kvablack/ddpo-pytorch

Issues

• In train.py, the sample order between sample["advantages"] and sample["timesteps", "latents", "next_latents", "log_probs"] does not match.

  #29 opened 4 months ago by YangSun22
  0

• Support for EulerAncestralDiscreteScheduler

  #28 opened 4 months ago by anonymous-atom
  2

• Questions about the reward curve and bert.

  #23 opened a year ago by zjuAIHz
  2

• Does this training process apply to latest SD, such as stabilityai/stable-diffusion-3-medium

  #27 opened 6 months ago by roywang021
  0

• How to save fine-tuned model properly

  #26 opened 6 months ago by shashankg7
  0

• training-code

  #25 opened 7 months ago by ParnyanAtaei
  0

• OOM when using "stabilityai/stable-diffusion-2-1" with batch size of 2

  #24 opened a year ago by xilongzhou
  1

• Finetuning on google colab

  #22 opened a year ago by alirezanobakht13
  0

• Batch size unrecogonized

  #20 opened a year ago by mao-code
  1

• SDXL Support?

  #17 opened a year ago by rdcoder33
  1

• Code logics, thanks

  #19 opened a year ago by junyongyou
  0

• unet keeps producing nan during training

  #18 opened a year ago by EYcab
  2

• Hello, when I trained an aesthetic model using the default configuration on 8 A800 cards, I found that the training process got stuck after completing one epoch, but it worked fine when using a single A800 card. May I ask what could be the cause of this situation?

  #13 opened 2 years ago by cjt222
  8

• About the training with prompt_image_alignment configuration which uses llava_bertscore reward function

  #11 opened 2 years ago by QZJ-2003
  5

• prompt-dependent value function optimization

  #15 opened 2 years ago by hkunzhe
  0

• About the large dataset and Unet Training

  #7 opened 2 years ago by liuyuemaicha
  3

• reproducing the aesthetic experiment

  #3 opened 2 years ago by seashell123
  7

• Support for other schedulers

  #14 opened 2 years ago by desaixie
  1

• Question about the optimized objective.

  #12 opened 2 years ago by JacobYuan7
  1

• Suggest to use larger gradient accumulation steps instead of multi GPUs

  #10 opened 2 years ago by hkunzhe
  3

• Prompt Alignment with LLaVA-server: Client-side prompt and image doesn't match server side reward

  #6 opened 2 years ago by desaixie
  4

• fp16 only if using lora?

  #8 opened 2 years ago by GiilDe
  2

• On reproducing LLaVA alignment experiments.

  #5 opened 2 years ago by bhattg
  2

• OOM despite using A100-80GB GPU and 128GB CPU memory (+16 CPUs per task)

  #4 opened 2 years ago by noshaba
  2

• Gif visualization

  #1 opened 2 years ago by SnowdenLee
  1

• On reproducibility and LoRA

  #2 opened 2 years ago by bhattg
  1