track dvcyaml file

.github/workflows/tests.yml

@@ -92,4 +92,4 @@ jobs:
           pip install -e '.[tests]'
 
       - name: Run tests
-        run: pytest -v tests --ignore=tests/test_frameworks
+        run: pytest -v tests --ignore=tests/frameworks

.gitignore

@@ -137,4 +137,6 @@ dmypy.json
 # Cython debug symbols
 cython_debug/
 
+.dvc/
+.dvcignore
 src/dvclive/_dvclive_version.py

README.md

@@ -54,7 +54,7 @@ from dvclive import Live
 
 params = {"learning_rate": 0.002, "optimizer": "Adam", "epochs": 20}
 
-with Live(save_dvc_exp=True) as live:
+with Live() as live:
 
     # log a parameters
     for param in params:

examples/DVCLive-HuggingFace.ipynb

@@ -229,7 +229,7 @@
     "        train_dataset=small_train_dataset,\n",
     "        eval_dataset=small_eval_dataset,\n",
     "        compute_metrics=compute_metrics,\n",
-    "        callbacks=[DVCLiveCallback(report=\"notebook\", save_dvc_exp=True, log_model=True)],\n",
+    "        callbacks=[DVCLiveCallback(log_model=True, report=\"notebook\")]\n",
     "    )\n",
     "    trainer.train()"
    ]

examples/DVCLive-PyTorch-Lightning.ipynb

@@ -173,7 +173,7 @@
         "        limit_train_batches=200,\n",
         "        limit_val_batches=100,\n",
         "        max_epochs=5,\n",
-        "        logger=DVCLiveLogger(save_dvc_exp=True, report=\"notebook\", log_model=True),\n",
+        "        logger=DVCLiveLogger(log_model=True, report=\"notebook\"),\n",
         "    )\n",
         "    trainer.fit(model, train_loader, validation_loader)\n"
       ]

examples/DVCLive-Quickstart.ipynb

@@ -220,7 +220,7 @@
     "\n",
     "best_test_acc = 0\n",
     "\n",
-    "with Live(save_dvc_exp=True, report=\"notebook\") as live:\n",
+    "with Live(report=\"notebook\") as live:\n",
     "\n",
     "    live.log_params(params)\n",
     "\n",

examples/DVCLive-scikit-learn.ipynb

@@ -96,7 +96,7 @@
     "\n",
     "for n_estimators in (10, 50, 100):\n",
     "\n",
-    "  with Live(report=None, save_dvc_exp=True) as live:\n",
+    "  with Live() as live:\n",
     "\n",
     "    live.log_param(\"n_estimators\", n_estimators)\n",
     "\n",

pyproject.toml

@@ -31,7 +31,7 @@ classifiers = [
 ]
 dynamic = ["version"]
 dependencies = [
-    "dvc>=3.17.0",
+    "dvc>=3.20.0",
     "dvc-render>=0.5.0,<1.0",
     "dvc-studio-client>=0.15.0,<1",
     "funcy",

src/dvclive/catalyst.py

@@ -1,16 +1,14 @@
 # ruff: noqa: ARG002
 from typing import Optional
 
-from catalyst import utils
 from catalyst.core.callback import Callback, CallbackOrder
 
 from dvclive import Live
 
 
 class DVCLiveCallback(Callback):
-    def __init__(self, model_file=None, live: Optional[Live] = None, **kwargs):
+    def __init__(self, live: Optional[Live] = None, **kwargs):
         super().__init__(order=CallbackOrder.external)
-        self.model_file = model_file
         self.live = live if live is not None else Live(**kwargs)
 
     def on_epoch_end(self, runner) -> None:
@@ -19,15 +17,6 @@ def on_epoch_end(self, runner) -> None:
                 self.live.log_metric(
                     f"{loader_key}/{key.replace('/', '_')}", float(value)
                 )
-
-        if self.model_file:
-            checkpoint = utils.pack_checkpoint(
-                model=runner.model,
-                criterion=runner.criterion,
-                optimizer=runner.optimizer,
-                scheduler=runner.scheduler,
-            )
-            utils.save_checkpoint(checkpoint, self.model_file)
         self.live.next_step()
 
     def on_experiment_end(self, runner):

src/dvclive/dvc.py

@@ -6,7 +6,7 @@
 
 from dvclive.plots import Image, Metric
 from dvclive.serialize import dump_yaml
-from dvclive.utils import StrPath
+from dvclive.utils import StrPath, rel_path
 
 if TYPE_CHECKING:
     from dvc.repo import Repo
@@ -51,38 +51,78 @@ def get_dvc_repo() -> Optional["Repo"]:
             return None
 
 
-def make_dvcyaml(live) -> None:
+def make_dvcyaml(live) -> None:  # noqa: C901
+    dvcyaml_dir = Path(live.dvc_file).parent.absolute().as_posix()
+
     dvcyaml = {}
     if live._params:
-        dvcyaml["params"] = [os.path.relpath(live.params_file, live.dir)]
+        dvcyaml["params"] = [rel_path(live.params_file, dvcyaml_dir)]
     if live._metrics or live.summary:
-        dvcyaml["metrics"] = [os.path.relpath(live.metrics_file, live.dir)]
+        dvcyaml["metrics"] = [rel_path(live.metrics_file, dvcyaml_dir)]
     plots: List[Any] = []
     plots_path = Path(live.plots_dir)
-    metrics_path = plots_path / Metric.subfolder
-    if metrics_path.exists():
-        metrics_relpath = metrics_path.relative_to(live.dir).as_posix()
-        metrics_config = {metrics_relpath: {"x": "step"}}
+    plots_metrics_path = plots_path / Metric.subfolder
+    if plots_metrics_path.exists():
+        metrics_config = {rel_path(plots_metrics_path, dvcyaml_dir): {"x": "step"}}
         plots.append(metrics_config)
     if live._images:
-        images_path = (plots_path / Image.subfolder).relative_to(live.dir)
-        plots.append(images_path.as_posix())
+        images_path = rel_path(plots_path / Image.subfolder, dvcyaml_dir)
+        plots.append(images_path)
     if live._plots:
         for plot in live._plots.values():
-            plot_path = plot.output_path.relative_to(live.dir)
-            plots.append({plot_path.as_posix(): plot.plot_config})
+            plot_path = rel_path(plot.output_path, dvcyaml_dir)
+            plots.append({plot_path: plot.plot_config})
     if plots:
         dvcyaml["plots"] = plots
 
     if live._artifacts:
         dvcyaml["artifacts"] = copy.deepcopy(live._artifacts)
         for artifact in dvcyaml["artifacts"].values():  # type: ignore
-            abs_path = os.path.abspath(artifact["path"])
-            abs_dir = os.path.realpath(live.dir)
-            relative_path = os.path.relpath(abs_path, abs_dir)
-            artifact["path"] = Path(relative_path).as_posix()
-
-    dump_yaml(dvcyaml, live.dvc_file)
+            artifact["path"] = rel_path(artifact["path"], dvcyaml_dir)
+
+    if not os.path.exists(live.dvc_file):
+        dump_yaml(dvcyaml, live.dvc_file)
+    else:
+        update_dvcyaml(live, dvcyaml)
+
+
+def update_dvcyaml(live, updates):  # noqa: C901
+    from dvc.utils.serialize import modify_yaml
+
+    dvcyaml_dir = os.path.abspath(os.path.dirname(live.dvc_file))
+    dvclive_dir = os.path.relpath(live.dir, dvcyaml_dir) + "/"
+
+    def _drop_stale_dvclive_entries(entries):
+        non_dvclive = []
+        for e in entries:
+            if isinstance(e, str):
+                if dvclive_dir not in e:
+                    non_dvclive.append(e)
+            elif isinstance(e, dict) and len(e) == 1:
+                if dvclive_dir not in next(iter(e.keys())):
+                    non_dvclive.append(e)
+            else:
+                non_dvclive.append(e)
+        return non_dvclive
+
+    def _update_entries(old, new, key):
+        keepers = _drop_stale_dvclive_entries(old.get(key, []))
+        old[key] = keepers + new.get(key, [])
+        if not old[key]:
+            del old[key]
+        return old
+
+    with modify_yaml(live.dvc_file) as orig:
+        orig = _update_entries(orig, updates, "params")  # noqa: PLW2901
+        orig = _update_entries(orig, updates, "metrics")  # noqa: PLW2901
+        orig = _update_entries(orig, updates, "plots")  # noqa: PLW2901
+        old_artifacts = {}
+        for name, meta in orig.get("artifacts", {}).items():
+            if dvclive_dir not in meta.get("path", dvclive_dir):
+                old_artifacts[name] = meta
+        orig["artifacts"] = {**old_artifacts, **updates.get("artifacts", {})}
+        if not orig["artifacts"]:
+            del orig["artifacts"]
 
 
 def get_random_exp_name(scm, baseline_rev) -> str:

src/dvclive/error.py

@@ -12,6 +12,11 @@ def __init__(self, name, val):
         super().__init__(f"Data '{name}' has not supported type {val}")
 
 
+class InvalidDvcyamlError(DvcLiveError):
+    def __init__(self):
+        super().__init__("`dvcyaml` path must have filename 'dvc.yaml'")
+
+
 class InvalidPlotTypeError(DvcLiveError):
     def __init__(self, name):
         from .plots import SKLEARN_PLOTS

src/dvclive/fastai.py

@@ -27,13 +27,11 @@ def _inside_fine_tune():
 class DVCLiveCallback(Callback):
     def __init__(
         self,
-        model_file: Optional[str] = None,
         with_opt: bool = False,
         live: Optional[Live] = None,
         **kwargs,
     ):
         super().__init__()
-        self.model_file = model_file
         self.with_opt = with_opt
         self.live = live if live is not None else Live(**kwargs)
         self.freeze_stage_ended = False
@@ -66,9 +64,6 @@ def after_epoch(self):
         # When resuming (i.e. passing `start_epoch` to learner)
         # fast.ai calls after_epoch but we don't want to increase the step.
         if logged_metrics:
-            if self.model_file:
-                file = self.learn.save(self.model_file, with_opt=self.with_opt)
-                self.live.log_artifact(str(file))
             self.live.next_step()
 
     def after_fit(self):

src/dvclive/huggingface.py

@@ -26,12 +26,6 @@ def __init__(
     ):
         super().__init__()
         self._log_model = log_model
-        self.model_file = kwargs.pop("model_file", None)
-        if self.model_file:
-            logger.warning(
-                "model_file is deprecated and will be removed"
-                " in the next major version, use log_model instead"
-            )
         self.live = live if live is not None else Live(**kwargs)
 
     def on_train_begin(
@@ -65,21 +59,6 @@ def on_save(
         if self._log_model == "all" and state.is_world_process_zero:
             self.live.log_artifact(args.output_dir)
 
-    def on_epoch_end(
-        self,
-        args: TrainingArguments,
-        state: TrainerState,
-        control: TrainerControl,
-        **kwargs,
-    ):
-        if self.model_file:
-            model = kwargs["model"]
-            model.save_pretrained(self.model_file)
-            tokenizer = kwargs.get("tokenizer")
-            if tokenizer:
-                tokenizer.save_pretrained(self.model_file)
-            self.live.log_artifact(self.model_file)
-
     def on_train_end(
         self,
         args: TrainingArguments,

src/dvclive/keras.py

@@ -1,5 +1,4 @@
 # ruff: noqa: ARG002
-import os
 from typing import Dict, Optional
 
 import tensorflow as tf
@@ -11,37 +10,18 @@
 class DVCLiveCallback(tf.keras.callbacks.Callback):
     def __init__(
         self,
-        model_file=None,
         save_weights_only: bool = False,
         live: Optional[Live] = None,
         **kwargs,
     ):
         super().__init__()
-        self.model_file = model_file
         self.save_weights_only = save_weights_only
         self.live = live if live is not None else Live(**kwargs)
 
-    def on_train_begin(self, logs=None):
-        if (
-            self.live._resume  # noqa: SLF001
-            and self.model_file is not None
-            and os.path.exists(self.model_file)
-        ):
-            if self.save_weights_only:
-                self.model.load_weights(self.model_file)
-            else:
-                self.model = tf.keras.models.load_model(self.model_file)
-
     def on_epoch_end(self, epoch: int, logs: Optional[Dict] = None):
         logs = logs or {}
         for metric, value in logs.items():
             self.live.log_metric(standardize_metric_name(metric, __name__), value)
-        if self.model_file:
-            if self.save_weights_only:
-                self.model.save_weights(self.model_file)
-            else:
-                self.model.save(self.model_file)
-            self.live.log_artifact(self.model_file)
         self.live.next_step()
 
     def on_train_end(self, logs: Optional[Dict] = None):

src/dvclive/lgbm.py

@@ -4,9 +4,8 @@
 
 
 class DVCLiveCallback:
-    def __init__(self, model_file=None, live: Optional[Live] = None, **kwargs):
+    def __init__(self, live: Optional[Live] = None, **kwargs):
         super().__init__()
-        self.model_file = model_file
         self.live = live if live is not None else Live(**kwargs)
 
     def __call__(self, env):
@@ -16,7 +15,4 @@ def __call__(self, env):
             self.live.log_metric(
                 f"{data_name}/{eval_name}" if multi_eval else eval_name, result
             )
-
-        if self.model_file:
-            env.model.save_model(self.model_file)
         self.live.next_step()

src/dvclive/lightning.py

@@ -58,7 +58,7 @@ def __init__(  # noqa: PLR0913
         experiment=None,
         dir: Optional[str] = None,  # noqa: A002
         resume: bool = False,
-        report: Optional[str] = "auto",
+        report: Optional[str] = None,
         save_dvc_exp: bool = False,
         dvcyaml: bool = True,
         cache_images: bool = False,