Reduce reliance on env var, use args placeholder instead

benchmarks/huggingface/prepare.py

@@ -8,3 +8,8 @@
     print(f"Preparing {args.model}")
     make_config = models[args.model]
     make_config()
+
+    # bert dataset
+    # t5 dataset
+    # reformer dataset
+    # whisper dataset

benchmarks/super-slomo/prepare.py

@@ -1,16 +1,9 @@
 #!/usr/bin/env python
 
 import torchvision
-
-
-
-def download_celebA():
-    # celebA use Google drive, and google drive wants to tell us that 
-    # they cant scan for virus so the download fails
-    # torchvision 0.17.1 might solve this issue though but we dont have it
-    pass
-
+from benchmate.datagen import generate_fakeimagenet
 
 if __name__ == "__main__":
     # This will download the weights for vgg16
+    generate_fakeimagenet()
     torchvision.models.vgg16(pretrained=True)

benchmarks/super-slomo/slomo/train.py

@@ -10,9 +10,11 @@
 import torchcompat.core as accelerator
 
 import model
-from giving import give
 import voir.wrapper
+import torchvision.transforms as transforms
+
 from synth import SyntheticData
+import dataloader
 
 
 def main():
@@ -75,6 +77,12 @@ def main():
         action="store_false",
         help="do not allow tf32",
     )
+    parser.add_argument(
+        "--loader",
+        type=str,
+        default="synthetic",
+        help="Dataloader to use",
+    )
 
     args = parser.parse_args()
 
@@ -96,35 +104,52 @@ def main():
     validationFlowBackWarp = validationFlowBackWarp.to(device)
 
     ###Load Datasets
+    def load_dataset():
+        if args.loader == "synthetic":
+            def igen():
+                sz = 352
+                f0 = torch.rand((3, sz, sz)) * 2 - 1
+                ft = torch.rand((3, sz, sz)) * 2 - 1
+                f1 = torch.rand((3, sz, sz)) * 2 - 1
+                return [f0, ft, f1]
+
+            def ogen():
+                return torch.randint(0, 7, ())
+
+            trainset = SyntheticData(
+                n=args.train_batch_size, 
+                repeat=10000, 
+                generators=[igen, ogen]
+            )
 
-    # # Channel wise mean calculated on adobe240-fps training dataset
-    # mean = [0.429, 0.431, 0.397]
-    # std  = [1, 1, 1]
-    # normalize = transforms.Normalize(mean=mean,
-    #                                 std=std)
-    # transform = transforms.Compose([transforms.ToTensor(), normalize])
+            return torch.utils.data.DataLoader(
+                trainset, 
+                batch_size=args.train_batch_size,
+                num_workers=8
+            )
 
-    # trainset = dataloader.SuperSloMo(root=args.dataset_root + '/train', transform=transform, train=True)
-    # trainloader = torch.utils.data.DataLoader(trainset, batch_size=args.train_batch_size, shuffle=False)
+        # Channel wise mean calculated on adobe240-fps training dataset
+        transform = transforms.Compose([
+            transforms.ToTensor(), 
+            transforms.Normalize(
+                mean=[0.429, 0.431, 0.397],
+                std=[1, 1, 1]
+            )
+        ])
 
-    def igen():
-        sz = 352
-        f0 = torch.rand((3, sz, sz)) * 2 - 1
-        ft = torch.rand((3, sz, sz)) * 2 - 1
-        f1 = torch.rand((3, sz, sz)) * 2 - 1
-        return [f0, ft, f1]
+        trainset = dataloader.SuperSloMo(root=args.dataset_root + '/train', transform=transform, train=True)
 
-    def ogen():
-        return torch.randint(0, 7, ())
+        too_small = []
+        for i, p in enumerate(trainset.framesPath):
+            if len(p) < 12:
+                too_small.append(i)
 
-    trainset = SyntheticData(
-        n=args.train_batch_size, repeat=10000, generators=[igen, ogen]
-    )
-    trainloader = torch.utils.data.DataLoader(
-        trainset, 
-        batch_size=args.train_batch_size,
-        num_workers=8
-    )
+        for i in reversed(too_small):
+            del trainset.framesPath[i]
+
+        return torch.utils.data.DataLoader(trainset, batch_size=args.train_batch_size, shuffle=False)
+
+    trainloader = load_dataset()
 
     ###Utils
 

benchmarks/timm/voirfile.py

@@ -32,10 +32,9 @@ def instrument_main(ov, options: Config):
 
     import os
     import torchcompat.core as accelerator
-    from voir.wrapper import DataloaderWrapper, Wrapper
+    from voir.wrapper import Wrapper
 
     from timm.utils.distributed import is_global_primary
-    from timm.data import create_loader
 
     wrapper = Wrapper(
         accelerator.Event, 

benchmate/benchmate/datagen.py

@@ -22,7 +22,7 @@ def write(args):
     offset, outdir, size  = args
 
     img = torch.randn(*size)
-    target = torch.randint(0, 1000, size=(1,), dtype=torch.long)[0]
+    target = offset % 1000 # torch.randint(0, 1000, size=(1,), dtype=torch.long)[0]
     img = transforms.ToPILImage()(img)
 
     class_val = int(target)

benchmate/benchmate/dataset.py

@@ -1,5 +1,6 @@
 
-
+import os
+from collections import defaultdict
 
 
 def no_transform(args):
@@ -30,3 +31,20 @@ def __len__(self):
 
     def __getitem__(self, item):
         return self.transforms(self.dataset[item])
+
+
+
+class ImageNetAsFrames:
+    def __init__(self, folder) -> None:
+        self.clip = defaultdict(list)
+        for root, _, files in os.walk(folder):
+            clip_id = root.split("/")[-1]
+            video = self.clip[clip_id]
+            for frame in files:
+                video.append(frame)
+
+    def __getitem__(self, item):
+        return self.clip[item]
+
+    def __len__(self):
+        return len(self.clip)

config/base.yaml

@@ -506,14 +506,15 @@ super-slomo:
     - video-interpolation
     - unet
     - convnet
-    - noio
   definition: ../benchmarks/super-slomo
   group: super-slomo
   install_group: torch
   plan:
     method: per_gpu
   argv:
     --train_batch_size: 32
+    --dataset_root: /tmp/milabench/cuda/results/data/FakeImageNet
+    --loader: pytorch
 
 ppo:
   inherits: _sb3