the accuracy test always returns none for the output

Summary:
As titled, we should fix it to have the actual output return and compared.

Before the fix, test will always pass because it is comparing None with None

After the fix, it can compare the actual outputs.

we detected output mismatch in cmf_10x, cmf_10x_batch_fusion and ads_dhen_5x and alexnet now (see v2), which we will disable and need investigation ore replacing with up-to-date models.

Reviewed By: xuzhao9, jackiexu1992

Differential Revision: D49077949

fbshipit-source-id: 2952bf007e522bf8cacc1415b92349a82d1fab9b

torchbenchmark/util/env_check.py

@@ -367,9 +367,9 @@ def forward_and_backward_pass(mod, inputs, contexts, optimizer, collect_outputs=
 def run_n_iterations(mod, inputs, contexts, optimizer=None, is_training=False, iterations=STABLENESS_CHECK_ROUNDS):
     def _model_iter_fn(mod, inputs, contexts, optimizer, collect_outputs):
         if is_training:
-            forward_and_backward_pass(mod, inputs, contexts, optimizer, collect_outputs)
+            return forward_and_backward_pass(mod, inputs, contexts, optimizer, collect_outputs)
         else:
-            forward_pass(mod, inputs, contexts, collect_outputs)
+            return forward_pass(mod, inputs, contexts, collect_outputs)
     for _ in range(iterations - 1):
         _model_iter_fn(mod, inputs, contexts, optimizer, collect_outputs=False)
     return _model_iter_fn(mod, inputs, contexts, optimizer, collect_outputs=True)