Change strategy

train_sa.py

@@ -70,7 +70,7 @@ def main():
 
     trainer = pl.Trainer(accelerator="gpu", devices=num_gpus, profiler="simple", 
                          default_root_dir="./logs", logger=WandbLogger(project="slot-att") if args.wandb else None,
-                         strategy="ddp" if num_gpus > 1 else "default", callbacks=[checkpoint_callback],
+                         strategy="ddp_find_unused_parameters_true" if num_gpus > 1 else "default", callbacks=[checkpoint_callback],
                          log_every_n_steps=100, max_steps=num_train_steps)
 
     trainer.fit(model, train_loader, val_loader)