Vestiq / deepfashion2_utils.py
Hashii1729's picture
Integrate DeepFashion2 dataset: add evaluation module, utilities, and API endpoints for dataset management and analysis
f8b306b
"""
DeepFashion2 Dataset Integration Utilities
Provides tools for loading, processing, and using the DeepFashion2 dataset
with the Vestiq fashion analysis system.
"""
import os
import json
import torch
import numpy as np
from PIL import Image
from torch.utils.data import Dataset, DataLoader
from pathlib import Path
from typing import Dict, List, Tuple, Optional, Union
import torchvision.transforms as transforms
from dataclasses import dataclass, field
import requests
import zipfile
import shutil
@dataclass
class DeepFashion2Config:
"""Configuration for DeepFashion2 dataset"""
dataset_root: str = "./data/deepfashion2"
download_url: str = "https://github.com/switchablenorms/DeepFashion2/releases/download/v1.0/deepfashion2.zip"
categories: List[str] = field(default_factory=list)
image_size: Tuple[int, int] = (224, 224)
batch_size: int = 32
num_workers: int = 4
def __post_init__(self):
if not self.categories:
# DeepFashion2 13 categories
self.categories = [
'short_sleeved_shirt', 'long_sleeved_shirt', 'short_sleeved_outwear',
'long_sleeved_outwear', 'vest', 'sling', 'shorts', 'trousers',
'skirt', 'short_sleeved_dress', 'long_sleeved_dress', 'vest_dress', 'sling_dress'
]
class DeepFashion2CategoryMapper:
"""Maps DeepFashion2 categories to yainage90 model categories"""
def __init__(self):
# Mapping from DeepFashion2 categories to yainage90 categories
self.df2_to_yainage90 = {
'short_sleeved_shirt': 'top',
'long_sleeved_shirt': 'top',
'short_sleeved_outwear': 'outer',
'long_sleeved_outwear': 'outer',
'vest': 'top',
'sling': 'top',
'shorts': 'bottom',
'trousers': 'bottom',
'skirt': 'bottom',
'short_sleeved_dress': 'dress',
'long_sleeved_dress': 'dress',
'vest_dress': 'dress',
'sling_dress': 'dress'
}
# Reverse mapping
self.yainage90_to_df2 = {}
for df2_cat, yainage_cat in self.df2_to_yainage90.items():
if yainage_cat not in self.yainage90_to_df2:
self.yainage90_to_df2[yainage_cat] = []
self.yainage90_to_df2[yainage_cat].append(df2_cat)
def map_to_yainage90(self, df2_category: str) -> str:
"""Map DeepFashion2 category to yainage90 category"""
return self.df2_to_yainage90.get(df2_category, 'unknown')
def map_from_yainage90(self, yainage_category: str) -> List[str]:
"""Map yainage90 category to DeepFashion2 categories"""
return self.yainage90_to_df2.get(yainage_category, [])
class DeepFashion2Dataset(Dataset):
"""PyTorch Dataset for DeepFashion2"""
def __init__(self,
root_dir: str,
split: str = 'train',
transform: Optional[transforms.Compose] = None,
load_annotations: bool = True):
"""
Initialize DeepFashion2 dataset
Args:
root_dir: Root directory of DeepFashion2 dataset
split: Dataset split ('train', 'validation', 'test')
transform: Image transformations
load_annotations: Whether to load bounding box annotations
"""
self.root_dir = Path(root_dir)
self.split = split
self.transform = transform
self.load_annotations = load_annotations
self.category_mapper = DeepFashion2CategoryMapper()
# Load dataset metadata
self.images_dir = self.root_dir / split / "image"
self.annos_dir = self.root_dir / split / "annos"
# Get all image files
self.image_files = []
if self.images_dir.exists():
self.image_files = list(self.images_dir.glob("*.jpg"))
print(f"Found {len(self.image_files)} images in {split} split")
def __len__(self):
return len(self.image_files)
def __getitem__(self, idx):
"""Get dataset item"""
image_path = self.image_files[idx]
image_name = image_path.stem
# Load image
image = Image.open(image_path).convert('RGB')
# Load annotations if requested
annotations = None
if self.load_annotations:
anno_path = self.annos_dir / f"{image_name}.json"
if anno_path.exists():
with open(anno_path, 'r') as f:
annotations = json.load(f)
# Apply transforms
if self.transform:
image = self.transform(image)
return {
'image': image,
'image_path': str(image_path),
'image_name': image_name,
'annotations': annotations
}
def get_categories_in_image(self, annotations: Dict) -> List[str]:
"""Extract categories from annotations"""
if not annotations or 'item' not in annotations:
return []
categories = []
for item_id, item_data in annotations['item'].items():
if 'category_name' in item_data:
categories.append(item_data['category_name'])
return list(set(categories))
class DeepFashion2Downloader:
"""Download and setup DeepFashion2 dataset"""
def __init__(self, config: DeepFashion2Config):
self.config = config
self.dataset_root = Path(config.dataset_root)
def download_dataset(self, force_download: bool = False) -> bool:
"""
Download DeepFashion2 dataset
Args:
force_download: Force re-download even if dataset exists
Returns:
True if successful, False otherwise
"""
if self.dataset_root.exists() and not force_download:
print(f"Dataset already exists at {self.dataset_root}")
return True
print("DeepFashion2 dataset download requires manual setup.")
print("Please follow these steps:")
print("1. Visit: https://github.com/switchablenorms/DeepFashion2")
print("2. Follow the dataset download instructions")
print("3. Extract the dataset to:", self.dataset_root)
print("4. Ensure the directory structure is:")
print(" deepfashion2/")
print(" β”œβ”€β”€ train/")
print(" β”‚ β”œβ”€β”€ image/")
print(" β”‚ └── annos/")
print(" β”œβ”€β”€ validation/")
print(" β”‚ β”œβ”€β”€ image/")
print(" β”‚ └── annos/")
print(" └── test/")
print(" β”œβ”€β”€ image/")
print(" └── annos/")
return False
def verify_dataset(self) -> bool:
"""Verify dataset structure"""
required_dirs = [
self.dataset_root / "train" / "image",
self.dataset_root / "train" / "annos",
self.dataset_root / "validation" / "image",
self.dataset_root / "validation" / "annos"
]
for dir_path in required_dirs:
if not dir_path.exists():
print(f"Missing required directory: {dir_path}")
return False
print("Dataset structure verified successfully")
return True
def create_deepfashion2_transforms(image_size: Tuple[int, int] = (224, 224)) -> transforms.Compose:
"""Create standard transforms for DeepFashion2 images"""
return transforms.Compose([
transforms.Resize(image_size),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
def create_deepfashion2_dataloader(config: DeepFashion2Config,
split: str = 'train',
shuffle: bool = True) -> DataLoader:
"""Create DataLoader for DeepFashion2 dataset"""
transform = create_deepfashion2_transforms(config.image_size)
dataset = DeepFashion2Dataset(
root_dir=config.dataset_root,
split=split,
transform=transform,
load_annotations=True
)
return DataLoader(
dataset,
batch_size=config.batch_size,
shuffle=shuffle,
num_workers=config.num_workers,
pin_memory=torch.cuda.is_available()
)
def get_deepfashion2_statistics(config: DeepFashion2Config) -> Dict:
"""Get statistics about the DeepFashion2 dataset"""
stats = {
'splits': {},
'total_images': 0,
'categories': config.categories,
'category_counts': {cat: 0 for cat in config.categories}
}
for split in ['train', 'validation', 'test']:
try:
dataset = DeepFashion2Dataset(
root_dir=config.dataset_root,
split=split,
transform=None,
load_annotations=True
)
split_stats = {
'num_images': len(dataset),
'categories_found': set()
}
# Sample a few images to get category statistics
sample_size = min(100, len(dataset))
for i in range(0, len(dataset), max(1, len(dataset) // sample_size)):
item = dataset[i]
if item['annotations']:
categories = dataset.get_categories_in_image(item['annotations'])
split_stats['categories_found'].update(categories)
for cat in categories:
if cat in stats['category_counts']:
stats['category_counts'][cat] += 1
split_stats['categories_found'] = list(split_stats['categories_found'])
stats['splits'][split] = split_stats
stats['total_images'] += split_stats['num_images']
except Exception as e:
print(f"Error processing {split} split: {e}")
stats['splits'][split] = {'error': str(e)}
return stats