Batch size + worker experiment

benchmarks/torchvision/main.py

@@ -100,6 +100,7 @@ def get_dali_pipeline():
             random_shuffle=True, 
             name="Reader",
         )
+
         # decode data on the GPU
         images = fn.decoders.image_random_crop(
             images, 
@@ -124,31 +125,37 @@ def get_dali_pipeline():
         reader_name='Reader'
     )
 
-    def iter():
-        for _ in range(args.epochs):
-            for data in train_data:
+    class Adapter:
+        def __init__(self, iter):
+            self.iter = iter
+
+        def __len__(self):
+            return len(self.iter)
+
+        def __iter__(self):
+            for data in self.iter:
                 x, y = data[0]['data'], data[0]['label']
                 yield x, torch.squeeze(y, dim=1).type(torch.LongTensor)
 
-    yield from iter()
+    return Adapter(train_data)
 
 
 def dataloader(args, model):
-    if args.loader == "dali":
-        return dali(args, args.data)
+    if not args.data:
+        data_directory = os.environ.get("MILABENCH_DIR_DATA", None)
+        if data_directory:
+            args.data = os.path.join(data_directory, "FakeImageNet")
 
     if args.fixed_batch:
         args.synthetic_data = True
 
     if args.synthetic_data:
         args.data = None
-    else:
-        if not args.data:
-            data_directory = os.environ.get("MILABENCH_DIR_DATA", None)
-            if data_directory:
-                args.data = os.path.join(data_directory, "FakeImageNet")
 
     if args.data:
+        if args.loader == "dali":
+            return dali(args, os.path.join(args.data, "train"))
+
         train = datasets.ImageFolder(os.path.join(args.data, "train"), data_transforms)
         return torch.utils.data.DataLoader(
             train,

config/matrix.yaml

@@ -39,3 +39,14 @@ resnet-batch:
       --batch-size: '{batch-size}'
       --num-workers: '{worker}'
       --loader: '{loader}'
+
+resnet-batch-noio:
+  matrix:
+    batch-size: [32, 64, 128, 256, 512, 1024]
+
+  job:
+    name: 'resnet50-noio-bs{batch-size}'
+    inherits: _resnet50
+    argv:
+      --batch-size: '{batch-size}'
+      --synthetic-data: true

scripts/article/run_cuda.sh

@@ -41,6 +41,11 @@ install_prepare() {
         pip install -e $MILABENCH_WORDIR/voir
         pip install -e $MILABENCH_WORDIR/torchcompat
         pip install torch torchvision torchaudio
+
+        # DALI stuff
+        pip install --extra-index-url https://pypi.nvidia.com --upgrade nvidia-dali-cuda120
+        pip install nvidia-pyindex
+        pip install nvidia-nvjpeg-cu12
     )
 
     #

scripts/batch/run_cuda.sh

@@ -0,0 +1,14 @@
+
+
+
+export MILABENCH_GPU_ARCH=cuda
+export MILABENCH_WORDIR="$(pwd)/$MILABENCH_GPU_ARCH"
+export MILABENCH_CONFIG="$MILABENCH_WORDIR/milabench/config/matrix.yaml"
+
+CUDA_VISIBLE_DEVICES=0 bash $MILABENCH_WORDIR/milabench/scripts/article/run_cuda.sh --config $MILABENCH_CONFIG --select resnet
+
+CUDA_VISIBLE_DEVICES=0,1 bash $MILABENCH_WORDIR/milabench/scripts/article/run_cuda.sh --config $MILABENCH_CONFIG --select resnet
+
+CUDA_VISIBLE_DEVICES=0,1,2,3 bash $MILABENCH_WORDIR/milabench/scripts/article/run_cuda.sh --config $MILABENCH_CONFIG --select resnet
+
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash $MILABENCH_WORDIR/milabench/scripts/article/run_cuda.sh --config $MILABENCH_CONFIG --select resnet