# Load a COCO format database from disk for DETR
**URL:**
**Author:**  [@Godouche](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Fu%2FGodouche)%5C
**Post date:** [May 6, 2025, 12:13pm UTC](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fload-a-coco-format-database-from-disk-for-detr%2F153752%2F1 "2025-05-06T12:13:56Z")
I have a COCO database in my disk (with a JSON in the annotations folder that contains image directions) and I would like to load it in HF dataset in orther to use CV models.
Is there a function that allows that?
---
**Author:**  [@John6666](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Fu%2FJohn6666)%5C
**Post date:** [May 7, 2025, 1:56am UTC](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fload-a-coco-format-database-from-disk-for-detr%2F153752%2F2 "2025-05-07T01:56:39Z")
Hmm… This?
>
>
> \## Adding a Dataset
> \- \*\*Name:\*\* COCO
> \- \*\*Description:\*\* COCO is a large-scale …object detection, segmentation, and captioning dataset.
> \- \*\*Paper + website:\*\* https://cocodataset.org/#home
> \- \*\*Data:\*\* https://cocodataset.org/#download
> \- \*\*Motivation:\*\* It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
>
> Instructions to add a new dataset can be found \[here\](https://github.com/huggingface/datasets/blob/master/ADD\_NEW\_DATASET.md).
---
**Author:**  [@lhoestq](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Fu%2Flhoestq)%5C
**Post date:** [May 7, 2025, 12:45pm UTC](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fload-a-coco-format-database-from-disk-for-detr%2F153752%2F3 "2025-05-07T12:45:42Z")
> [@Godouche](#):
>
> I have a COCO database in my disk (with a JSON in the annotations folder that contains image directions) and I would like to load it in HF dataset in orther to use CV models.
>
> Is there a function that allows that?
There is no COCO loader in the `datasets` library, but it would be a welcomed contribution in my opinion.
All the existing data modules are listed [here](https://github.com/huggingface/datasets/tree/main/src/datasets/packaged_modules)
---
**Author:**  [@Godouche](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Fu%2FGodouche)%5C
**Post date:** [May 14, 2025, 12:48pm UTC](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fload-a-coco-format-database-from-disk-for-detr%2F153752%2F4 "2025-05-14T12:48:46Z")
I wrote this code for loading COCO datasets in hugging face datasets that works with DETR,
Adaptations:
- features of your COCO JSON file
- path to COCO folder in local
```auto
import json
import os
import subprocess
from datasets import DatasetDict, Dataset, Features, Value, Sequence, ClassLabel, Image
# Ensure the datasets module is installed
subprocess.check_call(["pip", "install", "datasets"])
class CocoDatasetLoader:
def __init__ (self, coco_folder):
self.coco_folder = coco_folder
def group_by_key_id(self, data, key_id, category_id_to_index):
"""
Groups data by a specified key and maps category IDs to indices.
Args:
data (list): List of dictionaries containing the data.
key_id (str): The key to group by.
category_id_to_index (dict): Mapping from category IDs to indices.
Returns:
dict: Grouped data.
"""
grouped_data = {}
for item in data:
key_value = item[key_id]
if key_value not in grouped_data:
grouped_data[key_value] = {k: [] for k in item.keys() if k != key_id}
for k, v in item.items():
if k != key_id:
grouped_data[key_value][k].append(v)
grouped_data[key_value]['category'] = [category_id_to_index[x] for x in grouped_data[key_value]['category_id']]
return grouped_data
def load_coco_hf_dataset(self, split):
"""
Loads COCO dataset and processes it into a format suitable for Hugging Face datasets.
Args:
split (str): Dataset split (e.g., 'Train', 'Test', 'Validation').
Returns:
Dataset: HuggingFace Dataset of the split of COCO dataset.
"""
# Load the JSON file
json_file_path = os.path.join(self.coco_folder, f'annotations/instances_{split}.json')
try:
with open(json_file_path, 'r') as f:
coco_data = json.load(f)
except FileNotFoundError:
print(f"File not found: {json_file_path}")
return []
# Extract category names and create a mapping from category IDs to indices
category_names = [cat['name'] for cat in coco_data['categories']]
category_id_to_index = {cat['id']: idx for idx, cat in enumerate(coco_data['categories'])}
# Group annotations by 'image_id'
grouped_annotations = self.group_by_key_id(coco_data['annotations'], 'image_id', category_id_to_index)
# Create a dictionary of images
grouped_images = {item['id']: item for item in coco_data['images']}
# Initialize 'objects' field in grouped_images
annotations_keys = list(grouped_annotations.values())[0].keys()
for k, v in grouped_images.items():
grouped_images[k]['objects'] = {key: [] for key in annotations_keys}
# Populate 'objects' field with annotations
for k, v in grouped_annotations.items():
grouped_images[k]['objects'] = v
# Add image paths and IDs
for k, v in grouped_images.items():
v['image'] = os.path.join(self.coco_folder, 'images', split, v['file_name'])
v['image_id'] = v['id']
# Create a Hugging Face dataset from the custom data using from_list for efficiency
hf_dataset = Dataset.from_list(list(grouped_images.values()))
# Define the features for the main dataset
features = Features({
'id': Value('int64'),
'image_id': Value('int64'),
'image': Image(),
'file_name': Value('string'),
'license': Value('string'),
'flickr_url': Value('string'),
'coco_url': Value('string'),
'date_captured': Value('string'),
'width': Value('int64'),
'height': Value('int64'),
'objects': Sequence({
'id': Value('int64'),
'area': Value('float32'),
'bbox': Sequence(Value('float32')),
'category': ClassLabel(names=category_names),
'attributes': {'occluded': Value('bool')},
'category_id': Value('int64'),
'iscrowd': Value('int64'),
'segmentation': {
'counts': Sequence(Value('int64')),
'size': Sequence(Value('int64'))
}
})
})
# Cast the features for the Hugging Face dataset
hf_dataset = hf_dataset.cast(features)
return hf_dataset
# Initialize the CocoDatasetLoader class
coco_loader = CocoDatasetLoader('/path/to/coco/folder/')
hf_dataset_dict = DatasetDict()
for split in ['Train', 'Test', 'Validation']:
# Load the COCO dataset for each split
hf_dataset = coco_loader.load_coco_hf_dataset(split)
# Print the dataset
print(f"Dataset for {split} split:")
print(hf_dataset)
# Create a DatasetDict with the split
hf_dataset_dict[split.lower()] = hf_dataset
```
---
**Author:**  [@system](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Fu%2Fsystem)%5C
**Post date:** [May 15, 2025, 12:48am UTC](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fload-a-coco-format-database-from-disk-for-detr%2F153752%2F5 "2025-05-15T00:48:58Z")
Closed on May 15, 2025, 12:48am UTC
This topic was automatically closed 12 hours after the last reply. New replies are no longer allowed.