feat: changed DIR paths + weird optimization

Former-commit-id: c85c819a022a4cc9fffa88e833d1798b38d5a600 [formerly fb78d7be52badb87cc670bcb8bf0a83c5648d315]
Former-commit-id: 442152b8b84db736380cc17009f1bd4329bb6a22

.devcontainer.tmp/docker-compose.yml

@@ -19,6 +19,6 @@ services:
     container_name: wandb-local
     hostname: wandb-local
     volumes:
-      - /media/disk2/lfainsin/wandb-local/:/vol
+      - ./wandb-local/:/vol
     ports:
       - 8080:8080

src/data/dataloader.py

@@ -25,9 +25,9 @@ class Spheres(pl.LightningDataModule):
         # )
 
         # dataset = SyntheticDataset(image_dir=wandb.config.DIR_TRAIN_IMG, transform=transform)
-        # dataset = Subset(dataset, list(range(0, len(dataset), len(dataset) // 10000 + 1)))
 
-        dataset = LabeledDataset2(image_dir="/home/lilian/data_disk/lfainsin/prerender/")
+        dataset = LabeledDataset2(image_dir="/media/disk1/lfainsin/TRAIN_prerender/")
+        dataset = Subset(dataset, list(range(len(dataset))))  # somhow this allows to better utilize the gpu
 
         return DataLoader(
             dataset,
@@ -40,6 +40,7 @@ class Spheres(pl.LightningDataModule):
 
     def val_dataloader(self):
         dataset = LabeledDataset(image_dir=wandb.config.DIR_VALID_IMG)
+        dataset = Subset(dataset, list(range(len(dataset))))  # somhow this allows to better utilize the gpu
 
         return DataLoader(
             dataset,

src/data/dataset.py

@@ -81,18 +81,19 @@ class LabeledDataset(Dataset):
 
 class LabeledDataset2(Dataset):
     def __init__(self, image_dir):
-        self.images = list(Path(image_dir).glob("**/*.jpg"))
+        self.image_dir = Path(image_dir)
 
     def __len__(self):
-        return len(self.images)
+        return len(list(self.image_dir.iterdir()))
 
     def __getitem__(self, index):
+        path = self.image_dir / str(index)
+
         # open and convert image
-        image = np.array(Image.open(self.images[index]).convert("RGB"), dtype=np.uint8)
+        image = np.array(Image.open(path / "image.jpg").convert("RGB"), dtype=np.uint8)
 
         # open and convert mask
-        mask_path = self.images[index].parent.joinpath("MASK.PNG")
-        mask = np.array(Image.open(mask_path).convert("L"), dtype=np.uint8) // 255
+        mask = np.array(Image.open(path / "MASK.PNG").convert("L"), dtype=np.uint8) // 255
 
         # convert image & mask to Tensor float in [0, 1]
         post_process = A.Compose(

src/train.py

@@ -48,12 +48,13 @@ if __name__ == "__main__":
         max_epochs=wandb.config.EPOCHS,
         accelerator=wandb.config.DEVICE,
         benchmark=wandb.config.BENCHMARK,
-        # profiler="simple",
         precision=16,
         logger=logger,
         log_every_n_steps=1,
         val_check_interval=100,
         callbacks=[RichProgressBar(), ArtifactLog(), TableLog()],
+        # profiler="simple",
+        # num_sanity_val_steps=0,
     )
 
     # actually train the model

wandb.yaml

@@ -1,9 +1,9 @@
 DIR_TRAIN_IMG:
-  value: "/home/lilian/data_disk/lfainsin/train/"
+  value: "/media/disk1/lfainsin/BACKGROUND/"
 DIR_VALID_IMG:
-  value: "/home/lilian/data_disk/lfainsin/test_batched_fast/"
+  value: "/media/disk1/lfainsin/TEST_batched/"
 DIR_SPHERE:
-  value: "/home/lilian/data_disk/lfainsin/spheres+real/"
+  value: "/media/disk1/lfainsin/SPHERES/"
 
 FEATURES:
   value: [8, 16, 32, 64]
@@ -29,13 +29,13 @@ SPHERES:
   value: 3
 
 EPOCHS:
-  value: 20
+  value: 1
 TRAIN_BATCH_SIZE:
-  value: 64 # 100
+  value: 128 # 100
 VAL_BATCH_SIZE:
   value: 8 # 10
 PREFETCH_FACTOR:
-  value: 16
+  value: 2
 
 LEARNING_RATE:
   value: 1.0e-4