From 389dc1c31037fa57d4474bb74ac8288d8dd90aaa Mon Sep 17 00:00:00 2001 From: yinfeng Date: Wed, 9 Apr 2025 11:21:00 +0800 Subject: [PATCH] Fix bug: When resuming training from a checkpoint, model_avg was not assigned, resulting in a None error. --- egs/wenetspeech/KWS/zipformer/finetune.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/egs/wenetspeech/KWS/zipformer/finetune.py b/egs/wenetspeech/KWS/zipformer/finetune.py index cd437da4c..249209352 100755 --- a/egs/wenetspeech/KWS/zipformer/finetune.py +++ b/egs/wenetspeech/KWS/zipformer/finetune.py @@ -593,6 +593,9 @@ def run(rank, world_size, args): if params.continue_finetune: assert params.start_epoch > 0, params.start_epoch + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) checkpoints = load_checkpoint_if_available( params=params, model=model, model_avg=model_avg )