Update README.md
Browse files
README.md
CHANGED
|
@@ -102,22 +102,23 @@ instances["categories"]
|
|
| 102 |
## Build the dataset and upload to Hub
|
| 103 |
|
| 104 |
```py
|
| 105 |
-
!
|
| 106 |
-
!wget http://images.cocodataset.org/zips/val2017.zip
|
| 107 |
-
!wget http://images.cocodataset.org/annotations/annotations_trainval2017.zip
|
| 108 |
|
| 109 |
-
|
| 110 |
-
!
|
| 111 |
-
!
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 112 |
|
| 113 |
import json
|
|
|
|
| 114 |
from pathlib import Path
|
| 115 |
from tqdm import tqdm
|
| 116 |
-
from
|
| 117 |
-
from datasets import Dataset, DatasetDict, Features, Value, Sequence, Array2D
|
| 118 |
-
import shutil
|
| 119 |
|
| 120 |
-
# === Paths ===
|
| 121 |
base_dir = Path("/content")
|
| 122 |
splits = {
|
| 123 |
"train": {
|
|
@@ -130,11 +131,15 @@ splits = {
|
|
| 130 |
}
|
| 131 |
}
|
| 132 |
output_dir = base_dir / "coco_imagefolder"
|
|
|
|
| 133 |
|
| 134 |
-
|
| 135 |
-
if
|
| 136 |
-
|
| 137 |
-
|
|
|
|
|
|
|
|
|
|
| 138 |
|
| 139 |
def convert_coco_to_jsonl(image_dir, annotation_path, output_metadata_path):
|
| 140 |
with open(annotation_path) as f:
|
|
@@ -145,42 +150,46 @@ def convert_coco_to_jsonl(image_dir, annotation_path, output_metadata_path):
|
|
| 145 |
|
| 146 |
for ann in data['annotations']:
|
| 147 |
img_id = ann['image_id']
|
| 148 |
-
bbox = ann['bbox']
|
| 149 |
category = ann['category_id']
|
|
|
|
|
|
|
|
|
|
|
|
|
| 150 |
|
| 151 |
if img_id not in annotations_by_image:
|
| 152 |
annotations_by_image[img_id] = {
|
| 153 |
"file_name": id_to_filename[img_id],
|
| 154 |
"objects": {
|
| 155 |
"bbox": [],
|
| 156 |
-
"
|
|
|
|
| 157 |
}
|
| 158 |
}
|
| 159 |
|
| 160 |
annotations_by_image[img_id]["objects"]["bbox"].append(bbox)
|
|
|
|
| 161 |
annotations_by_image[img_id]["objects"]["categories"].append(category)
|
| 162 |
|
| 163 |
with open(output_metadata_path, "w") as f:
|
| 164 |
-
for
|
| 165 |
json.dump(metadata, f)
|
| 166 |
f.write("\n")
|
| 167 |
|
| 168 |
-
#
|
| 169 |
for split, info in splits.items():
|
| 170 |
split_dir = output_dir / split
|
| 171 |
-
split_dir.mkdir(parents=True)
|
| 172 |
|
| 173 |
# Copy images
|
| 174 |
for img_path in tqdm(info["image_dir"].glob("*.jpg"), desc=f"Copying {split} images"):
|
| 175 |
shutil.copy(img_path, split_dir / img_path.name)
|
| 176 |
|
| 177 |
-
#
|
| 178 |
metadata_path = split_dir / "metadata.jsonl"
|
| 179 |
convert_coco_to_jsonl(split_dir, info["annotation_file"], metadata_path)
|
| 180 |
|
| 181 |
-
#
|
| 182 |
-
|
| 183 |
-
|
| 184 |
-
dataset = load_dataset("imagefolder", data_dir="/content/coco_imagefolder")
|
| 185 |
dataset.push_to_hub("ariG23498/coco2017")
|
| 186 |
```
|
|
|
|
| 102 |
## Build the dataset and upload to Hub
|
| 103 |
|
| 104 |
```py
|
| 105 |
+
!pip install -U -q datasets
|
|
|
|
|
|
|
| 106 |
|
| 107 |
+
# Download and unzip COCO 2017
|
| 108 |
+
!wget -q http://images.cocodataset.org/zips/train2017.zip
|
| 109 |
+
!wget -q http://images.cocodataset.org/zips/val2017.zip
|
| 110 |
+
!wget -q http://images.cocodataset.org/annotations/annotations_trainval2017.zip
|
| 111 |
+
|
| 112 |
+
!unzip -q train2017.zip
|
| 113 |
+
!unzip -q val2017.zip
|
| 114 |
+
!unzip -q annotations_trainval2017.zip
|
| 115 |
|
| 116 |
import json
|
| 117 |
+
import shutil
|
| 118 |
from pathlib import Path
|
| 119 |
from tqdm import tqdm
|
| 120 |
+
from datasets import load_dataset
|
|
|
|
|
|
|
| 121 |
|
|
|
|
| 122 |
base_dir = Path("/content")
|
| 123 |
splits = {
|
| 124 |
"train": {
|
|
|
|
| 131 |
}
|
| 132 |
}
|
| 133 |
output_dir = base_dir / "coco_imagefolder"
|
| 134 |
+
output_dir.mkdir(parents=True, exist_ok=True)
|
| 135 |
|
| 136 |
+
def normalize_segmentation(segmentation):
|
| 137 |
+
if isinstance(segmentation, list):
|
| 138 |
+
if all(isinstance(poly, list) for poly in segmentation):
|
| 139 |
+
return segmentation # already a list of polygons
|
| 140 |
+
elif all(isinstance(pt, (int, float)) for pt in segmentation):
|
| 141 |
+
return [segmentation] # wrap single polygon
|
| 142 |
+
return [] # skip RLE or malformed segmentations
|
| 143 |
|
| 144 |
def convert_coco_to_jsonl(image_dir, annotation_path, output_metadata_path):
|
| 145 |
with open(annotation_path) as f:
|
|
|
|
| 150 |
|
| 151 |
for ann in data['annotations']:
|
| 152 |
img_id = ann['image_id']
|
| 153 |
+
bbox = ann['bbox']
|
| 154 |
category = ann['category_id']
|
| 155 |
+
segmentation = normalize_segmentation(ann['segmentation'])
|
| 156 |
+
|
| 157 |
+
if not segmentation:
|
| 158 |
+
continue # skip if malformed or RLE
|
| 159 |
|
| 160 |
if img_id not in annotations_by_image:
|
| 161 |
annotations_by_image[img_id] = {
|
| 162 |
"file_name": id_to_filename[img_id],
|
| 163 |
"objects": {
|
| 164 |
"bbox": [],
|
| 165 |
+
"segmentation": [],
|
| 166 |
+
"categories": [],
|
| 167 |
}
|
| 168 |
}
|
| 169 |
|
| 170 |
annotations_by_image[img_id]["objects"]["bbox"].append(bbox)
|
| 171 |
+
annotations_by_image[img_id]["objects"]["segmentation"].append(segmentation)
|
| 172 |
annotations_by_image[img_id]["objects"]["categories"].append(category)
|
| 173 |
|
| 174 |
with open(output_metadata_path, "w") as f:
|
| 175 |
+
for metadata in annotations_by_image.values():
|
| 176 |
json.dump(metadata, f)
|
| 177 |
f.write("\n")
|
| 178 |
|
| 179 |
+
# Build imagefolder structure
|
| 180 |
for split, info in splits.items():
|
| 181 |
split_dir = output_dir / split
|
| 182 |
+
split_dir.mkdir(parents=True, exist_ok=True)
|
| 183 |
|
| 184 |
# Copy images
|
| 185 |
for img_path in tqdm(info["image_dir"].glob("*.jpg"), desc=f"Copying {split} images"):
|
| 186 |
shutil.copy(img_path, split_dir / img_path.name)
|
| 187 |
|
| 188 |
+
# Write JSONL metadata
|
| 189 |
metadata_path = split_dir / "metadata.jsonl"
|
| 190 |
convert_coco_to_jsonl(split_dir, info["annotation_file"], metadata_path)
|
| 191 |
|
| 192 |
+
# Load and push
|
| 193 |
+
dataset = load_dataset("imagefolder", data_dir=str(output_dir))
|
|
|
|
|
|
|
| 194 |
dataset.push_to_hub("ariG23498/coco2017")
|
| 195 |
```
|