From b5c041cc909fd9f171ea8e61c7aa5039734452dc Mon Sep 17 00:00:00 2001
From: smisra <sanchit.misra@intel.com>
Date: Wed, 21 Aug 2024 04:57:33 -0700
Subject: [PATCH] Added support for PyTorch Profiler

---
 tutorials/examples/train_hypergrid.py | 30 +++++++++++++++++++++++++++
 1 file changed, 30 insertions(+)

diff --git a/tutorials/examples/train_hypergrid.py b/tutorials/examples/train_hypergrid.py
index b585a9c7..2e66f93f 100644
--- a/tutorials/examples/train_hypergrid.py
+++ b/tutorials/examples/train_hypergrid.py
@@ -30,6 +30,7 @@
 from gfn.utils.common import set_seed
 from gfn.utils.modules import DiscreteUniform, NeuralNet, Tabular
 from gfn.utils.training import validate
+from torch.profiler import profile, ProfilerActivity
 
 DEFAULT_SEED = 4444
 
@@ -237,7 +238,20 @@ def main(args):  # noqa: C901
     discovered_modes = set()
     is_on_policy = args.replay_buffer_size == 0
 
+    if (args.profile):
+        keep_active = args.trajectories_to_profile//args.batch_size
+        prof = profile(
+            schedule=torch.profiler.schedule(wait=1, warmup=1, active=keep_active, repeat=1),
+            activities=[ProfilerActivity.CPU],
+            record_shapes=True,
+            with_stack=True
+            )
+        prof.start()
     for iteration in trange(n_iterations):
+        if (args.profile):
+            prof.step()
+            if iteration >= 1 + 1 + keep_active:
+                break
         trajectories = gflownet.sample_trajectories(
             env,
             n_samples=args.batch_size,
@@ -279,6 +293,10 @@ def main(args):  # noqa: C901
             to_log.update(validation_info)
             tqdm.write(f"{iteration}: {to_log}")
 
+    if (args.profile):
+        prof.stop()
+        print(prof.key_averages().table(sort_by="cpu_time_total", row_limit=20))
+        prof.export_chrome_trace("trace.json")
     try:
         return validation_info["l1_dist"]
     except KeyError:
@@ -456,6 +474,18 @@ def validate_hypergrid(
         action="store_true",
         help="Calculates the true partition function.",
     )
+    parser.add_argument(
+        "--profile",
+        action="store_true",
+        help="Profiles the execution using PyTorch Profiler.",
+    )
+    parser.add_argument(
+        "--trajectories_to_profile",
+        type=int,
+        default=2048,
+        help="Number of trajectories to profile using the Pytorch Profiler." +
+        " Preferably, a multiple of batch size.",
+    )
 
     args = parser.parse_args()