Refactors dataset and datamodules
Browse files- src/data_module.py +117 -0
- src/dataset.py +1 -79
src/data_module.py
ADDED
|
@@ -0,0 +1,117 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import lightning as L
|
| 2 |
+
import numpy as np
|
| 3 |
+
import torch
|
| 4 |
+
from sklearn.utils.class_weight import compute_class_weight
|
| 5 |
+
from torch.utils.data import DataLoader, WeightedRandomSampler
|
| 6 |
+
from torchvision.transforms import v2 as T
|
| 7 |
+
|
| 8 |
+
from src.dataset import DRDataset
|
| 9 |
+
|
| 10 |
+
|
| 11 |
+
class DRDataModule(L.LightningDataModule):
|
| 12 |
+
def __init__(
|
| 13 |
+
self,
|
| 14 |
+
train_csv_path,
|
| 15 |
+
val_csv_path,
|
| 16 |
+
image_size: int = 224,
|
| 17 |
+
batch_size: int = 8,
|
| 18 |
+
num_workers: int = 4,
|
| 19 |
+
use_class_weighting: bool = False,
|
| 20 |
+
use_weighted_sampler: bool = False,
|
| 21 |
+
):
|
| 22 |
+
super().__init__()
|
| 23 |
+
self.batch_size = batch_size
|
| 24 |
+
self.num_workers = num_workers
|
| 25 |
+
|
| 26 |
+
# Ensure mutual exclusivity between use_class_weighting and use_weighted_sampler
|
| 27 |
+
if use_class_weighting and use_weighted_sampler:
|
| 28 |
+
raise ValueError(
|
| 29 |
+
"use_class_weighting and use_weighted_sampler cannot both be True"
|
| 30 |
+
)
|
| 31 |
+
|
| 32 |
+
self.train_csv_path = train_csv_path
|
| 33 |
+
self.val_csv_path = val_csv_path
|
| 34 |
+
self.use_class_weighting = use_class_weighting
|
| 35 |
+
self.use_weighted_sampler = use_weighted_sampler
|
| 36 |
+
|
| 37 |
+
# Define the transformations
|
| 38 |
+
self.train_transform = T.Compose(
|
| 39 |
+
[
|
| 40 |
+
T.Resize((image_size, image_size), antialias=True),
|
| 41 |
+
T.RandomAffine(degrees=10, translate=(0.01, 0.01), scale=(0.99, 1.01)),
|
| 42 |
+
T.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.2, hue=0.01),
|
| 43 |
+
T.RandomHorizontalFlip(p=0.5),
|
| 44 |
+
T.ToDtype(torch.float32, scale=True),
|
| 45 |
+
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
|
| 46 |
+
]
|
| 47 |
+
)
|
| 48 |
+
|
| 49 |
+
self.val_transform = T.Compose(
|
| 50 |
+
[
|
| 51 |
+
T.Resize((image_size, image_size), antialias=True),
|
| 52 |
+
T.ToDtype(torch.float32, scale=True),
|
| 53 |
+
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
|
| 54 |
+
]
|
| 55 |
+
)
|
| 56 |
+
|
| 57 |
+
def setup(self, stage=None):
|
| 58 |
+
"""Set up datasets for training and validation."""
|
| 59 |
+
# Initialize datasets with specified transformations
|
| 60 |
+
self.train_dataset = DRDataset(
|
| 61 |
+
self.train_csv_path, transform=self.train_transform
|
| 62 |
+
)
|
| 63 |
+
self.val_dataset = DRDataset(self.val_csv_path, transform=self.val_transform)
|
| 64 |
+
|
| 65 |
+
# Compute number of classes and class weights
|
| 66 |
+
labels = self.train_dataset.labels.numpy()
|
| 67 |
+
self.num_classes = len(np.unique(labels))
|
| 68 |
+
self.class_weights = (
|
| 69 |
+
self._compute_class_weights(labels) if self.use_class_weighting else None
|
| 70 |
+
)
|
| 71 |
+
|
| 72 |
+
def train_dataloader(self):
|
| 73 |
+
"""Returns a DataLoader for training data."""
|
| 74 |
+
if self.use_weighted_sampler:
|
| 75 |
+
sampler = self._get_weighted_sampler(self.train_dataset.labels.numpy())
|
| 76 |
+
shuffle = False # Sampler will handle shuffling
|
| 77 |
+
else:
|
| 78 |
+
sampler = None
|
| 79 |
+
shuffle = True
|
| 80 |
+
|
| 81 |
+
return DataLoader(
|
| 82 |
+
self.train_dataset,
|
| 83 |
+
batch_size=self.batch_size,
|
| 84 |
+
sampler=sampler,
|
| 85 |
+
shuffle=shuffle,
|
| 86 |
+
num_workers=self.num_workers,
|
| 87 |
+
)
|
| 88 |
+
|
| 89 |
+
def val_dataloader(self):
|
| 90 |
+
return DataLoader(
|
| 91 |
+
self.val_dataset, batch_size=self.batch_size, num_workers=self.num_workers
|
| 92 |
+
)
|
| 93 |
+
|
| 94 |
+
def _compute_class_weights(self, labels):
|
| 95 |
+
class_weights = compute_class_weight(
|
| 96 |
+
class_weight="balanced", classes=np.unique(labels), y=labels
|
| 97 |
+
)
|
| 98 |
+
return torch.tensor(class_weights, dtype=torch.float32)
|
| 99 |
+
|
| 100 |
+
def _get_weighted_sampler(self, labels: np.ndarray) -> WeightedRandomSampler:
|
| 101 |
+
"""Returns a WeightedRandomSampler based on class weights.
|
| 102 |
+
|
| 103 |
+
The weights tensor should contain a weight for each sample, not the class weights.
|
| 104 |
+
Have a look at this post for an example: https://discuss.pytorch.org/t/how-to-handle-imbalanced-classes/11264/2
|
| 105 |
+
https://www.maskaravivek.com/post/pytorch-weighted-random-sampler/
|
| 106 |
+
"""
|
| 107 |
+
|
| 108 |
+
class_sample_count = np.array(
|
| 109 |
+
[len(np.where(labels == label)[0]) for label in np.unique(labels)]
|
| 110 |
+
)
|
| 111 |
+
weight = 1.0 / class_sample_count
|
| 112 |
+
samples_weight = np.array([weight[label] for label in labels])
|
| 113 |
+
samples_weight = torch.from_numpy(samples_weight)
|
| 114 |
+
|
| 115 |
+
return WeightedRandomSampler(
|
| 116 |
+
weights=samples_weight, num_samples=len(labels), replacement=True
|
| 117 |
+
)
|
src/dataset.py
CHANGED
|
@@ -1,13 +1,9 @@
|
|
| 1 |
import os
|
| 2 |
|
| 3 |
-
import lightning as L
|
| 4 |
-
import numpy as np
|
| 5 |
import pandas as pd
|
| 6 |
import torch
|
| 7 |
-
from
|
| 8 |
-
from torch.utils.data import DataLoader, Dataset, WeightedRandomSampler
|
| 9 |
from torchvision.io import read_image
|
| 10 |
-
from torchvision.transforms import v2 as T
|
| 11 |
|
| 12 |
|
| 13 |
class DRDataset(Dataset):
|
|
@@ -68,77 +64,3 @@ class DRDataset(Dataset):
|
|
| 68 |
|
| 69 |
return image, label
|
| 70 |
|
| 71 |
-
|
| 72 |
-
class DRDataModule(L.LightningDataModule):
|
| 73 |
-
def __init__(self, batch_size: int = 8, num_workers: int = 4):
|
| 74 |
-
super().__init__()
|
| 75 |
-
self.batch_size = batch_size
|
| 76 |
-
self.num_workers = num_workers
|
| 77 |
-
|
| 78 |
-
# Define the transformations
|
| 79 |
-
self.train_transform = T.Compose(
|
| 80 |
-
[
|
| 81 |
-
T.Resize((224, 224), antialias=True),
|
| 82 |
-
T.RandomAffine(degrees=10, translate=(0.01, 0.01), scale=(0.99, 1.01)),
|
| 83 |
-
T.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.2, hue=0.01),
|
| 84 |
-
T.RandomHorizontalFlip(p=0.5),
|
| 85 |
-
T.ToDtype(torch.float32, scale=True),
|
| 86 |
-
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
|
| 87 |
-
]
|
| 88 |
-
)
|
| 89 |
-
|
| 90 |
-
self.val_transform = T.Compose(
|
| 91 |
-
[
|
| 92 |
-
T.Resize((224, 224), antialias=True),
|
| 93 |
-
T.ToDtype(torch.float32, scale=True),
|
| 94 |
-
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
|
| 95 |
-
]
|
| 96 |
-
)
|
| 97 |
-
|
| 98 |
-
self.num_classes = 5
|
| 99 |
-
|
| 100 |
-
def setup(self, stage=None):
|
| 101 |
-
self.train_dataset = DRDataset("data/train.csv", transform=self.train_transform)
|
| 102 |
-
self.val_dataset = DRDataset("data/val.csv", transform=self.val_transform)
|
| 103 |
-
|
| 104 |
-
# compute class weights
|
| 105 |
-
labels = self.train_dataset.labels.numpy()
|
| 106 |
-
self.class_weights = None # self.compute_class_weights(labels)
|
| 107 |
-
|
| 108 |
-
def train_dataloader(self):
|
| 109 |
-
return DataLoader(
|
| 110 |
-
self.train_dataset,
|
| 111 |
-
batch_size=self.batch_size,
|
| 112 |
-
sampler=self._get_weighted_sampler(self.train_dataset.labels.numpy()),
|
| 113 |
-
# shuffle=True,
|
| 114 |
-
num_workers=self.num_workers,
|
| 115 |
-
)
|
| 116 |
-
|
| 117 |
-
def val_dataloader(self):
|
| 118 |
-
return DataLoader(
|
| 119 |
-
self.val_dataset, batch_size=self.batch_size, num_workers=self.num_workers
|
| 120 |
-
)
|
| 121 |
-
|
| 122 |
-
def compute_class_weights(self, labels):
|
| 123 |
-
class_weights = compute_class_weight(
|
| 124 |
-
class_weight="balanced", classes=np.unique(labels), y=labels
|
| 125 |
-
)
|
| 126 |
-
return torch.tensor(class_weights, dtype=torch.float32)
|
| 127 |
-
|
| 128 |
-
def _get_weighted_sampler(self, labels: np.ndarray) -> WeightedRandomSampler:
|
| 129 |
-
"""Returns a WeightedRandomSampler based on class weights.
|
| 130 |
-
|
| 131 |
-
The weights tensor should contain a weight for each sample, not the class weights.
|
| 132 |
-
Have a look at this post for an example: https://discuss.pytorch.org/t/how-to-handle-imbalanced-classes/11264/2
|
| 133 |
-
https://www.maskaravivek.com/post/pytorch-weighted-random-sampler/
|
| 134 |
-
"""
|
| 135 |
-
|
| 136 |
-
|
| 137 |
-
class_sample_count = np.array([len(np.where(labels == label)[0]) for label in np.unique(labels)])
|
| 138 |
-
weight = 1. / class_sample_count
|
| 139 |
-
samples_weight = np.array([weight[label] for label in labels])
|
| 140 |
-
samples_weight = torch.from_numpy(samples_weight)
|
| 141 |
-
|
| 142 |
-
# class_weights = compute_class_weight("balanced", classes=np.unique(labels), y=labels)
|
| 143 |
-
# class_weights_tensor = torch.tensor(class_weights, dtype=torch.float32)
|
| 144 |
-
return WeightedRandomSampler(weights=samples_weight, num_samples=len(labels), replacement=True)
|
|
|
|
| 1 |
import os
|
| 2 |
|
|
|
|
|
|
|
| 3 |
import pandas as pd
|
| 4 |
import torch
|
| 5 |
+
from torch.utils.data import Dataset
|
|
|
|
| 6 |
from torchvision.io import read_image
|
|
|
|
| 7 |
|
| 8 |
|
| 9 |
class DRDataset(Dataset):
|
|
|
|
| 64 |
|
| 65 |
return image, label
|
| 66 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|