From 8ee4d711f4f3476bff33d16f3f9aede9f925fccf Mon Sep 17 00:00:00 2001 From: Nachos Date: Mon, 13 Nov 2023 10:06:56 +0100 Subject: [PATCH 1/3] Update utils.py - Fixed save checkpoint with scheduler set to None --- src/optim/utils.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/src/optim/utils.py b/src/optim/utils.py index dd00621..957fad8 100755 --- a/src/optim/utils.py +++ b/src/optim/utils.py @@ -115,7 +115,7 @@ def save_checkpoint(distributed_backend, model, opt, scheduler, itr, ckpt_path, checkpoint = dict({ 'model': distributed_backend.get_raw_model(model).state_dict(), 'optimizer': opt.state_dict(), - 'scheduler': scheduler.state_dict(), + 'scheduler': scheduler.state_dict() if scheduler != None else None, 'itr': itr, }, **extra_args) From c015add977076a7f05348fbf5e0fdf90b9ae8419 Mon Sep 17 00:00:00 2001 From: Nachos Date: Mon, 13 Nov 2023 12:14:11 +0100 Subject: [PATCH 2/3] Update base.py - Fixed scheduler set to None causes error on scheduler.step() during training --- src/optim/base.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/src/optim/base.py b/src/optim/base.py index fa176f0..018934a 100755 --- a/src/optim/base.py +++ b/src/optim/base.py @@ -40,7 +40,8 @@ def train_base(model, opt, data, scheduler, iterations, acc_steps, batch_size, s if extra_args.grad_clip != 0.0: torch.nn.utils.clip_grad_norm_(model.parameters(), extra_args.grad_clip) opt.step() - scheduler.step() + if scheduler != None: + scheduler.step() opt.zero_grad(set_to_none=True) itr += 1 From c4b9ae8410e7b79c0b7eebb1714b919a9c4cd760 Mon Sep 17 00:00:00 2001 From: Nachos Date: Mon, 13 Nov 2023 12:16:49 +0100 Subject: [PATCH 3/3] Update sparse.py - same for sparse.py --- src/optim/sparse.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/src/optim/sparse.py b/src/optim/sparse.py index b223f17..6669174 100755 --- a/src/optim/sparse.py +++ b/src/optim/sparse.py @@ -37,7 +37,8 @@ def train_sparse(model, opt, data, scheduler, iterations, acc_steps, batch_size, substep += 1 opt.step() - scheduler.step() + if scheduler != None: + scheduler.step() opt.zero_grad(set_to_none=True) itr += 1