expose interface of pin memory and modify cnn configuration (#75)

examples/benchmarks/pytorch_cnn.py

@@ -24,7 +24,9 @@
     # Specify the model name and benchmark parameters.
     # For example, resnet50, resnet101, resnet152, densenet169, densenet201, vgg11, vgg13, vgg16, vgg19.
     model_name = 'resnet101'
-    parameters = '--batch_size 32 --precision float32 float16 --num_warmup 64 --num_steps 2048 --sample_count 8192'
+    parameters = '--batch_size 192 --precision float32 float16 --num_warmup 64 --num_steps 512 \
+        --sample_count 8192 --pin_memory'
+
     if args.distributed:
         parameters += ' --distributed_impl ddp --distributed_backend nccl'
 

superbench/benchmarks/model_benchmarks/model_base.py

@@ -131,6 +131,13 @@ def add_parser_arguments(self):
             help='Disable GPU training.',
         )
 
+        self._parser.add_argument(
+            '--pin_memory',
+            action='store_true',
+            default=False,
+            help='Enable option to pin memory in data loader.',
+        )
+
     @abstractmethod
     def _judge_gpu_availability(self):
         """Judge GPUs' availability according to arguments and running environment."""

superbench/benchmarks/model_benchmarks/pytorch_base.py

@@ -115,7 +115,8 @@ def _init_dataloader(self):
             shuffle=False,
             num_workers=8,
             sampler=train_sampler,
-            drop_last=True
+            drop_last=True,
+            pin_memory=self._args.pin_memory
         )
 
         return True

tests/benchmarks/model_benchmarks/test_model_base.py

@@ -160,6 +160,7 @@ def test_arguments_related_interfaces():
   --distributed_backend DistributedBackend
                         Distributed backends. E.g. nccl mpi gloo.
   --no_gpu              Disable GPU training.
+  --pin_memory          Enable option to pin memory in data loader.
   --hidden_size int     Hidden size.
   --seq_len int         Sequence length."""
     )
@@ -192,6 +193,7 @@ def test_preprocess():
   --distributed_backend DistributedBackend
                         Distributed backends. E.g. nccl mpi gloo.
   --no_gpu              Disable GPU training.
+  --pin_memory          Enable option to pin memory in data loader.
   --hidden_size int     Hidden size.
   --seq_len int         Sequence length."""
     )