diff --git a/tests/test_engine.py b/tests/test_engine.py index 67502093..18c47fb2 100644 --- a/tests/test_engine.py +++ b/tests/test_engine.py @@ -163,13 +163,17 @@ def test_nan_recovery(): assert nan_injected[0], "NaN injection failed" -@pytest.mark.parametrize("pretrained", [True, False, MODEL]) -def test_train_reuses_loaded_checkpoint_config(monkeypatch, pretrained): - """Test training reuses an already-loaded checkpoint config without forcing pretrained weights.""" +@pytest.mark.parametrize( + "kwargs,uses_weights", + [({}, True), ({"pretrained": True}, True), ({"pretrained": False}, False), ({"pretrained": MODEL}, True)], +) +def test_train_reuses_loaded_checkpoint_model(monkeypatch, kwargs, uses_weights): + """Test training reuses loaded checkpoint config while respecting the pretrained argument.""" model = YOLO("yolo26n.yaml") model.ckpt = {"checkpoint": True} model.ckpt_path = "/tmp/fake.pt" model.overrides["model"] = "ul://glenn-jocher/m2/exp-14" + model.overrides["pretrained"] = False original_model = model.model captured = {} @@ -198,15 +202,15 @@ def test_train_reuses_loaded_checkpoint_config(monkeypatch, pretrained): lambda path: (original_model, {"checkpoint": True}), ) - model.train(data="coco8.yaml", epochs=1, pretrained=pretrained) + model.train(data="coco8.yaml", epochs=1, **kwargs) assert captured["trainer"].model is original_model, "Trainer model does not match original" assert captured["cfg"] == original_model.yaml, f"Config mismatch: {captured['cfg']} != {original_model.yaml}" - assert captured["weights"] is (None if pretrained is False else original_model), "Unexpected weights loaded" + assert captured["weights"] is (original_model if uses_weights else None), "Unexpected weights loaded" -@pytest.mark.parametrize("pretrained", [False, MODEL]) -def test_setup_model_respects_pretrained_arg_for_pt_models(monkeypatch, pretrained): +@pytest.mark.parametrize("pretrained,uses_weights", [(True, True), (False, False), (MODEL, True)]) +def test_setup_model_respects_pretrained_arg_for_pt_models(monkeypatch, pretrained, uses_weights): """Test .pt models use checkpoint config while respecting the pretrained argument.""" captured = {} checkpoint_model = SimpleNamespace(yaml={"nc": 80}) @@ -228,4 +232,4 @@ def test_setup_model_respects_pretrained_arg_for_pt_models(monkeypatch, pretrain trainer.setup_model() assert captured["cfg"] == checkpoint_model.yaml, "Checkpoint config was not used" - assert captured["weights"] is (None if pretrained is False else checkpoint_model), "Unexpected weights loaded" + assert captured["weights"] is (checkpoint_model if uses_weights else None), "Unexpected weights loaded" diff --git a/ultralytics/__init__.py b/ultralytics/__init__.py index 207310f4..071f2c58 100644 --- a/ultralytics/__init__.py +++ b/ultralytics/__init__.py @@ -1,6 +1,6 @@ # Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license -__version__ = "8.4.44" +__version__ = "8.4.45" import importlib import os diff --git a/ultralytics/engine/model.py b/ultralytics/engine/model.py index 71e3d821..c899ba65 100644 --- a/ultralytics/engine/model.py +++ b/ultralytics/engine/model.py @@ -767,6 +767,7 @@ class Model(torch.nn.Module): "task": self.task, } # method defaults args = {**overrides, **custom, **kwargs, "mode": "train", "session": self.session} # prioritizes rightmost args + pretrained = kwargs.get("pretrained", overrides.get("pretrained", True) if kwargs.get("cfg") else True) if args.get("resume"): if args["resume"] is True: # resume=True (boolean) uses current model as checkpoint if self.ckpt and self.ckpt.get("epoch", -1) >= 0 and self.ckpt.get("optimizer") is not None: @@ -781,10 +782,10 @@ class Model(torch.nn.Module): self.trainer = (trainer or self._smart_load("trainer"))(overrides=args, _callbacks=self.callbacks) if not args.get("resume") and self.ckpt: - # Reuse the already-loaded checkpoint config to avoid re-resolving remote sources during trainer setup. - weights = self.model if args.get("pretrained") else None - if isinstance(args.get("pretrained"), (str, Path)): - weights, _ = load_checkpoint(args["pretrained"]) + # Reuse the already-loaded checkpoint model to avoid re-resolving remote weight sources during trainer setup. + weights = None if pretrained is False else self.model + if isinstance(pretrained, (str, Path)): + weights, _ = load_checkpoint(pretrained) self.trainer.model = self.trainer.get_model(weights=weights, cfg=self.model.yaml) self.model = self.trainer.model diff --git a/ultralytics/engine/trainer.py b/ultralytics/engine/trainer.py index 6ffe449d..fb23bbd9 100644 --- a/ultralytics/engine/trainer.py +++ b/ultralytics/engine/trainer.py @@ -718,12 +718,10 @@ class BaseTrainer: if str(self.model).endswith(".pt"): weights, ckpt = load_checkpoint(self.model) cfg = weights.yaml - if isinstance(self.args.pretrained, (str, Path)): - weights, _ = load_checkpoint(self.args.pretrained) - elif not self.args.pretrained and not self.resume: - weights = None - elif isinstance(self.args.pretrained, (str, Path)): + if isinstance(self.args.pretrained, (str, Path)): weights, _ = load_checkpoint(self.args.pretrained) + elif self.args.pretrained is False and not self.resume: + weights = None self.model = self.get_model(cfg=cfg, weights=weights, verbose=RANK in {-1, 0}) # calls Model(cfg, weights) return ckpt diff --git a/ultralytics/models/yolo/classify/train.py b/ultralytics/models/yolo/classify/train.py index 6985f2a2..34049416 100644 --- a/ultralytics/models/yolo/classify/train.py +++ b/ultralytics/models/yolo/classify/train.py @@ -84,6 +84,8 @@ class ClassificationTrainer(BaseTrainer): model.load(weights) for m in model.modules(): + if self.args.pretrained is False and hasattr(m, "reset_parameters"): + m.reset_parameters() if isinstance(m, torch.nn.Dropout) and self.args.dropout: m.p = self.args.dropout # set dropout for p in model.parameters():