05212111

diffusion.py

@@ -468,7 +468,7 @@ class DDPM21CM:
             # sampler=DistributedSampler(dataset),
             )
         if len(self.dataloader) % self.config.gradient_accumulation_steps != 0:
-            raise ValueError(f"len(self.dataloader) % self.config.gradient_accumulation_steps = {len(self.dataloader) % self.config.gradient_accumulation_steps} instead of 0")
+            raise ValueError(f"len(self.dataloader) % self.config.gradient_accumulation_steps = {len(self.dataloader) % self.config.gradient_accumulation_steps} instead of 0. Make sure len(dataloader)={len(dataloader)} is dividable by gradient_accumulation_steps={self.config.gradient_accumulation_steps}.")
 
         dataloader_end = time()
         print(f"cuda:{torch.cuda.current_device()}/{self.config.global_rank} dataloader costs {dataloader_end-dataloader_start:.3f}s")