Spaces:
Build error
Build error
File size: 8,727 Bytes
95848bc | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 | """
Utility functions for dataset handling, splitting, and subset creation.
"""
import os
import zipfile
from collections import Counter
from glob import glob
import numpy as np
import torch
from sklearn.model_selection import train_test_split
def ensure_dataset_extracted(path):
"""
Ensures that if the path points to a zip file or a directory containing a zip file,
it is extracted. Returns the path to the directory containing the actual data.
"""
target_path = path
# Case 1: path is a zip file
if os.path.isfile(path) and path.lower().endswith(".zip"):
extract_dir = os.path.splitext(path)[0]
if not os.path.exists(extract_dir):
print(f"Extracting {path}...")
with zipfile.ZipFile(path, "r") as zip_ref:
zip_ref.extractall(extract_dir)
target_path = extract_dir
# Case 2: path is a directory that might contain one or more zip parts
elif os.path.isdir(path):
# Check if it already looks like a dataset (has 'color' folder)
if os.path.exists(os.path.join(path, "color")):
return path
# Look for zip files (including chunked parts)
files = os.listdir(path)
zip_files = sorted([f for f in files if f.lower().endswith(".zip")])
if zip_files:
# Use the base name of the first zip part for extraction folder
base_name = os.path.splitext(zip_files[0])[0].replace("_part_001", "")
extract_dir = os.path.join(path, base_name)
if not os.path.exists(extract_dir):
os.makedirs(extract_dir, exist_ok=True)
print(f"Extracting {len(zip_files)} zip parts to {extract_dir}...")
for zip_file in zip_files:
zip_path = os.path.join(path, zip_file)
print(f" Extracting {zip_file}...")
with zipfile.ZipFile(zip_path, "r") as zip_ref:
zip_ref.extractall(extract_dir)
target_path = extract_dir
# After extraction, check if the data is nested (e.g. extracted_folder/dataset_name/color)
# We look for the 'color' folder
for root, dirs, files in os.walk(target_path):
if "color" in dirs:
return root
return target_path
def build_class_mapping(data_dir, modality="color"):
"""
Build a mapping from class names to integer IDs.
Args:
data_dir: root directory containing modality subfolders
modality: which modality folder to scan for class names (default: "color")
Returns:
tuple: (class_names, class_to_idx)
- class_names: sorted list of class names
- class_to_idx: dict mapping class name to integer ID
"""
modality_path = os.path.join(data_dir, modality)
class_names = sorted(next(os.walk(modality_path))[1])
class_to_idx = {cls: i for i, cls in enumerate(class_names)}
return class_names, class_to_idx
def gather_samples(data_dir, modalities, class_to_idx):
"""
Gather all image samples from the dataset.
Args:
data_dir: root directory containing modality subfolders
modalities: list of modality names (e.g., ["color", "grayscale", "segmented"])
class_to_idx: dict mapping class name to integer ID
Returns:
list: samples as (img_path, label_id, modality_name) tuples
"""
samples = []
for modality in modalities:
for cls, idx in class_to_idx.items():
folder = os.path.join(data_dir, modality, cls)
seen_paths = set()
# Handle both .jpg and .JPG extensions
for pattern in ["*.jpg", "*.JPG"]:
for img_path in glob(os.path.join(folder, pattern)):
if img_path in seen_paths:
continue
seen_paths.add(img_path)
samples.append((img_path, idx, modality))
return samples
def split_dataset(samples, test_size=0.15, val_size=0.18, random_state=42):
"""
Split dataset into train, validation, and test sets.
Args:
samples: list of (img_path, label_id, modality_name) tuples
test_size: proportion of data for test set (default: 0.15 = 15%)
val_size: proportion of remaining data for validation (default: 0.18 ≈ 15% of total)
random_state: random seed for reproducibility
Returns:
tuple: (train_samples, val_samples, test_samples)
Note:
With default values: ~70% train, ~15% val, ~15% test
"""
# First split: separate test set
train_val, test = train_test_split(
samples,
test_size=test_size,
shuffle=True,
stratify=[s[1] for s in samples],
random_state=random_state,
)
# Second split: separate train and validation
train, val = train_test_split(
train_val,
test_size=val_size,
shuffle=True,
stratify=[s[1] for s in train_val],
random_state=random_state,
)
return train, val, test
def make_subset(samples, ratio, seed=42):
"""
Create a stratified subset of samples.
Useful for quick prototyping or hyperparameter tuning on smaller data.
Args:
samples: list of (img_path, label_id, modality_name) tuples
ratio: proportion of samples to keep (e.g., 0.05 = 5%, 0.3 = 30%)
seed: random seed for reproducibility
Returns:
list: subset of samples maintaining class distribution
"""
subset, _ = train_test_split(
samples, train_size=ratio, stratify=[s[1] for s in samples], random_state=seed
)
return subset
def get_class_distribution(samples):
"""
Get the distribution of classes in the dataset.
Args:
samples: list of (img_path, label_id, modality_name) tuples
Returns:
Counter: class_id -> count mapping
"""
labels = [s[1] for s in samples]
return Counter(labels)
def balance_dataset_uniform(samples, seed=42):
"""
Balance dataset by uniform sampling - take same number from each class.
Uses the size of the smallest class as the target count for all classes.
Args:
samples: list of (img_path, label_id, modality_name) tuples
seed: random seed for reproducibility
Returns:
list: balanced samples with equal number per class
"""
np.random.seed(seed)
# Group samples by class
class_samples = {}
for sample in samples:
label = sample[1]
if label not in class_samples:
class_samples[label] = []
class_samples[label].append(sample)
# Find minimum class size
min_count = min(len(samples_list) for samples_list in class_samples.values())
# Randomly sample min_count from each class
balanced = []
for label, samples_list in class_samples.items():
sampled = np.random.choice(len(samples_list), size=min_count, replace=False)
balanced.extend([samples_list[i] for i in sampled])
# Shuffle the balanced dataset
np.random.shuffle(balanced)
return balanced
def calculate_class_weights(samples, num_classes=None):
"""
Calculate class weights for imbalanced datasets.
Uses inverse frequency weighting: weight = 1 / frequency
Normalized so the weights sum to num_classes.
Args:
samples: list of (img_path, label_id, modality_name) tuples
num_classes: total number of classes (if None, inferred from samples)
Returns:
torch.Tensor: weight for each class (length = num_classes)
"""
labels = [s[1] for s in samples]
if num_classes is None:
num_classes = max(labels) + 1
# Count samples per class
class_counts = Counter(labels)
# Calculate weights: inverse frequency
weights = torch.zeros(num_classes)
for class_id in range(num_classes):
count = class_counts.get(class_id, 0)
if count > 0:
weights[class_id] = 1.0 / count
else:
weights[class_id] = 0.0
# Normalize weights so they sum to num_classes
weights = weights / weights.sum() * num_classes
return weights
def get_sample_weights(samples):
"""
Calculate per-sample weights for WeightedRandomSampler.
Each sample gets weight = 1 / (count of its class)
Args:
samples: list of (img_path, label_id, modality_name) tuples
Returns:
torch.Tensor: weight for each sample (length = len(samples))
"""
labels = [s[1] for s in samples]
class_counts = Counter(labels)
# Assign weight to each sample based on its class frequency
sample_weights = torch.tensor([1.0 / class_counts[label] for label in labels])
return sample_weights
|