├── LICENSE.txt
├── MODEL_LICENSE.txt
├── README.md
├── environment.yml
├── figs
├── model_framework.png
└── results.png
├── guided_diffusion
├── __init__.py
├── __pycache__
│ ├── __init__.cpython-37.pyc
│ ├── __init__.cpython-38.pyc
│ ├── __init__.cpython-39.pyc
│ ├── dist_util.cpython-37.pyc
│ ├── dist_util.cpython-38.pyc
│ ├── dist_util.cpython-39.pyc
│ ├── fp16_util.cpython-37.pyc
│ ├── fp16_util.cpython-39.pyc
│ ├── gaussian_diffusion.cpython-37.pyc
│ ├── gaussian_diffusion.cpython-39.pyc
│ ├── image_datasets.cpython-37.pyc
│ ├── image_datasets.cpython-39.pyc
│ ├── logger.cpython-37.pyc
│ ├── logger.cpython-39.pyc
│ ├── losses.cpython-37.pyc
│ ├── losses.cpython-39.pyc
│ ├── nn.cpython-37.pyc
│ ├── nn.cpython-39.pyc
│ ├── resample.cpython-37.pyc
│ ├── resample.cpython-39.pyc
│ ├── respace.cpython-37.pyc
│ ├── respace.cpython-39.pyc
│ ├── script_util.cpython-37.pyc
│ ├── script_util.cpython-39.pyc
│ ├── train_util.cpython-37.pyc
│ ├── train_util.cpython-39.pyc
│ ├── unet.cpython-37.pyc
│ └── unet.cpython-39.pyc
├── dist_util.py
├── fp16_util.py
├── gaussian_diffusion.py
├── image_datasets.py
├── logger.py
├── losses.py
├── nn.py
├── resample.py
├── respace.py
├── script_util.py
├── train_util.py
└── unet.py
├── scripts
├── test.py
└── train.py
└── test_DDPM_3d_mpi.sh
/LICENSE.txt:
--------------------------------------------------------------------------------
1 | MIT License
2 |
3 | Copyright (c) 2025 Boxiao Yu and Kuang Gong
4 |
5 | Permission is hereby granted, free of charge, to any person obtaining a copy
6 | of this software and associated documentation files (the "Software"), to deal
7 | in the Software without restriction, including without limitation the rights
8 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9 | copies of the Software, and to permit persons to whom the Software is
10 | furnished to do so, subject to the following conditions:
11 |
12 | The above copyright notice and this permission notice shall be included in all
13 | copies or substantial portions of the Software.
14 |
15 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 | SOFTWARE.
22 |
--------------------------------------------------------------------------------
/MODEL_LICENSE.txt:
--------------------------------------------------------------------------------
1 | Attribution-NonCommercial-ShareAlike 4.0 International
2 |
3 | =======================================================================
4 |
5 | Creative Commons Corporation ("Creative Commons") is not a law firm and
6 | does not provide legal services or legal advice. Distribution of
7 | Creative Commons public licenses does not create a lawyer-client or
8 | other relationship. Creative Commons makes its licenses and related
9 | information available on an "as-is" basis. Creative Commons gives no
10 | warranties regarding its licenses, any material licensed under their
11 | terms and conditions, or any related information. Creative Commons
12 | disclaims all liability for damages resulting from their use to the
13 | fullest extent possible.
14 |
15 | Using Creative Commons Public Licenses
16 |
17 | Creative Commons public licenses provide a standard set of terms and
18 | conditions that creators and other rights holders may use to share
19 | original works of authorship and other material subject to copyright
20 | and certain other rights specified in the public license below. The
21 | following considerations are for informational purposes only, are not
22 | exhaustive, and do not form part of our licenses.
23 |
24 | Considerations for licensors: Our public licenses are
25 | intended for use by those authorized to give the public
26 | permission to use material in ways otherwise restricted by
27 | copyright and certain other rights. Our licenses are
28 | irrevocable. Licensors should read and understand the terms
29 | and conditions of the license they choose before applying it.
30 | Licensors should also secure all rights necessary before
31 | applying our licenses so that the public can reuse the
32 | material as expected. Licensors should clearly mark any
33 | material not subject to the license. This includes other CC-
34 | licensed material, or material used under an exception or
35 | limitation to copyright. More considerations for licensors:
36 | wiki.creativecommons.org/Considerations_for_licensors
37 |
38 | Considerations for the public: By using one of our public
39 | licenses, a licensor grants the public permission to use the
40 | licensed material under specified terms and conditions. If
41 | the licensor's permission is not necessary for any reason--for
42 | example, because of any applicable exception or limitation to
43 | copyright--then that use is not regulated by the license. Our
44 | licenses grant only permissions under copyright and certain
45 | other rights that a licensor has authority to grant. Use of
46 | the licensed material may still be restricted for other
47 | reasons, including because others have copyright or other
48 | rights in the material. A licensor may make special requests,
49 | such as asking that all changes be marked or described.
50 | Although not required by our licenses, you are encouraged to
51 | respect those requests where reasonable. More considerations
52 | for the public:
53 | wiki.creativecommons.org/Considerations_for_licensees
54 |
55 | =======================================================================
56 |
57 | Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International
58 | Public License
59 |
60 | By exercising the Licensed Rights (defined below), You accept and agree
61 | to be bound by the terms and conditions of this Creative Commons
62 | Attribution-NonCommercial-ShareAlike 4.0 International Public License
63 | ("Public License"). To the extent this Public License may be
64 | interpreted as a contract, You are granted the Licensed Rights in
65 | consideration of Your acceptance of these terms and conditions, and the
66 | Licensor grants You such rights in consideration of benefits the
67 | Licensor receives from making the Licensed Material available under
68 | these terms and conditions.
69 |
70 |
71 | Section 1 -- Definitions.
72 |
73 | a. Adapted Material means material subject to Copyright and Similar
74 | Rights that is derived from or based upon the Licensed Material
75 | and in which the Licensed Material is translated, altered,
76 | arranged, transformed, or otherwise modified in a manner requiring
77 | permission under the Copyright and Similar Rights held by the
78 | Licensor. For purposes of this Public License, where the Licensed
79 | Material is a musical work, performance, or sound recording,
80 | Adapted Material is always produced where the Licensed Material is
81 | synched in timed relation with a moving image.
82 |
83 | b. Adapter's License means the license You apply to Your Copyright
84 | and Similar Rights in Your contributions to Adapted Material in
85 | accordance with the terms and conditions of this Public License.
86 |
87 | c. BY-NC-SA Compatible License means a license listed at
88 | creativecommons.org/compatiblelicenses, approved by Creative
89 | Commons as essentially the equivalent of this Public License.
90 |
91 | d. Copyright and Similar Rights means copyright and/or similar rights
92 | closely related to copyright including, without limitation,
93 | performance, broadcast, sound recording, and Sui Generis Database
94 | Rights, without regard to how the rights are labeled or
95 | categorized. For purposes of this Public License, the rights
96 | specified in Section 2(b)(1)-(2) are not Copyright and Similar
97 | Rights.
98 |
99 | e. Effective Technological Measures means those measures that, in the
100 | absence of proper authority, may not be circumvented under laws
101 | fulfilling obligations under Article 11 of the WIPO Copyright
102 | Treaty adopted on December 20, 1996, and/or similar international
103 | agreements.
104 |
105 | f. Exceptions and Limitations means fair use, fair dealing, and/or
106 | any other exception or limitation to Copyright and Similar Rights
107 | that applies to Your use of the Licensed Material.
108 |
109 | g. License Elements means the license attributes listed in the name
110 | of a Creative Commons Public License. The License Elements of this
111 | Public License are Attribution, NonCommercial, and ShareAlike.
112 |
113 | h. Licensed Material means the artistic or literary work, database,
114 | or other material to which the Licensor applied this Public
115 | License.
116 |
117 | i. Licensed Rights means the rights granted to You subject to the
118 | terms and conditions of this Public License, which are limited to
119 | all Copyright and Similar Rights that apply to Your use of the
120 | Licensed Material and that the Licensor has authority to license.
121 |
122 | j. Licensor means the individual(s) or entity(ies) granting rights
123 | under this Public License.
124 |
125 | k. NonCommercial means not primarily intended for or directed towards
126 | commercial advantage or monetary compensation. For purposes of
127 | this Public License, the exchange of the Licensed Material for
128 | other material subject to Copyright and Similar Rights by digital
129 | file-sharing or similar means is NonCommercial provided there is
130 | no payment of monetary compensation in connection with the
131 | exchange.
132 |
133 | l. Share means to provide material to the public by any means or
134 | process that requires permission under the Licensed Rights, such
135 | as reproduction, public display, public performance, distribution,
136 | dissemination, communication, or importation, and to make material
137 | available to the public including in ways that members of the
138 | public may access the material from a place and at a time
139 | individually chosen by them.
140 |
141 | m. Sui Generis Database Rights means rights other than copyright
142 | resulting from Directive 96/9/EC of the European Parliament and of
143 | the Council of 11 March 1996 on the legal protection of databases,
144 | as amended and/or succeeded, as well as other essentially
145 | equivalent rights anywhere in the world.
146 |
147 | n. You means the individual or entity exercising the Licensed Rights
148 | under this Public License. Your has a corresponding meaning.
149 |
150 |
151 | Section 2 -- Scope.
152 |
153 | a. License grant.
154 |
155 | 1. Subject to the terms and conditions of this Public License,
156 | the Licensor hereby grants You a worldwide, royalty-free,
157 | non-sublicensable, non-exclusive, irrevocable license to
158 | exercise the Licensed Rights in the Licensed Material to:
159 |
160 | a. reproduce and Share the Licensed Material, in whole or
161 | in part, for NonCommercial purposes only; and
162 |
163 | b. produce, reproduce, and Share Adapted Material for
164 | NonCommercial purposes only.
165 |
166 | 2. Exceptions and Limitations. For the avoidance of doubt, where
167 | Exceptions and Limitations apply to Your use, this Public
168 | License does not apply, and You do not need to comply with
169 | its terms and conditions.
170 |
171 | 3. Term. The term of this Public License is specified in Section
172 | 6(a).
173 |
174 | 4. Media and formats; technical modifications allowed. The
175 | Licensor authorizes You to exercise the Licensed Rights in
176 | all media and formats whether now known or hereafter created,
177 | and to make technical modifications necessary to do so. The
178 | Licensor waives and/or agrees not to assert any right or
179 | authority to forbid You from making technical modifications
180 | necessary to exercise the Licensed Rights, including
181 | technical modifications necessary to circumvent Effective
182 | Technological Measures. For purposes of this Public License,
183 | simply making modifications authorized by this Section 2(a)
184 | (4) never produces Adapted Material.
185 |
186 | 5. Downstream recipients.
187 |
188 | a. Offer from the Licensor -- Licensed Material. Every
189 | recipient of the Licensed Material automatically
190 | receives an offer from the Licensor to exercise the
191 | Licensed Rights under the terms and conditions of this
192 | Public License.
193 |
194 | b. Additional offer from the Licensor -- Adapted Material.
195 | Every recipient of Adapted Material from You
196 | automatically receives an offer from the Licensor to
197 | exercise the Licensed Rights in the Adapted Material
198 | under the conditions of the Adapter's License You apply.
199 |
200 | c. No downstream restrictions. You may not offer or impose
201 | any additional or different terms or conditions on, or
202 | apply any Effective Technological Measures to, the
203 | Licensed Material if doing so restricts exercise of the
204 | Licensed Rights by any recipient of the Licensed
205 | Material.
206 |
207 | 6. No endorsement. Nothing in this Public License constitutes or
208 | may be construed as permission to assert or imply that You
209 | are, or that Your use of the Licensed Material is, connected
210 | with, or sponsored, endorsed, or granted official status by,
211 | the Licensor or others designated to receive attribution as
212 | provided in Section 3(a)(1)(A)(i).
213 |
214 | b. Other rights.
215 |
216 | 1. Moral rights, such as the right of integrity, are not
217 | licensed under this Public License, nor are publicity,
218 | privacy, and/or other similar personality rights; however, to
219 | the extent possible, the Licensor waives and/or agrees not to
220 | assert any such rights held by the Licensor to the limited
221 | extent necessary to allow You to exercise the Licensed
222 | Rights, but not otherwise.
223 |
224 | 2. Patent and trademark rights are not licensed under this
225 | Public License.
226 |
227 | 3. To the extent possible, the Licensor waives any right to
228 | collect royalties from You for the exercise of the Licensed
229 | Rights, whether directly or through a collecting society
230 | under any voluntary or waivable statutory or compulsory
231 | licensing scheme. In all other cases the Licensor expressly
232 | reserves any right to collect such royalties, including when
233 | the Licensed Material is used other than for NonCommercial
234 | purposes.
235 |
236 |
237 | Section 3 -- License Conditions.
238 |
239 | Your exercise of the Licensed Rights is expressly made subject to the
240 | following conditions.
241 |
242 | a. Attribution.
243 |
244 | 1. If You Share the Licensed Material (including in modified
245 | form), You must:
246 |
247 | a. retain the following if it is supplied by the Licensor
248 | with the Licensed Material:
249 |
250 | i. identification of the creator(s) of the Licensed
251 | Material and any others designated to receive
252 | attribution, in any reasonable manner requested by
253 | the Licensor (including by pseudonym if
254 | designated);
255 |
256 | ii. a copyright notice;
257 |
258 | iii. a notice that refers to this Public License;
259 |
260 | iv. a notice that refers to the disclaimer of
261 | warranties;
262 |
263 | v. a URI or hyperlink to the Licensed Material to the
264 | extent reasonably practicable;
265 |
266 | b. indicate if You modified the Licensed Material and
267 | retain an indication of any previous modifications; and
268 |
269 | c. indicate the Licensed Material is licensed under this
270 | Public License, and include the text of, or the URI or
271 | hyperlink to, this Public License.
272 |
273 | 2. You may satisfy the conditions in Section 3(a)(1) in any
274 | reasonable manner based on the medium, means, and context in
275 | which You Share the Licensed Material. For example, it may be
276 | reasonable to satisfy the conditions by providing a URI or
277 | hyperlink to a resource that includes the required
278 | information.
279 | 3. If requested by the Licensor, You must remove any of the
280 | information required by Section 3(a)(1)(A) to the extent
281 | reasonably practicable.
282 |
283 | b. ShareAlike.
284 |
285 | In addition to the conditions in Section 3(a), if You Share
286 | Adapted Material You produce, the following conditions also apply.
287 |
288 | 1. The Adapter's License You apply must be a Creative Commons
289 | license with the same License Elements, this version or
290 | later, or a BY-NC-SA Compatible License.
291 |
292 | 2. You must include the text of, or the URI or hyperlink to, the
293 | Adapter's License You apply. You may satisfy this condition
294 | in any reasonable manner based on the medium, means, and
295 | context in which You Share Adapted Material.
296 |
297 | 3. You may not offer or impose any additional or different terms
298 | or conditions on, or apply any Effective Technological
299 | Measures to, Adapted Material that restrict exercise of the
300 | rights granted under the Adapter's License You apply.
301 |
302 |
303 | Section 4 -- Sui Generis Database Rights.
304 |
305 | Where the Licensed Rights include Sui Generis Database Rights that
306 | apply to Your use of the Licensed Material:
307 |
308 | a. for the avoidance of doubt, Section 2(a)(1) grants You the right
309 | to extract, reuse, reproduce, and Share all or a substantial
310 | portion of the contents of the database for NonCommercial purposes
311 | only;
312 |
313 | b. if You include all or a substantial portion of the database
314 | contents in a database in which You have Sui Generis Database
315 | Rights, then the database in which You have Sui Generis Database
316 | Rights (but not its individual contents) is Adapted Material,
317 | including for purposes of Section 3(b); and
318 |
319 | c. You must comply with the conditions in Section 3(a) if You Share
320 | all or a substantial portion of the contents of the database.
321 |
322 | For the avoidance of doubt, this Section 4 supplements and does not
323 | replace Your obligations under this Public License where the Licensed
324 | Rights include other Copyright and Similar Rights.
325 |
326 |
327 | Section 5 -- Disclaimer of Warranties and Limitation of Liability.
328 |
329 | a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE
330 | EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS
331 | AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF
332 | ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,
333 | IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,
334 | WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR
335 | PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,
336 | ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT
337 | KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT
338 | ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.
339 |
340 | b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE
341 | TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,
342 | NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,
343 | INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,
344 | COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR
345 | USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN
346 | ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR
347 | DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR
348 | IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.
349 |
350 | c. The disclaimer of warranties and limitation of liability provided
351 | above shall be interpreted in a manner that, to the extent
352 | possible, most closely approximates an absolute disclaimer and
353 | waiver of all liability.
354 |
355 |
356 | Section 6 -- Term and Termination.
357 |
358 | a. This Public License applies for the term of the Copyright and
359 | Similar Rights licensed here. However, if You fail to comply with
360 | this Public License, then Your rights under this Public License
361 | terminate automatically.
362 |
363 | b. Where Your right to use the Licensed Material has terminated under
364 | Section 6(a), it reinstates:
365 |
366 | 1. automatically as of the date the violation is cured, provided
367 | it is cured within 30 days of Your discovery of the
368 | violation; or
369 |
370 | 2. upon express reinstatement by the Licensor.
371 |
372 | For the avoidance of doubt, this Section 6(b) does not affect any
373 | right the Licensor may have to seek remedies for Your violations
374 | of this Public License.
375 |
376 | c. For the avoidance of doubt, the Licensor may also offer the
377 | Licensed Material under separate terms or conditions or stop
378 | distributing the Licensed Material at any time; however, doing so
379 | will not terminate this Public License.
380 |
381 | d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
382 | License.
383 |
384 |
385 | Section 7 -- Other Terms and Conditions.
386 |
387 | a. The Licensor shall not be bound by any additional or different
388 | terms or conditions communicated by You unless expressly agreed.
389 |
390 | b. Any arrangements, understandings, or agreements regarding the
391 | Licensed Material not stated herein are separate from and
392 | independent of the terms and conditions of this Public License.
393 |
394 |
395 | Section 8 -- Interpretation.
396 |
397 | a. For the avoidance of doubt, this Public License does not, and
398 | shall not be interpreted to, reduce, limit, restrict, or impose
399 | conditions on any use of the Licensed Material that could lawfully
400 | be made without permission under this Public License.
401 |
402 | b. To the extent possible, if any provision of this Public License is
403 | deemed unenforceable, it shall be automatically reformed to the
404 | minimum extent necessary to make it enforceable. If the provision
405 | cannot be reformed, it shall be severed from this Public License
406 | without affecting the enforceability of the remaining terms and
407 | conditions.
408 |
409 | c. No term or condition of this Public License will be waived and no
410 | failure to comply consented to unless expressly agreed to by the
411 | Licensor.
412 |
413 | d. Nothing in this Public License constitutes or may be interpreted
414 | as a limitation upon, or waiver of, any privileges and immunities
415 | that apply to the Licensor or You, including from the legal
416 | processes of any jurisdiction or authority.
417 |
418 | =======================================================================
419 |
420 | Creative Commons is not a party to its public
421 | licenses. Notwithstanding, Creative Commons may elect to apply one of
422 | its public licenses to material it publishes and in those instances
423 | will be considered the “Licensor.” The text of the Creative Commons
424 | public licenses is dedicated to the public domain under the CC0 Public
425 | Domain Dedication. Except for the limited purpose of indicating that
426 | material is shared under a Creative Commons public license or as
427 | otherwise permitted by the Creative Commons policies published at
428 | creativecommons.org/policies, Creative Commons does not authorize the
429 | use of the trademark "Creative Commons" or any other trademark or logo
430 | of Creative Commons without its prior written consent including,
431 | without limitation, in connection with any unauthorized modifications
432 | to any of its public licenses or any other arrangements,
433 | understandings, or agreements concerning use of licensed material. For
434 | the avoidance of doubt, this paragraph does not form part of the
435 | public licenses.
436 |
437 | Creative Commons may be contacted at creativecommons.org.
438 |
--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
1 |
2 |
3 |
4 | Boxiao Yu1,
5 |
6 | Savas Ozdemir2,
7 |
8 | Yafei Dong3,
9 |
10 | Wei Shao4,
11 |
12 | Tinsu Pan5,
13 |
14 | Kuangyu Shi6,
15 |
16 | Kuang Gong1
17 |
18 |
19 | 1J. Crayton Pruitt Family Department of Biomedical Engineering,
20 | University of Florida;
21 | 2Department of Raiology, University of Florida;
22 | 3Yale School of Medicine, Yale University;
23 | 4Department of Medicine, University of Florida;
24 | 5Department of Imaging Physics, University of Texas MD Anderson Cancer Center;
25 | 6Department of Nuclear Medicine, University of Bern
26 |
27 |
28 | ## Purpose
29 |
30 | Whole-body PET imaging plays an essential role in cancer diagnosis and treatment but suffers from low image quality. Traditional deep learning-based denoising methods work well for a specific acquisition but are less effective in handling diverse PET protocols. In this study, we proposed and validated a 3D Denoising Diffusion Probabilistic Model (3D DDPM) as a robust and universal solution for whole-body PET image denoising.
31 |
32 | ## Method
33 |
34 |
35 |
36 |
37 | The proposed 3D DDPM gradually injected noise into the images during the forward diffusion phase, allowing the model to learn to reconstruct the clean data during the reverse diffusion process. A 3D convolutional network was trained using high-quality data from the Biograph Vision Quadra PET/CT scanner to generate the score function, enabling the model to capture accurate PET distribution information extracted from the total-body datasets. The trained 3D DDPM was evaluated on datasets from four scanners, four tracer types, and six dose levels representing a broad spectrum of clinical scenarios.
38 |
39 | ## Result
40 |
41 |
42 |
43 |
44 | The proposed 3D DDPM consistently outperformed 2D DDPM, 3D UNet, and 3D GAN, demonstrating its superior denoising performance across all tested conditions. Additionally, the model’s uncertainty maps exhibited lower variance, reflecting its higher confidence in its outputs.
45 |
46 | ## Installation
47 | ### Step 1: Clone the Repository
48 |
49 | git clone https://github.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model.git
50 | cd PET-Image-Denoising-Using-3D-Diffusion-Model
51 |
52 | ### Step 2: Create and activate the conda environment from the environment.yml file:
53 |
54 | conda env create -f environment.yml
55 | conda activate PET-3D-DDPM
56 |
57 | ### Step 3: Download Pre-trained Models
58 | Download the pre-trained model files from this [link](https://www.dropbox.com/scl/fo/nj52fz7p23icnkxo3v5y2/AAsggV-0DAuJjd4ILYAE1m4?rlkey=uivlrx0oi68l7n34fkbmamkdj&st=fztnoohh&dl=0) and place them into the `./checkpoint/` folder.
59 |
60 | **Note**: This model is licensed under CC BY-NC-SA 4.0. **Commercial use is prohibited.**
61 |
62 |
63 | ## Testing
64 |
65 | ### Data Preparation
66 |
67 | Before running the denoising script, modify the `load_data_for_worker` function in `./scripts/test.py` to align with your data format and dimensions. This function is responsible for loading your low-dose PET data into the model.
68 |
69 | ### Running the Denoising Script
70 |
71 | We provide a shell script `test_DDPM_3d_mpi.sh` to facilitate the testing process.
72 |
73 | #### Usage
74 |
75 | - `--base_samples`: Path to the `.npz` files containing your low-dose PET images.
76 | - `--num_samples`: Total number of samples you wish to process.
77 | - `-n`: Number of GPUs to utilize for parallel processing.
78 | - `--save_dir`: Path to the directory where you want to save the denoised images.
79 |
80 | ## License
81 |
82 | - The **code** in this repository is licensed under the **MIT License**.
83 | - The **model weights** are licensed under **CC BY-NC-SA 4.0**, meaning:
84 | - You **can** share and modify the model weights, but **must** use the same license.
85 | - You **cannot** use it for **commercial purposes**.
86 |
87 | For details, check:
88 | - [LICENSE](./LICENSE.txt) (MIT for code)
89 | - [MODEL_LICENSE](./MODEL_LICENSE.txt) (CC BY-NC-SA 4.0 for model weights)
90 |
91 | ## Citation
92 | If you find our work is useful in your research or applications, please consider giving us a star 🌟 and citing it by the following BibTeX entry.
93 |
94 | ```bibtex
95 | @article{yu2025robust,
96 | title={Robust whole-body PET image denoising using 3D diffusion models: evaluation across various scanners, tracers, and dose levels},
97 | author={Yu, Boxiao and Ozdemir, Savas and Dong, Yafei and Shao, Wei and Pan, Tinsu and Shi, Kuangyu and Gong, Kuang},
98 | journal={European Journal of Nuclear Medicine and Molecular Imaging},
99 | pages={1--14},
100 | year={2025},
101 | publisher={Springer}
102 | }
103 | ```
104 |
105 | ## Contact
106 |
107 | For any questions or inquiries, please contact boxiao.yu@ufl.edu.
108 |
109 |
--------------------------------------------------------------------------------
/environment.yml:
--------------------------------------------------------------------------------
1 | name: PET-3D-DDPM
2 | channels:
3 | - conda-forge
4 | - anaconda
5 | - defaults
6 | dependencies:
7 | - _libgcc_mutex=0.1=main
8 | - _openmp_mutex=5.1=1_gnu
9 | - blas=1.0=mkl
10 | - ca-certificates=2023.08.22=h06a4308_0
11 | - certifi=2022.12.7=py37h06a4308_0
12 | - freetype=2.12.1=h4a9f257_0
13 | - giflib=5.2.1=h5eee18b_3
14 | - intel-openmp=2021.4.0=h06a4308_3561
15 | - jpeg=9e=h5eee18b_1
16 | - lcms2=2.12=h3be6417_0
17 | - ld_impl_linux-64=2.38=h1181459_1
18 | - lerc=3.0=h295c915_0
19 | - libdeflate=1.17=h5eee18b_1
20 | - libffi=3.4.4=h6a678d5_0
21 | - libgcc-ng=11.2.0=h1234567_1
22 | - libgfortran-ng=7.5.0=h14aa051_20
23 | - libgfortran4=7.5.0=h14aa051_20
24 | - libgomp=11.2.0=h1234567_1
25 | - libpng=1.6.39=h5eee18b_0
26 | - libstdcxx-ng=11.2.0=h1234567_1
27 | - libtiff=4.5.1=h6a678d5_0
28 | - libwebp=1.2.4=h11a3e52_1
29 | - libwebp-base=1.2.4=h5eee18b_1
30 | - lz4-c=1.9.4=h6a678d5_0
31 | - mkl=2021.4.0=h06a4308_640
32 | - mkl-service=2.4.0=py37h7f8727e_0
33 | - mkl_fft=1.3.1=py37hd3c417c_0
34 | - mkl_random=1.2.2=py37h51133e4_0
35 | - mpi=1.0=openmpi
36 | - mpi4py=3.1.4=py37h3e5f7c9_0
37 | - ncurses=6.4=h6a678d5_0
38 | - numpy=1.21.5=py37h6c91a56_3
39 | - numpy-base=1.21.5=py37ha15fc14_3
40 | - openmpi=4.0.4=hdf1f1ad_0
41 | - openssl=1.1.1w=h7f8727e_0
42 | - pillow=9.4.0=py37h6a678d5_0
43 | - python=3.7.16=h7a1cb2a_0
44 | - readline=8.2=h5eee18b_0
45 | - setuptools=65.6.3=py37h06a4308_0
46 | - six=1.16.0=pyhd3eb1b0_1
47 | - sqlite=3.41.2=h5eee18b_0
48 | - tk=8.6.12=h1ccaba5_0
49 | - wheel=0.38.4=py37h06a4308_0
50 | - xz=5.4.2=h5eee18b_0
51 | - zlib=1.2.13=h5eee18b_0
52 | - zstd=1.5.5=hc292b87_0
53 | - pip:
54 | - blobfile==2.0.2
55 | - filelock==3.12.2
56 | - lxml==4.9.3
57 | - nvidia-cublas-cu11==11.10.3.66
58 | - nvidia-cuda-nvrtc-cu11==11.7.99
59 | - nvidia-cuda-runtime-cu11==11.7.99
60 | - nvidia-cudnn-cu11==8.5.0.96
61 | - pip==23.3
62 | - pycryptodomex==3.19.0
63 | - torch==1.13.1
64 | - tqdm==4.66.1
65 | - typing-extensions==4.7.1
66 | - urllib3==2.0.7
67 |
68 |
--------------------------------------------------------------------------------
/figs/model_framework.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/figs/model_framework.png
--------------------------------------------------------------------------------
/figs/results.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/figs/results.png
--------------------------------------------------------------------------------
/guided_diffusion/__init__.py:
--------------------------------------------------------------------------------
1 | """
2 | Codebase for "Improved Denoising Diffusion Probabilistic Models".
3 | """
4 |
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/__init__.cpython-37.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/__init__.cpython-37.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/__init__.cpython-38.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/__init__.cpython-38.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/__init__.cpython-39.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/__init__.cpython-39.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/dist_util.cpython-37.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/dist_util.cpython-37.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/dist_util.cpython-38.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/dist_util.cpython-38.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/dist_util.cpython-39.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/dist_util.cpython-39.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/fp16_util.cpython-37.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/fp16_util.cpython-37.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/fp16_util.cpython-39.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/fp16_util.cpython-39.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/gaussian_diffusion.cpython-37.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/gaussian_diffusion.cpython-37.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/gaussian_diffusion.cpython-39.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/gaussian_diffusion.cpython-39.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/image_datasets.cpython-37.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/image_datasets.cpython-37.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/image_datasets.cpython-39.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/image_datasets.cpython-39.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/logger.cpython-37.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/logger.cpython-37.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/logger.cpython-39.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/logger.cpython-39.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/losses.cpython-37.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/losses.cpython-37.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/losses.cpython-39.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/losses.cpython-39.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/nn.cpython-37.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/nn.cpython-37.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/nn.cpython-39.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/nn.cpython-39.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/resample.cpython-37.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/resample.cpython-37.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/resample.cpython-39.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/resample.cpython-39.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/respace.cpython-37.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/respace.cpython-37.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/respace.cpython-39.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/respace.cpython-39.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/script_util.cpython-37.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/script_util.cpython-37.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/script_util.cpython-39.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/script_util.cpython-39.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/train_util.cpython-37.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/train_util.cpython-37.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/train_util.cpython-39.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/train_util.cpython-39.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/unet.cpython-37.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/unet.cpython-37.pyc
--------------------------------------------------------------------------------
/guided_diffusion/__pycache__/unet.cpython-39.pyc:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/Miche11eU/PET-Image-Denoising-Using-3D-Diffusion-Model/dd4bf6232e7adc9a998eafa769e1482c8075b1fb/guided_diffusion/__pycache__/unet.cpython-39.pyc
--------------------------------------------------------------------------------
/guided_diffusion/dist_util.py:
--------------------------------------------------------------------------------
1 | """
2 | Helpers for distributed training.
3 | """
4 |
5 | import io
6 | import os
7 | import socket
8 |
9 | import blobfile as bf
10 | from mpi4py import MPI
11 | import torch as th
12 | import torch.distributed as dist
13 |
14 | # Change this to reflect your cluster layout.
15 | # The GPU for a given rank is (rank % GPUS_PER_NODE).
16 | GPUS_PER_NODE = 8
17 |
18 | SETUP_RETRY_COUNT = 3
19 |
20 |
21 | def setup_dist():
22 | """
23 | Setup a distributed process group.
24 | """
25 | if dist.is_initialized():
26 | return
27 | os.environ["CUDA_VISIBLE_DEVICES"] = f"{MPI.COMM_WORLD.Get_rank() % GPUS_PER_NODE}"
28 | # os.environ['CUDA_VISIBLE_DEVICES'] = "3"
29 |
30 | comm = MPI.COMM_WORLD
31 | backend = "gloo" if not th.cuda.is_available() else "nccl"
32 |
33 | if backend == "gloo":
34 | hostname = "localhost"
35 | else:
36 | hostname = socket.gethostbyname(socket.getfqdn())
37 | os.environ["MASTER_ADDR"] = comm.bcast(hostname, root=0)
38 | os.environ["RANK"] = str(comm.rank)
39 | os.environ["WORLD_SIZE"] = str(comm.size)
40 |
41 | port = comm.bcast(_find_free_port(), root=0)
42 | os.environ["MASTER_PORT"] = str(port)
43 | dist.init_process_group(backend=backend, init_method="env://")
44 |
45 |
46 | def dev():
47 | """
48 | Get the device to use for torch.distributed.
49 | """
50 | if th.cuda.is_available():
51 | return th.device(f"cuda")
52 | return th.device("cpu")
53 |
54 |
55 | def load_state_dict(path, **kwargs):
56 | """
57 | Load a PyTorch file without redundant fetches across MPI ranks.
58 | """
59 | chunk_size = 2 ** 30 # MPI has a relatively small size limit
60 | if MPI.COMM_WORLD.Get_rank() == 0:
61 | with bf.BlobFile(path, "rb") as f:
62 | data = f.read()
63 | num_chunks = len(data) // chunk_size
64 | if len(data) % chunk_size:
65 | num_chunks += 1
66 | MPI.COMM_WORLD.bcast(num_chunks)
67 | for i in range(0, len(data), chunk_size):
68 | MPI.COMM_WORLD.bcast(data[i : i + chunk_size])
69 | else:
70 | num_chunks = MPI.COMM_WORLD.bcast(None)
71 | data = bytes()
72 | for _ in range(num_chunks):
73 | data += MPI.COMM_WORLD.bcast(None)
74 |
75 | return th.load(io.BytesIO(data), **kwargs)
76 |
77 |
78 | def sync_params(params):
79 | """
80 | Synchronize a sequence of Tensors across ranks from rank 0.
81 | """
82 | for p in params:
83 | with th.no_grad():
84 | dist.broadcast(p, 0)
85 |
86 |
87 | def _find_free_port():
88 | try:
89 | s = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
90 | s.bind(("", 0))
91 | s.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
92 | return s.getsockname()[1]
93 | finally:
94 | s.close()
95 |
--------------------------------------------------------------------------------
/guided_diffusion/fp16_util.py:
--------------------------------------------------------------------------------
1 | """
2 | Helpers to train with 16-bit precision.
3 | """
4 |
5 | import numpy as np
6 | import torch as th
7 | import torch.nn as nn
8 | from torch._utils import _flatten_dense_tensors, _unflatten_dense_tensors
9 |
10 | from . import logger
11 |
12 | INITIAL_LOG_LOSS_SCALE = 20.0
13 |
14 |
15 | def convert_module_to_f16(l):
16 | """
17 | Convert primitive modules to float16.
18 | """
19 | if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Conv3d)):
20 | l.weight.data = l.weight.data.half()
21 | if l.bias is not None:
22 | l.bias.data = l.bias.data.half()
23 |
24 |
25 | def convert_module_to_f32(l):
26 | """
27 | Convert primitive modules to float32, undoing convert_module_to_f16().
28 | """
29 | if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Conv3d)):
30 | l.weight.data = l.weight.data.float()
31 | if l.bias is not None:
32 | l.bias.data = l.bias.data.float()
33 |
34 |
35 | def make_master_params(param_groups_and_shapes):
36 | """
37 | Copy model parameters into a (differently-shaped) list of full-precision
38 | parameters.
39 | """
40 | master_params = []
41 | for param_group, shape in param_groups_and_shapes:
42 | master_param = nn.Parameter(
43 | _flatten_dense_tensors(
44 | [param.detach().float() for (_, param) in param_group]
45 | ).view(shape)
46 | )
47 | master_param.requires_grad = True
48 | master_params.append(master_param)
49 | return master_params
50 |
51 |
52 | def model_grads_to_master_grads(param_groups_and_shapes, master_params):
53 | """
54 | Copy the gradients from the model parameters into the master parameters
55 | from make_master_params().
56 | """
57 | for master_param, (param_group, shape) in zip(
58 | master_params, param_groups_and_shapes
59 | ):
60 | master_param.grad = _flatten_dense_tensors(
61 | [param_grad_or_zeros(param) for (_, param) in param_group]
62 | ).view(shape)
63 |
64 |
65 | def master_params_to_model_params(param_groups_and_shapes, master_params):
66 | """
67 | Copy the master parameter data back into the model parameters.
68 | """
69 | # Without copying to a list, if a generator is passed, this will
70 | # silently not copy any parameters.
71 | for master_param, (param_group, _) in zip(master_params, param_groups_and_shapes):
72 | for (_, param), unflat_master_param in zip(
73 | param_group, unflatten_master_params(param_group, master_param.view(-1))
74 | ):
75 | param.detach().copy_(unflat_master_param)
76 |
77 |
78 | def unflatten_master_params(param_group, master_param):
79 | return _unflatten_dense_tensors(master_param, [param for (_, param) in param_group])
80 |
81 |
82 | def get_param_groups_and_shapes(named_model_params):
83 | named_model_params = list(named_model_params)
84 | scalar_vector_named_params = (
85 | [(n, p) for (n, p) in named_model_params if p.ndim <= 1],
86 | (-1),
87 | )
88 | matrix_named_params = (
89 | [(n, p) for (n, p) in named_model_params if p.ndim > 1],
90 | (1, -1),
91 | )
92 | return [scalar_vector_named_params, matrix_named_params]
93 |
94 |
95 | def master_params_to_state_dict(
96 | model, param_groups_and_shapes, master_params, use_fp16
97 | ):
98 | if use_fp16:
99 | state_dict = model.state_dict()
100 | for master_param, (param_group, _) in zip(
101 | master_params, param_groups_and_shapes
102 | ):
103 | for (name, _), unflat_master_param in zip(
104 | param_group, unflatten_master_params(param_group, master_param.view(-1))
105 | ):
106 | assert name in state_dict
107 | state_dict[name] = unflat_master_param
108 | else:
109 | state_dict = model.state_dict()
110 | for i, (name, _value) in enumerate(model.named_parameters()):
111 | assert name in state_dict
112 | state_dict[name] = master_params[i]
113 | return state_dict
114 |
115 |
116 | def state_dict_to_master_params(model, state_dict, use_fp16):
117 | if use_fp16:
118 | named_model_params = [
119 | (name, state_dict[name]) for name, _ in model.named_parameters()
120 | ]
121 | param_groups_and_shapes = get_param_groups_and_shapes(named_model_params)
122 | master_params = make_master_params(param_groups_and_shapes)
123 | else:
124 | master_params = [state_dict[name] for name, _ in model.named_parameters()]
125 | return master_params
126 |
127 |
128 | def zero_master_grads(master_params):
129 | for param in master_params:
130 | param.grad = None
131 |
132 |
133 | def zero_grad(model_params):
134 | for param in model_params:
135 | # Taken from https://pytorch.org/docs/stable/_modules/torch/optim/optimizer.html#Optimizer.add_param_group
136 | if param.grad is not None:
137 | param.grad.detach_()
138 | param.grad.zero_()
139 |
140 |
141 | def param_grad_or_zeros(param):
142 | if param.grad is not None:
143 | return param.grad.data.detach()
144 | else:
145 | return th.zeros_like(param)
146 |
147 |
148 | class MixedPrecisionTrainer:
149 | def __init__(
150 | self,
151 | *,
152 | model,
153 | use_fp16=False,
154 | fp16_scale_growth=1e-3,
155 | initial_lg_loss_scale=INITIAL_LOG_LOSS_SCALE,
156 | ):
157 | self.model = model
158 | self.use_fp16 = use_fp16
159 | self.fp16_scale_growth = fp16_scale_growth
160 |
161 | self.model_params = list(self.model.parameters())
162 | self.master_params = self.model_params
163 | self.param_groups_and_shapes = None
164 | self.lg_loss_scale = initial_lg_loss_scale
165 |
166 | if self.use_fp16:
167 | self.param_groups_and_shapes = get_param_groups_and_shapes(
168 | self.model.named_parameters()
169 | )
170 | self.master_params = make_master_params(self.param_groups_and_shapes)
171 | self.model.convert_to_fp16()
172 |
173 | def zero_grad(self):
174 | zero_grad(self.model_params)
175 |
176 | def backward(self, loss: th.Tensor):
177 | if self.use_fp16:
178 | loss_scale = 2 ** self.lg_loss_scale
179 | (loss * loss_scale).backward()
180 | else:
181 | loss.backward()
182 |
183 | def optimize(self, opt: th.optim.Optimizer):
184 | if self.use_fp16:
185 | return self._optimize_fp16(opt)
186 | else:
187 | return self._optimize_normal(opt)
188 |
189 | def _optimize_fp16(self, opt: th.optim.Optimizer):
190 | logger.logkv_mean("lg_loss_scale", self.lg_loss_scale)
191 | model_grads_to_master_grads(self.param_groups_and_shapes, self.master_params)
192 | grad_norm, param_norm = self._compute_norms(grad_scale=2 ** self.lg_loss_scale)
193 | if check_overflow(grad_norm):
194 | self.lg_loss_scale -= 1
195 | logger.log(f"Found NaN, decreased lg_loss_scale to {self.lg_loss_scale}")
196 | zero_master_grads(self.master_params)
197 | return False
198 |
199 | logger.logkv_mean("grad_norm", grad_norm)
200 | logger.logkv_mean("param_norm", param_norm)
201 |
202 | for p in self.master_params:
203 | p.grad.mul_(1.0 / (2 ** self.lg_loss_scale))
204 | opt.step()
205 | zero_master_grads(self.master_params)
206 | master_params_to_model_params(self.param_groups_and_shapes, self.master_params)
207 | self.lg_loss_scale += self.fp16_scale_growth
208 | return True
209 |
210 | def _optimize_normal(self, opt: th.optim.Optimizer):
211 | grad_norm, param_norm = self._compute_norms()
212 | logger.logkv_mean("grad_norm", grad_norm)
213 | logger.logkv_mean("param_norm", param_norm)
214 | opt.step()
215 | return True
216 |
217 | def _compute_norms(self, grad_scale=1.0):
218 | grad_norm = 0.0
219 | param_norm = 0.0
220 | for p in self.master_params:
221 | with th.no_grad():
222 | param_norm += th.norm(p, p=2, dtype=th.float32).item() ** 2
223 | if p.grad is not None:
224 | grad_norm += th.norm(p.grad, p=2, dtype=th.float32).item() ** 2
225 | return np.sqrt(grad_norm) / grad_scale, np.sqrt(param_norm)
226 |
227 | def master_params_to_state_dict(self, master_params):
228 | return master_params_to_state_dict(
229 | self.model, self.param_groups_and_shapes, master_params, self.use_fp16
230 | )
231 |
232 | def state_dict_to_master_params(self, state_dict):
233 | return state_dict_to_master_params(self.model, state_dict, self.use_fp16)
234 |
235 |
236 | def check_overflow(value):
237 | return (value == float("inf")) or (value == -float("inf")) or (value != value)
238 |
--------------------------------------------------------------------------------
/guided_diffusion/gaussian_diffusion.py:
--------------------------------------------------------------------------------
1 | """
2 | This code started out as a PyTorch port of Ho et al's diffusion models:
3 | https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/diffusion_utils_2.py
4 |
5 | Docstrings have been added, as well as DDIM sampling and a new collection of beta schedules.
6 | """
7 |
8 | import enum
9 | import math
10 |
11 | import numpy as np
12 | import torch as th
13 |
14 | from .nn import mean_flat
15 | from .losses import normal_kl, discretized_gaussian_log_likelihood
16 |
17 |
18 | def get_named_beta_schedule(schedule_name, num_diffusion_timesteps):
19 | """
20 | Get a pre-defined beta schedule for the given name.
21 |
22 | The beta schedule library consists of beta schedules which remain similar
23 | in the limit of num_diffusion_timesteps.
24 | Beta schedules may be added, but should not be removed or changed once
25 | they are committed to maintain backwards compatibility.
26 | """
27 | if schedule_name == "linear":
28 | # Linear schedule from Ho et al, extended to work for any number of
29 | # diffusion steps.
30 | scale = 1000 / num_diffusion_timesteps
31 | beta_start = scale * 0.0001
32 | beta_end = scale * 0.02
33 | return np.linspace(
34 | beta_start, beta_end, num_diffusion_timesteps, dtype=np.float64
35 | )
36 | elif schedule_name == "cosine":
37 | return betas_for_alpha_bar(
38 | num_diffusion_timesteps,
39 | lambda t: math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2,
40 | )
41 | else:
42 | raise NotImplementedError(f"unknown beta schedule: {schedule_name}")
43 |
44 |
45 | def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999):
46 | """
47 | Create a beta schedule that discretizes the given alpha_t_bar function,
48 | which defines the cumulative product of (1-beta) over time from t = [0,1].
49 |
50 | :param num_diffusion_timesteps: the number of betas to produce.
51 | :param alpha_bar: a lambda that takes an argument t from 0 to 1 and
52 | produces the cumulative product of (1-beta) up to that
53 | part of the diffusion process.
54 | :param max_beta: the maximum beta to use; use values lower than 1 to
55 | prevent singularities.
56 | """
57 | betas = []
58 | for i in range(num_diffusion_timesteps):
59 | t1 = i / num_diffusion_timesteps
60 | t2 = (i + 1) / num_diffusion_timesteps
61 | betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
62 | return np.array(betas)
63 |
64 |
65 | class ModelMeanType(enum.Enum):
66 | """
67 | Which type of output the model predicts.
68 | """
69 |
70 | PREVIOUS_X = enum.auto() # the model predicts x_{t-1}
71 | START_X = enum.auto() # the model predicts x_0
72 | EPSILON = enum.auto() # the model predicts epsilon
73 |
74 |
75 | class ModelVarType(enum.Enum):
76 | """
77 | What is used as the model's output variance.
78 |
79 | The LEARNED_RANGE option has been added to allow the model to predict
80 | values between FIXED_SMALL and FIXED_LARGE, making its job easier.
81 | """
82 |
83 | LEARNED = enum.auto()
84 | FIXED_SMALL = enum.auto()
85 | FIXED_LARGE = enum.auto()
86 | LEARNED_RANGE = enum.auto()
87 |
88 |
89 | class LossType(enum.Enum):
90 | MSE = enum.auto() # use raw MSE loss (and KL when learning variances)
91 | RESCALED_MSE = (
92 | enum.auto()
93 | ) # use raw MSE loss (with RESCALED_KL when learning variances)
94 | KL = enum.auto() # use the variational lower-bound
95 | RESCALED_KL = enum.auto() # like KL, but rescale to estimate the full VLB
96 |
97 | def is_vb(self):
98 | return self == LossType.KL or self == LossType.RESCALED_KL
99 |
100 |
101 | class GaussianDiffusion:
102 | """
103 | Utilities for training and sampling diffusion models.
104 |
105 | Ported directly from here, and then adapted over time to further experimentation.
106 | https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/diffusion_utils_2.py#L42
107 |
108 | :param betas: a 1-D numpy array of betas for each diffusion timestep,
109 | starting at T and going to 1.
110 | :param model_mean_type: a ModelMeanType determining what the model outputs.
111 | :param model_var_type: a ModelVarType determining how variance is output.
112 | :param loss_type: a LossType determining the loss function to use.
113 | :param rescale_timesteps: if True, pass floating point timesteps into the
114 | model so that they are always scaled like in the
115 | original paper (0 to 1000).
116 | """
117 |
118 | def __init__(
119 | self,
120 | *,
121 | betas,
122 | model_mean_type,
123 | model_var_type,
124 | loss_type,
125 | rescale_timesteps=False,
126 | ):
127 | self.model_mean_type = model_mean_type
128 | self.model_var_type = model_var_type
129 | self.loss_type = loss_type
130 | self.rescale_timesteps = rescale_timesteps
131 |
132 | # Use float64 for accuracy.
133 | betas = np.array(betas, dtype=np.float64)
134 | self.betas = betas
135 | assert len(betas.shape) == 1, "betas must be 1-D"
136 | assert (betas > 0).all() and (betas <= 1).all()
137 |
138 | self.num_timesteps = int(betas.shape[0])
139 |
140 | alphas = 1.0 - betas
141 | self.alphas_cumprod = np.cumprod(alphas, axis=0)
142 | self.alphas_cumprod_prev = np.append(1.0, self.alphas_cumprod[:-1])
143 | self.alphas_cumprod_next = np.append(self.alphas_cumprod[1:], 0.0)
144 | assert self.alphas_cumprod_prev.shape == (self.num_timesteps,)
145 |
146 | # calculations for diffusion q(x_t | x_{t-1}) and others
147 | self.sqrt_alphas_cumprod = np.sqrt(self.alphas_cumprod)
148 | self.sqrt_one_minus_alphas_cumprod = np.sqrt(1.0 - self.alphas_cumprod)
149 | self.log_one_minus_alphas_cumprod = np.log(1.0 - self.alphas_cumprod)
150 | self.sqrt_recip_alphas_cumprod = np.sqrt(1.0 / self.alphas_cumprod)
151 | self.sqrt_recipm1_alphas_cumprod = np.sqrt(1.0 / self.alphas_cumprod - 1)
152 |
153 | # calculations for posterior q(x_{t-1} | x_t, x_0)
154 | self.posterior_variance = (
155 | betas * (1.0 - self.alphas_cumprod_prev) / (1.0 - self.alphas_cumprod)
156 | )
157 | # log calculation clipped because the posterior variance is 0 at the
158 | # beginning of the diffusion chain.
159 | self.posterior_log_variance_clipped = np.log(
160 | np.append(self.posterior_variance[1], self.posterior_variance[1:])
161 | )
162 | self.posterior_mean_coef1 = (
163 | betas * np.sqrt(self.alphas_cumprod_prev) / (1.0 - self.alphas_cumprod)
164 | )
165 | self.posterior_mean_coef2 = (
166 | (1.0 - self.alphas_cumprod_prev)
167 | * np.sqrt(alphas)
168 | / (1.0 - self.alphas_cumprod)
169 | )
170 |
171 | def q_mean_variance(self, x_start, t):
172 | """
173 | Get the distribution q(x_t | x_0).
174 |
175 | :param x_start: the [N x C x ...] tensor of noiseless inputs.
176 | :param t: the number of diffusion steps (minus 1). Here, 0 means one step.
177 | :return: A tuple (mean, variance, log_variance), all of x_start's shape.
178 | """
179 | mean = (
180 | _extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start
181 | )
182 | variance = _extract_into_tensor(1.0 - self.alphas_cumprod, t, x_start.shape)
183 | log_variance = _extract_into_tensor(
184 | self.log_one_minus_alphas_cumprod, t, x_start.shape
185 | )
186 | return mean, variance, log_variance
187 |
188 | def q_sample(self, x_start, t, noise=None):
189 | """
190 | Diffuse the data for a given number of diffusion steps.
191 |
192 | In other words, sample from q(x_t | x_0).
193 |
194 | :param x_start: the initial data batch.
195 | :param t: the number of diffusion steps (minus 1). Here, 0 means one step.
196 | :param noise: if specified, the split-out normal noise.
197 | :return: A noisy version of x_start.
198 | """
199 | if noise is None:
200 | noise = th.randn_like(x_start)
201 | assert noise.shape == x_start.shape
202 | return (
203 | _extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start
204 | + _extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape)
205 | * noise
206 | )
207 |
208 | def q_posterior_mean_variance(self, x_start, x_t, t):
209 | """
210 | Compute the mean and variance of the diffusion posterior:
211 |
212 | q(x_{t-1} | x_t, x_0)
213 |
214 | """
215 | assert x_start.shape == x_t.shape
216 | posterior_mean = (
217 | _extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * x_start
218 | + _extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t
219 | )
220 | posterior_variance = _extract_into_tensor(self.posterior_variance, t, x_t.shape)
221 | posterior_log_variance_clipped = _extract_into_tensor(
222 | self.posterior_log_variance_clipped, t, x_t.shape
223 | )
224 | assert (
225 | posterior_mean.shape[0]
226 | == posterior_variance.shape[0]
227 | == posterior_log_variance_clipped.shape[0]
228 | == x_start.shape[0]
229 | )
230 | return posterior_mean, posterior_variance, posterior_log_variance_clipped
231 |
232 | def p_mean_variance(
233 | self, model, x, t, clip_denoised=True, denoised_fn=None, model_kwargs=None
234 | ):
235 | """
236 | Apply the model to get p(x_{t-1} | x_t), as well as a prediction of
237 | the initial x, x_0.
238 |
239 | :param model: the model, which takes a signal and a batch of timesteps
240 | as input.
241 | :param x: the [N x C x ...] tensor at time t.
242 | :param t: a 1-D Tensor of timesteps.
243 | :param clip_denoised: if True, clip the denoised signal into [-1, 1].
244 | :param denoised_fn: if not None, a function which applies to the
245 | x_start prediction before it is used to sample. Applies before
246 | clip_denoised.
247 | :param model_kwargs: if not None, a dict of extra keyword arguments to
248 | pass to the model. This can be used for conditioning.
249 | :return: a dict with the following keys:
250 | - 'mean': the model mean output.
251 | - 'variance': the model variance output.
252 | - 'log_variance': the log of 'variance'.
253 | - 'pred_xstart': the prediction for x_0.
254 | """
255 | if model_kwargs is None:
256 | model_kwargs = {}
257 |
258 | B, C = x.shape[:2]
259 | assert t.shape == (B,)
260 | model_output = model(x, self._scale_timesteps(t), **model_kwargs)
261 |
262 | if self.model_var_type in [ModelVarType.LEARNED, ModelVarType.LEARNED_RANGE]:
263 | assert model_output.shape == (B, C * 2, *x.shape[2:])
264 | model_output, model_var_values = th.split(model_output, C, dim=1)
265 | if self.model_var_type == ModelVarType.LEARNED:
266 | model_log_variance = model_var_values
267 | model_variance = th.exp(model_log_variance)
268 | else:
269 | min_log = _extract_into_tensor(
270 | self.posterior_log_variance_clipped, t, x.shape
271 | )
272 | max_log = _extract_into_tensor(np.log(self.betas), t, x.shape)
273 | # The model_var_values is [-1, 1] for [min_var, max_var].
274 | frac = (model_var_values + 1) / 2
275 | model_log_variance = frac * max_log + (1 - frac) * min_log
276 | model_variance = th.exp(model_log_variance)
277 | else:
278 | model_variance, model_log_variance = {
279 | # for fixedlarge, we set the initial (log-)variance like so
280 | # to get a better decoder log likelihood.
281 | ModelVarType.FIXED_LARGE: (
282 | np.append(self.posterior_variance[1], self.betas[1:]),
283 | np.log(np.append(self.posterior_variance[1], self.betas[1:])),
284 | ),
285 | ModelVarType.FIXED_SMALL: (
286 | self.posterior_variance,
287 | self.posterior_log_variance_clipped,
288 | ),
289 | }[self.model_var_type]
290 | model_variance = _extract_into_tensor(model_variance, t, x.shape)
291 | model_log_variance = _extract_into_tensor(model_log_variance, t, x.shape)
292 |
293 | def process_xstart(x):
294 | if denoised_fn is not None:
295 | x = denoised_fn(x)
296 | if clip_denoised:
297 | return x.clamp(-1, 1)
298 | return x
299 |
300 | if self.model_mean_type == ModelMeanType.PREVIOUS_X:
301 | pred_xstart = process_xstart(
302 | self._predict_xstart_from_xprev(x_t=x, t=t, xprev=model_output)
303 | )
304 | model_mean = model_output
305 | elif self.model_mean_type in [ModelMeanType.START_X, ModelMeanType.EPSILON]:
306 | if self.model_mean_type == ModelMeanType.START_X:
307 | pred_xstart = process_xstart(model_output)
308 | else:
309 | pred_xstart = process_xstart(
310 | self._predict_xstart_from_eps(x_t=x, t=t, eps=model_output)
311 | )
312 | model_mean, _, _ = self.q_posterior_mean_variance(
313 | x_start=pred_xstart, x_t=x, t=t
314 | )
315 | else:
316 | raise NotImplementedError(self.model_mean_type)
317 |
318 | assert (
319 | model_mean.shape == model_log_variance.shape == pred_xstart.shape == x.shape
320 | )
321 | return {
322 | "mean": model_mean,
323 | "variance": model_variance,
324 | "log_variance": model_log_variance,
325 | "pred_xstart": pred_xstart,
326 | }
327 |
328 | def _predict_xstart_from_eps(self, x_t, t, eps):
329 | assert x_t.shape == eps.shape
330 | return (
331 | _extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t
332 | - _extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * eps
333 | )
334 |
335 | def _predict_xstart_from_xprev(self, x_t, t, xprev):
336 | assert x_t.shape == xprev.shape
337 | return ( # (xprev - coef2*x_t) / coef1
338 | _extract_into_tensor(1.0 / self.posterior_mean_coef1, t, x_t.shape) * xprev
339 | - _extract_into_tensor(
340 | self.posterior_mean_coef2 / self.posterior_mean_coef1, t, x_t.shape
341 | )
342 | * x_t
343 | )
344 |
345 | def _predict_eps_from_xstart(self, x_t, t, pred_xstart):
346 | return (
347 | _extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t
348 | - pred_xstart
349 | ) / _extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape)
350 |
351 | def _scale_timesteps(self, t):
352 | if self.rescale_timesteps:
353 | return t.float() * (1000.0 / self.num_timesteps)
354 | return t
355 |
356 | def condition_mean(self, cond_fn, p_mean_var, x, t, model_kwargs=None):
357 | """
358 | Compute the mean for the previous step, given a function cond_fn that
359 | computes the gradient of a conditional log probability with respect to
360 | x. In particular, cond_fn computes grad(log(p(y|x))), and we want to
361 | condition on y.
362 |
363 | This uses the conditioning strategy from Sohl-Dickstein et al. (2015).
364 | """
365 | gradient = cond_fn(x, self._scale_timesteps(t), **model_kwargs)
366 | new_mean = (
367 | p_mean_var["mean"].float() + p_mean_var["variance"] * gradient.float()
368 | )
369 | return new_mean
370 |
371 | def condition_score(self, cond_fn, p_mean_var, x, t, model_kwargs=None):
372 | """
373 | Compute what the p_mean_variance output would have been, should the
374 | model's score function be conditioned by cond_fn.
375 |
376 | See condition_mean() for details on cond_fn.
377 |
378 | Unlike condition_mean(), this instead uses the conditioning strategy
379 | from Song et al (2020).
380 | """
381 | alpha_bar = _extract_into_tensor(self.alphas_cumprod, t, x.shape)
382 |
383 | eps = self._predict_eps_from_xstart(x, t, p_mean_var["pred_xstart"])
384 | eps = eps - (1 - alpha_bar).sqrt() * cond_fn(
385 | x, self._scale_timesteps(t), **model_kwargs
386 | )
387 |
388 | out = p_mean_var.copy()
389 | out["pred_xstart"] = self._predict_xstart_from_eps(x, t, eps)
390 | out["mean"], _, _ = self.q_posterior_mean_variance(
391 | x_start=out["pred_xstart"], x_t=x, t=t
392 | )
393 | return out
394 |
395 | def p_sample(
396 | self,
397 | model,
398 | x,
399 | t,
400 | clip_denoised=True,
401 | denoised_fn=None,
402 | cond_fn=None,
403 | model_kwargs=None,
404 | ):
405 | """
406 | Sample x_{t-1} from the model at the given timestep.
407 |
408 | :param model: the model to sample from.
409 | :param x: the current tensor at x_{t-1}.
410 | :param t: the value of t, starting at 0 for the first diffusion step.
411 | :param clip_denoised: if True, clip the x_start prediction to [-1, 1].
412 | :param denoised_fn: if not None, a function which applies to the
413 | x_start prediction before it is used to sample.
414 | :param cond_fn: if not None, this is a gradient function that acts
415 | similarly to the model.
416 | :param model_kwargs: if not None, a dict of extra keyword arguments to
417 | pass to the model. This can be used for conditioning.
418 | :return: a dict containing the following keys:
419 | - 'sample': a random sample from the model.
420 | - 'pred_xstart': a prediction of x_0.
421 | """
422 | out = self.p_mean_variance(
423 | model,
424 | x,
425 | t,
426 | clip_denoised=clip_denoised,
427 | denoised_fn=denoised_fn,
428 | model_kwargs=model_kwargs,
429 | )
430 | noise = th.randn_like(x)
431 | nonzero_mask = (
432 | (t != 0).float().view(-1, *([1] * (len(x.shape) - 1)))
433 | ) # no noise when t == 0
434 | if cond_fn is not None:
435 | out["mean"] = self.condition_mean(
436 | cond_fn, out, x, t, model_kwargs=model_kwargs
437 | )
438 | sample = out["mean"] + nonzero_mask * th.exp(0.5 * out["log_variance"]) * noise
439 | return {"sample": sample, "pred_xstart": out["pred_xstart"]}
440 |
441 | def p_sample_loop(
442 | self,
443 | model,
444 | shape,
445 | noise=None,
446 | clip_denoised=True,
447 | denoised_fn=None,
448 | cond_fn=None,
449 | model_kwargs=None,
450 | device=None,
451 | progress=False,
452 | ):
453 | """
454 | Generate samples from the model.
455 |
456 | :param model: the model module.
457 | :param shape: the shape of the samples, (N, C, H, W).
458 | :param noise: if specified, the noise from the encoder to sample.
459 | Should be of the same shape as `shape`.
460 | :param clip_denoised: if True, clip x_start predictions to [-1, 1].
461 | :param denoised_fn: if not None, a function which applies to the
462 | x_start prediction before it is used to sample.
463 | :param cond_fn: if not None, this is a gradient function that acts
464 | similarly to the model.
465 | :param model_kwargs: if not None, a dict of extra keyword arguments to
466 | pass to the model. This can be used for conditioning.
467 | :param device: if specified, the device to create the samples on.
468 | If not specified, use a model parameter's device.
469 | :param progress: if True, show a tqdm progress bar.
470 | :return: a non-differentiable batch of samples.
471 | """
472 | final = None
473 | for sample in self.p_sample_loop_progressive(
474 | model,
475 | shape,
476 | noise=noise,
477 | clip_denoised=clip_denoised,
478 | denoised_fn=denoised_fn,
479 | cond_fn=cond_fn,
480 | model_kwargs=model_kwargs,
481 | device=device,
482 | progress=progress,
483 | ):
484 | final = sample
485 | return final["sample"]
486 |
487 | def p_sample_loop_progressive(
488 | self,
489 | model,
490 | shape,
491 | noise=None,
492 | clip_denoised=True,
493 | denoised_fn=None,
494 | cond_fn=None,
495 | model_kwargs=None,
496 | device=None,
497 | progress=False,
498 | ):
499 | """
500 | Generate samples from the model and yield intermediate samples from
501 | each timestep of diffusion.
502 |
503 | Arguments are the same as p_sample_loop().
504 | Returns a generator over dicts, where each dict is the return value of
505 | p_sample().
506 | """
507 | if device is None:
508 | device = next(model.parameters()).device
509 | assert isinstance(shape, (tuple, list))
510 | if noise is not None:
511 | img = noise #here
512 | else:
513 | img = th.randn(*shape, device=device)
514 | indices = list(range(self.num_timesteps))[::-1]
515 |
516 | if progress:
517 | # Lazy import so that we don't depend on tqdm.
518 | from tqdm.auto import tqdm
519 |
520 | indices = tqdm(indices)
521 |
522 | for i in indices:
523 | t = th.tensor([i] * shape[0], device=device)
524 | with th.no_grad():
525 | out = self.p_sample(
526 | model,
527 | img,
528 | t,
529 | clip_denoised=clip_denoised,
530 | denoised_fn=denoised_fn,
531 | cond_fn=cond_fn,
532 | model_kwargs=model_kwargs,
533 | )
534 | yield out
535 | img = out["sample"]
536 |
537 | def ddim_sample(
538 | self,
539 | model,
540 | x,
541 | t,
542 | clip_denoised=True,
543 | denoised_fn=None,
544 | cond_fn=None,
545 | model_kwargs=None,
546 | eta=0.0,
547 | ):
548 | """
549 | Sample x_{t-1} from the model using DDIM.
550 |
551 | Same usage as p_sample().
552 | """
553 | out = self.p_mean_variance(
554 | model,
555 | x,
556 | t,
557 | clip_denoised=clip_denoised,
558 | denoised_fn=denoised_fn,
559 | model_kwargs=model_kwargs,
560 | )
561 | if cond_fn is not None:
562 | out = self.condition_score(cond_fn, out, x, t, model_kwargs=model_kwargs)
563 |
564 | # Usually our model outputs epsilon, but we re-derive it
565 | # in case we used x_start or x_prev prediction.
566 | eps = self._predict_eps_from_xstart(x, t, out["pred_xstart"])
567 |
568 | alpha_bar = _extract_into_tensor(self.alphas_cumprod, t, x.shape)
569 | alpha_bar_prev = _extract_into_tensor(self.alphas_cumprod_prev, t, x.shape)
570 | sigma = (
571 | eta
572 | * th.sqrt((1 - alpha_bar_prev) / (1 - alpha_bar))
573 | * th.sqrt(1 - alpha_bar / alpha_bar_prev)
574 | )
575 | # Equation 12.
576 | noise = th.randn_like(x)
577 | mean_pred = (
578 | out["pred_xstart"] * th.sqrt(alpha_bar_prev)
579 | + th.sqrt(1 - alpha_bar_prev - sigma ** 2) * eps
580 | )
581 | nonzero_mask = (
582 | (t != 0).float().view(-1, *([1] * (len(x.shape) - 1)))
583 | ) # no noise when t == 0
584 | sample = mean_pred + nonzero_mask * sigma * noise
585 | return {"sample": sample, "pred_xstart": out["pred_xstart"]}
586 |
587 | def ddim_reverse_sample(
588 | self,
589 | model,
590 | x,
591 | t,
592 | clip_denoised=True,
593 | denoised_fn=None,
594 | model_kwargs=None,
595 | eta=0.0,
596 | ):
597 | """
598 | Sample x_{t+1} from the model using DDIM reverse ODE.
599 | """
600 | assert eta == 0.0, "Reverse ODE only for deterministic path"
601 | out = self.p_mean_variance(
602 | model,
603 | x,
604 | t,
605 | clip_denoised=clip_denoised,
606 | denoised_fn=denoised_fn,
607 | model_kwargs=model_kwargs,
608 | )
609 | # Usually our model outputs epsilon, but we re-derive it
610 | # in case we used x_start or x_prev prediction.
611 | eps = (
612 | _extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x.shape) * x
613 | - out["pred_xstart"]
614 | ) / _extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x.shape)
615 | alpha_bar_next = _extract_into_tensor(self.alphas_cumprod_next, t, x.shape)
616 |
617 | # Equation 12. reversed
618 | mean_pred = (
619 | out["pred_xstart"] * th.sqrt(alpha_bar_next)
620 | + th.sqrt(1 - alpha_bar_next) * eps
621 | )
622 |
623 | return {"sample": mean_pred, "pred_xstart": out["pred_xstart"]}
624 |
625 | def ddim_sample_loop(
626 | self,
627 | model,
628 | shape,
629 | noise=None,
630 | clip_denoised=True,
631 | denoised_fn=None,
632 | cond_fn=None,
633 | model_kwargs=None,
634 | device=None,
635 | progress=False,
636 | eta=0.0,
637 | ):
638 | """
639 | Generate samples from the model using DDIM.
640 |
641 | Same usage as p_sample_loop().
642 | """
643 | final = None
644 | for sample in self.ddim_sample_loop_progressive(
645 | model,
646 | shape,
647 | noise=noise,
648 | clip_denoised=clip_denoised,
649 | denoised_fn=denoised_fn,
650 | cond_fn=cond_fn,
651 | model_kwargs=model_kwargs,
652 | device=device,
653 | progress=progress,
654 | eta=eta,
655 | ):
656 | final = sample
657 | return final["sample"]
658 |
659 | def ddim_sample_loop_progressive(
660 | self,
661 | model,
662 | shape,
663 | noise=None,
664 | clip_denoised=True,
665 | denoised_fn=None,
666 | cond_fn=None,
667 | model_kwargs=None,
668 | device=None,
669 | progress=False,
670 | eta=0.0,
671 | ):
672 | """
673 | Use DDIM to sample from the model and yield intermediate samples from
674 | each timestep of DDIM.
675 |
676 | Same usage as p_sample_loop_progressive().
677 | """
678 | if device is None:
679 | device = next(model.parameters()).device
680 | assert isinstance(shape, (tuple, list))
681 | if noise is not None:
682 | img = noise
683 | else:
684 | img = th.randn(*shape, device=device)
685 | indices = list(range(self.num_timesteps))[::-1]
686 |
687 | if progress:
688 | # Lazy import so that we don't depend on tqdm.
689 | from tqdm.auto import tqdm
690 |
691 | indices = tqdm(indices)
692 |
693 | for i in indices:
694 | t = th.tensor([i] * shape[0], device=device)
695 | with th.no_grad():
696 | out = self.ddim_sample(
697 | model,
698 | img,
699 | t,
700 | clip_denoised=clip_denoised,
701 | denoised_fn=denoised_fn,
702 | cond_fn=cond_fn,
703 | model_kwargs=model_kwargs,
704 | eta=eta,
705 | )
706 | yield out
707 | img = out["sample"]
708 |
709 | def _vb_terms_bpd(
710 | self, model, x_start, x_t, t, clip_denoised=True, model_kwargs=None
711 | ):
712 | """
713 | Get a term for the variational lower-bound.
714 |
715 | The resulting units are bits (rather than nats, as one might expect).
716 | This allows for comparison to other papers.
717 |
718 | :return: a dict with the following keys:
719 | - 'output': a shape [N] tensor of NLLs or KLs.
720 | - 'pred_xstart': the x_0 predictions.
721 | """
722 | true_mean, _, true_log_variance_clipped = self.q_posterior_mean_variance(
723 | x_start=x_start, x_t=x_t, t=t
724 | )
725 | out = self.p_mean_variance(
726 | model, x_t, t, clip_denoised=clip_denoised, model_kwargs=model_kwargs
727 | )
728 | kl = normal_kl(
729 | true_mean, true_log_variance_clipped, out["mean"], out["log_variance"]
730 | )
731 | kl = mean_flat(kl) / np.log(2.0)
732 |
733 | decoder_nll = -discretized_gaussian_log_likelihood(
734 | x_start, means=out["mean"], log_scales=0.5 * out["log_variance"]
735 | )
736 | assert decoder_nll.shape == x_start.shape
737 | decoder_nll = mean_flat(decoder_nll) / np.log(2.0)
738 |
739 | # At the first timestep return the decoder NLL,
740 | # otherwise return KL(q(x_{t-1}|x_t,x_0) || p(x_{t-1}|x_t))
741 | output = th.where((t == 0), decoder_nll, kl)
742 | return {"output": output, "pred_xstart": out["pred_xstart"]}
743 |
744 | def training_losses(self, model, x_start, t, model_kwargs=None, noise=None):
745 | """
746 | Compute training losses for a single timestep.
747 |
748 | :param model: the model to evaluate loss on.
749 | :param x_start: the [N x C x ...] tensor of inputs.
750 | :param t: a batch of timestep indices.
751 | :param model_kwargs: if not None, a dict of extra keyword arguments to
752 | pass to the model. This can be used for conditioning.
753 | :param noise: if specified, the specific Gaussian noise to try to remove.
754 | :return: a dict with the key "loss" containing a tensor of shape [N].
755 | Some mean or variance settings may also have other keys.
756 | """
757 | if model_kwargs is None:
758 | model_kwargs = {}
759 | if noise is None:
760 | noise = th.randn_like(x_start)
761 | x_t = self.q_sample(x_start, t, noise=noise)
762 |
763 | terms = {}
764 |
765 | if self.loss_type == LossType.KL or self.loss_type == LossType.RESCALED_KL:
766 | terms["loss"] = self._vb_terms_bpd(
767 | model=model,
768 | x_start=x_start,
769 | x_t=x_t,
770 | t=t,
771 | clip_denoised=False,
772 | model_kwargs=model_kwargs,
773 | )["output"]
774 | if self.loss_type == LossType.RESCALED_KL:
775 | terms["loss"] *= self.num_timesteps
776 | elif self.loss_type == LossType.MSE or self.loss_type == LossType.RESCALED_MSE:
777 | model_output = model(x_t, self._scale_timesteps(t), **model_kwargs)
778 | # print(model_output.shape)
779 | # print(x_t.shape)
780 |
781 | if self.model_var_type in [
782 | ModelVarType.LEARNED,
783 | ModelVarType.LEARNED_RANGE,
784 | ]:
785 | B, C = x_t.shape[:2]
786 | assert model_output.shape == (B, C * 2, *x_t.shape[2:])
787 | model_output, model_var_values = th.split(model_output, C, dim=1)
788 | # Learn the variance using the variational bound, but don't let
789 | # it affect our mean prediction.
790 | frozen_out = th.cat([model_output.detach(), model_var_values], dim=1)
791 | terms["vb"] = self._vb_terms_bpd(
792 | model=lambda *args, r=frozen_out: r,
793 | x_start=x_start,
794 | x_t=x_t,
795 | t=t,
796 | clip_denoised=False,
797 | )["output"]
798 | if self.loss_type == LossType.RESCALED_MSE:
799 | # Divide by 1000 for equivalence with initial implementation.
800 | # Without a factor of 1/1000, the VB term hurts the MSE term.
801 | terms["vb"] *= self.num_timesteps / 1000.0
802 |
803 | target = {
804 | ModelMeanType.PREVIOUS_X: self.q_posterior_mean_variance(
805 | x_start=x_start, x_t=x_t, t=t
806 | )[0],
807 | ModelMeanType.START_X: x_start,
808 | ModelMeanType.EPSILON: noise,
809 | }[self.model_mean_type]
810 | assert model_output.shape == target.shape == x_start.shape
811 | terms["mse"] = mean_flat((target - model_output) ** 2)
812 | if "vb" in terms:
813 | terms["loss"] = terms["mse"] + terms["vb"]
814 | else:
815 | terms["loss"] = terms["mse"]
816 | else:
817 | raise NotImplementedError(self.loss_type)
818 |
819 | return terms
820 |
821 | def _prior_bpd(self, x_start):
822 | """
823 | Get the prior KL term for the variational lower-bound, measured in
824 | bits-per-dim.
825 |
826 | This term can't be optimized, as it only depends on the encoder.
827 |
828 | :param x_start: the [N x C x ...] tensor of inputs.
829 | :return: a batch of [N] KL values (in bits), one per batch element.
830 | """
831 | batch_size = x_start.shape[0]
832 | t = th.tensor([self.num_timesteps - 1] * batch_size, device=x_start.device)
833 | qt_mean, _, qt_log_variance = self.q_mean_variance(x_start, t)
834 | kl_prior = normal_kl(
835 | mean1=qt_mean, logvar1=qt_log_variance, mean2=0.0, logvar2=0.0
836 | )
837 | return mean_flat(kl_prior) / np.log(2.0)
838 |
839 | def calc_bpd_loop(self, model, x_start, clip_denoised=True, model_kwargs=None):
840 | """
841 | Compute the entire variational lower-bound, measured in bits-per-dim,
842 | as well as other related quantities.
843 |
844 | :param model: the model to evaluate loss on.
845 | :param x_start: the [N x C x ...] tensor of inputs.
846 | :param clip_denoised: if True, clip denoised samples.
847 | :param model_kwargs: if not None, a dict of extra keyword arguments to
848 | pass to the model. This can be used for conditioning.
849 |
850 | :return: a dict containing the following keys:
851 | - total_bpd: the total variational lower-bound, per batch element.
852 | - prior_bpd: the prior term in the lower-bound.
853 | - vb: an [N x T] tensor of terms in the lower-bound.
854 | - xstart_mse: an [N x T] tensor of x_0 MSEs for each timestep.
855 | - mse: an [N x T] tensor of epsilon MSEs for each timestep.
856 | """
857 | device = x_start.device
858 | batch_size = x_start.shape[0]
859 |
860 | vb = []
861 | xstart_mse = []
862 | mse = []
863 | for t in list(range(self.num_timesteps))[::-1]:
864 | t_batch = th.tensor([t] * batch_size, device=device)
865 | noise = th.randn_like(x_start)
866 | x_t = self.q_sample(x_start=x_start, t=t_batch, noise=noise)
867 | # Calculate VLB term at the current timestep
868 | with th.no_grad():
869 | out = self._vb_terms_bpd(
870 | model,
871 | x_start=x_start,
872 | x_t=x_t,
873 | t=t_batch,
874 | clip_denoised=clip_denoised,
875 | model_kwargs=model_kwargs,
876 | )
877 | vb.append(out["output"])
878 | xstart_mse.append(mean_flat((out["pred_xstart"] - x_start) ** 2))
879 | eps = self._predict_eps_from_xstart(x_t, t_batch, out["pred_xstart"])
880 | mse.append(mean_flat((eps - noise) ** 2))
881 |
882 | vb = th.stack(vb, dim=1)
883 | xstart_mse = th.stack(xstart_mse, dim=1)
884 | mse = th.stack(mse, dim=1)
885 |
886 | prior_bpd = self._prior_bpd(x_start)
887 | total_bpd = vb.sum(dim=1) + prior_bpd
888 | return {
889 | "total_bpd": total_bpd,
890 | "prior_bpd": prior_bpd,
891 | "vb": vb,
892 | "xstart_mse": xstart_mse,
893 | "mse": mse,
894 | }
895 |
896 |
897 | def _extract_into_tensor(arr, timesteps, broadcast_shape):
898 | """
899 | Extract values from a 1-D numpy array for a batch of indices.
900 |
901 | :param arr: the 1-D numpy array.
902 | :param timesteps: a tensor of indices into the array to extract.
903 | :param broadcast_shape: a larger shape of K dimensions with the batch
904 | dimension equal to the length of timesteps.
905 | :return: a tensor of shape [batch_size, 1, ...] where the shape has K dims.
906 | """
907 | res = th.from_numpy(arr).to(device=timesteps.device)[timesteps].float()
908 | while len(res.shape) < len(broadcast_shape):
909 | res = res[..., None]
910 | return res.expand(broadcast_shape)
911 |
--------------------------------------------------------------------------------
/guided_diffusion/image_datasets.py:
--------------------------------------------------------------------------------
1 | import math
2 | import random
3 | import os
4 |
5 | from PIL import Image
6 | import blobfile as bf
7 | from mpi4py import MPI
8 | import numpy as np
9 | from torch.utils.data import DataLoader, Dataset
10 |
11 | def load_data(
12 | *,
13 | data_dir,
14 | batch_size,
15 | image_size,
16 | class_cond=False,
17 | deterministic=False,
18 | random_crop=False,
19 | random_flip=False,
20 | ):
21 | """
22 | For a dataset, create a generator over (images, kwargs) pairs.
23 |
24 | Each images is an NCHW float tensor, and the kwargs dict contains zero or
25 | more keys, each of which map to a batched Tensor of their own.
26 | The kwargs dict can be used for class labels, in which case the key is "y"
27 | and the values are integer tensors of class labels.
28 |
29 | :param data_dir: a dataset directory.
30 | :param batch_size: the batch size of each returned pair.
31 | :param image_size: the size to which images are resized.
32 | :param class_cond: if True, include a "y" key in returned dicts for class
33 | label. If classes are not available and this is true, an
34 | exception will be raised.
35 | :param deterministic: if True, yield results in a deterministic order.
36 | :param random_crop: if True, randomly crop the images for augmentation.
37 | :param random_flip: if True, randomly flip the images for augmentation.
38 | """
39 | if not data_dir:
40 | raise ValueError("unspecified data directory")
41 | all_files = _list_image_files_recursively(data_dir)
42 | classes = None
43 | if class_cond:
44 | # Assume classes are the first part of the filename,
45 | # before an underscore.
46 | class_names = [bf.basename(path).split("_")[0] for path in all_files]
47 | sorted_classes = {x: i for i, x in enumerate(sorted(set(class_names)))}
48 | classes = [sorted_classes[x] for x in class_names]
49 | dataset = ImageDataset(
50 | image_size,
51 | all_files,
52 | classes=classes,
53 | shard=MPI.COMM_WORLD.Get_rank(),
54 | num_shards=MPI.COMM_WORLD.Get_size(),
55 | random_crop=random_crop,
56 | random_flip=random_flip,
57 | )
58 | if deterministic:
59 | loader = DataLoader(
60 | dataset, batch_size=batch_size, shuffle=False, num_workers=1, drop_last=True
61 | )
62 | else:
63 | loader = DataLoader(
64 | dataset, batch_size=batch_size, shuffle=True, num_workers=1, drop_last=True
65 | )
66 | while True:
67 | yield from loader
68 |
69 |
70 | def _list_image_files_recursively(data_dir):
71 | results = []
72 | for entry in sorted(bf.listdir(data_dir)):
73 | full_path = bf.join(data_dir, entry)
74 | ext = entry.split(".")[-1]
75 | if "." in entry and ext.lower() in ["npz"]:
76 | results.append(full_path)
77 | elif bf.isdir(full_path):
78 | results.extend(_list_image_files_recursively(full_path))
79 | return results
80 |
81 |
82 |
83 | # pet and ct data 3D
84 | # normalize to (-1,1)
85 | class ImageDataset(Dataset):
86 | def __init__(
87 | self,
88 | resolution,
89 | image_paths,
90 | classes=None,
91 | shard=0,
92 | num_shards=1,
93 | random_crop=False,
94 | random_flip=False,
95 | ):
96 | super().__init__()
97 | self.resolution = resolution
98 | self.local_images = image_paths[shard:][::num_shards]
99 | self.local_classes = None if classes is None else classes[shard:][::num_shards]
100 | self.random_crop = random_crop
101 | self.random_flip = random_flip
102 |
103 | def __len__(self):
104 | return len(self.local_images)
105 |
106 | def __getitem__(self, idx):
107 | path = self.local_images[idx]
108 |
109 | pet_data = np.load(path)['arr_0']
110 | # pet_data = pet_data.astype(np.float32) * 2.0 - 1.0
111 | pet_data = pet_data.astype(np.float32)
112 |
113 | pet_data = pet_data/4 # add nmlz
114 | # pet_data = pet_data * 2.0 - 1.0
115 |
116 | # input 96*96*96
117 | # size = 96 # patch size
118 | # rand_xyz = np.random.randint(0, 144-size+1, 3)
119 | # pet_low = pet_data[0:2,rand_xyz[0]:rand_xyz[0]+size,rand_xyz[1]:rand_xyz[1]+size,rand_xyz[2]:rand_xyz[2]+size].copy()
120 | # label = pet_data[2,rand_xyz[0]:rand_xyz[0]+size,rand_xyz[1]:rand_xyz[1]+size,rand_xyz[2]:rand_xyz[2]+size].copy()
121 | # while label.max() == 0:
122 | # rand_xyz = np.random.randint(0, 144-size+1, 3)
123 | # pet_low = pet_data[0:2,rand_xyz[0]:rand_xyz[0]+size,rand_xyz[1]:rand_xyz[1]+size,rand_xyz[2]:rand_xyz[2]+size].copy()
124 | # label = pet_data[2,rand_xyz[0]:rand_xyz[0]+size,rand_xyz[1]:rand_xyz[1]+size,rand_xyz[2]:rand_xyz[2]+size].copy()
125 | # C, H, W, T = pet_low.shape
126 |
127 | # input 96*96*32
128 | size_xy = 96 # patch size
129 | size_z = 96 #32
130 | rand_x = np.random.randint(0, 180-size_xy+1)
131 | rand_y = np.random.randint(0, 280-size_xy+1)
132 | rand_z = np.random.randint(0, 520-size_z+1)
133 | pet_low = pet_data[0, rand_x:rand_x+size_xy, rand_y:rand_y+size_xy, rand_z:rand_z+size_z].copy()
134 | label = pet_data[1, rand_x:rand_x+size_xy, rand_y:rand_y+size_xy, rand_z:rand_z+size_z].copy()
135 | # while label.max() == label.min():
136 | # rand_xy = np.random.randint(0, 144-size_xy+1, 2)
137 | # rand_z = np.random.randint(0, 144-size_z+1)
138 | # pet_low = pet_data[0:2, rand_xy[0]:rand_xy[0]+size_xy, rand_xy[1]:rand_xy[1]+size_xy, rand_z:rand_z+size_z].copy()
139 | # label = pet_data[2, rand_xy[0]:rand_xy[0]+size_xy, rand_xy[1]:rand_xy[1]+size_xy, rand_z:rand_z+size_z].copy()
140 | H, W, T = pet_low.shape
141 |
142 | out_dict = {}
143 | if self.local_classes is not None:
144 | out_dict["y"] = np.array(self.local_classes[idx], dtype=np.int64)
145 | return np.transpose(pet_low.reshape((1, H, W, T)), [0, 3, 1, 2]), np.transpose(label.reshape((1, H, W, T)), [0, 3, 1, 2]), out_dict
146 |
147 |
148 |
149 | def center_crop_arr(pil_image, image_size):
150 | # We are not on a new enough PIL to support the `reducing_gap`
151 | # argument, which uses BOX downsampling at powers of two first.
152 | # Thus, we do it by hand to improve downsample quality.
153 | while min(*pil_image.size) >= 2 * image_size:
154 | pil_image = pil_image.resize(
155 | tuple(x // 2 for x in pil_image.size), resample=Image.BOX
156 | )
157 |
158 | scale = image_size / min(*pil_image.size)
159 | pil_image = pil_image.resize(
160 | tuple(round(x * scale) for x in pil_image.size), resample=Image.BICUBIC
161 | )
162 |
163 | arr = np.array(pil_image)
164 | crop_y = (arr.shape[0] - image_size) // 2
165 | crop_x = (arr.shape[1] - image_size) // 2
166 | return arr[crop_y : crop_y + image_size, crop_x : crop_x + image_size]
167 |
168 |
169 | def random_crop_arr(pil_image, image_size, min_crop_frac=0.8, max_crop_frac=1.0):
170 | min_smaller_dim_size = math.ceil(image_size / max_crop_frac)
171 | max_smaller_dim_size = math.ceil(image_size / min_crop_frac)
172 | smaller_dim_size = random.randrange(min_smaller_dim_size, max_smaller_dim_size + 1)
173 |
174 | # We are not on a new enough PIL to support the `reducing_gap`
175 | # argument, which uses BOX downsampling at powers of two first.
176 | # Thus, we do it by hand to improve downsample quality.
177 | while min(*pil_image.size) >= 2 * smaller_dim_size:
178 | pil_image = pil_image.resize(
179 | tuple(x // 2 for x in pil_image.size), resample=Image.BOX
180 | )
181 |
182 | scale = smaller_dim_size / min(*pil_image.size)
183 | pil_image = pil_image.resize(
184 | tuple(round(x * scale) for x in pil_image.size), resample=Image.BICUBIC
185 | )
186 |
187 | arr = np.array(pil_image)
188 | crop_y = random.randrange(arr.shape[0] - image_size + 1)
189 | crop_x = random.randrange(arr.shape[1] - image_size + 1)
190 | return arr[crop_y : crop_y + image_size, crop_x : crop_x + image_size]
191 |
--------------------------------------------------------------------------------
/guided_diffusion/logger.py:
--------------------------------------------------------------------------------
1 | """
2 | Logger copied from OpenAI baselines to avoid extra RL-based dependencies:
3 | https://github.com/openai/baselines/blob/ea25b9e8b234e6ee1bca43083f8f3cf974143998/baselines/logger.py
4 | """
5 |
6 | import os
7 | import sys
8 | import shutil
9 | import os.path as osp
10 | import json
11 | import time
12 | import datetime
13 | import tempfile
14 | import warnings
15 | from collections import defaultdict
16 | from contextlib import contextmanager
17 |
18 | DEBUG = 10
19 | INFO = 20
20 | WARN = 30
21 | ERROR = 40
22 |
23 | DISABLED = 50
24 |
25 |
26 | class KVWriter(object):
27 | def writekvs(self, kvs):
28 | raise NotImplementedError
29 |
30 |
31 | class SeqWriter(object):
32 | def writeseq(self, seq):
33 | raise NotImplementedError
34 |
35 |
36 | class HumanOutputFormat(KVWriter, SeqWriter):
37 | def __init__(self, filename_or_file):
38 | if isinstance(filename_or_file, str):
39 | self.file = open(filename_or_file, "wt")
40 | self.own_file = True
41 | else:
42 | assert hasattr(filename_or_file, "read"), (
43 | "expected file or str, got %s" % filename_or_file
44 | )
45 | self.file = filename_or_file
46 | self.own_file = False
47 |
48 | def writekvs(self, kvs):
49 | # Create strings for printing
50 | key2str = {}
51 | for (key, val) in sorted(kvs.items()):
52 | if hasattr(val, "__float__"):
53 | valstr = "%-8.3g" % val
54 | else:
55 | valstr = str(val)
56 | key2str[self._truncate(key)] = self._truncate(valstr)
57 |
58 | # Find max widths
59 | if len(key2str) == 0:
60 | print("WARNING: tried to write empty key-value dict")
61 | return
62 | else:
63 | keywidth = max(map(len, key2str.keys()))
64 | valwidth = max(map(len, key2str.values()))
65 |
66 | # Write out the data
67 | dashes = "-" * (keywidth + valwidth + 7)
68 | lines = [dashes]
69 | for (key, val) in sorted(key2str.items(), key=lambda kv: kv[0].lower()):
70 | lines.append(
71 | "| %s%s | %s%s |"
72 | % (key, " " * (keywidth - len(key)), val, " " * (valwidth - len(val)))
73 | )
74 | lines.append(dashes)
75 | self.file.write("\n".join(lines) + "\n")
76 |
77 | # Flush the output to the file
78 | self.file.flush()
79 |
80 | def _truncate(self, s):
81 | maxlen = 30
82 | return s[: maxlen - 3] + "..." if len(s) > maxlen else s
83 |
84 | def writeseq(self, seq):
85 | seq = list(seq)
86 | for (i, elem) in enumerate(seq):
87 | self.file.write(elem)
88 | if i < len(seq) - 1: # add space unless this is the last one
89 | self.file.write(" ")
90 | self.file.write("\n")
91 | self.file.flush()
92 |
93 | def close(self):
94 | if self.own_file:
95 | self.file.close()
96 |
97 |
98 | class JSONOutputFormat(KVWriter):
99 | def __init__(self, filename):
100 | self.file = open(filename, "wt")
101 |
102 | def writekvs(self, kvs):
103 | for k, v in sorted(kvs.items()):
104 | if hasattr(v, "dtype"):
105 | kvs[k] = float(v)
106 | self.file.write(json.dumps(kvs) + "\n")
107 | self.file.flush()
108 |
109 | def close(self):
110 | self.file.close()
111 |
112 |
113 | class CSVOutputFormat(KVWriter):
114 | def __init__(self, filename):
115 | self.file = open(filename, "w+t")
116 | self.keys = []
117 | self.sep = ","
118 |
119 | def writekvs(self, kvs):
120 | # Add our current row to the history
121 | extra_keys = list(kvs.keys() - self.keys)
122 | extra_keys.sort()
123 | if extra_keys:
124 | self.keys.extend(extra_keys)
125 | self.file.seek(0)
126 | lines = self.file.readlines()
127 | self.file.seek(0)
128 | for (i, k) in enumerate(self.keys):
129 | if i > 0:
130 | self.file.write(",")
131 | self.file.write(k)
132 | self.file.write("\n")
133 | for line in lines[1:]:
134 | self.file.write(line[:-1])
135 | self.file.write(self.sep * len(extra_keys))
136 | self.file.write("\n")
137 | for (i, k) in enumerate(self.keys):
138 | if i > 0:
139 | self.file.write(",")
140 | v = kvs.get(k)
141 | if v is not None:
142 | self.file.write(str(v))
143 | self.file.write("\n")
144 | self.file.flush()
145 |
146 | def close(self):
147 | self.file.close()
148 |
149 |
150 | class TensorBoardOutputFormat(KVWriter):
151 | """
152 | Dumps key/value pairs into TensorBoard's numeric format.
153 | """
154 |
155 | def __init__(self, dir):
156 | os.makedirs(dir, exist_ok=True)
157 | self.dir = dir
158 | self.step = 1
159 | prefix = "events"
160 | path = osp.join(osp.abspath(dir), prefix)
161 | import tensorflow as tf
162 | from tensorflow.python import pywrap_tensorflow
163 | from tensorflow.core.util import event_pb2
164 | from tensorflow.python.util import compat
165 |
166 | self.tf = tf
167 | self.event_pb2 = event_pb2
168 | self.pywrap_tensorflow = pywrap_tensorflow
169 | self.writer = pywrap_tensorflow.EventsWriter(compat.as_bytes(path))
170 |
171 | def writekvs(self, kvs):
172 | def summary_val(k, v):
173 | kwargs = {"tag": k, "simple_value": float(v)}
174 | return self.tf.Summary.Value(**kwargs)
175 |
176 | summary = self.tf.Summary(value=[summary_val(k, v) for k, v in kvs.items()])
177 | event = self.event_pb2.Event(wall_time=time.time(), summary=summary)
178 | event.step = (
179 | self.step
180 | ) # is there any reason why you'd want to specify the step?
181 | self.writer.WriteEvent(event)
182 | self.writer.Flush()
183 | self.step += 1
184 |
185 | def close(self):
186 | if self.writer:
187 | self.writer.Close()
188 | self.writer = None
189 |
190 |
191 | def make_output_format(format, ev_dir, log_suffix=""):
192 | os.makedirs(ev_dir, exist_ok=True)
193 | if format == "stdout":
194 | return HumanOutputFormat(sys.stdout)
195 | elif format == "log":
196 | return HumanOutputFormat(osp.join(ev_dir, "log%s.txt" % log_suffix))
197 | elif format == "json":
198 | return JSONOutputFormat(osp.join(ev_dir, "progress%s.json" % log_suffix))
199 | elif format == "csv":
200 | return CSVOutputFormat(osp.join(ev_dir, "progress%s.csv" % log_suffix))
201 | elif format == "tensorboard":
202 | return TensorBoardOutputFormat(osp.join(ev_dir, "tb%s" % log_suffix))
203 | else:
204 | raise ValueError("Unknown format specified: %s" % (format,))
205 |
206 |
207 | # ================================================================
208 | # API
209 | # ================================================================
210 |
211 |
212 | def logkv(key, val):
213 | """
214 | Log a value of some diagnostic
215 | Call this once for each diagnostic quantity, each iteration
216 | If called many times, last value will be used.
217 | """
218 | get_current().logkv(key, val)
219 |
220 |
221 | def logkv_mean(key, val):
222 | """
223 | The same as logkv(), but if called many times, values averaged.
224 | """
225 | get_current().logkv_mean(key, val)
226 |
227 |
228 | def logkvs(d):
229 | """
230 | Log a dictionary of key-value pairs
231 | """
232 | for (k, v) in d.items():
233 | logkv(k, v)
234 |
235 |
236 | def dumpkvs():
237 | """
238 | Write all of the diagnostics from the current iteration
239 | """
240 | return get_current().dumpkvs()
241 |
242 |
243 | def getkvs():
244 | return get_current().name2val
245 |
246 |
247 | def log(*args, level=INFO):
248 | """
249 | Write the sequence of args, with no separators, to the console and output files (if you've configured an output file).
250 | """
251 | get_current().log(*args, level=level)
252 |
253 |
254 | def debug(*args):
255 | log(*args, level=DEBUG)
256 |
257 |
258 | def info(*args):
259 | log(*args, level=INFO)
260 |
261 |
262 | def warn(*args):
263 | log(*args, level=WARN)
264 |
265 |
266 | def error(*args):
267 | log(*args, level=ERROR)
268 |
269 |
270 | def set_level(level):
271 | """
272 | Set logging threshold on current logger.
273 | """
274 | get_current().set_level(level)
275 |
276 |
277 | def set_comm(comm):
278 | get_current().set_comm(comm)
279 |
280 |
281 | def get_dir():
282 | """
283 | Get directory that log files are being written to.
284 | will be None if there is no output directory (i.e., if you didn't call start)
285 | """
286 | return get_current().get_dir()
287 |
288 |
289 | record_tabular = logkv
290 | dump_tabular = dumpkvs
291 |
292 |
293 | @contextmanager
294 | def profile_kv(scopename):
295 | logkey = "wait_" + scopename
296 | tstart = time.time()
297 | try:
298 | yield
299 | finally:
300 | get_current().name2val[logkey] += time.time() - tstart
301 |
302 |
303 | def profile(n):
304 | """
305 | Usage:
306 | @profile("my_func")
307 | def my_func(): code
308 | """
309 |
310 | def decorator_with_name(func):
311 | def func_wrapper(*args, **kwargs):
312 | with profile_kv(n):
313 | return func(*args, **kwargs)
314 |
315 | return func_wrapper
316 |
317 | return decorator_with_name
318 |
319 |
320 | # ================================================================
321 | # Backend
322 | # ================================================================
323 |
324 |
325 | def get_current():
326 | if Logger.CURRENT is None:
327 | _configure_default_logger()
328 |
329 | return Logger.CURRENT
330 |
331 |
332 | class Logger(object):
333 | DEFAULT = None # A logger with no output files. (See right below class definition)
334 | # So that you can still log to the terminal without setting up any output files
335 | CURRENT = None # Current logger being used by the free functions above
336 |
337 | def __init__(self, dir, output_formats, comm=None):
338 | self.name2val = defaultdict(float) # values this iteration
339 | self.name2cnt = defaultdict(int)
340 | self.level = INFO
341 | self.dir = dir
342 | self.output_formats = output_formats
343 | self.comm = comm
344 |
345 | # Logging API, forwarded
346 | # ----------------------------------------
347 | def logkv(self, key, val):
348 | self.name2val[key] = val
349 |
350 | def logkv_mean(self, key, val):
351 | oldval, cnt = self.name2val[key], self.name2cnt[key]
352 | self.name2val[key] = oldval * cnt / (cnt + 1) + val / (cnt + 1)
353 | self.name2cnt[key] = cnt + 1
354 |
355 | def dumpkvs(self):
356 | if self.comm is None:
357 | d = self.name2val
358 | else:
359 | d = mpi_weighted_mean(
360 | self.comm,
361 | {
362 | name: (val, self.name2cnt.get(name, 1))
363 | for (name, val) in self.name2val.items()
364 | },
365 | )
366 | if self.comm.rank != 0:
367 | d["dummy"] = 1 # so we don't get a warning about empty dict
368 | out = d.copy() # Return the dict for unit testing purposes
369 | for fmt in self.output_formats:
370 | if isinstance(fmt, KVWriter):
371 | fmt.writekvs(d)
372 | self.name2val.clear()
373 | self.name2cnt.clear()
374 | return out
375 |
376 | def log(self, *args, level=INFO):
377 | if self.level <= level:
378 | self._do_log(args)
379 |
380 | # Configuration
381 | # ----------------------------------------
382 | def set_level(self, level):
383 | self.level = level
384 |
385 | def set_comm(self, comm):
386 | self.comm = comm
387 |
388 | def get_dir(self):
389 | return self.dir
390 |
391 | def close(self):
392 | for fmt in self.output_formats:
393 | fmt.close()
394 |
395 | # Misc
396 | # ----------------------------------------
397 | def _do_log(self, args):
398 | for fmt in self.output_formats:
399 | if isinstance(fmt, SeqWriter):
400 | fmt.writeseq(map(str, args))
401 |
402 |
403 | def get_rank_without_mpi_import():
404 | # check environment variables here instead of importing mpi4py
405 | # to avoid calling MPI_Init() when this module is imported
406 | for varname in ["PMI_RANK", "OMPI_COMM_WORLD_RANK"]:
407 | if varname in os.environ:
408 | return int(os.environ[varname])
409 | return 0
410 |
411 |
412 | def mpi_weighted_mean(comm, local_name2valcount):
413 | """
414 | Copied from: https://github.com/openai/baselines/blob/ea25b9e8b234e6ee1bca43083f8f3cf974143998/baselines/common/mpi_util.py#L110
415 | Perform a weighted average over dicts that are each on a different node
416 | Input: local_name2valcount: dict mapping key -> (value, count)
417 | Returns: key -> mean
418 | """
419 | all_name2valcount = comm.gather(local_name2valcount)
420 | if comm.rank == 0:
421 | name2sum = defaultdict(float)
422 | name2count = defaultdict(float)
423 | for n2vc in all_name2valcount:
424 | for (name, (val, count)) in n2vc.items():
425 | try:
426 | val = float(val)
427 | except ValueError:
428 | if comm.rank == 0:
429 | warnings.warn(
430 | "WARNING: tried to compute mean on non-float {}={}".format(
431 | name, val
432 | )
433 | )
434 | else:
435 | name2sum[name] += val * count
436 | name2count[name] += count
437 | return {name: name2sum[name] / name2count[name] for name in name2sum}
438 | else:
439 | return {}
440 |
441 |
442 | def configure(dir=None, format_strs=None, comm=None, log_suffix=""):
443 | """
444 | If comm is provided, average all numerical stats across that comm
445 | """
446 | if dir is None:
447 | dir = os.getenv("OPENAI_LOGDIR")
448 | if dir is None:
449 | dir = osp.join(
450 | tempfile.gettempdir(),
451 | datetime.datetime.now().strftime("openai-%Y-%m-%d-%H-%M-%S-%f"),
452 | )
453 | assert isinstance(dir, str)
454 | dir = os.path.expanduser(dir)
455 | os.makedirs(os.path.expanduser(dir), exist_ok=True)
456 |
457 | rank = get_rank_without_mpi_import()
458 | if rank > 0:
459 | log_suffix = log_suffix + "-rank%03i" % rank
460 |
461 | if format_strs is None:
462 | if rank == 0:
463 | format_strs = os.getenv("OPENAI_LOG_FORMAT", "stdout,log,csv").split(",")
464 | else:
465 | format_strs = os.getenv("OPENAI_LOG_FORMAT_MPI", "log").split(",")
466 | format_strs = filter(None, format_strs)
467 | output_formats = [make_output_format(f, dir, log_suffix) for f in format_strs]
468 |
469 | Logger.CURRENT = Logger(dir=dir, output_formats=output_formats, comm=comm)
470 | if output_formats:
471 | log("Logging to %s" % dir)
472 |
473 |
474 | def _configure_default_logger():
475 | configure()
476 | Logger.DEFAULT = Logger.CURRENT
477 |
478 |
479 | def reset():
480 | if Logger.CURRENT is not Logger.DEFAULT:
481 | Logger.CURRENT.close()
482 | Logger.CURRENT = Logger.DEFAULT
483 | log("Reset logger")
484 |
485 |
486 | @contextmanager
487 | def scoped_configure(dir=None, format_strs=None, comm=None):
488 | prevlogger = Logger.CURRENT
489 | configure(dir=dir, format_strs=format_strs, comm=comm)
490 | try:
491 | yield
492 | finally:
493 | Logger.CURRENT.close()
494 | Logger.CURRENT = prevlogger
495 |
496 |
--------------------------------------------------------------------------------
/guided_diffusion/losses.py:
--------------------------------------------------------------------------------
1 | """
2 | Helpers for various likelihood-based losses. These are ported from the original
3 | Ho et al. diffusion models codebase:
4 | https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/utils.py
5 | """
6 |
7 | import numpy as np
8 |
9 | import torch as th
10 |
11 |
12 | def normal_kl(mean1, logvar1, mean2, logvar2):
13 | """
14 | Compute the KL divergence between two gaussians.
15 |
16 | Shapes are automatically broadcasted, so batches can be compared to
17 | scalars, among other use cases.
18 | """
19 | tensor = None
20 | for obj in (mean1, logvar1, mean2, logvar2):
21 | if isinstance(obj, th.Tensor):
22 | tensor = obj
23 | break
24 | assert tensor is not None, "at least one argument must be a Tensor"
25 |
26 | # Force variances to be Tensors. Broadcasting helps convert scalars to
27 | # Tensors, but it does not work for th.exp().
28 | logvar1, logvar2 = [
29 | x if isinstance(x, th.Tensor) else th.tensor(x).to(tensor)
30 | for x in (logvar1, logvar2)
31 | ]
32 |
33 | return 0.5 * (
34 | -1.0
35 | + logvar2
36 | - logvar1
37 | + th.exp(logvar1 - logvar2)
38 | + ((mean1 - mean2) ** 2) * th.exp(-logvar2)
39 | )
40 |
41 |
42 | def approx_standard_normal_cdf(x):
43 | """
44 | A fast approximation of the cumulative distribution function of the
45 | standard normal.
46 | """
47 | return 0.5 * (1.0 + th.tanh(np.sqrt(2.0 / np.pi) * (x + 0.044715 * th.pow(x, 3))))
48 |
49 |
50 | def discretized_gaussian_log_likelihood(x, *, means, log_scales):
51 | """
52 | Compute the log-likelihood of a Gaussian distribution discretizing to a
53 | given image.
54 |
55 | :param x: the target images. It is assumed that this was uint8 values,
56 | rescaled to the range [-1, 1].
57 | :param means: the Gaussian mean Tensor.
58 | :param log_scales: the Gaussian log stddev Tensor.
59 | :return: a tensor like x of log probabilities (in nats).
60 | """
61 | assert x.shape == means.shape == log_scales.shape
62 | centered_x = x - means
63 | inv_stdv = th.exp(-log_scales)
64 | plus_in = inv_stdv * (centered_x + 1.0 / 255.0)
65 | cdf_plus = approx_standard_normal_cdf(plus_in)
66 | min_in = inv_stdv * (centered_x - 1.0 / 255.0)
67 | cdf_min = approx_standard_normal_cdf(min_in)
68 | log_cdf_plus = th.log(cdf_plus.clamp(min=1e-12))
69 | log_one_minus_cdf_min = th.log((1.0 - cdf_min).clamp(min=1e-12))
70 | cdf_delta = cdf_plus - cdf_min
71 | log_probs = th.where(
72 | x < -0.999,
73 | log_cdf_plus,
74 | th.where(x > 0.999, log_one_minus_cdf_min, th.log(cdf_delta.clamp(min=1e-12))),
75 | )
76 | assert log_probs.shape == x.shape
77 | return log_probs
78 |
--------------------------------------------------------------------------------
/guided_diffusion/nn.py:
--------------------------------------------------------------------------------
1 | """
2 | Various utilities for neural networks.
3 | """
4 |
5 | import math
6 |
7 | import torch as th
8 | import torch.nn as nn
9 |
10 |
11 | # PyTorch 1.7 has SiLU, but we support PyTorch 1.5.
12 | class SiLU(nn.Module):
13 | def forward(self, x):
14 | return x * th.sigmoid(x)
15 |
16 |
17 | class GroupNorm32(nn.GroupNorm):
18 | def forward(self, x):
19 | return super().forward(x.float()).type(x.dtype)
20 |
21 |
22 | def conv_nd(dims, *args, **kwargs):
23 | """
24 | Create a 1D, 2D, or 3D convolution module.
25 | """
26 | if dims == 1:
27 | return nn.Conv1d(*args, **kwargs)
28 | elif dims == 2:
29 | return nn.Conv2d(*args, **kwargs)
30 | elif dims == 3:
31 | return nn.Conv3d(*args, **kwargs)
32 | raise ValueError(f"unsupported dimensions: {dims}")
33 |
34 |
35 | def linear(*args, **kwargs):
36 | """
37 | Create a linear module.
38 | """
39 | return nn.Linear(*args, **kwargs)
40 |
41 |
42 | def avg_pool_nd(dims, *args, **kwargs):
43 | """
44 | Create a 1D, 2D, or 3D average pooling module.
45 | """
46 | if dims == 1:
47 | return nn.AvgPool1d(*args, **kwargs)
48 | elif dims == 2:
49 | return nn.AvgPool2d(*args, **kwargs)
50 | elif dims == 3:
51 | return nn.AvgPool3d(*args, **kwargs)
52 | raise ValueError(f"unsupported dimensions: {dims}")
53 |
54 |
55 | def update_ema(target_params, source_params, rate=0.99):
56 | """
57 | Update target parameters to be closer to those of source parameters using
58 | an exponential moving average.
59 |
60 | :param target_params: the target parameter sequence.
61 | :param source_params: the source parameter sequence.
62 | :param rate: the EMA rate (closer to 1 means slower).
63 | """
64 | for targ, src in zip(target_params, source_params):
65 | targ.detach().mul_(rate).add_(src, alpha=1 - rate)
66 |
67 |
68 | def zero_module(module):
69 | """
70 | Zero out the parameters of a module and return it.
71 | """
72 | for p in module.parameters():
73 | p.detach().zero_()
74 | return module
75 |
76 |
77 | def scale_module(module, scale):
78 | """
79 | Scale the parameters of a module and return it.
80 | """
81 | for p in module.parameters():
82 | p.detach().mul_(scale)
83 | return module
84 |
85 |
86 | def mean_flat(tensor):
87 | """
88 | Take the mean over all non-batch dimensions.
89 | """
90 | return tensor.mean(dim=list(range(1, len(tensor.shape))))
91 |
92 |
93 | def normalization(channels):
94 | """
95 | Make a standard normalization layer.
96 |
97 | :param channels: number of input channels.
98 | :return: an nn.Module for normalization.
99 | """
100 | return GroupNorm32(32, channels)
101 |
102 |
103 | def timestep_embedding(timesteps, dim, max_period=10000):
104 | """
105 | Create sinusoidal timestep embeddings.
106 |
107 | :param timesteps: a 1-D Tensor of N indices, one per batch element.
108 | These may be fractional.
109 | :param dim: the dimension of the output.
110 | :param max_period: controls the minimum frequency of the embeddings.
111 | :return: an [N x dim] Tensor of positional embeddings.
112 | """
113 | half = dim // 2
114 | freqs = th.exp(
115 | -math.log(max_period) * th.arange(start=0, end=half, dtype=th.float32) / half
116 | ).to(device=timesteps.device)
117 | args = timesteps[:, None].float() * freqs[None]
118 | embedding = th.cat([th.cos(args), th.sin(args)], dim=-1)
119 | if dim % 2:
120 | embedding = th.cat([embedding, th.zeros_like(embedding[:, :1])], dim=-1)
121 | return embedding
122 |
123 |
124 | def checkpoint(func, inputs, params, flag):
125 | """
126 | Evaluate a function without caching intermediate activations, allowing for
127 | reduced memory at the expense of extra compute in the backward pass.
128 |
129 | :param func: the function to evaluate.
130 | :param inputs: the argument sequence to pass to `func`.
131 | :param params: a sequence of parameters `func` depends on but does not
132 | explicitly take as arguments.
133 | :param flag: if False, disable gradient checkpointing.
134 | """
135 | if flag:
136 | args = tuple(inputs) + tuple(params)
137 | return CheckpointFunction.apply(func, len(inputs), *args)
138 | else:
139 | return func(*inputs)
140 |
141 |
142 | class CheckpointFunction(th.autograd.Function):
143 | @staticmethod
144 | def forward(ctx, run_function, length, *args):
145 | ctx.run_function = run_function
146 | ctx.input_tensors = list(args[:length])
147 | ctx.input_params = list(args[length:])
148 | with th.no_grad():
149 | output_tensors = ctx.run_function(*ctx.input_tensors)
150 | return output_tensors
151 |
152 | @staticmethod
153 | def backward(ctx, *output_grads):
154 | ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors]
155 | with th.enable_grad():
156 | # Fixes a bug where the first op in run_function modifies the
157 | # Tensor storage in place, which is not allowed for detach()'d
158 | # Tensors.
159 | shallow_copies = [x.view_as(x) for x in ctx.input_tensors]
160 | output_tensors = ctx.run_function(*shallow_copies)
161 | input_grads = th.autograd.grad(
162 | output_tensors,
163 | ctx.input_tensors + ctx.input_params,
164 | output_grads,
165 | allow_unused=True,
166 | )
167 | del ctx.input_tensors
168 | del ctx.input_params
169 | del output_tensors
170 | return (None, None) + input_grads
171 |
--------------------------------------------------------------------------------
/guided_diffusion/resample.py:
--------------------------------------------------------------------------------
1 | from abc import ABC, abstractmethod
2 |
3 | import numpy as np
4 | import torch as th
5 | import torch.distributed as dist
6 |
7 |
8 | def create_named_schedule_sampler(name, diffusion):
9 | """
10 | Create a ScheduleSampler from a library of pre-defined samplers.
11 |
12 | :param name: the name of the sampler.
13 | :param diffusion: the diffusion object to sample for.
14 | """
15 | if name == "uniform":
16 | return UniformSampler(diffusion)
17 | elif name == "loss-second-moment":
18 | return LossSecondMomentResampler(diffusion)
19 | else:
20 | raise NotImplementedError(f"unknown schedule sampler: {name}")
21 |
22 |
23 | class ScheduleSampler(ABC):
24 | """
25 | A distribution over timesteps in the diffusion process, intended to reduce
26 | variance of the objective.
27 |
28 | By default, samplers perform unbiased importance sampling, in which the
29 | objective's mean is unchanged.
30 | However, subclasses may override sample() to change how the resampled
31 | terms are reweighted, allowing for actual changes in the objective.
32 | """
33 |
34 | @abstractmethod
35 | def weights(self):
36 | """
37 | Get a numpy array of weights, one per diffusion step.
38 |
39 | The weights needn't be normalized, but must be positive.
40 | """
41 |
42 | def sample(self, batch_size, device):
43 | """
44 | Importance-sample timesteps for a batch.
45 |
46 | :param batch_size: the number of timesteps.
47 | :param device: the torch device to save to.
48 | :return: a tuple (timesteps, weights):
49 | - timesteps: a tensor of timestep indices.
50 | - weights: a tensor of weights to scale the resulting losses.
51 | """
52 | w = self.weights()
53 | p = w / np.sum(w)
54 | indices_np = np.random.choice(len(p), size=(batch_size,), p=p)
55 | indices = th.from_numpy(indices_np).long().to(device)
56 | weights_np = 1 / (len(p) * p[indices_np])
57 | weights = th.from_numpy(weights_np).float().to(device)
58 | return indices, weights
59 |
60 |
61 | class UniformSampler(ScheduleSampler):
62 | def __init__(self, diffusion):
63 | self.diffusion = diffusion
64 | self._weights = np.ones([diffusion.num_timesteps])
65 |
66 | def weights(self):
67 | return self._weights
68 |
69 |
70 | class LossAwareSampler(ScheduleSampler):
71 | def update_with_local_losses(self, local_ts, local_losses):
72 | """
73 | Update the reweighting using losses from a model.
74 |
75 | Call this method from each rank with a batch of timesteps and the
76 | corresponding losses for each of those timesteps.
77 | This method will perform synchronization to make sure all of the ranks
78 | maintain the exact same reweighting.
79 |
80 | :param local_ts: an integer Tensor of timesteps.
81 | :param local_losses: a 1D Tensor of losses.
82 | """
83 | batch_sizes = [
84 | th.tensor([0], dtype=th.int32, device=local_ts.device)
85 | for _ in range(dist.get_world_size())
86 | ]
87 | dist.all_gather(
88 | batch_sizes,
89 | th.tensor([len(local_ts)], dtype=th.int32, device=local_ts.device),
90 | )
91 |
92 | # Pad all_gather batches to be the maximum batch size.
93 | batch_sizes = [x.item() for x in batch_sizes]
94 | max_bs = max(batch_sizes)
95 |
96 | timestep_batches = [th.zeros(max_bs).to(local_ts) for bs in batch_sizes]
97 | loss_batches = [th.zeros(max_bs).to(local_losses) for bs in batch_sizes]
98 | dist.all_gather(timestep_batches, local_ts)
99 | dist.all_gather(loss_batches, local_losses)
100 | timesteps = [
101 | x.item() for y, bs in zip(timestep_batches, batch_sizes) for x in y[:bs]
102 | ]
103 | losses = [x.item() for y, bs in zip(loss_batches, batch_sizes) for x in y[:bs]]
104 | self.update_with_all_losses(timesteps, losses)
105 |
106 | @abstractmethod
107 | def update_with_all_losses(self, ts, losses):
108 | """
109 | Update the reweighting using losses from a model.
110 |
111 | Sub-classes should override this method to update the reweighting
112 | using losses from the model.
113 |
114 | This method directly updates the reweighting without synchronizing
115 | between workers. It is called by update_with_local_losses from all
116 | ranks with identical arguments. Thus, it should have deterministic
117 | behavior to maintain state across workers.
118 |
119 | :param ts: a list of int timesteps.
120 | :param losses: a list of float losses, one per timestep.
121 | """
122 |
123 |
124 | class LossSecondMomentResampler(LossAwareSampler):
125 | def __init__(self, diffusion, history_per_term=10, uniform_prob=0.001):
126 | self.diffusion = diffusion
127 | self.history_per_term = history_per_term
128 | self.uniform_prob = uniform_prob
129 | self._loss_history = np.zeros(
130 | [diffusion.num_timesteps, history_per_term], dtype=np.float64
131 | )
132 | self._loss_counts = np.zeros([diffusion.num_timesteps], dtype=np.int)
133 |
134 | def weights(self):
135 | if not self._warmed_up():
136 | return np.ones([self.diffusion.num_timesteps], dtype=np.float64)
137 | weights = np.sqrt(np.mean(self._loss_history ** 2, axis=-1))
138 | weights /= np.sum(weights)
139 | weights *= 1 - self.uniform_prob
140 | weights += self.uniform_prob / len(weights)
141 | return weights
142 |
143 | def update_with_all_losses(self, ts, losses):
144 | for t, loss in zip(ts, losses):
145 | if self._loss_counts[t] == self.history_per_term:
146 | # Shift out the oldest loss term.
147 | self._loss_history[t, :-1] = self._loss_history[t, 1:]
148 | self._loss_history[t, -1] = loss
149 | else:
150 | self._loss_history[t, self._loss_counts[t]] = loss
151 | self._loss_counts[t] += 1
152 |
153 | def _warmed_up(self):
154 | return (self._loss_counts == self.history_per_term).all()
155 |
--------------------------------------------------------------------------------
/guided_diffusion/respace.py:
--------------------------------------------------------------------------------
1 | import numpy as np
2 | import torch as th
3 |
4 | from .gaussian_diffusion import GaussianDiffusion
5 |
6 |
7 | def space_timesteps(num_timesteps, section_counts):
8 | """
9 | Create a list of timesteps to use from an original diffusion process,
10 | given the number of timesteps we want to take from equally-sized portions
11 | of the original process.
12 |
13 | For example, if there's 300 timesteps and the section counts are [10,15,20]
14 | then the first 100 timesteps are strided to be 10 timesteps, the second 100
15 | are strided to be 15 timesteps, and the final 100 are strided to be 20.
16 |
17 | If the stride is a string starting with "ddim", then the fixed striding
18 | from the DDIM paper is used, and only one section is allowed.
19 |
20 | :param num_timesteps: the number of diffusion steps in the original
21 | process to divide up.
22 | :param section_counts: either a list of numbers, or a string containing
23 | comma-separated numbers, indicating the step count
24 | per section. As a special case, use "ddimN" where N
25 | is a number of steps to use the striding from the
26 | DDIM paper.
27 | :return: a set of diffusion steps from the original process to use.
28 | """
29 | if isinstance(section_counts, str):
30 | if section_counts.startswith("ddim"):
31 | desired_count = int(section_counts[len("ddim") :])
32 | for i in range(1, num_timesteps):
33 | if len(range(0, num_timesteps, i)) == desired_count:
34 | return set(range(0, num_timesteps, i))
35 | raise ValueError(
36 | f"cannot create exactly {num_timesteps} steps with an integer stride"
37 | )
38 | section_counts = [int(x) for x in section_counts.split(",")]
39 | size_per = num_timesteps // len(section_counts)
40 | extra = num_timesteps % len(section_counts)
41 | start_idx = 0
42 | all_steps = []
43 | for i, section_count in enumerate(section_counts):
44 | size = size_per + (1 if i < extra else 0)
45 | if size < section_count:
46 | raise ValueError(
47 | f"cannot divide section of {size} steps into {section_count}"
48 | )
49 | if section_count <= 1:
50 | frac_stride = 1
51 | else:
52 | frac_stride = (size - 1) / (section_count - 1)
53 | cur_idx = 0.0
54 | taken_steps = []
55 | for _ in range(section_count):
56 | taken_steps.append(start_idx + round(cur_idx))
57 | cur_idx += frac_stride
58 | all_steps += taken_steps
59 | start_idx += size
60 | return set(all_steps)
61 |
62 |
63 | class SpacedDiffusion(GaussianDiffusion):
64 | """
65 | A diffusion process which can skip steps in a base diffusion process.
66 |
67 | :param use_timesteps: a collection (sequence or set) of timesteps from the
68 | original diffusion process to retain.
69 | :param kwargs: the kwargs to create the base diffusion process.
70 | """
71 |
72 | def __init__(self, use_timesteps, **kwargs):
73 | self.use_timesteps = set(use_timesteps)
74 | self.timestep_map = []
75 | self.original_num_steps = len(kwargs["betas"])
76 |
77 | base_diffusion = GaussianDiffusion(**kwargs) # pylint: disable=missing-kwoa
78 | last_alpha_cumprod = 1.0
79 | new_betas = []
80 | for i, alpha_cumprod in enumerate(base_diffusion.alphas_cumprod):
81 | if i in self.use_timesteps:
82 | new_betas.append(1 - alpha_cumprod / last_alpha_cumprod)
83 | last_alpha_cumprod = alpha_cumprod
84 | self.timestep_map.append(i)
85 | kwargs["betas"] = np.array(new_betas)
86 | super().__init__(**kwargs)
87 |
88 | def p_mean_variance(
89 | self, model, *args, **kwargs
90 | ): # pylint: disable=signature-differs
91 | return super().p_mean_variance(self._wrap_model(model), *args, **kwargs)
92 |
93 | def training_losses(
94 | self, model, *args, **kwargs
95 | ): # pylint: disable=signature-differs
96 | return super().training_losses(self._wrap_model(model), *args, **kwargs)
97 |
98 | def condition_mean(self, cond_fn, *args, **kwargs):
99 | return super().condition_mean(self._wrap_model(cond_fn), *args, **kwargs)
100 |
101 | def condition_score(self, cond_fn, *args, **kwargs):
102 | return super().condition_score(self._wrap_model(cond_fn), *args, **kwargs)
103 |
104 | def _wrap_model(self, model):
105 | if isinstance(model, _WrappedModel):
106 | return model
107 | return _WrappedModel(
108 | model, self.timestep_map, self.rescale_timesteps, self.original_num_steps
109 | )
110 |
111 | def _scale_timesteps(self, t):
112 | # Scaling is done by the wrapped model.
113 | return t
114 |
115 |
116 | class _WrappedModel:
117 | def __init__(self, model, timestep_map, rescale_timesteps, original_num_steps):
118 | self.model = model
119 | self.timestep_map = timestep_map
120 | self.rescale_timesteps = rescale_timesteps
121 | self.original_num_steps = original_num_steps
122 |
123 | def __call__(self, x, ts, **kwargs):
124 | map_tensor = th.tensor(self.timestep_map, device=ts.device, dtype=ts.dtype)
125 | new_ts = map_tensor[ts]
126 | if self.rescale_timesteps:
127 | new_ts = new_ts.float() * (1000.0 / self.original_num_steps)
128 | return self.model(x, new_ts, **kwargs)
129 |
--------------------------------------------------------------------------------
/guided_diffusion/script_util.py:
--------------------------------------------------------------------------------
1 | import argparse
2 | import inspect
3 |
4 | from . import gaussian_diffusion as gd
5 | from .respace import SpacedDiffusion, space_timesteps
6 | from .unet import SuperResModel, UNetModel, EncoderUNetModel, SegModelv2, SegModelv2_6c, SegModelv3_6c, SuperResModel_noatt, SegModelv2_3d_noatt, SegModel_3d_noatt_midcat
7 |
8 | NUM_CLASSES = 1000
9 |
10 |
11 | def diffusion_defaults():
12 | """
13 | Defaults for image and classifier training.
14 | """
15 | return dict(
16 | learn_sigma=False,
17 | diffusion_steps=1000,
18 | noise_schedule="linear",
19 | timestep_respacing="",
20 | use_kl=False,
21 | predict_xstart=False,
22 | rescale_timesteps=False,
23 | rescale_learned_sigmas=False,
24 | )
25 |
26 |
27 | def classifier_defaults():
28 | """
29 | Defaults for classifier models.
30 | """
31 | return dict(
32 | image_size=64,
33 | classifier_use_fp16=False,
34 | classifier_width=128,
35 | classifier_depth=2,
36 | classifier_attention_resolutions="32,16,8", # 16
37 | classifier_use_scale_shift_norm=True, # False
38 | classifier_resblock_updown=True, # False
39 | classifier_pool="attention",
40 | )
41 |
42 |
43 | def model_and_diffusion_defaults():
44 | """
45 | Defaults for image training.
46 | """
47 | res = dict(
48 | image_size=64,
49 | num_channels=128,
50 | num_res_blocks=2,
51 | num_heads=4,
52 | num_heads_upsample=-1,
53 | num_head_channels=-1,
54 | attention_resolutions="16,8",
55 | channel_mult="",
56 | dropout=0.0,
57 | class_cond=False,
58 | use_checkpoint=False,
59 | use_scale_shift_norm=True,
60 | resblock_updown=False,
61 | use_fp16=False,
62 | use_new_attention_order=False,
63 | )
64 | res.update(diffusion_defaults())
65 | return res
66 |
67 |
68 | def classifier_and_diffusion_defaults():
69 | res = classifier_defaults()
70 | res.update(diffusion_defaults())
71 | return res
72 |
73 |
74 | def create_model_and_diffusion(
75 | image_size,
76 | class_cond,
77 | learn_sigma,
78 | num_channels,
79 | num_res_blocks,
80 | channel_mult,
81 | num_heads,
82 | num_head_channels,
83 | num_heads_upsample,
84 | attention_resolutions,
85 | dropout,
86 | diffusion_steps,
87 | noise_schedule,
88 | timestep_respacing,
89 | use_kl,
90 | predict_xstart,
91 | rescale_timesteps,
92 | rescale_learned_sigmas,
93 | use_checkpoint,
94 | use_scale_shift_norm,
95 | resblock_updown,
96 | use_fp16,
97 | use_new_attention_order,
98 | ):
99 | model = create_model(
100 | image_size,
101 | num_channels,
102 | num_res_blocks,
103 | channel_mult=channel_mult,
104 | learn_sigma=learn_sigma,
105 | class_cond=class_cond,
106 | use_checkpoint=use_checkpoint,
107 | attention_resolutions=attention_resolutions,
108 | num_heads=num_heads,
109 | num_head_channels=num_head_channels,
110 | num_heads_upsample=num_heads_upsample,
111 | use_scale_shift_norm=use_scale_shift_norm,
112 | dropout=dropout,
113 | resblock_updown=resblock_updown,
114 | use_fp16=use_fp16,
115 | use_new_attention_order=use_new_attention_order,
116 | )
117 | diffusion = create_gaussian_diffusion(
118 | steps=diffusion_steps,
119 | learn_sigma=learn_sigma,
120 | noise_schedule=noise_schedule,
121 | use_kl=use_kl,
122 | predict_xstart=predict_xstart,
123 | rescale_timesteps=rescale_timesteps,
124 | rescale_learned_sigmas=rescale_learned_sigmas,
125 | timestep_respacing=timestep_respacing,
126 | )
127 | return model, diffusion
128 |
129 |
130 | def create_model(
131 | image_size,
132 | num_channels,
133 | num_res_blocks,
134 | channel_mult="",
135 | learn_sigma=False,
136 | class_cond=False,
137 | use_checkpoint=False,
138 | attention_resolutions="16",
139 | num_heads=1,
140 | num_head_channels=-1,
141 | num_heads_upsample=-1,
142 | use_scale_shift_norm=False,
143 | dropout=0,
144 | resblock_updown=False,
145 | use_fp16=False,
146 | use_new_attention_order=False,
147 | ):
148 | if channel_mult == "":
149 | if image_size == 512:
150 | channel_mult = (0.5, 1, 1, 2, 2, 4, 4)
151 | elif image_size == 256:
152 | channel_mult = (1, 1, 2, 2, 4, 4)
153 | elif image_size == 128:
154 | channel_mult = (1, 1, 2, 3, 4)
155 | elif image_size == 64:
156 | channel_mult = (1, 2, 3, 4)
157 | else:
158 | raise ValueError(f"unsupported image size: {image_size}")
159 | else:
160 | channel_mult = tuple(int(ch_mult) for ch_mult in channel_mult.split(","))
161 |
162 | attention_ds = []
163 | for res in attention_resolutions.split(","):
164 | attention_ds.append(image_size // int(res))
165 |
166 | return UNetModel(
167 | image_size=image_size,
168 | in_channels=3,
169 | model_channels=num_channels,
170 | out_channels=(3 if not learn_sigma else 6),
171 | num_res_blocks=num_res_blocks,
172 | attention_resolutions=tuple(attention_ds),
173 | dropout=dropout,
174 | channel_mult=channel_mult,
175 | num_classes=(NUM_CLASSES if class_cond else None),
176 | use_checkpoint=use_checkpoint,
177 | use_fp16=use_fp16,
178 | num_heads=num_heads,
179 | num_head_channels=num_head_channels,
180 | num_heads_upsample=num_heads_upsample,
181 | use_scale_shift_norm=use_scale_shift_norm,
182 | resblock_updown=resblock_updown,
183 | use_new_attention_order=use_new_attention_order,
184 | )
185 |
186 |
187 | def create_classifier_and_diffusion(
188 | image_size,
189 | classifier_use_fp16,
190 | classifier_width,
191 | classifier_depth,
192 | classifier_attention_resolutions,
193 | classifier_use_scale_shift_norm,
194 | classifier_resblock_updown,
195 | classifier_pool,
196 | learn_sigma,
197 | diffusion_steps,
198 | noise_schedule,
199 | timestep_respacing,
200 | use_kl,
201 | predict_xstart,
202 | rescale_timesteps,
203 | rescale_learned_sigmas,
204 | ):
205 | classifier = create_classifier(
206 | image_size,
207 | classifier_use_fp16,
208 | classifier_width,
209 | classifier_depth,
210 | classifier_attention_resolutions,
211 | classifier_use_scale_shift_norm,
212 | classifier_resblock_updown,
213 | classifier_pool,
214 | )
215 | diffusion = create_gaussian_diffusion(
216 | steps=diffusion_steps,
217 | learn_sigma=learn_sigma,
218 | noise_schedule=noise_schedule,
219 | use_kl=use_kl,
220 | predict_xstart=predict_xstart,
221 | rescale_timesteps=rescale_timesteps,
222 | rescale_learned_sigmas=rescale_learned_sigmas,
223 | timestep_respacing=timestep_respacing,
224 | )
225 | return classifier, diffusion
226 |
227 |
228 | def create_classifier(
229 | image_size,
230 | classifier_use_fp16,
231 | classifier_width,
232 | classifier_depth,
233 | classifier_attention_resolutions,
234 | classifier_use_scale_shift_norm,
235 | classifier_resblock_updown,
236 | classifier_pool,
237 | ):
238 | if image_size == 512:
239 | channel_mult = (0.5, 1, 1, 2, 2, 4, 4)
240 | elif image_size == 256:
241 | channel_mult = (1, 1, 2, 2, 4, 4)
242 | elif image_size == 128:
243 | channel_mult = (1, 1, 2, 3, 4)
244 | elif image_size == 64:
245 | channel_mult = (1, 2, 3, 4)
246 | else:
247 | raise ValueError(f"unsupported image size: {image_size}")
248 |
249 | attention_ds = []
250 | for res in classifier_attention_resolutions.split(","):
251 | attention_ds.append(image_size // int(res))
252 |
253 | return EncoderUNetModel(
254 | image_size=image_size,
255 | in_channels=3,
256 | model_channels=classifier_width,
257 | out_channels=1000,
258 | num_res_blocks=classifier_depth,
259 | attention_resolutions=tuple(attention_ds),
260 | channel_mult=channel_mult,
261 | use_fp16=classifier_use_fp16,
262 | num_head_channels=64,
263 | use_scale_shift_norm=classifier_use_scale_shift_norm,
264 | resblock_updown=classifier_resblock_updown,
265 | pool=classifier_pool,
266 | )
267 |
268 |
269 | def sr_model_and_diffusion_defaults():
270 | res = model_and_diffusion_defaults()
271 | res["large_size"] = 256
272 | res["small_size"] = 64
273 | arg_names = inspect.getfullargspec(sr_create_model_and_diffusion)[0]
274 | for k in res.copy().keys():
275 | if k not in arg_names:
276 | del res[k]
277 | return res
278 |
279 |
280 | def sr_create_model_and_diffusion(
281 | large_size,
282 | small_size,
283 | class_cond,
284 | learn_sigma,
285 | num_channels,
286 | num_res_blocks,
287 | num_heads,
288 | num_head_channels,
289 | num_heads_upsample,
290 | attention_resolutions,
291 | dropout,
292 | diffusion_steps,
293 | noise_schedule,
294 | timestep_respacing,
295 | use_kl,
296 | predict_xstart,
297 | rescale_timesteps,
298 | rescale_learned_sigmas,
299 | use_checkpoint,
300 | use_scale_shift_norm,
301 | resblock_updown,
302 | use_fp16,
303 | ):
304 | model = sr_create_model(
305 | large_size,
306 | small_size,
307 | num_channels,
308 | num_res_blocks,
309 | learn_sigma=learn_sigma,
310 | class_cond=class_cond,
311 | use_checkpoint=use_checkpoint,
312 | attention_resolutions=attention_resolutions,
313 | num_heads=num_heads,
314 | num_head_channels=num_head_channels,
315 | num_heads_upsample=num_heads_upsample,
316 | use_scale_shift_norm=use_scale_shift_norm,
317 | dropout=dropout,
318 | resblock_updown=resblock_updown,
319 | use_fp16=use_fp16,
320 | )
321 | diffusion = create_gaussian_diffusion(
322 | steps=diffusion_steps,
323 | learn_sigma=learn_sigma,
324 | noise_schedule=noise_schedule,
325 | use_kl=use_kl,
326 | predict_xstart=predict_xstart,
327 | rescale_timesteps=rescale_timesteps,
328 | rescale_learned_sigmas=rescale_learned_sigmas,
329 | timestep_respacing=timestep_respacing,
330 | )
331 | return model, diffusion
332 |
333 |
334 | def sr_create_model(
335 | large_size,
336 | small_size,
337 | num_channels,
338 | num_res_blocks,
339 | learn_sigma,
340 | class_cond,
341 | use_checkpoint,
342 | attention_resolutions,
343 | num_heads,
344 | num_head_channels,
345 | num_heads_upsample,
346 | use_scale_shift_norm,
347 | dropout,
348 | resblock_updown,
349 | use_fp16,
350 | ):
351 | _ = small_size # hack to prevent unused variable
352 |
353 | if large_size == 512:
354 | channel_mult = (1, 1, 2, 2, 4, 4)
355 | elif large_size == 256:
356 | channel_mult = (1, 1, 2, 2, 4, 4)
357 | elif large_size == 64:
358 | channel_mult = (1, 2, 3, 4)
359 | else:
360 | channel_mult = (1, 1, 2, 3, 4)
361 | # raise ValueError(f"unsupported large size: {large_size}")
362 |
363 | attention_ds = []
364 | for res in attention_resolutions.split(","):
365 | attention_ds.append(large_size // int(res))
366 |
367 | # return SuperResModel(
368 | # image_size=large_size,
369 | # in_channels=3,
370 | # model_channels=num_channels,
371 | # out_channels=(3 if not learn_sigma else 6),
372 | # num_res_blocks=num_res_blocks,
373 | # attention_resolutions=tuple(attention_ds),
374 | # dropout=dropout,
375 | # channel_mult=channel_mult,
376 | # num_classes=(NUM_CLASSES if class_cond else None),
377 | # use_checkpoint=use_checkpoint,
378 | # num_heads=num_heads,
379 | # num_head_channels=num_head_channels,
380 | # num_heads_upsample=num_heads_upsample,
381 | # use_scale_shift_norm=use_scale_shift_norm,
382 | # resblock_updown=resblock_updown,
383 | # use_fp16=use_fp16,
384 | # )
385 |
386 |
387 | # 3D input PET CT, input channel is 1.5*2, output channel is 1 or 2
388 | # return SuperResModel(
389 | # image_size=large_size,
390 | # in_channels=1.5,
391 | # model_channels=num_channels,
392 | # out_channels=(1 if not learn_sigma else 2),
393 | # num_res_blocks=num_res_blocks,
394 | # attention_resolutions=tuple(attention_ds),
395 | # dropout=dropout,
396 | # channel_mult=channel_mult,
397 | # dims=3, # 3D conv
398 | # num_classes=(NUM_CLASSES if class_cond else None),
399 | # use_checkpoint=use_checkpoint,
400 | # num_heads=num_heads,
401 | # num_head_channels=num_head_channels,
402 | # num_heads_upsample=num_heads_upsample,
403 | # use_scale_shift_norm=use_scale_shift_norm,
404 | # resblock_updown=resblock_updown,
405 | # use_fp16=use_fp16,
406 | # )
407 |
408 |
409 | # # 3D input PET CT, input channel is 1.5*2, output channel is 1 or 2
410 | # return SuperResModel_noatt(
411 | # image_size=large_size,
412 | # in_channels=1.5,
413 | # model_channels=num_channels,
414 | # out_channels=(1 if not learn_sigma else 2),
415 | # num_res_blocks=num_res_blocks,
416 | # attention_resolutions=tuple(attention_ds),
417 | # dropout=dropout,
418 | # channel_mult=channel_mult,
419 | # dims=3, # 3D conv
420 | # num_classes=(NUM_CLASSES if class_cond else None),
421 | # use_checkpoint=use_checkpoint,
422 | # num_heads=num_heads,
423 | # num_head_channels=num_head_channels,
424 | # num_heads_upsample=num_heads_upsample,
425 | # use_scale_shift_norm=use_scale_shift_norm,
426 | # resblock_updown=resblock_updown,
427 | # use_fp16=use_fp16,
428 | # )
429 |
430 |
431 | # 3D input PET or CT only, input channel is 1*2, output channel is 1 or 2
432 | return SuperResModel_noatt(
433 | image_size=large_size,
434 | in_channels=1,
435 | model_channels=num_channels,
436 | out_channels=(1 if not learn_sigma else 2),
437 | num_res_blocks=num_res_blocks,
438 | attention_resolutions=tuple(attention_ds),
439 | dropout=dropout,
440 | channel_mult=channel_mult,
441 | dims=3, # 3D conv
442 | num_classes=(NUM_CLASSES if class_cond else None),
443 | use_checkpoint=use_checkpoint,
444 | num_heads=num_heads,
445 | num_head_channels=num_head_channels,
446 | num_heads_upsample=num_heads_upsample,
447 | use_scale_shift_norm=use_scale_shift_norm,
448 | resblock_updown=resblock_updown,
449 | use_fp16=use_fp16,
450 | )
451 |
452 |
453 | # 3D input PET CT, input channel is 1.5*2, output channel is 1 or 2
454 | # return SegModelv2_3d_noatt(
455 | # image_size=large_size,
456 | # in_channels=1.5,
457 | # model_channels=num_channels,
458 | # out_channels=(1 if not learn_sigma else 2),
459 | # num_res_blocks=num_res_blocks,
460 | # attention_resolutions=tuple(attention_ds),
461 | # dropout=dropout,
462 | # channel_mult=channel_mult,
463 | # dims=3, # 3D conv
464 | # num_classes=(NUM_CLASSES if class_cond else None),
465 | # use_checkpoint=use_checkpoint,
466 | # num_heads=num_heads,
467 | # num_head_channels=num_head_channels,
468 | # num_heads_upsample=num_heads_upsample,
469 | # use_scale_shift_norm=use_scale_shift_norm,
470 | # resblock_updown=resblock_updown,
471 | # use_fp16=use_fp16,
472 | # )
473 |
474 |
475 | # 3D input PET CT, input channel is 1.5*2, output channel is 1 or 2
476 | # return SegModel_3d_noatt_midcat(
477 | # image_size=large_size,
478 | # in_channels=1.5,
479 | # model_channels=num_channels,
480 | # out_channels=(1 if not learn_sigma else 2),
481 | # num_res_blocks=num_res_blocks,
482 | # attention_resolutions=tuple(attention_ds),
483 | # dropout=dropout,
484 | # channel_mult=channel_mult,
485 | # dims=3, # 3D conv
486 | # num_classes=(NUM_CLASSES if class_cond else None),
487 | # use_checkpoint=use_checkpoint,
488 | # num_heads=num_heads,
489 | # num_head_channels=num_head_channels,
490 | # num_heads_upsample=num_heads_upsample,
491 | # use_scale_shift_norm=use_scale_shift_norm,
492 | # resblock_updown=resblock_updown,
493 | # use_fp16=use_fp16,
494 | # )
495 |
496 |
497 | # input channel 4.5*2 (pet and ct and noise)
498 | # return SuperResModel(
499 | # image_size=large_size,
500 | # in_channels=4.5,
501 | # model_channels=num_channels,
502 | # out_channels=(3 if not learn_sigma else 6),
503 | # num_res_blocks=num_res_blocks,
504 | # attention_resolutions=tuple(attention_ds),
505 | # dropout=dropout,
506 | # channel_mult=channel_mult,
507 | # num_classes=(NUM_CLASSES if class_cond else None),
508 | # use_checkpoint=use_checkpoint,
509 | # num_heads=num_heads,
510 | # num_head_channels=num_head_channels,
511 | # num_heads_upsample=num_heads_upsample,
512 | # use_scale_shift_norm=use_scale_shift_norm,
513 | # resblock_updown=resblock_updown,
514 | # use_fp16=use_fp16,
515 | # )
516 |
517 | # return SegModelv2(
518 | # image_size=large_size,
519 | # in_channels=3,
520 | # model_channels=num_channels,
521 | # out_channels=(3 if not learn_sigma else 6),
522 | # num_res_blocks=num_res_blocks,
523 | # attention_resolutions=tuple(attention_ds),
524 | # dropout=dropout,
525 | # channel_mult=channel_mult,
526 | # num_classes=(NUM_CLASSES if class_cond else None),
527 | # use_checkpoint=use_checkpoint,
528 | # num_heads=num_heads,
529 | # num_head_channels=num_head_channels,
530 | # num_heads_upsample=num_heads_upsample,
531 | # use_scale_shift_norm=use_scale_shift_norm,
532 | # resblock_updown=resblock_updown,
533 | # use_fp16=use_fp16,
534 | # )
535 |
536 | # input channel 6 (pet and ct)
537 | # return SegModelv2_6c(
538 | # image_size=large_size,
539 | # in_channels=6,
540 | # model_channels=num_channels,
541 | # out_channels=(3 if not learn_sigma else 6),
542 | # num_res_blocks=num_res_blocks,
543 | # attention_resolutions=tuple(attention_ds),
544 | # dropout=dropout,
545 | # channel_mult=channel_mult,
546 | # num_classes=(NUM_CLASSES if class_cond else None),
547 | # use_checkpoint=use_checkpoint,
548 | # num_heads=num_heads,
549 | # num_head_channels=num_head_channels,
550 | # num_heads_upsample=num_heads_upsample,
551 | # use_scale_shift_norm=use_scale_shift_norm,
552 | # resblock_updown=resblock_updown,
553 | # use_fp16=use_fp16,
554 | # )
555 |
556 | # input channel 6 (pet and ct)
557 | # use concat to fuse and a 1*1 conv to reduce channel
558 | # return SegModelv3_6c(
559 | # image_size=large_size,
560 | # in_channels=6,
561 | # model_channels=num_channels,
562 | # out_channels=(3 if not learn_sigma else 6),
563 | # num_res_blocks=num_res_blocks,
564 | # attention_resolutions=tuple(attention_ds),
565 | # dropout=dropout,
566 | # channel_mult=channel_mult,
567 | # num_classes=(NUM_CLASSES if class_cond else None),
568 | # use_checkpoint=use_checkpoint,
569 | # num_heads=num_heads,
570 | # num_head_channels=num_head_channels,
571 | # num_heads_upsample=num_heads_upsample,
572 | # use_scale_shift_norm=use_scale_shift_norm,
573 | # resblock_updown=resblock_updown,
574 | # use_fp16=use_fp16,
575 | # )
576 |
577 |
578 | def create_gaussian_diffusion(
579 | *,
580 | steps=1000,
581 | learn_sigma=False,
582 | sigma_small=False,
583 | noise_schedule="linear",
584 | use_kl=False,
585 | predict_xstart=False,
586 | rescale_timesteps=False,
587 | rescale_learned_sigmas=False,
588 | timestep_respacing="",
589 | ):
590 | betas = gd.get_named_beta_schedule(noise_schedule, steps)
591 | if use_kl:
592 | loss_type = gd.LossType.RESCALED_KL
593 | elif rescale_learned_sigmas:
594 | loss_type = gd.LossType.RESCALED_MSE
595 | else:
596 | loss_type = gd.LossType.MSE
597 | if not timestep_respacing:
598 | timestep_respacing = [steps]
599 | return SpacedDiffusion(
600 | use_timesteps=space_timesteps(steps, timestep_respacing),
601 | betas=betas,
602 | model_mean_type=(
603 | gd.ModelMeanType.EPSILON if not predict_xstart else gd.ModelMeanType.START_X
604 | ),
605 | model_var_type=(
606 | (
607 | gd.ModelVarType.FIXED_LARGE
608 | if not sigma_small
609 | else gd.ModelVarType.FIXED_SMALL
610 | )
611 | if not learn_sigma
612 | else gd.ModelVarType.LEARNED_RANGE
613 | ),
614 | loss_type=loss_type,
615 | rescale_timesteps=rescale_timesteps,
616 | )
617 |
618 |
619 | def add_dict_to_argparser(parser, default_dict):
620 | for k, v in default_dict.items():
621 | v_type = type(v)
622 | if v is None:
623 | v_type = str
624 | elif isinstance(v, bool):
625 | v_type = str2bool
626 | parser.add_argument(f"--{k}", default=v, type=v_type)
627 |
628 |
629 | def args_to_dict(args, keys):
630 | return {k: getattr(args, k) for k in keys}
631 |
632 |
633 | def str2bool(v):
634 | """
635 | https://stackoverflow.com/questions/15008758/parsing-boolean-values-with-argparse
636 | """
637 | if isinstance(v, bool):
638 | return v
639 | if v.lower() in ("yes", "true", "t", "y", "1"):
640 | return True
641 | elif v.lower() in ("no", "false", "f", "n", "0"):
642 | return False
643 | else:
644 | raise argparse.ArgumentTypeError("boolean value expected")
645 |
--------------------------------------------------------------------------------
/guided_diffusion/train_util.py:
--------------------------------------------------------------------------------
1 | import copy
2 | import functools
3 | import os
4 |
5 | import blobfile as bf
6 | import torch as th
7 | import torch.distributed as dist
8 | from torch.nn.parallel.distributed import DistributedDataParallel as DDP
9 | from torch.optim import AdamW
10 |
11 | from . import dist_util, logger
12 | from .fp16_util import MixedPrecisionTrainer
13 | from .nn import update_ema
14 | from .resample import LossAwareSampler, UniformSampler
15 |
16 | # For ImageNet experiments, this was a good default value.
17 | # We found that the lg_loss_scale quickly climbed to
18 | # 20-21 within the first ~1K steps of training.
19 | INITIAL_LOG_LOSS_SCALE = 20.0
20 |
21 |
22 | class TrainLoop:
23 | def __init__(
24 | self,
25 | *,
26 | model,
27 | diffusion,
28 | data,
29 | batch_size,
30 | microbatch,
31 | lr,
32 | ema_rate,
33 | log_interval,
34 | save_interval,
35 | resume_checkpoint,
36 | use_fp16=False,
37 | fp16_scale_growth=1e-3,
38 | schedule_sampler=None,
39 | weight_decay=0.0,
40 | lr_anneal_steps=0,
41 | ):
42 | self.model = model
43 | self.diffusion = diffusion
44 | self.data = data
45 | self.batch_size = batch_size
46 | self.microbatch = microbatch if microbatch > 0 else batch_size
47 | self.lr = lr
48 | self.ema_rate = (
49 | [ema_rate]
50 | if isinstance(ema_rate, float)
51 | else [float(x) for x in ema_rate.split(",")]
52 | )
53 | self.log_interval = log_interval
54 | self.save_interval = save_interval
55 | self.resume_checkpoint = resume_checkpoint
56 | self.use_fp16 = use_fp16
57 | self.fp16_scale_growth = fp16_scale_growth
58 | self.schedule_sampler = schedule_sampler or UniformSampler(diffusion)
59 | self.weight_decay = weight_decay
60 | self.lr_anneal_steps = lr_anneal_steps
61 |
62 | self.step = 0
63 | self.resume_step = 0
64 | self.global_batch = self.batch_size * dist.get_world_size()
65 |
66 | self.sync_cuda = th.cuda.is_available()
67 |
68 | self._load_and_sync_parameters()
69 | self.mp_trainer = MixedPrecisionTrainer(
70 | model=self.model,
71 | use_fp16=self.use_fp16,
72 | fp16_scale_growth=fp16_scale_growth,
73 | )
74 |
75 | self.opt = AdamW(
76 | self.mp_trainer.master_params, lr=self.lr, weight_decay=self.weight_decay
77 | )
78 | if self.resume_step:
79 | self._load_optimizer_state()
80 | # Model was resumed, either due to a restart or a checkpoint
81 | # being specified at the command line.
82 | self.ema_params = [
83 | self._load_ema_parameters(rate) for rate in self.ema_rate
84 | ]
85 | else:
86 | self.ema_params = [
87 | copy.deepcopy(self.mp_trainer.master_params)
88 | for _ in range(len(self.ema_rate))
89 | ]
90 |
91 | if th.cuda.is_available():
92 | self.use_ddp = True
93 | self.ddp_model = DDP(
94 | self.model,
95 | device_ids=[dist_util.dev()],
96 | output_device=dist_util.dev(),
97 | broadcast_buffers=False,
98 | bucket_cap_mb=128,
99 | find_unused_parameters=False,
100 | )
101 | else:
102 | if dist.get_world_size() > 1:
103 | logger.warn(
104 | "Distributed training requires CUDA. "
105 | "Gradients will not be synchronized properly!"
106 | )
107 | self.use_ddp = False
108 | self.ddp_model = self.model
109 |
110 | def _load_and_sync_parameters(self):
111 | resume_checkpoint = find_resume_checkpoint() or self.resume_checkpoint
112 |
113 | if resume_checkpoint:
114 | self.resume_step = parse_resume_step_from_filename(resume_checkpoint)
115 | #if dist.get_rank() == 0:
116 | logger.log(f"loading model from checkpoint: {resume_checkpoint}...")
117 | self.model.load_state_dict(
118 | dist_util.load_state_dict(
119 | resume_checkpoint, map_location=dist_util.dev()
120 | )
121 | )
122 |
123 | dist_util.sync_params(self.model.parameters())
124 |
125 | def _load_ema_parameters(self, rate):
126 | ema_params = copy.deepcopy(self.mp_trainer.master_params)
127 |
128 | main_checkpoint = find_resume_checkpoint() or self.resume_checkpoint
129 | ema_checkpoint = find_ema_checkpoint(main_checkpoint, self.resume_step, rate)
130 | if ema_checkpoint:
131 | #if dist.get_rank() == 0:
132 | logger.log(f"loading EMA from checkpoint: {ema_checkpoint}...")
133 | state_dict = dist_util.load_state_dict(
134 | ema_checkpoint, map_location=dist_util.dev()
135 | )
136 | ema_params = self.mp_trainer.state_dict_to_master_params(state_dict)
137 |
138 | dist_util.sync_params(ema_params)
139 | return ema_params
140 |
141 | def _load_optimizer_state(self):
142 | main_checkpoint = find_resume_checkpoint() or self.resume_checkpoint
143 | opt_checkpoint = bf.join(
144 | bf.dirname(main_checkpoint), f"opt{self.resume_step:06}.pt"
145 | )
146 | if bf.exists(opt_checkpoint):
147 | logger.log(f"loading optimizer state from checkpoint: {opt_checkpoint}")
148 | state_dict = dist_util.load_state_dict(
149 | opt_checkpoint, map_location=dist_util.dev()
150 | )
151 | self.opt.load_state_dict(state_dict)
152 |
153 | def run_loop(self):
154 | while (
155 | not self.lr_anneal_steps
156 | or self.step + self.resume_step < self.lr_anneal_steps
157 | ):
158 | batch, cond = next(self.data)
159 | # print(batch.shape)
160 | # print(cond["low_res"].shape)
161 | self.run_step(batch, cond)
162 | if self.step % self.log_interval == 0:
163 | logger.dumpkvs()
164 | if self.step % self.save_interval == 0:
165 | self.save()
166 | # Run for a finite amount of time in integration tests.
167 | if os.environ.get("DIFFUSION_TRAINING_TEST", "") and self.step > 0:
168 | return
169 | self.step += 1
170 | # Save the last checkpoint if it wasn't already saved.
171 | if (self.step - 1) % self.save_interval != 0:
172 | self.save()
173 |
174 | def run_step(self, batch, cond):
175 | self.forward_backward(batch, cond)
176 | took_step = self.mp_trainer.optimize(self.opt)
177 | if took_step:
178 | self._update_ema()
179 | self._anneal_lr()
180 | self.log_step()
181 |
182 | def forward_backward(self, batch, cond):
183 | self.mp_trainer.zero_grad()
184 | for i in range(0, batch.shape[0], self.microbatch):
185 | micro = batch[i : i + self.microbatch].to(dist_util.dev())
186 | micro_cond = {
187 | k: v[i : i + self.microbatch].to(dist_util.dev())
188 | for k, v in cond.items()
189 | }
190 | last_batch = (i + self.microbatch) >= batch.shape[0]
191 | t, weights = self.schedule_sampler.sample(micro.shape[0], dist_util.dev())
192 |
193 | compute_losses = functools.partial(
194 | self.diffusion.training_losses,
195 | self.ddp_model,
196 | micro,
197 | t,
198 | model_kwargs=micro_cond,
199 | )
200 |
201 | if last_batch or not self.use_ddp:
202 | losses = compute_losses()
203 | else:
204 | with self.ddp_model.no_sync():
205 | losses = compute_losses()
206 |
207 | if isinstance(self.schedule_sampler, LossAwareSampler):
208 | self.schedule_sampler.update_with_local_losses(
209 | t, losses["loss"].detach()
210 | )
211 |
212 | loss = (losses["loss"] * weights).mean()
213 | log_loss_dict(
214 | self.diffusion, t, {k: v * weights for k, v in losses.items()}
215 | )
216 | self.mp_trainer.backward(loss)
217 |
218 | def _update_ema(self):
219 | for rate, params in zip(self.ema_rate, self.ema_params):
220 | update_ema(params, self.mp_trainer.master_params, rate=rate)
221 |
222 | def _anneal_lr(self):
223 | if not self.lr_anneal_steps:
224 | return
225 | frac_done = (self.step + self.resume_step) / self.lr_anneal_steps
226 | lr = self.lr * (1 - frac_done)
227 | for param_group in self.opt.param_groups:
228 | param_group["lr"] = lr
229 |
230 | def log_step(self):
231 | logger.logkv("step", self.step + self.resume_step)
232 | logger.logkv("samples", (self.step + self.resume_step + 1) * self.global_batch)
233 |
234 | def save(self):
235 | def save_checkpoint(rate, params):
236 | state_dict = self.mp_trainer.master_params_to_state_dict(params)
237 | if dist.get_rank() == 0:
238 | logger.log(f"saving model {rate}...")
239 | if not rate:
240 | filename = f"model{(self.step+self.resume_step):06d}.pt"
241 | else:
242 | filename = f"ema_{rate}_{(self.step+self.resume_step):06d}.pt"
243 | with bf.BlobFile(bf.join(get_blob_logdir(), filename), "wb") as f:
244 | th.save(state_dict, f)
245 |
246 | save_checkpoint(0, self.mp_trainer.master_params)
247 | for rate, params in zip(self.ema_rate, self.ema_params):
248 | save_checkpoint(rate, params)
249 |
250 | if dist.get_rank() == 0:
251 | with bf.BlobFile(
252 | bf.join(get_blob_logdir(), f"opt{(self.step+self.resume_step):06d}.pt"),
253 | "wb",
254 | ) as f:
255 | th.save(self.opt.state_dict(), f)
256 |
257 | dist.barrier()
258 |
259 |
260 | def parse_resume_step_from_filename(filename):
261 | """
262 | Parse filenames of the form path/to/modelNNNNNN.pt, where NNNNNN is the
263 | checkpoint's number of steps.
264 | """
265 | split = filename.split("model")
266 | if len(split) < 2:
267 | return 0
268 | split1 = split[-1].split(".")[0]
269 | try:
270 | return int(split1)
271 | except ValueError:
272 | return 0
273 |
274 |
275 | def get_blob_logdir():
276 | # You can change this to be a separate path to save checkpoints to
277 | # a blobstore or some external drive.
278 | return logger.get_dir()
279 | # return '/autofs/vast/kggp/ydong/seg_ddpm/save_models/3d/pet_ct/ddpm96_32_noatt2_-1to1_2'
280 |
281 |
282 | def find_resume_checkpoint():
283 | # On your infrastructure, you may want to override this to automatically
284 | # discover the latest checkpoint on your blob storage, etc.
285 | return None
286 |
287 |
288 | def find_ema_checkpoint(main_checkpoint, step, rate):
289 | if main_checkpoint is None:
290 | return None
291 | filename = f"ema_{rate}_{(step):06d}.pt"
292 | path = bf.join(bf.dirname(main_checkpoint), filename)
293 | if bf.exists(path):
294 | return path
295 | return None
296 |
297 |
298 | def log_loss_dict(diffusion, ts, losses):
299 | for key, values in losses.items():
300 | logger.logkv_mean(key, values.mean().item())
301 | # Log the quantiles (four quartiles, in particular).
302 | for sub_t, sub_loss in zip(ts.cpu().numpy(), values.detach().cpu().numpy()):
303 | quartile = int(4 * sub_t / diffusion.num_timesteps)
304 | logger.logkv_mean(f"{key}_q{quartile}", sub_loss)
305 |
--------------------------------------------------------------------------------
/scripts/test.py:
--------------------------------------------------------------------------------
1 | import argparse
2 | import os
3 | import sys
4 | sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
5 |
6 | import blobfile as bf
7 | import numpy as np
8 | import torch as th
9 | import torch.distributed as dist
10 | import datetime
11 |
12 | from guided_diffusion import dist_util, logger
13 | from guided_diffusion.script_util import (
14 | sr_model_and_diffusion_defaults,
15 | sr_create_model_and_diffusion,
16 | args_to_dict,
17 | add_dict_to_argparser,
18 | )
19 |
20 |
21 | def main():
22 | args = create_argparser().parse_args()
23 |
24 | dist_util.setup_dist()
25 | logger.configure(dir=args.save_dir)
26 |
27 | logger.log("creating model...")
28 | model, diffusion = sr_create_model_and_diffusion(
29 | **args_to_dict(args, sr_model_and_diffusion_defaults().keys())
30 | )
31 | model.load_state_dict(
32 | dist_util.load_state_dict(args.model_path, map_location="cpu")
33 | )
34 | model.to(dist_util.dev())
35 | if args.use_fp16:
36 | model.convert_to_fp16()
37 | model.eval()
38 |
39 | logger.log("loading data...")
40 | data = load_data_for_worker(args.base_samples, args.batch_size, args.class_cond)
41 | logger.log("creating samples...")
42 | all_images = []
43 |
44 | device = th.device("cuda" if th.cuda.is_available() else "cpu")
45 |
46 |
47 | while len(all_images) * args.batch_size < args.num_samples:
48 | model_kwargs = next(data)
49 | model_kwargs = {k: v.to(dist_util.dev()) for k, v in model_kwargs.items()}
50 |
51 | shape = (args.batch_size, 1, model_kwargs['low_res'].shape[2], model_kwargs['low_res'].shape[3], model_kwargs['low_res'].shape[4])
52 |
53 | if device == "cuda":
54 | th.cuda.manual_seed_all(10)
55 | else:
56 | th.manual_seed(10)
57 | noise = th.randn(*shape, device=device)
58 |
59 | sample = diffusion.p_sample_loop(
60 | model,
61 | shape,
62 | noise,
63 | clip_denoised=args.clip_denoised,
64 | model_kwargs=model_kwargs,
65 | )
66 |
67 | sample = sample.permute(0, 1, 3, 4, 2)
68 | sample = sample.contiguous()
69 |
70 | all_samples = [th.zeros_like(sample) for _ in range(dist.get_world_size())]
71 | dist.all_gather(all_samples, sample)
72 | for sample in all_samples:
73 | all_images.append(sample.cpu().numpy())
74 | logger.log(f"created {len(all_images) * args.batch_size} samples\n")
75 |
76 | arr = np.concatenate(all_images, axis=0)
77 | arr = arr[: args.num_samples]
78 | arr = arr.reshape((arr.shape[0],arr.shape[2],arr.shape[3],arr.shape[4]))
79 |
80 | arr_result = np.zeros((192,288,576))
81 | index = 0
82 | for i in range(6):
83 | arr_result[:, :, i*96:(i+1)*96] = arr[index,:,:,:]
84 | index += 1
85 |
86 | if dist.get_rank() == 0:
87 | shape_str = "x".join([str(x) for x in arr_result.shape])
88 | out_path = os.path.join(logger.get_dir(), f"samples_{shape_str}_{datetime.datetime.now().strftime('%H%M%S%f')}.npz")
89 | logger.log(f"saving to {out_path}")
90 | np.savez(out_path, arr_result)
91 |
92 | dist.barrier()
93 | logger.log("sampling complete")
94 |
95 |
96 | def load_data_for_worker(base_samples, batch_size, class_cond):
97 | with bf.BlobFile(base_samples, "rb") as f:
98 | obj = np.load(f)
99 | low_pet = obj["arr_0"][0]
100 | image_arr = np.zeros((6,192,288,96))
101 | index = 0
102 | for i in (0, 86, 172, 258, 344, 424):
103 | image_arr[index,:,:,:] = low_pet[:, :, i:i+96]
104 | index += 1
105 |
106 | image_arr[image_arr>4] = 4
107 | image_arr = image_arr/4
108 |
109 | rank = dist.get_rank()
110 | logger.log('rank:{%d}' % (rank))
111 |
112 | num_ranks = dist.get_world_size()
113 | logger.log('num_ranks:{%d}' % (num_ranks))
114 |
115 | buffer = []
116 | label_buffer = []
117 | while True:
118 | for i in range(rank, len(image_arr), num_ranks):
119 | buffer.append(image_arr[i])
120 | logger.log('rank:{%d}, i:{%d}, buffer_len:{%d}' % (rank, i, len(buffer)))
121 |
122 | if len(buffer) == batch_size:
123 | batch = th.from_numpy(np.stack(buffer)).float()
124 | batch = batch.unsqueeze(0)
125 | batch = batch.permute(0, 1, 4, 2, 3)
126 | logger.log('batch_shape:{%d,%d,%d,%d,%d}' % (batch.shape[0], batch.shape[1], batch.shape[2], batch.shape[3], batch.shape[4]))
127 | res = dict(low_res=batch)
128 | yield res
129 | buffer, label_buffer = [], []
130 |
131 |
132 | def create_argparser():
133 | defaults = dict(
134 | save_dir='',
135 | clip_denoised=True,
136 | num_samples=10000,
137 | batch_size=16,
138 | use_ddim=False,
139 | base_samples="",
140 | model_path="",
141 | )
142 | defaults.update(sr_model_and_diffusion_defaults())
143 | parser = argparse.ArgumentParser()
144 | add_dict_to_argparser(parser, defaults)
145 | return parser
146 |
147 |
148 | if __name__ == "__main__":
149 | main()
150 |
--------------------------------------------------------------------------------
/scripts/train.py:
--------------------------------------------------------------------------------
1 | import argparse
2 | import os
3 | import torch.nn.functional as F
4 | import sys
5 | sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
6 |
7 | from guided_diffusion import dist_util, logger
8 | from guided_diffusion.image_datasets import load_data
9 | from guided_diffusion.resample import create_named_schedule_sampler
10 | from guided_diffusion.script_util import (
11 | sr_model_and_diffusion_defaults,
12 | sr_create_model_and_diffusion,
13 | args_to_dict,
14 | add_dict_to_argparser,
15 | )
16 | from guided_diffusion.train_util import TrainLoop
17 |
18 |
19 | def main():
20 | args = create_argparser().parse_args()
21 |
22 | dist_util.setup_dist()
23 | # logger.configure(dir='/autofs/vast/kggp/ydong/seg_ddpm/logs/3d/pet_ct/ddpm96_32_noatt2_-1to1_2')
24 | logger.configure(args.result_folder)
25 |
26 | logger.log("creating model...")
27 | model, diffusion = sr_create_model_and_diffusion(
28 | **args_to_dict(args, sr_model_and_diffusion_defaults().keys())
29 | )
30 | model.to(dist_util.dev())
31 | schedule_sampler = create_named_schedule_sampler(args.schedule_sampler, diffusion)
32 | for name, param in model.named_parameters():
33 | logger.log('%s:%s' % (name, str(param.shape)))
34 | logger.log('parameters:{%d}' % (sum(param.numel() for param in model.parameters())))
35 | logger.log('attention_resolutions:{%s}' % args.attention_resolutions)
36 | logger.log('num_channels:{%s}' % str(args.num_channels))
37 | logger.log('num_res_blocks:{%s}' % str(args.num_res_blocks))
38 | logger.log('num_head_channels:{%s}' % str(args.num_head_channels))
39 | logger.log('channel_mult:{%s}' % str(model.channel_mult))
40 |
41 | logger.log("creating data loader...")
42 | data = load_superres_data(
43 | args.data_dir,
44 | args.batch_size,
45 | large_size=args.large_size,
46 | small_size=args.small_size,
47 | class_cond=args.class_cond,
48 | )
49 |
50 | logger.log("training...")
51 |
52 | TrainLoop(
53 | model=model,
54 | diffusion=diffusion,
55 | data=data,
56 | batch_size=args.batch_size,
57 | microbatch=args.microbatch,
58 | lr=args.lr,
59 | ema_rate=args.ema_rate,
60 | log_interval=args.log_interval,
61 | save_interval=args.save_interval,
62 | resume_checkpoint=args.resume_checkpoint,
63 | use_fp16=args.use_fp16,
64 | fp16_scale_growth=args.fp16_scale_growth,
65 | schedule_sampler=schedule_sampler,
66 | weight_decay=args.weight_decay,
67 | lr_anneal_steps=args.lr_anneal_steps,
68 | ).run_loop()
69 |
70 |
71 | def load_superres_data(data_dir, batch_size, large_size, small_size, class_cond=False):
72 | data = load_data(
73 | data_dir=data_dir,
74 | batch_size=batch_size,
75 | image_size=large_size,
76 | class_cond=class_cond,
77 | )
78 | for pet_batch, label_batch, model_kwargs in data:
79 | model_kwargs["low_res"] = pet_batch.clone()
80 | yield label_batch, model_kwargs
81 |
82 | '''
83 | def load_superres_data(data_dir, batch_size, large_size, small_size, class_cond=False):
84 | data = load_data(
85 | data_dir=data_dir,
86 | batch_size=batch_size,
87 | image_size=large_size,
88 | class_cond=class_cond,
89 | )
90 | for large_batch, model_kwargs in data:
91 | model_kwargs["low_res"] = F.interpolate(large_batch, small_size, mode="area")
92 | yield large_batch, model_kwargs
93 | '''
94 |
95 | def create_argparser():
96 | defaults = dict(
97 | data_dir="",
98 | schedule_sampler="uniform",
99 | lr=1e-4,
100 | weight_decay=0.0,
101 | lr_anneal_steps=0,
102 | batch_size=1,
103 | microbatch=-1,
104 | ema_rate="0.9999",
105 | log_interval=10,
106 | save_interval=10000,
107 | resume_checkpoint="",
108 | use_fp16=False,
109 | fp16_scale_growth=1e-3,
110 | result_folder = None,
111 | )
112 | defaults.update(sr_model_and_diffusion_defaults())
113 | parser = argparse.ArgumentParser()
114 | add_dict_to_argparser(parser, defaults)
115 | return parser
116 |
117 |
118 | if __name__ == "__main__":
119 | main()
120 |
--------------------------------------------------------------------------------
/test_DDPM_3d_mpi.sh:
--------------------------------------------------------------------------------
1 | SAMPLE_FLAGS="--batch_size 1 --num_samples 6"
2 | MODEL_FLAGS="--attention_resolutions 1000 --large_size 96 --small_size 96 --num_channels 128 --use_fp16 True --num_head_channels 64 --learn_sigma True --resblock_updown True --use_scale_shift_norm True"
3 | DIFFUSION_FLAGS="--diffusion_steps 1000 --noise_schedule linear --rescale_learned_sigmas False --rescale_timesteps False"
4 |
5 | mpiexec -n 6 python ./scripts/test.py $MODEL_FLAGS --model_path ./checkpoints/model.pt --base_samples sample_PET.npz --save_dir ./results/ $SAMPLE_FLAGS
6 |
7 |
--------------------------------------------------------------------------------