Denis

How to calculate denotation accuracy for queries for Overnight?

6

Various papers report denotation accuracy on Overnight, which requires executing the logical form on a database. (The Overnight paper specifies that "We created a database for each domain by randomly...

Wrote extended DDIM scheduler that takes pairs of timesteps

1

Re-implemented #1236 a bit differently after a suggestion from @patrickvonplaten . Now the original DDIMScheduler is untouched but a new class is introduced with extra featuers (see below) and DDIMPipeline...

Scale and shift in Resnet block

6

Implemented optional scale and shift (instead of currently only shift) based on timestep embedding, similarly to OpenAI's improved diffusion code. However, it is currently not being used and I would...

Added sampling script for unconditional training example

7

Added sampling script that supports both DDPM and DDIM sampling for unconditional training example

Improve DDIM sampling pipeline (support for predict x0 and uneven jumps)

1

* support for predict_epsilon = False in DDIM sampler * changed timestep selection such that it is more uniform if number of sampling steps doesn't cleanly divide number of training...

Remove unethically trained diffusion models from model hub

10

@patrickvonplaten @patil-suraj @anton-l I wonder what your thoughts are on the arguments the artists are making on Artstation, Twitter and Instagram right now. The main point is that the artists...

storing original images from dataset

4

Script to store original images from unconditional training datasets as a collection of images. This script can be used before computing FID scores with tools like torch-fidelity.

added support for cifar10 and automatic resolution selection in training example

5

Added support for CIFAR10 in training example and automatic resolution selection + printing of param counts

FID on CIFAR10 does not match literature

4

### Describe the bug I used the unconditional training script from examples to train a model on CIFAR10, using default parameters (except using 1024 steps). After around 250k updates (batch...

Added --save_every option in dreambooth script.

1

Added --save_every option to the dreambooth training script. In contast to --checkpointing_steps, it saves the model where the final model will be stored and does so independently of checkpointing settings....