Merge branch 'master' into fix-deprecated-statement

Author: vfdev-5

.github/workflows/gpu-tests.yml

@@ -124,7 +124,7 @@ jobs:
         uses: nick-fields/[email protected]
         with:
           max_attempts: 5
-          timeout_minutes: 25
+          timeout_minutes: 45
           shell: bash
           command: docker exec -t pthd /bin/bash -xec 'bash tests/run_gpu_tests.sh 2'
           new_command_on_retry: docker exec -e USE_LAST_FAILED=1 -t pthd /bin/bash -xec 'bash tests/run_gpu_tests.sh 2'

ignite/contrib/engines/common.py

@@ -78,7 +78,7 @@ def setup_common_training_handlers(
         lr_scheduler: learning rate scheduler
             as native torch LRScheduler or ignite's parameter scheduler.
         with_gpu_stats: if True, :class:`~ignite.metrics.GpuInfo` is attached to the
-            trainer. This requires `pynvml` package to be installed.
+            trainer. This requires `pynvml<12` package to be installed.
         output_names: list of names associated with `update_function` output dictionary.
         with_pbars: if True, two progress bars on epochs and optionally on iterations are attached.
             Default, True.

ignite/engine/engine.py

@@ -140,6 +140,7 @@ def __init__(self, process_function: Callable[["Engine", Any], Any]):
         self._process_function = process_function
         self.last_event_name: Optional[Events] = None
         self.should_terminate = False
+        self.skip_completed_after_termination = False
         self.should_terminate_single_epoch = False
         self.should_interrupt = False
         self.state = State()
@@ -538,7 +539,7 @@ def call_interrupt():
         self.logger.info("interrupt signaled. Engine will interrupt the run after current iteration is finished.")
         self.should_interrupt = True
 
-    def terminate(self) -> None:
+    def terminate(self, skip_completed: bool = False) -> None:
         """Sends terminate signal to the engine, so that it terminates completely the run. The run is
         terminated after the event on which ``terminate`` method was called. The following events are triggered:
 
@@ -547,6 +548,9 @@ def terminate(self) -> None:
         - :attr:`~ignite.engine.events.Events.TERMINATE`
         - :attr:`~ignite.engine.events.Events.COMPLETED`
 
+        Args:
+            skip_completed: if True, the event :attr:`~ignite.engine.events.Events.COMPLETED` is not fired after
+                :attr:`~ignite.engine.events.Events.TERMINATE`. Default is False.
 
         Examples:
             .. testcode::
@@ -617,9 +621,12 @@ def terminate():
         .. versionchanged:: 0.4.10
             Behaviour changed, for details see https://github.com/pytorch/ignite/issues/2669
 
+        .. versionchanged:: 0.5.2
+            Added `skip_completed` flag
         """
         self.logger.info("Terminate signaled. Engine will stop after current iteration is finished.")
         self.should_terminate = True
+        self.skip_completed_after_termination = skip_completed
 
     def terminate_epoch(self) -> None:
         """Sends terminate signal to the engine, so that it terminates the current epoch. The run
@@ -993,13 +1000,17 @@ def _internal_run_as_gen(self) -> Generator[Any, None, State]:
             time_taken = time.time() - start_time
             # time is available for handlers but must be updated after fire
             self.state.times[Events.COMPLETED.name] = time_taken
-            handlers_start_time = time.time()
-            self._fire_event(Events.COMPLETED)
-            time_taken += time.time() - handlers_start_time
-            # update time wrt handlers
-            self.state.times[Events.COMPLETED.name] = time_taken
+
+            # do not fire Events.COMPLETED if we terminated the run with flag `skip_completed=True`
+            if not (self.should_terminate and self.skip_completed_after_termination):
+                handlers_start_time = time.time()
+                self._fire_event(Events.COMPLETED)
+                time_taken += time.time() - handlers_start_time
+                # update time wrt handlers
+                self.state.times[Events.COMPLETED.name] = time_taken
+
             hours, mins, secs = _to_hours_mins_secs(time_taken)
-            self.logger.info(f"Engine run complete. Time taken: {hours:02d}:{mins:02d}:{secs:06.3f}")
+            self.logger.info(f"Engine run finished. Time taken: {hours:02d}:{mins:02d}:{secs:06.3f}")
 
         except BaseException as e:
             self._dataloader_iter = None
@@ -1174,13 +1185,17 @@ def _internal_run_legacy(self) -> State:
             time_taken = time.time() - start_time
             # time is available for handlers but must be updated after fire
             self.state.times[Events.COMPLETED.name] = time_taken
-            handlers_start_time = time.time()
-            self._fire_event(Events.COMPLETED)
-            time_taken += time.time() - handlers_start_time
-            # update time wrt handlers
-            self.state.times[Events.COMPLETED.name] = time_taken
+
+            # do not fire Events.COMPLETED if we terminated the run with flag `skip_completed=True`
+            if not (self.should_terminate and self.skip_completed_after_termination):
+                handlers_start_time = time.time()
+                self._fire_event(Events.COMPLETED)
+                time_taken += time.time() - handlers_start_time
+                # update time wrt handlers
+                self.state.times[Events.COMPLETED.name] = time_taken
+
             hours, mins, secs = _to_hours_mins_secs(time_taken)
-            self.logger.info(f"Engine run complete. Time taken: {hours:02d}:{mins:02d}:{secs:06.3f}")
+            self.logger.info(f"Engine run finished. Time taken: {hours:02d}:{mins:02d}:{secs:06.3f}")
 
         except BaseException as e:
             self._dataloader_iter = None

ignite/engine/events.py

@@ -259,36 +259,47 @@ class Events(EventEnum):
     - TERMINATE_SINGLE_EPOCH : triggered when the run is about to end the current epoch,
       after receiving a :meth:`~ignite.engine.engine.Engine.terminate_epoch()` or
       :meth:`~ignite.engine.engine.Engine.terminate()` call.
+    - EPOCH_COMPLETED : triggered when the epoch is ended. Note that this is triggered even
+      when :meth:`~ignite.engine.engine.Engine.terminate_epoch()` is called.
 
     - TERMINATE : triggered when the run is about to end completely,
       after receiving :meth:`~ignite.engine.engine.Engine.terminate()` call.
 
-    - EPOCH_COMPLETED : triggered when the epoch is ended. Note that this is triggered even
-      when :meth:`~ignite.engine.engine.Engine.terminate_epoch()` is called.
-    - COMPLETED : triggered when engine's run is completed
+    - COMPLETED : triggered when engine's run is completed or terminated with
+      :meth:`~ignite.engine.engine.Engine.terminate()`, unless the flag
+      `skip_completed` is set to True.
 
     The table below illustrates which events are triggered when various termination methods are called.
 
     .. list-table::
-       :widths: 24 25 33 18
+       :widths: 35 38 28 20 20
        :header-rows: 1
 
        * - Method
-         - EVENT_COMPLETED
          - TERMINATE_SINGLE_EPOCH
+         - EPOCH_COMPLETED
          - TERMINATE
+         - COMPLETED
        * - no termination
-         - ✔
          - ✗
+         - ✔
          - ✗
+         - ✔
        * - :meth:`~ignite.engine.engine.Engine.terminate_epoch()`
          - ✔
          - ✔
          - ✗
+         - ✔
        * - :meth:`~ignite.engine.engine.Engine.terminate()`
          - ✗
          - ✔
          - ✔
+         - ✔
+       * - :meth:`~ignite.engine.engine.Engine.terminate()` with `skip_completed=True`
+         - ✗
+         - ✔
+         - ✔
+         - ✗
 
     Since v0.3.0, Events become more flexible and allow to pass an event filter to the Engine:
 
@@ -357,7 +368,7 @@ class CustomEvents(EventEnum):
     STARTED = "started"
     """triggered when engine's run is started."""
     COMPLETED = "completed"
-    """triggered when engine's run is completed"""
+    """triggered when engine's run is completed, or after receiving terminate() call."""
 
     ITERATION_STARTED = "iteration_started"
     """triggered when an iteration is started."""

ignite/metrics/clustering/calinski_harabasz_score.py

@@ -11,8 +11,8 @@
 def _calinski_harabasz_score(features: Tensor, labels: Tensor) -> float:
     from sklearn.metrics import calinski_harabasz_score
 
-    np_features = features.numpy()
-    np_labels = labels.numpy()
+    np_features = features.cpu().numpy()
+    np_labels = labels.cpu().numpy()
     score = calinski_harabasz_score(np_features, np_labels)
     return score
 

ignite/metrics/clustering/davies_bouldin_score.py

@@ -11,8 +11,8 @@
 def _davies_bouldin_score(features: Tensor, labels: Tensor) -> float:
     from sklearn.metrics import davies_bouldin_score
 
-    np_features = features.numpy()
-    np_labels = labels.numpy()
+    np_features = features.cpu().numpy()
+    np_labels = labels.cpu().numpy()
     score = davies_bouldin_score(np_features, np_labels)
     return score
 

ignite/metrics/clustering/silhouette_score.py

@@ -111,7 +111,7 @@ def __init__(
     def _silhouette_score(self, features: Tensor, labels: Tensor) -> float:
         from sklearn.metrics import silhouette_score
 
-        np_features = features.numpy()
-        np_labels = labels.numpy()
+        np_features = features.cpu().numpy()
+        np_labels = labels.cpu().numpy()
         score = silhouette_score(np_features, np_labels, **self._silhouette_kwargs)
         return score

ignite/metrics/gpu_info.py

@@ -10,7 +10,7 @@
 
 class GpuInfo(Metric):
     """Provides GPU information: a) used memory percentage, b) gpu utilization percentage values as Metric
-    on each iterations.
+    on each iterations. This metric requires `pynvml <https://pypi.org/project/pynvml/>`_ package of version `<12`.
 
     .. Note ::
 
@@ -39,7 +39,7 @@ def __init__(self) -> None:
         except ImportError:
             raise ModuleNotFoundError(
                 "This contrib module requires pynvml to be installed. "
-                "Please install it with command: \n pip install pynvml"
+                "Please install it with command: \n pip install 'pynvml<12'"
             )
             # Let's check available devices
         if not torch.cuda.is_available():

ignite/metrics/regression/kendall_correlation.py

@@ -16,8 +16,8 @@ def _get_kendall_tau(variant: str = "b") -> Callable[[Tensor, Tensor], float]:
         raise ValueError(f"variant accepts 'b' or 'c', got {variant!r}.")
 
     def _tau(predictions: Tensor, targets: Tensor) -> float:
-        np_preds = predictions.flatten().numpy()
-        np_targets = targets.flatten().numpy()
+        np_preds = predictions.flatten().cpu().numpy()
+        np_targets = targets.flatten().cpu().numpy()
         r = kendalltau(np_preds, np_targets, variant=variant).statistic
         return r
 

ignite/metrics/regression/spearman_correlation.py

@@ -12,8 +12,8 @@
 def _spearman_r(predictions: Tensor, targets: Tensor) -> float:
     from scipy.stats import spearmanr
 
-    np_preds = predictions.flatten().numpy()
-    np_targets = targets.flatten().numpy()
+    np_preds = predictions.flatten().cpu().numpy()
+    np_targets = targets.flatten().cpu().numpy()
     r = spearmanr(np_preds, np_targets).statistic
     return r