Open ChangyaoTian opened 1 week ago
Thank you for reporting the issues. We have rechecked the evaluation scripts and believe the issue may be due to the setting of top-k sampling. It should be set to 10,000 (k=10% of the vocabulary size) rather than 100,000 (without top-k sampling). We have made the necessary modifications.
Hey, when I directly run the eval scripts of GPT here with the provided pre-trained ckpts here, I can only get 24.6 FID Score rather than the claimed 15.4. Is something wrong here?