IDDM v1.1.5

What's Changed

• FID calculator used for evaluating generated images. by @egoist945402376 in #64
• Update: Add Pytorch_fid in requirements.txt by @EdwardTj in #65
• Update: Refactor train.py format. by @chairc in #66
• Separate get dataset methods into data files; Add initialization operation. by @chairc in #67
• Update: Modify repository structure, parameter explanation and citation. by @chairc in #68
• Add: Add better FID calculator to verify image quality. by @chairc in #69
• Update: Modify the citation link. by @chairc in #70
• Add: Add Evaluation in README. by @chairc in #71
• Update: Modify type description error in README. by @chairc in #72

New Contributors

• @EdwardTj made their contribution in #65
• @egoist945402376 made their contribution in #64

Weights

Note: The weight include model, ema_model and optimizer.

• celebahq-120-weight.pt: Trained on a dataset of 30,000 people face, and image size is 120
• animate-ganyu-120-weight.pt: Trained on a dataset of 500 animate ganyu face, and image size is 120
• neu-120-weight.pt: Trained on a dataset of 1,800 defect, and image size is 120
• cifar-64-weight.pt: Trained on a dataset of 60,000 images, and image size is 64
• animate-face-64-weight.pt: Trained on a dataset of 63,565 animate face, and image size is 64

Full Changelog: v1.1.4...v1.1.5

IDDM v1.1.4

What's Changed

• Add PLMS sampler; Add PLMS sample initializer. by @chairc in #51
• Removed forced checking of samplers, now free choice of sample generation; Add output image format; Add choices.py; Modify the choices attribute in parsers. by @chairc in #52
• Update README.md by @chairc in #53
• Add version.py. by @chairc in #54
• Add --save_model_interval_epochs parameter. Save model interval and save it every X epochs. by @chairc in #55
• Modify the mean, std and random resized crop settings in torchvision.transforms.Compose; Modify the choices in parser and adjust the order of parts. by @chairc in #56
• Fix: Fix unable to load unconditional model. by @chairc in #57
• Modify the config in repository structure; Added "--noise_schedule" training parameters, this method is a model noise adding method. by @chairc in #59
• Added unetv2.py, replace nn.Upsample with nn.ConvTranspose2d; Add DDIM code comments; Fix the problem that bool in parser cannot be correctly recognized by the console. by @chairc in #61
• Modify the get_dataset; Add check.py; Add classes_initializer function by @chairc in #62

Weights

• celebahq-120-weight.pt: Trained on a dataset of 30,000 people face, and image size is 120
• animate-ganyu-120-weight.pt: Trained on a dataset of 500 animate ganyu face, and image size is 120
• neu-120-weight.pt: Trained on a dataset of 1,800 defect, and image size is 120

Full Changelog: v1.1.3...v1.1.4

IDDM v1.1.3

What's Changed

• Add the setting to run GPU commands; Add the tag and update acknowledgements. by @chairc in #37
• Add: Add Visual webui. by @chairc in #38
• Modify the GPU settings. CPU training is not supported after version 1.1.2; Modify automatic mixed precision setting, optimize GradScaler. by @chairc in #39
• Modify information prompt problems; Modify beta_end; Add sqrt_linear and sqrt schedules. by @chairc in #41
• Add: Add super resolution model (The implement of RDN), low images to high. by @chairc in #43
• Generation image format by @egoist945402376 in #45
• Add: Add option to load EMA model. by @chairc in #46
• Update: Modify Chinese introduction errors and add --image_format parameter introduction. by @chairc in #48

New Contributors

• @egoist945402376 made their first contribution in #45

Full Changelog: v1.1.2...v1.1.3

Weights

• celebahq-weight.pt: Trained on a dataset of 30,000 people face, and image size is 120.

IDDM v1.1.2-stable

What's Changed

• Update: Modify automatic mixed precision training name; reconstruct automatic mixed precision training structure. by @chairc in #28
• Update: Modify project's name. by @chairc in #29
• Add: Add current epoch average loss log. by @chairc in #30
• Update: Update the introduction in README. by @chairc in #31
• Major update about generating images by @chairc in #33
• Modify some parameters and formats; Generate images in new folder. by @chairc in #34
• Modify comment; Modify README, add issue link and correct some text. by @chairc in #35

Note

Due to changes in the model saving structure, the following parameters in 'generate.py' may not be entered after this version.

Parameter Name	Conditional	Usage	Type	Description
--sample		Sampling method	str	Set the sampling method type, currently supporting DDPM and DDIM. (No need to set for models after version 1.1.1)
--network		Training network	str	Set the training network, currently supporting UNet, CSPDarkUNet. (No need to set for models after version 1.1.1)
--act		Activation function	str	Activation function selection. Currently supports gelu, silu, relu, relu6 and lrelu. If you do not set the same activation function as the model, mosaic phenomenon will occur. (No need to set for models after version 1.1.1)
--num_classes	✓	Number of classes	int	Number of classes for classification (No need to set for models after version 1.1.1)

Full Changelog: v1.1.1...v1.1.2

IDDM v1.1.1

What's Changed

• Update: Update README. by @chairc in #20
• Add: Add separate checkpoint weights function. by @chairc in #21
• Add: Add pretrain ckeckpoint method. by @chairc in #22
• Update: The utils.initializer encapsulation method is used in distributed training. by @chairc in #23
• Update: Add "Base on the 64×64 model to generate every size images". by @chairc in #24
• Add: Add 160×160 NEU-DET generate images. by @chairc in #25

Full Changelog: v1.1.0...v1.1.1

IDDM v1.1.0-stable

Note

In this version, The storage format of our model has completely changed, and the model needs to be retrained if the latest code is pulled.
If you don‘t like this version, you can download this tag (v1.0.2).

What's Changed

• Major update: model layer refactored; New network selection (unet and cspdarkunet); Updated the README by @chairc in #11
• Update: Encapsulate code to improve reusability. by @chairc in #12
• Update: Run GPU with custom settings. by @chairc in #13
• Add: Add show and save image in images. by @chairc in #14
• Update: Reconstruct model/networks and other included files. by @chairc in #15
• Update: Add function comments; Update model.parameters() and add SGD optimizer. by @chairc in #16
• Major update: Reconstruct modules.py, and update package path; Rewrite checkpoint storage and loading functions, add checkpoint files by @chairc in #18

New Contributors

• @chairc made their first contribution in #11

Full Changelog: v1.0.2...v1.1.0

About weight model

NEU weight parameters is {"conditional"=True, "image_size"=64, "sample"=ddim, "network"=unet, "act"=gelu, "num_classes"=6}
Cifar10 weight parameters is {"conditional"=True, "image_size"=64, "sample"=ddim, "network"=unet, "act"=gelu, "num_classes"=10}
This weight include model , ema model and optimizer checkpoints.

IDDM v1.0.2

We are delighted to announce that version 1.0.2 of IDDM (Industrial Defect Diffusion Model) has been officially released.
In this version, we have made the following changes.

• Update: Modify tqdm progress bar problem. #b1ea7c3
• Add: Add a parameter for the number of custom visualization images generated. #b48b61b
• Update: Modify the training setting in 4 datasets. #5f20649
• Add: Add fast deployment and API on cloud servers, and send message test function. #f48c468
• Add: Add activation function setting. #8b34de2
• Update: Optimize some code. #cd78a21
• Add: Add saving train log method. #5c4b49a
• Fix: Fix the problem of different training times and modified the distributed training synchronization lock. #c0157d9
• Add: Add selection of activation functions in the generated model. #f8767d2
• Add: Add dataset example. #d3ec97a
• Add: Add results README. #0cd6f57
• Update: Modify README. #526134b
• Add: Add animate face images. #256393c
• Update: Modify the subtitle style. #49dfcf2
• Add: Add the image of IDDM training. #662b4c3
• Update: Add training demo. #8f1ade9
• Update: Modify --distributed initial parameter default value. #6297617
• Fix: Fix the expression type error and changge in to float. #2ca7449
• Update: Modify test error. #4b43d32
• Update: This is a method to display the results of each model during training and can be commented out. #3443662
• Add: Add the 'check_and_create_dir' function. #7f88e91
• Add: Add the activation function in deploy.py. #6529f80
• Fix: Fix no folder error after generate. #5b15de6
• Add: When adding conditional generation, all class images can be generated at once. Multiple images of a single class can also be generated.#a4c77b9
• Update: Add parameter explanation. #dfbe132
• Update: Add model construction explanation. #8a9e121
• Update: Modify model construction explanation error. #b38864c

IDDM v1.0.1

We are delighted to announce that version 1.0.1 of IDDM (Industrial Defect Diffusion Model) has been officially released.
In this version, we have made the following changes.

• Fix the bug that the model does not save when a single card or CPU is used. #5ee6010
• Fix resuming training model and optimizer parameters loading. #e58cd4e
• Modify the resume training model loading and cancel the broadcast method. #19bc4d9
• Add seed initializer. #9189913
• Add learning rate function such as default, cosine and warmup_cosine. #35d024e
• Add network summary test function. #58f0694
• Modify some formatting issues. #f35040b
• Add requirements.txt #0fb164f
• Refactoring the sampling model structure by creating a base class for common parts and having other sampling methods inherit from it.#278debd
• Add DDIM sample method. #c469e4f
• Add sample selector method during training. #6acf6cc
• Update README #329ded9
• Add image results generated in README. #5df2947
• Modify Repository Structure. #6c21b9b
• Optimizing loading weight method, optimizing generate.py and Check for issues where distributed weight names are not the same as network weights. #24a03c3
• Edited to english in 'train.py' and 'test_module.py'. #788138d
• Edited to english in 'train.py' and 'generate.py'. #f2e9d7c
• Edited to english in 'network.py' and 'modules.py'. #8c0bb55
• Edited to english in 'base.py', 'ddpm.py' and 'ddim.py'. #c92fd82
• Edit english in 'generate.py'. #970ec3c
• Edit untranslated parts in 'utils.py'. #3089bb8

IDDM v1.0.0

We are delighted to announce that version 0.0.1 of IDDM (Industrial Defect Diffusion Model) has been officially released. In this version, we have added a distributed training method, allowing you to make full use of multiple GPUs for larger batch sizes and faster training. Additionally, we have provided pre-trained models on multiple industrial defect datasets, namely NRSD-pretrain and NEUDET-pretrain.

Thank you for your attention, and enjoy using our IDDM!

IDDM v0.0.1alpha

Update: Update README.