adds dataloader and optimiser

dataloading.py

@@ -0,0 +1,50 @@
+from pathlib import Path
+import numpy as np
+import torch as tt
+import sys
+
+class Dataset(tt.utils.data.Dataset):
+
+	def __init__(self,images,labels):
+		self.images = images  # (Batch, Height, Width)
+		self.images = self.images[:, None]  # (Batch, 1, Height, Width)
+		self.images = tt.from_numpy(self.images)  # Convert to tensor
+
+		self.labels = tt.from_numpy(labels)  # (Batch)
+		self.labels = tt.nn.functional.one_hot(self.labels.long(), num_classes=10)  # (Batch, 10)
+
+	def __len__(self):
+		return len(self.images)
+
+	def __getitem__(self, index): # get image-label pair
+		return dict(image=self.images[index], label=self.labels[index])
+
+if __name__ == "__main__":
+
+	here = Path(__file__).resolve().parent
+	path_splits = here/"data"/"datasplits"
+	path_processed = here/"data"/"processed"
+
+	# split contains test data + train data + train data-->(5 * (train_##,val_##) overlapping)
+	split = np.load(path_splits/"datasplit.npz")
+
+	# load processed images and labels
+	images = np.load(path_processed/"images.npy")
+	labels = np.load(path_processed/"labels.npy")
+
+	# use indices from split to create data for splits
+	split_images_data = images[split["train_00"]]
+	split_labels_data = labels[split["train_00"]]
+
+	#  create Dataset for specific split
+	dataset_train = Dataset(split_images_data,split_labels_data)
+
+	# pass to dataset loader
+	loader_train = tt.utils.data.DataLoader(dataset_train, batch_size=5, shuffle=True)
+
+	for data in loader_train:
+		image = data["image"]
+		label = data["label"]
+		print(f"Image shape: {image.shape}, Label shape: {label.shape}")
+		print(f"Image min: {image.min()}, Image max: {image.max()}")
+		sys.exit()  # Exit after first iteration

datasplitting.py

@@ -15,21 +15,26 @@
 	images = np.load(path_images)
 	labels = np.load(path_labels)
 
+	# Create index lists for splitting data
 	split = dict()
 	idx_train, idx_test, _, _ = train_test_split(
         range(len(labels)),
         labels,
         test_size=0.2,
         stratify=labels,
     )
+
+	# Indices in images & labels of training and test data
 	split["train"] = idx_train
 	split["test"] = idx_test
 
+	# Use indexes with training data to create 5-fold split of training data into 5 new train+val sets (new sets overlap in data)
 	skf_train_val = StratifiedKFold(n_splits=5, shuffle=True)
 	for i, (idx_train, idx_va) in enumerate(skf_train_val.split(split["train"], labels[split["train"]])):
 		split[f"train_{i:02d}"] = idx_train
 		split[f"val_{i:02d}"] = idx_va
 
+	# save test data, train data and 5 splits of the train data (train_## + val_##)
 	np.savez(path_split/"datasplit.npz", **split)
 
 	"""Print datasplit information"""

optimiser.py

@@ -0,0 +1,9 @@
+import torch.optim
+
+
+class Solver(torch.optim):
+	def __init__(lr):
+		self.optimizer = super().AdamW(
+			model.parameters(),
+			lr=configuration["lr"],
+			)