Add _get_checkpoint_mapping function (#831)

Summary: Pull Request resolved: #831

Reviewed By: JKSenthil

Differential Revision: D57118392

fbshipit-source-id: 319ff905786a181ff302ecb91a61af76561fc942

Author: diego-urgell

tests/framework/callbacks/test_checkpoint_utils.py

@@ -8,11 +8,21 @@
 
 import unittest
 
-from torchtnt.framework._test_utils import DummyTrainUnit, get_dummy_train_state
+from torch import nn
+
+from torchtnt.framework._test_utils import (
+    DummyAutoUnit,
+    DummyTrainUnit,
+    get_dummy_eval_state,
+    get_dummy_fit_state,
+    get_dummy_train_state,
+)
 
 from torchtnt.framework.callbacks._checkpoint_utils import (
+    _get_step_phase_mapping,
     _prepare_app_state_for_checkpoint,
 )
+from torchtnt.utils.checkpoint import Phase
 
 
 class CheckpointUtilsTest(unittest.TestCase):
@@ -26,3 +36,20 @@ def test_get_app_state(self) -> None:
             app_state.keys(),
             ["module", "optimizer", "loss_fn", "train_progress"],
         )
+
+    def test_get_step_phase_mapping(self) -> None:
+        unit = DummyAutoUnit(module=nn.Linear(2, 2))
+        unit.train_progress._num_steps_completed = 5
+        unit.eval_progress._num_steps_completed = 7
+
+        fit_state = get_dummy_fit_state()
+        self.assertEqual(
+            {Phase.TRAIN: 5, Phase.EVALUATE: 7},
+            _get_step_phase_mapping(fit_state, unit),
+        )
+
+        train_state = get_dummy_train_state()
+        self.assertEqual({Phase.TRAIN: 5}, _get_step_phase_mapping(train_state, unit))
+
+        eval_state = get_dummy_eval_state()
+        self.assertEqual({Phase.EVALUATE: 7}, _get_step_phase_mapping(eval_state, unit))

torchtnt/framework/_test_utils.py

@@ -33,6 +33,19 @@ def get_dummy_train_state(dataloader: Optional[Iterable[object]] = None) -> Stat
     )
 
 
+def get_dummy_eval_state(dataloader: Optional[Iterable[object]] = None) -> State:
+    return State(
+        entry_point=EntryPoint.EVALUATE,
+        eval_state=PhaseState(
+            dataloader=dataloader or [1, 2, 3, 4],
+            max_epochs=1,
+            max_steps=1,
+            max_steps_per_epoch=1,
+        ),
+        timer=None,
+    )
+
+
 def get_dummy_fit_state() -> State:
     return State(
         entry_point=EntryPoint.FIT,

torchtnt/framework/callbacks/_checkpoint_utils.py

@@ -7,12 +7,13 @@
 # pyre-strict
 
 
-from typing import Any, Dict
+from typing import Any, cast, Dict, Union
 
 from pyre_extensions import none_throws
 from torchtnt.framework.callbacks.checkpointer_types import RestoreOptions
-from torchtnt.framework.state import State
-from torchtnt.framework.unit import AppStateMixin
+from torchtnt.framework.state import EntryPoint, State
+from torchtnt.framework.unit import AppStateMixin, TEvalUnit, TTrainUnit
+from torchtnt.utils.checkpoint import Phase
 
 from torchtnt.utils.stateful import Stateful
 
@@ -23,6 +24,26 @@
 _EVAL_PROGRESS_STATE_KEY = "eval_progress"
 
 
+def _get_step_phase_mapping(
+    state: State, unit: Union[TTrainUnit, TEvalUnit]
+) -> Dict[Phase, int]:
+    """
+    Returns a mapping of phase to step, depending on the entrypoint.
+    For FIT, it always includes train and eval progress.
+    """
+    step_mapping = {}
+
+    if state.entry_point in (EntryPoint.TRAIN, EntryPoint.FIT):
+        train_unit = cast(TTrainUnit, unit)
+        step_mapping[Phase.TRAIN] = train_unit.train_progress.num_steps_completed
+
+    if state.entry_point in (EntryPoint.EVALUATE, EntryPoint.FIT):
+        eval_unit = cast(TEvalUnit, unit)
+        step_mapping[Phase.EVALUATE] = eval_unit.eval_progress.num_steps_completed
+
+    return step_mapping
+
+
 def _prepare_app_state(unit: AppStateMixin) -> Dict[str, Any]:
     """Join together all of the tracked stateful entities to simplify registration of snapshottable states, deals with FSDP case"""
     app_state = unit.app_state()