Awesome

Extended Analytic-DPM

• This is the official implementation for Estimating the Optimal Covariance with Imperfect Mean in Diffusion Probabilistic Models (Accepted in ICML 2022). It extends Analytic-DPM under the following two settings:

  • The reverse process adpots complicated covariance matrices dependent to states, instead of simple scalar variances (which motivates the SN-DPM in the paper).
  • The score-based model has some error w.r.t. the exact score function (which motivates NPR-DPM in the paper).

• This codebase also reimplements Analytic-DPM and reproduces its most results. The pretrained DPMs used in the Analytic-DPM paper are provided <a href="#pretrained_dpm">here</a>, and have already been converted to a format that can be directly used for this codebase. We also additionally applies Analytic-DPM to score-based SDE.

• Models and FID statistics are available <a href="#model">here</a> to reproduce results in this paper.

Dependencies

The codebase is based on pytorch. The dependencies are listed below.

pip install pytorch>=1.9.0 torchvision ml-collections ninja tensorboard

Basic usage

The basic usage for training is

python run_train.py --pretrained_path path/to/pretrained_dpm --dataset dataset --workspace path/to/working_directory $train_hparams

• pretrained_path is the path to a pretrained diffusion probabilistic model (DPM). <a href="#pretrained_dpm">Here</a> provide all pretrained DPMs used in this work.
• dataset represents the training dataset, one of <cifar10|celeba64|imagenet64|lsun_bedroom>.
• workspace is the place to put training outputs, e.g., logs and middle checkpoints.
• train_hparams specify other hyperparameters used in training. <a href="#model">Here</a> lists train_hparams for all models.

The basic usage for evaluation is

python run_eval.py --pretrained_path path/to/evaluated_model --dataset dataset --workspace path/to/working_directory \
    --phase phase --sample_steps sample_steps --batch_size batch_size --method method $eval_hparams

• pretrained_path is the path to a model to evaluate. <a href="#model">Here</a> provide all models evaluated in this work.
• dataset represents the dataset the model is trained on, one of <cifar10|celeba64|imagenet64|lsun_bedroom>.
• workspace is the place to put evaluation outputs, e.g., logs, samples and bpd values.
• phase specifies running sampling or likelihood evaluation, one of <sample4test|nll4test>.
• sample_steps is the number of steps to run during inference, the samller this value the faster the inference.
• batch_size is the batch size, e.g., 500.
• method specifies the type of the model, one of:
  • pred_eps the original DPM (i.e., a noise prediction model) with discrete timesteps
  • pred_eps_eps2_pretrained the SN-DPM with discrete timesteps
  • pred_eps_epsc_pretrained the NPR-DPM with discrete timesteps
  • pred_eps_ct2dt the original (i.e., a noise prediction model) with continuous timesteps (i.e., a score-based SDE)
  • pred_eps_eps2_pretrained_ct2dt the SN-DPM with continuous timesteps
  • pred_eps_epsc_pretrained_ct2dt the NPR-DPM with continuous timesteps
• eval_hparams specifies other optional hyperparameters used in evaluation.
• <a href="#evaluation">Here</a> lists method and eval_hparams for NPR/SN-DPM and Analytic-DPM results in this paper.

Models and FID statistics

Here is the list of NPR-DPMs and SN-DPMs trained in this work. These models only train an additional prediction head in the last layer of a pretrained diffusion probabilistic model (DPM).

Here is the list of pretrained DPMs, collected from prior works. They are converted to a format that can be directly used for this codebase. <a id="pretrained_dpm"/>

This link provides precalculated FID statistics on CIFAR10, CelebA 64x64, ImageNet 64x64 and LSUN Bedroom. They are computed following Appendix F.2 in Analytic-DPM.

Evaluation Hyperparamters for NPR/SN-DPM and Analytic-DPM

Note: Analytic-DPM needs to precalculate the expected mean squared norm of noise prediction model (ms_eps), which is provided <a href="#pretrained_dpm">here</a>. Specify their path by --ms_eps_path.

• Sampling experiments on CIFAR10 (LS) or CelebA 64x64, Table 1 in the paper:

• Sampling experiments on CIFAR10 (CS), Table 1 in the paper:

• Sampling experiments on CIFAR10 (VP SDE), Table 1 in the paper:

• Sampling experiments on ImageNet 64x64, Table 1 in the paper:

• Likelihood experiments on CIFAR10 (LS) or CelebA 64x64, Table 3 in the paper:

• Likelihood experiments on CIFAR10 (CS), Table 3 in the paper:

• Likelihood experiments on ImageNet 64x64, Table 3 in the paper:

This implementation is based on / inspired by

• Analytic-DPM (provide the code structure)

• pytorch_diffusion (provide codes of models for CelebA64x64 and LSUN Bedroom)

• Improved DDPM (provide codes of models for CIFAR10 and Imagenet64x64)

• score-sde (provide codes of models for CIFAR10)

• pytorch-fid (provide the official implementation of FID to PyTorch)