run inference less often

2024-07-02 04:25:26 -07:00
parent 3b41479aa1
commit 2c0939a733
1 changed files with 2 additions and 2 deletions
@@ -685,13 +685,13 @@ def main(config: TrainingConfig, wandb_log = False):
                print(f"Reached max steps ({config.max_train_steps}), stopping training!")
                break

-            if global_step % 5 == 0:
+            if global_step % 25 == 0:
                inference_device = "cuda:1"
                torch.cuda.empty_cache()
                pipeline = pipeline.to(inference_device)
                pipeline.transformer = transformer.to(inference_device)
                prompt_embeds, pooled_prompt_embeds = compute_text_embeddings(
-                    prompt = ['<s0><s1>, hanging out on mars'], 
+                    prompt = ["<s0><s1>, there is a cartoon banana that is eating popcorn while holding a knife"], 
                    text_encoders = text_encoders, 
                    tokenizers = tokenizers,
                    device=inference_device