Classifier-Free Diffusion Guidance

This implementation is adapted from coderpiaobozhe/classifier-free-diffusion-guidance-Pytorch

.
├─logs
├─notes
│  └─pics
├─samples
│  └─MultiheadAttn
└─src

logs: Tensorboard log directory

notes: My notes about key formulas in CFG

pt: Saved-model directory

samples: Some images sampled by the model with self-attention block

samples/MultiheadAttn: Some images sampled by the model with multihead-attention block

src: Source code directory

How to Run

Ensure the current directory in your terminal is the root of the repository, and then cd src.

Datasets: CIFAR-10

Train: python train.py [args]

usage: train.py [-h] [--batchsize BATCHSIZE] [--lr LR] [--epoch EPOCH] [--datadir DATADIR] [--modeldir MODELDIR] [--sampledir SAMPLEDIR] [--uncondrate UNCONDRATE] [--numgen NUMGEN] [--numlabel NUMLABEL] [--logdir LOGDIR] [--T T] [--betamode BETAMODE] [--w W] [--ddim DDIM] [--eta ETA] [--taumode TAUMODE] [--ddimsteps DDIMSTEPS] [--modelch MODELCH] [--chmul CHMUL] [--numres NUMRES] [--dropout DROPOUT]

diffusion model training

options:

-h, --help show this help message and exit

--batchsize batch size for training

--lr learning rate

--epoch epoch

--datadir data directory

--modeldir saved-model directory

--sampledir sample directory

--uncondrate rate of training without condition for classifier-free guidance

--numgen the number of samples for each label

--numlabel num of labels

--logdir log directory

--T total timesteps for diffusion

--betamode noise schedule mode: cosine or linear

--w hyperparameters for classifier-free guidance strength

--ddim whether DDIM is used in sampling. If not, DDPM is used.

--eta eta for variance during DDIM sampling process

--taumode sub-sequence selection for reverse process in DDIM, which should be 'linear' or 'quadratic'

--ddimsteps sampling steps for DDIM

--modelch model channels for UNet

--chmul architecture parameters training UNet

--numres number of ResBlock+AttnBlock for each block in UNet

--dropout dropout rate for ResBlock

Eval: python eval.py [args]

Specify your model path by joining --modeldir and --modelfilename

usage: eval.py [-h] [--modeldir MODELDIR] --modelfilename MODELFILENAME [--sampledir SAMPLEDIR] [--numgen NUMGEN] [--numlabel NUMLABEL] [--T T] [--betamode BETAMODE] [--w W] [--ddim DDIM] [--eta ETA] [--taumode TAUMODE] [--ddimsteps DDIMSTEPS] [--modelch MODELCH] [--chmul CHMUL] [--numres NUMRES] [--dropout DROPOUT]

diffusion model evaluation

options:

-h, --help show this help message and exit

--modeldir saved-model directory

--modelfilename saved-model filename

--sampledir sample directory

--numgen the number of samples for each label

--numlabel num of labels

--T total timesteps for diffusion

--betamode noise schedule mode: cosine or linear

--w hyperparameters for classifier-free guidance strength

--ddim whether DDIM is used in sample, if not DDPM is used

--eta eta for variance during DDIM sampling process

--taumode sub-sequence selection for reverse process in DDIM, which should be 'linear' or 'quadratic'

--ddimsteps DDIMSTEPS sampling steps for DDIM

--modelch model channels for UNet

--chmul architecture parameters training UNet

--numres number of ResBlock+AttnBlock for each block in UNet

--dropout dropout rate for ResBlock

Some Results

Using the model with self-attention block:

• Using DDPM with T=1000, betamode='linear', taumode='linear':

  

• Using DDIM with T=50, betamode='linear', eta=0.5, taumode='linear':

  

Using the model with self-attention block:

• Using DDPM with T=1000, betamode='linear', taumode='linear':

  

• Using DDIM with T=50, betamode='linear', eta=0.5, taumode='linear':