Add multi_gpu test for ghost clipping (#665)

Summary:
Pull Request resolved: #665

Modify the existing `multigpu_gradcheck.py` test to check gradient correctness for ghost clipping in a distributed setting.

Differential Revision: D60840755

Author: iden-kalemaj

Diff for: opacus/tests/multigpu_gradcheck.py

@@ -13,6 +13,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import itertools
 import os
 import sys
 import unittest
@@ -29,6 +30,10 @@
     SimpleDistributedPerLayerOptimizer,
 )
 from opacus.optimizers.ddpoptimizer import DistributedDPOptimizer
+from opacus.optimizers.ddpoptimizer_fast_gradient_clipping import (
+    DistributedDPOptimizerFastGradientClipping,
+)
+from opacus.utils.fast_gradient_clipping_utils import double_backward
 from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.utils.data import DataLoader, TensorDataset
 from torch.utils.data.distributed import DistributedSampler
@@ -84,8 +89,10 @@ def demo_basic(rank, weight, world_size, dp, clipping, grad_sample_mode):
 
     dataset = TensorDataset(data, labels)
 
-    loss_fn = nn.MSELoss()
-    if dp and clipping == "flat":
+    reduction = "none" if dp and clipping == "ghost" else "mean"
+    loss_fn = nn.CrossEntropyLoss(reduction=reduction)
+
+    if dp and clipping in ["flat", "ghost"]:
         ddp_model = DPDDP(model)
     else:
         ddp_model = DDP(model, device_ids=[rank])
@@ -115,15 +122,24 @@ def demo_basic(rank, weight, world_size, dp, clipping, grad_sample_mode):
                 optimizer,
                 (DistributedPerLayerOptimizer, SimpleDistributedPerLayerOptimizer),
             )
+        elif clipping == "ghost":
+            assert isinstance(optimizer, DistributedDPOptimizerFastGradientClipping)
         else:
             assert isinstance(optimizer, DistributedDPOptimizer)
 
     for x, y in data_loader:
-        outputs = ddp_model(x.to(rank))
-        loss = loss_fn(outputs, y)
-        optimizer.zero_grad()
-        loss.backward()
-        optimizer.step()
+        if dp and clipping == "ghost":
+            ddp_model.enable_hooks()
+            outputs = ddp_model(x.to(rank))
+            loss_per_sample = loss_fn(outputs, y)
+            double_backward(ddp_model, optimizer, loss_per_sample)
+            optimizer.step()
+        else:
+            outputs = ddp_model(x.to(rank))
+            loss = loss_fn(outputs, y)
+            optimizer.zero_grad()
+            loss.backward()
+            optimizer.step()
         break
 
     weight.copy_(model.net1.weight.data.cpu())
@@ -141,33 +157,38 @@ def run_demo(demo_fn, weight, world_size, dp, clipping, grad_sample_mode):
 
 class GradientComputationTest(unittest.TestCase):
     def test_gradient_correct(self) -> None:
-        # Tests that gradient is the same with DP or with DDP
+        # Tests that gradient is the same with DP or without DDP
         n_gpus = torch.cuda.device_count()
         self.assertTrue(
             n_gpus >= 2, f"Need at least 2 gpus but was provided only {n_gpus}."
         )
 
-        for clipping in ["flat", "per_layer"]:
-            for grad_sample_mode in ["hooks", "ew"]:
-                weight_dp, weight_nodp = torch.zeros(10, 10), torch.zeros(10, 10)
-
-                run_demo(
-                    demo_basic,
-                    weight_dp,
-                    2,
-                    dp=True,
-                    clipping=clipping,
-                    grad_sample_mode=grad_sample_mode,
-                )
-                run_demo(
-                    demo_basic,
-                    weight_nodp,
-                    2,
-                    dp=False,
-                    clipping=None,
-                    grad_sample_mode=None,
-                )
-
-                self.assertTrue(
-                    torch.allclose(weight_dp, weight_nodp, atol=1e-5, rtol=1e-3)
-                )
+        clipping_grad_sample_pairs = list(
+            itertools.product(["flat", "per_layer"], ["hooks", "ew"])
+        )
+        clipping_grad_sample_pairs.append(("ghost", "ghost"))
+
+        for clipping, grad_sample_mode in clipping_grad_sample_pairs:
+
+            weight_dp, weight_nodp = torch.zeros(10, 10), torch.zeros(10, 10)
+
+            run_demo(
+                demo_basic,
+                weight_dp,
+                2,
+                dp=True,
+                clipping=clipping,
+                grad_sample_mode=grad_sample_mode,
+            )
+            run_demo(
+                demo_basic,
+                weight_nodp,
+                2,
+                dp=False,
+                clipping=None,
+                grad_sample_mode=None,
+            )
+
+            self.assertTrue(
+                torch.allclose(weight_dp, weight_nodp, atol=1e-5, rtol=1e-3)
+            )