Adding a hook to profile memory usage (#175)

Summary:
Pull Request resolved: #175

recently nvidia-smi was removed from fb cluster making it hard to see what the memory utilization was. using pytorch, we extract the information at various steps of training. the profiling hook is currently experimental and we will use it and adapt it for more better usability. down the line, we can move it to classy vision when the hook is trusted to be useful + changed accordingly

Reviewed By: min-xu-ai

Differential Revision: D26284304

fbshipit-source-id: 1a8c3cd12a498fc55999e982a6c072723d54d144

vissl/config/defaults.yaml

@@ -83,6 +83,21 @@ config:
   # valid for some systems.
   LOG_GPU_STATS: True
 
+  # ----------------------------------------------------------------------------------- #
+  # HOOKS
+  # ----------------------------------------------------------------------------------- #
+  HOOKS:
+    # ----------------------------------------------------------------------------------- #
+    # torch.cuda.memory_summary()
+    # ----------------------------------------------------------------------------------- #
+    MEMORY_SUMMARY:
+      # set this to true if you want to print memory summary. useful for profiling
+      # memory consumption of model
+      PRINT_MEMORY_SUMMARY: False
+      # at what iteration number should the memory summary be printed. usually
+      # set to 1 for very large models
+      LOG_ITERATION_NUM: 0
+
   # ----------------------------------------------------------------------------------- #
   # DATA
   # ----------------------------------------------------------------------------------- #

vissl/hooks/__init__.py

@@ -7,6 +7,7 @@
 from vissl.hooks.deepclusterv2_hooks import ClusterMemoryHook, InitMemoryHook  # noqa
 from vissl.hooks.log_hooks import (  # noqa
     LogGpuStatsHook,
+    LogGpuMemoryHook,
     LogLossLrEtaHook,
     LogLossMetricsCheckpointHook,
     LogPerfTimeMetricsHook,
@@ -100,6 +101,8 @@ def default_hook_generator(cfg: AttrDict) -> List[ClassyHook]:
         hooks.extend([SSLModelComplexityHook()])
     if cfg.LOG_GPU_STATS:
         hooks.extend([LogGpuStatsHook()])
+    if cfg.HOOKS.MEMORY_SUMMARY.PRINT_MEMORY_SUMMARY:
+        hooks.extend([LogGpuMemoryHook(cfg.HOOKS.MEMORY_SUMMARY.LOG_ITERATION_NUM)])
     if cfg.TENSORBOARD_SETUP.USE_TENSORBOARD:
         assert is_tensorboard_available(), "Tensorboard must be installed to use it."
         tb_hook = get_tensorboard_hook(cfg)

vissl/hooks/log_hooks.py

@@ -22,6 +22,61 @@
 from vissl.utils.perf_stats import PerfStats
 
 
+class LogGpuMemoryHook(ClassyHook):
+    """
+    Hook executed at a specified iteration number and prints the
+    memory summary for the primary device at several steps of training.
+    """
+
+    on_start = ClassyHook._noop
+    on_loss_and_meter = ClassyHook._noop
+    on_step = ClassyHook._noop
+    on_phase_end = ClassyHook._noop
+    on_end = ClassyHook._noop
+
+    def __init__(
+        self,
+        log_iteration_num: int = 1,
+    ) -> None:
+        super().__init__()
+        self.log_iteration_num = log_iteration_num
+
+    def on_phase_start(self, task: "tasks.ClassyTask") -> None:
+        """
+        Print the stats just before the training epoch starts
+        """
+        self._print_memory_summary(task, "on_phase_start")
+
+    def on_forward(self, task: "tasks.ClassyTask") -> None:
+        """
+        Print the stats after the model forward pass is done
+        """
+        self._print_memory_summary(task, "on_forward")
+
+    def on_backward(self, task: "tasks.ClassyTask") -> None:
+        """
+        Print the stats just after model.backward() is done
+        """
+        self._print_memory_summary(task, "on_backward")
+
+    def on_update(self, task: "tasks.ClassyTask") -> None:
+        """
+        Print the stats just after model params are updated
+        """
+        self._print_memory_summary(task, "on_update")
+
+    def _print_memory_summary(self, task: "tasks.ClassyTask", stage_name: str) -> None:
+        if (
+            is_primary()
+            and (task.device.type == "cuda")
+            and task.local_iteration_num == self.log_iteration_num
+        ):
+            logging.info(
+                f"========= Memory Summary at {stage_name} ======="
+                f"\n{torch.cuda.memory_summary()}\n"
+            )
+
+
 class LogGpuStatsHook(ClassyHook):
     """
     Hook executed at the start of training and after every training iteration is done.
@@ -92,8 +147,8 @@ def on_update(self, task: "tasks.ClassyTask") -> None:
         monitoring the stats (optionally) for every N iterations to get better
         idea about the batch time and training eta.
 
-        Set the btime_freq input using cfg.PERF_STAT_FREQUENCY=N ensuring that
-        cfg.MONITOR_PERF_STATS = True.
+        Set the btime_freq input using cfg.HOOKS.PERF_STATS.PERF_STAT_FREQUENCY=N
+        ensuring that cfg.HOOKS.PERF_STATS.MONITOR_PERF_STATS = True.
         """
         phase_type = "train" if task.train else "test"
         if is_primary() and phase_type == "train":