TS bug fix

Summary:
TS bug fix: numerical errors leads to deltaA to be non symmetric (order small <1e-8). When summing this it may result with a non-symmetric matrix A which leads to an error in the type check: when sampling from a Gaussian distribution the covariance matrix is required to be both PD and symmetric.

The offered solution is to explicitly symmetrize it, hence guarantee deltaA is symmetric.

Reviewed By: danielrjiang

Differential Revision: D56936024

fbshipit-source-id: 8219ca1e1a1577b0a241c0a2d7848ab02f427c2c

pearl/neural_networks/contextual_bandit/linear_regression.py

@@ -151,6 +151,7 @@ def learn_batch(
         x, y, weight = self._validate_train_inputs(x, y, weight)
 
         delta_A = torch.matmul(x.t(), x * weight)
+        delta_A = (delta_A + delta_A.t()) / 2  # symmetrize to avoid numerical errors
         delta_b = torch.matmul(x.t(), y * weight).squeeze(-1)
         delta_sum_weight = weight.sum()
 

test/unit/test_tutorials/test_cb_tutorial.py

@@ -36,8 +36,6 @@
 
 set_seed(0)
 
-device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-device_id = 0 if torch.cuda.is_available() else -1
 
 """
 This is a unit test version of the CB tutorial.
@@ -56,7 +54,7 @@ def setUp(self) -> None:
 
     def test_cb_tutorials(self) -> None:
         # load environment
-        device = -1
+        device_id = 0 if torch.cuda.is_available() else -1
 
         # Download UCI dataset if doesn't exist
         uci_data_path = "./utils/instantiations/environments/uci_datasets"
@@ -75,8 +73,8 @@ def test_cb_tutorials(self) -> None:
         env = SLCBEnvironment(**pendigits_uci_dict)  # pyre-ignore
 
         # experiment code
-        number_of_steps = 200
-        record_period = 400
+        number_of_steps = 300
+        record_period = 300
 
         """
         SquareCB
@@ -98,7 +96,7 @@ def test_cb_tutorials(self) -> None:
                 ),
             ),
             replay_buffer=FIFOOffPolicyReplayBuffer(100_000),
-            device_id=device,
+            device_id=device_id,
         )
 
         _ = online_learning(
@@ -126,7 +124,7 @@ def test_cb_tutorials(self) -> None:
                 exploration_module=UCBExploration(alpha=1.0),
             ),
             replay_buffer=FIFOOffPolicyReplayBuffer(100_000),
-            device_id=device,
+            device_id=device_id,
         )
 
         _ = online_learning(
@@ -155,7 +153,7 @@ def test_cb_tutorials(self) -> None:
                 exploration_module=ThompsonSamplingExplorationLinear(),
             ),
             replay_buffer=FIFOOffPolicyReplayBuffer(100_000),
-            device_id=-1,
+            device_id=device_id,
         )
 
         _ = online_learning(