search

openai / guided-diffusion

MIT License
6.03k stars 803 forks source link

Sampling details for FID evaluation #56

Closed zhihongp closed 2 years ago

zhihongp commented 2 years ago

While it's mentioned that the full training set was used as reference for FID, I couldn't find details about evaluation so I wonder if I have missed them. Noticed the 10K samples (instead of 50K) used in ablation, but no other specifications about sampling. Besides, for guided sampling, specifics like samples per class would be useful too. Any insight is appreciated, either from author or readers.

zhihongp commented 2 years ago

Found in evaluation code that it's 50K but #samples/class is random

xcharxlie commented 1 year ago

Found in evaluation code that it's 50K but #samples/class is random

@zhihongp I also have the same question. Do you know why they use random int to represent the output class they desire to generate in classifier-guided sampling? I believe this "y" is the class you want to see in the outputs, right? classes = th.randint( low=0, high=NUM_CLASSES, size=(args.batch_size,), device=dist_util.dev() ) model_kwargs["y"] = classes

wwj723452017 commented 3 weeks ago

Found in evaluation code that it's 50K but #samples/class is random

@zhihongp I also have the same question. Do you know why they use random int to represent the output class they desire to generate in classifier-guided sampling? I believe this "y" is the class you want to see in the outputs, right? classes = th.randint( low=0, high=NUM_CLASSES, size=(args.batch_size,), device=dist_util.dev() ) model_kwargs["y"] = classes

@zhihongp @xcharxlie hi,have you solved this question?I tried to look into image_sample.py script,and in muti-class situation,I want to know how the pics matching with the lables. According to (model_kwargs["y"] = classes) , classes are something like these:682 380 174 125 533 27 133 421 581 985(tensor),AND I can't figure out the numbers and the relationship with the labels.Thanks for your reply.