Quickstart

Install required packages first:

pip install -r requirements.txt

Please download the CIFAR10 model via wget and put it at model/ddpm_ema_cifar10:

wget https://github.com/VainF/Diff-Pruning/releases/download/v0.0.1/ddpm_ema_cifar10.zip

PS: Note this model is only supported by old diffusers.

To train RL, run:

python train.py

To evaluate FID, run:

bash eval_fid.sh

PS. Please modify --save_path in eval_fid.sh, this means the path of the RL model produced by train.py.

(Note all these codes may not be 100% accurate)

Experimental Results

CIFAR 10

	T=5	T=10	T=20	T=100
DDIM	68.28	20.76	11.46	5.71
RL	34.67	18.13	11.02	-

Thresholds of sparse reward are set 0.75, 0.89, 0.93 for T of 5, 10, 20. The setting of thresholds impacts a lot. For example, experiment of T=5 with threshold of 0.8 only gets FID of 66.33.

LSUN-Church

	T=5	T=10	T=20	T=100
DDIM	49.84	19.10	12.04	10.55
RL	52.97	21.24	12.60	-

Thresholds of sparse reward are set 0.55, 0.76, 0.89 for T of 5, 10, 20. The results of RL are worse than DDIM, which are likely caused by the discrepancy between FID and SSIM. Moreover, prior work has not implemented on high-resolution (256x256) images, which are more difficult tasks.