models.utils.schemas import Checkpoint, Timestep
|
|
15
|
+
|
|
13
16
|
|
|
14
17
|
class DenoisingMode(str, Enum):
|
|
15
18
|
Linear = "linear"
|
|
16
19
|
Quadratic = "quadratic"
|
|
17
20
|
|
|
21
|
+
|
|
18
22
|
class DdimDiffuser(BaseDiffuser):
|
|
19
|
-
def __init__(
|
|
23
|
+
def __init__(
|
|
24
|
+
self,
|
|
25
|
+
beta_scheduler: BaseBetaScheduler,
|
|
26
|
+
mode: DenoisingMode = DenoisingMode.Quadratic,
|
|
27
|
+
number_of_steps: int = 20,
|
|
28
|
+
):
|
|
20
29
|
"""Initializes the class instance.
|
|
21
30
|
|
|
22
31
|
Args:
|
|
@@ -24,9 +33,44 @@ class DdimDiffuser(BaseDiffuser):
|
|
|
24
33
|
|
|
25
34
|
"""
|
|
26
35
|
super().__init__(beta_scheduler)
|
|
36
|
+
|
|
37
|
+
self.number_of_steps = number_of_steps
|
|
38
|
+
"""Number of steps to use in the denoising process."""
|
|
39
|
+
|
|
40
|
+
self.mode = mode
|
|
41
|
+
"""Linear or Quadratic sampling."""
|
|
42
|
+
|
|
27
43
|
self.device: str = "cpu"
|
|
28
44
|
"""The device to use. Defaults to cpu."""
|
|
29
45
|
|
|
46
|
+
@property
|
|
47
|
+
def steps(self) -> List[int]:
|
|
48
|
+
if self.mode == DenoisingMode.Linear:
|
|
49
|
+
a = self.beta_scheduler.steps // self.number_of_steps
|
|
50
|
+
time_steps = np.asarray(list(range(0, self.beta_scheduler.steps, a)))
|
|
51
|
+
else:
|
|
52
|
+
time_steps = (
|
|
53
|
+
np.linspace(
|
|
54
|
+
0, np.sqrt(self.beta_scheduler.steps * 0.8), self.number_of_steps
|
|
55
|
+
)
|
|
56
|
+
** 2
|
|
57
|
+
).astype(int)
|
|
58
|
+
self._time_steps = time_steps + 1
|
|
59
|
+
self._time_steps_prev = np.concatenate([[0], time_steps[:-1]])
|
|
60
|
+
return list(range(self.number_of_steps))[::-1]
|
|
61
|
+
|
|
62
|
+
def get_timestep(self, number_of_images: int, idx: int) -> Timestep:
|
|
63
|
+
timestep = torch.full(
|
|
64
|
+
(number_of_images,), self._time_steps[idx], device=self.device
|
|
65
|
+
)
|
|
66
|
+
timestep_prev = torch.full(
|
|
67
|
+
(number_of_images,), self._time_steps_prev[idx], device=self.device
|
|
68
|
+
)
|
|
69
|
+
return Timestep(
|
|
70
|
+
current=timestep,
|
|
71
|
+
previous=timestep_prev,
|
|
72
|
+
)
|
|
73
|
+
|
|
30
74
|
@classmethod
|
|
31
75
|
def from_checkpoint(cls, checkpoint: Checkpoint) -> "DdimDiffuser":
|
|
32
76
|
"""Instantiate a DDIM Diffuser from a training checkpoint.
|
|
@@ -107,18 +151,20 @@ class DdimDiffuser(BaseDiffuser):
|
|
|
107
151
|
self,
|
|
108
152
|
images: torch.Tensor,
|
|
109
153
|
model: torch.nn.Module,
|
|
110
|
-
timestep:
|
|
111
|
-
timestep_prev: torch.Tensor,
|
|
154
|
+
timestep: Timestep,
|
|
112
155
|
eta: float = 0.0,
|
|
113
156
|
) -> torch.Tensor:
|
|
114
|
-
|
|
157
|
+
current_timestep = timestep.current
|
|
158
|
+
previous_timestep = timestep.previous
|
|
159
|
+
|
|
160
|
+
epsilon_theta = model(images, current_timestep)
|
|
115
161
|
|
|
116
162
|
alpha_bar_t = self.beta_scheduler.alpha_bars.gather(
|
|
117
|
-
dim=0, index=
|
|
163
|
+
dim=0, index=current_timestep
|
|
118
164
|
).reshape(-1, 1, 1, 1)
|
|
119
165
|
|
|
120
166
|
alpha_bar_t_prev = self.beta_scheduler.alpha_bars.gather(
|
|
121
|
-
dim=0, index=
|
|
167
|
+
dim=0, index=previous_timestep
|
|
122
168
|
).reshape(-1, 1, 1, 1)
|
|
123
169
|
|
|
124
170
|
sigma = eta * torch.sqrt(
|
|
@@ -134,7 +180,8 @@ class DdimDiffuser(BaseDiffuser):
|
|
|
134
180
|
+ (
|
|
135
181
|
torch.sqrt(1 - alpha_bar_t_prev - sigma**2)
|
|
136
182
|
- torch.sqrt((alpha_bar_t_prev * (1 - alpha_bar_t)) / alpha_bar_t)
|
|
137
|
-
)
|
|
183
|
+
)
|
|
184
|
+
* epsilon_theta
|
|
138
185
|
)
|
|
139
186
|
return mu + sigma * epsilon_t
|
|
140
187
|
|
|
@@ -142,8 +189,6 @@ class DdimDiffuser(BaseDiffuser):
|
|
|
142
189
|
self,
|
|
143
190
|
images: torch.Tensor,
|
|
144
191
|
model: "BaseDiffusionModel",
|
|
145
|
-
number_of_steps: int = 50,
|
|
146
|
-
mode: DenoisingMode = DenoisingMode.Linear
|
|
147
192
|
) -> List[torch.Tensor]:
|
|
148
193
|
"""Denoise a batch of images.
|
|
149
194
|
|
|
@@ -152,35 +197,18 @@ class DdimDiffuser(BaseDiffuser):
|
|
|
152
197
|
Args:
|
|
153
198
|
images: A batch of noisy images.
|
|
154
199
|
model: The model to be used for denoising.
|
|
155
|
-
number_of_steps: Number of steps used in the denoising process.
|
|
156
|
-
mode: Linear or Quadratic sampling.
|
|
157
200
|
|
|
158
201
|
Returns:
|
|
159
202
|
A list of tensors containing a batch of denoised images.
|
|
160
203
|
"""
|
|
161
|
-
if mode == DenoisingMode.Linear:
|
|
162
|
-
a = self.beta_scheduler.steps // number_of_steps
|
|
163
|
-
time_steps = np.asarray(list(range(0, self.beta_scheduler.steps, a)))
|
|
164
|
-
else:
|
|
165
|
-
time_steps = (
|
|
166
|
-
np.linspace(0, np.sqrt(self.beta_scheduler.steps * 0.8), number_of_steps)
|
|
167
|
-
** 2
|
|
168
|
-
).astype(int)
|
|
169
|
-
|
|
170
|
-
time_steps = time_steps + 1
|
|
171
|
-
time_steps_prev = np.concatenate([[0], time_steps[:-1]])
|
|
172
|
-
|
|
173
204
|
denoised_images = []
|
|
174
|
-
for i in tqdm(
|
|
175
|
-
timestep =
|
|
176
|
-
(images.shape[0],), time_steps[i], device=self.device
|
|
177
|
-
)
|
|
178
|
-
timestep_prev = torch.full(
|
|
179
|
-
(images.shape[0],), time_steps_prev[i], device=self.device
|
|
180
|
-
)
|
|
205
|
+
for i in tqdm(self.steps, desc="Denoising"):
|
|
206
|
+
timestep = self.get_timestep(images.shape[0], idx=i)
|
|
181
207
|
|
|
182
208
|
images = self._denoise_step(
|
|
183
|
-
images,
|
|
209
|
+
images,
|
|
210
|
+
model=model,
|
|
211
|
+
timestep=timestep,
|
|
184
212
|
)
|
|
185
213
|
images = torch.clamp(images, -1.0, 1.0)
|
|
186
214
|
denoised_images.append(images)
|
|
@@ -1,8 +1,5 @@
|
|
|
1
|
-
from typing import TYPE_CHECKING
|
|
2
|
-
from typing import List
|
|
3
|
-
from typing import Tuple
|
|
1
|
+
from typing import TYPE_CHECKING, List, Tuple
|
|
4
2
|
|
|
5
|
-
import numpy as np
|
|
6
3
|
import torch
|
|
7
4
|
from tqdm import tqdm
|
|
8
5
|
|
|
@@ -10,8 +7,7 @@ from diffusion_models.gaussian_diffusion.base_diffuser import BaseDiffuser
|
|
|
10
7
|
from diffusion_models.gaussian_diffusion.beta_schedulers import (
|
|
11
8
|
BaseBetaScheduler,
|
|
12
9
|
)
|
|
13
|
-
from diffusion_models.utils.schemas import Checkpoint
|
|
14
|
-
|
|
10
|
+
from diffusion_models.utils.schemas import Checkpoint, Timestep
|
|
15
11
|
|
|
16
12
|
if TYPE_CHECKING:
|
|
17
13
|
from diffusion_models.models.base_diffusion_model import BaseDiffusionModel
|
|
@@ -29,6 +25,16 @@ class GaussianDiffuser(BaseDiffuser):
|
|
|
29
25
|
self.device: str = "cpu"
|
|
30
26
|
"""The device to use. Defaults to cpu."""
|
|
31
27
|
|
|
28
|
+
@property
|
|
29
|
+
def steps(self) -> List[int]:
|
|
30
|
+
return list(range(self.beta_scheduler.steps))[::-1]
|
|
31
|
+
|
|
32
|
+
def get_timestep(self, number_of_images: int, idx: int):
|
|
33
|
+
timestep = torch.full((number_of_images,), idx, device=self.device)
|
|
34
|
+
return Timestep(
|
|
35
|
+
current=timestep,
|
|
36
|
+
)
|
|
37
|
+
|
|
32
38
|
@classmethod
|
|
33
39
|
def from_checkpoint(cls, checkpoint: Checkpoint) -> "GaussianDiffuser":
|
|
34
40
|
"""Instantiate a Gaussian Diffuser from a training checkpoint.
|
|
@@ -106,18 +112,21 @@ class GaussianDiffuser(BaseDiffuser):
|
|
|
106
112
|
|
|
107
113
|
@torch.no_grad()
|
|
108
114
|
def _denoise_step(
|
|
109
|
-
self, images: torch.Tensor, model: torch.nn.Module, timestep:
|
|
115
|
+
self, images: torch.Tensor, model: torch.nn.Module, timestep: Timestep
|
|
110
116
|
) -> torch.Tensor:
|
|
111
|
-
|
|
117
|
+
current_timestep = timestep.current
|
|
118
|
+
beta_t = self.beta_scheduler.betas[current_timestep].reshape(-1, 1, 1, 1)
|
|
112
119
|
alpha_t = 1 - beta_t
|
|
113
120
|
alpha_bar_t = self.beta_scheduler.alpha_bars.gather(
|
|
114
|
-
dim=0, index=
|
|
121
|
+
dim=0, index=current_timestep
|
|
115
122
|
).reshape(-1, 1, 1, 1)
|
|
116
123
|
mu = (1 / torch.sqrt(alpha_t)) * (
|
|
117
|
-
images
|
|
124
|
+
images
|
|
125
|
+
- model(images, current_timestep)
|
|
126
|
+
* (beta_t / torch.sqrt(1 - alpha_bar_t))
|
|
118
127
|
)
|
|
119
128
|
|
|
120
|
-
if
|
|
129
|
+
if current_timestep[0] == 0:
|
|
121
130
|
return mu
|
|
122
131
|
else:
|
|
123
132
|
sigma = torch.sqrt(beta_t) * torch.randn_like(images)
|
|
@@ -140,8 +149,8 @@ class GaussianDiffuser(BaseDiffuser):
|
|
|
140
149
|
A list of tensors containing a batch of denoised images.
|
|
141
150
|
"""
|
|
142
151
|
denoised_images = []
|
|
143
|
-
for i in tqdm(
|
|
144
|
-
timestep =
|
|
152
|
+
for i in tqdm(self.steps, desc="Denoising"):
|
|
153
|
+
timestep = self.get_timestep(images.shape[0], idx=i)
|
|
145
154
|
images = self._denoise_step(images, model=model, timestep=timestep)
|
|
146
155
|
images = torch.clamp(images, -1.0, 1.0)
|
|
147
156
|
denoised_images.append(images)
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
__version__ = "1.0.0-rc.1"
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|