Deci-AI/super-gradients

...

You have to be logged in to leave a comment.

Computer Vision Datasets Setup

SuperGradients provides multiple Datasets implementations.

Classification Datasets

Cifar10

Supports download

from super_gradients.training.datasets import Cifar10
dataset = Cifar10(..., download=True)

Imagenet

Download imagenet dataset:

https://image-net.org/download.php

Unzip:

    Imagenet
     ├──train
     │  ├──n02093991
     │  │   ├──n02093991_1001.JPEG
     │  │   ├──n02093991_1004.JPEG
     │  │   └──...
     │  ├──n02093992
     │  └──...
     └──val
        ├──n02093991
        ├──n02093992
        └──...

Instantiate the dataset:

from super_gradients.training.datasets import ImageNetDataset
train_set = ImageNetDataset(root='.../Imagenet/train', ...)
valid_set = ImageNetDataset(root='.../Imagenet/val', ...)

Detection Datasets

Coco

Download coco dataset:

annotations: http://images.cocodataset.org/annotations/annotations_trainval2017.zip
train2017: http://images.cocodataset.org/zips/train2017.zip
val2017: http://images.cocodataset.org/zips/val2017.zip

Unzip and organize it as below:

    coco
    ├── annotations
    │      ├─ instances_train2017.json
    │      ├─ instances_val2017.json
    │      └─ ...
    └── images
        ├── train2017
        │   ├─ 000000000001.jpg
        │   └─ ...
        └── val2017
            └─ ...

Instantiate the dataset:

from super_gradients.training.datasets import COCODetectionDataset
train_set = COCODetectionDataset(data_dir='.../coco', subdir='images/train2017', json_file='instances_train2017.json', ...)
valid_set = COCODetectionDataset(data_dir='.../coco', subdir='images/val2017', json_file='instances_val2017.json', ...)

PascalVOC 2007 & 2012

Supports download

from super_gradients.training.datasets import PascalVOCDetectionDataset
train_set = PascalVOCDetectionDataset(download=True, ...)

Dataset Structure:

    Dataset structure:
        ├─images
        │   ├─ train2012
        │   ├─ val2012
        │   ├─ VOCdevkit
        │   │    ├─ VOC2007
        │   │    │  ├──JPEGImages
        │   │    │  ├──SegmentationClass
        │   │    │  ├──ImageSets
        │   │    │  ├──ImageSets/Segmentation
        │   │    │  ├──ImageSets/Main
        │   │    │  ├──ImageSets/Layout
        │   │    │  ├──Annotations
        │   │    │  └──SegmentationObject
        │   │    └──VOC2012
        │   │       ├──JPEGImages
        │   │       ├──SegmentationClass
        │   │       ├──ImageSets
        │   │       ├──ImageSets/Segmentation
        │   │       ├──ImageSets/Main
        │   │       ├──ImageSets/Action
        │   │       ├──ImageSets/Layout
        │   │       ├──Annotations
        │   │       └──SegmentationObject
        │   ├─train2007
        │   ├─test2007
        │   └─val2007
        └─labels
            ├─train2012
            ├─val2012
            ├─train2007
            ├─test2007
            └─val2007

Yolo/Darknet format

Download your dataset (can be from https://roboflow.com/universe)
You should have a structure similar to this.

     data_dir
     └── train/test/val
         ├── images
         │      ├─ 0001.jpg
         │      ├─ 0002.jpg
         │      └─ ...
         └── labels
                ├─ 0001.txt
                ├─ 0002.txt
                └─ ...

Note: train/test/val folders are not required, any folder structure is supported.

Instantiate the dataset:

from super_gradients.training.datasets import YoloDarknetFormatDetectionDataset
data_set = YoloDarknetFormatDetectionDataset(data_dir='<path-to>/data_dir', images_dir="<train/test/val>/images", labels_dir="<train/test/val>/labels", classes=["<to-fill>"])

Segmentation Datasets

Cityscapes

Download dataset:

a. Cityscapes dataset:
- gtFine: https://www.cityscapes-dataset.com/file-handling/?packageID=1
- leftImg8bit: https://www.cityscapes-dataset.com/file-handling/?packageID=3
b. metadata folder: https://deci-pretrained-models.s3.amazonaws.com/cityscape_lists.zip

a. Unzip and organize cityscapes dataset as below:

  root_dir (in recipe default to /data/cityscapes)
      ├─── gtFine
      │       ├── test
      │       │     ├── berlin
      │       │     │   ├── berlin_000000_000019_gtFine_color.png
      │       │     │   ├── berlin_000000_000019_gtFine_instanceIds.png
      │       │     │   └── ...
      │       │     ├── bielefeld
      │       │     │   └── ...
      │       │     └── ...
      │       ├─── train
      │       │     └── ...
      │       └─── val
      │             └── ...
      └─── leftImg8bit
              ├── test
              │     └── ...
              ├─── train
              │     └── ...
              └─── val
                    └── ...

b. Unzip and organize metadata folder as below:

  lists
      ├── labels.csv
      ├── test.lst
      ├── train.lst
      ├── trainval.lst
      ├── val.lst
      └── auto_labelling.lst

c. Move Metadata folder to the Cityscapes folder

  root_dir (in recipe default to /data/cityscapes)
      ├─── gtFine
      │      └── ...
      ├─── leftImg8bit
      │      └── ...
      └─── lists
             └── ...

Instantiate the dataset:

from super_gradients.training.datasets import CityscapesDataset

train_set = CityscapesDataset(root_dir='.../root_dir', list_file='lists/train.lst', labels_csv_path='lists/labels.csv', ...)

AutoLabelling dataset [Optional]

Cityscapes AutoLabelled dataset were introduced by NVIDIA research group in the paper: "Hierarchical Multi-Scale Attention for Semantic Segmentation".

AutoLabelled refer to the refinement of the Cityscapes coarse data and pseudo labels generation using their suggested Hierarchical multi-scale attention model.

To download the AutoLabelled labels please refer to the original repo. Unzip and rename the folder to AutoLabelling as described bellow.
Download the coarse RGB images from cityscapes official site, leftImg8bit_train_extra: https://www.cityscapes-dataset.com/file-handling/?packageID=4

  root_dir (in recipe default to /data/cityscapes)
      ├─── gtFine
      │       ├── test
      │       │     └── ...
      │       ├─── train
      │       │     └── ...
      │       └─── val
      │             └── ...
      ├─── leftImg8bit
      │       ├── test
      │       │     └── ...
      │       ├─── train
      │       │     └── ...
      │       └─── val
      │             └── ...
      ├─── AutoLabelling
      │       └─── train_extra
      │             └── ...
      └─── leftImg8bit
              └─── train_extra
                    └── ...

Coco

Download coco dataset:

annotations: http://images.cocodataset.org/annotations/annotations_trainval2017.zip
train2017: http://images.cocodataset.org/zips/train2017.zip
val2017: http://images.cocodataset.org/zips/val2017.zip

Unzip and organize it as below:

    coco
    ├── annotations
    │      ├─ instances_train2017.json
    │      ├─ instances_val2017.json
    │      └─ ...
    └── images
        ├── train2017
        │   ├─ 000000000001.jpg
        │   └─ ...
        └── val2017
            └─ ...

Instantiate the dataset:

from super_gradients.training.datasets import CoCoSegmentationDataSet
train_set = CoCoSegmentationDataSet(data_dir='.../coco', subdir='images/train2017', json_file='instances_train2017.json', ...)
valid_set = CoCoSegmentationDataSet(data_dir='.../coco', subdir='images/val2017', json_file='instances_val2017.json', ...)

Pascal VOC 2012

Download pascal datasets:

VOC 2012: http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar

Unzip and organize it as below:

   pascal_voc_2012
       └──VOCdevkit
             └──VOC2012
                ├──JPEGImages
                ├──SegmentationClass
                ├──ImageSets
                │    ├──Segmentation
                │    │   └── train.txt
                │    ├──Main
                │    ├──Action
                │    └──Layout
                ├──Annotations
                └──SegmentationObject

Instantiate the dataset:

from super_gradients.training.datasets import PascalVOC2012SegmentationDataSet

train_set = PascalVOC2012SegmentationDataSet(
     root='.../pascal_voc_2012',
     list_file='VOCdevkit/VOC2012/ImageSets/Segmentation/train.txt',
     samples_sub_directory='VOCdevkit/VOC2012/JPEGImages',
     targets_sub_directory='VOCdevkit/VOC2012/SegmentationClass',
     ...
 )
valid_set = PascalVOC2012SegmentationDataSet(
     root='.../pascal_voc_2012',
     list_file='VOCdevkit/VOC2012/ImageSets/Segmentation/val.txt',
     samples_sub_directory='VOCdevkit/VOC2012/JPEGImages',
     targets_sub_directory='VOCdevkit/VOC2012/SegmentationClass',
     ...
 )

Pascal AUG 2012

Download pascal dataset

AUG 2012: https://www2.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/semantic_contours/benchmark.tgz

Unzip and organize it as below:

   pascal_voc_2012
       └──VOCaug
           ├── aug.txt
           └── dataset
                 ├──inst
                 ├──img
                 └──cls

Instantiate the dataset:

from super_gradients.training.datasets import PascalAUG2012SegmentationDataSet

train_set = PascalAUG2012SegmentationDataSet(
     root='.../pascal_voc_2012',
     list_file='VOCaug/dataset/aug.txt',
     samples_sub_directory='VOCaug/dataset/img',
     targets_sub_directory='VOCaug/dataset/cls',
     ...
 )

NOTE: this dataset is only available for training. To test, please use PascalVOC2012SegmentationDataSet.

Pascal AUG & VOC 2012

Download pascal datasets:

Unzip and organize it as below:

   pascal_voc_2012
       ├─VOCdevkit
       │ └──VOC2012
       │    ├──JPEGImages
       │    ├──SegmentationClass
       │    ├──ImageSets
       │    │    ├──Segmentation
       │    │    │   └── train.txt
       │    │    ├──Main
       │    │    ├──Action
       │    │    └──Layout
       │    ├──Annotations
       │    └──SegmentationObject
       └──VOCaug
           ├── aug.txt
           └── dataset
                 ├──inst
                 ├──img
                 └──cls

Instantiate the dataset:

from super_gradients.training.datasets import PascalVOCAndAUGUnifiedDataset
train_set = PascalVOCAndAUGUnifiedDataset(root='.../pascal_voc_2012', ...)

NOTE: this dataset is only available for training. To test, please use PascalVOC2012SegmentationDataSet.

Supervisely Persons

Download supervisely dataset:

https://deci-pretrained-models.s3.amazonaws.com/supervisely-persons.zip

Unzip:

   supervisely-persons
    ├──images
    │    ├──image-name.png
    │    └──...
    ├──images_600x800
    │    ├──image-name.png
    │    └──...
    ├──masks
    └──masks_600x800

Instantiate the dataset:

from super_gradients.training.datasets import SuperviselyPersonsDataset
train_set = SuperviselyPersonsDataset(root_dir='.../supervisely-persons', list_file='train.csv', ...)
valid_set = SuperviselyPersonsDataset(root_dir='.../supervisely-persons', list_file='val.csv', ...)

NOTE: this dataset is only available for training. To test, please use PascalVOC2012SegmentationDataSet.

Pose Estimation Datasets

COCO 2017

Download coco dataset:

annotations: http://images.cocodataset.org/annotations/annotations_trainval2017.zip
train2017: http://images.cocodataset.org/zips/train2017.zip
val2017: http://images.cocodataset.org/zips/val2017.zip

Unzip and organize it as below:

    coco
    ├── annotations
    │      ├─ person_keypoints_train2017.json
    │      ├─ person_keypoints_val2017.json
    │      └─ ...
    └── images
        ├── train2017
        │   ├─ 000000000001.jpg
        │   └─ ...
        └── val2017
            └─ ...

Instantiate the dataset:

from super_gradients.training.datasets import COCOKeypointsDataset
train_set = COCOKeypointsDataset(data_dir='.../coco', images_dir='images/train2017', json_file='annotations/instances_train2017.json', ...)
valid_set = COCOKeypointsDataset(data_dir='.../coco', images_dir='images/val2017', json_file='annotations/instances_val2017.json', ...)

Oriented Box Detection Datasets

DOTA 2.0

Download DOTA dataset: https://captain-whu.github.io/DOTA/dataset.html
Unzip and organize it as below:

    dota
    └── train
        ├── images
        │   ├─ 000000000001.jpg
        │   └─ ...
        └── ann
            └─ 000000000001.txt
    └── val
        ├── images
        │   ├─ 000000000002.jpg
        │   └─ ...
        └── ann
            └─ 000000000002.txt

Run script to slice the dataset into tiles:

python src/super_gradients/examples/dota_prepare_dataset/dota_prepare_dataset.py --data_dir <path-to>/dota --output_dir <path-to>/dota_tiles

Specify path to the sliced dataset in the dataset (CLI):

python -m super_gradients.train_from_recipe --config-name yolo_nas_r_s_dota dataset_params.data_dir=<path-to>/dota_tiles

Specify path to the sliced dataset in the dataset (YAML):

dataset_params:
  train_dataset_params:
    data_dir: <path-to>/dota_tiles/train
  val_dataset_params:
    data_dir: <path-to>/dota_tiles/train

Specify path to the sliced dataset in the dataset (CODE):


from super_gradients.training.datasets import DOTAOBBDataset

train_loader = DOTAOBBDataset(data_dir="<path-to>/dota_tiles/train", ...)

Tip!

Press p or to see the previous file or, n or to see the next file

Dataset_Setup_Instructions.md 15 KB

Permalink History Raw

Computer Vision Datasets Setup

Classification Datasets

Detection Datasets

Segmentation Datasets

Pose Estimation Datasets

Oriented Box Detection Datasets

Comments

Use AWS S3 as storage!

Specify your S3 bucket

Access key (If needed)

Congratulations!

Use Google Cloud Storage!

Specify your Google Storage bucket

Service Account Key

Congratulations!

Use Azure Cloud Storage!

Specify your Azure Storage bucket

Access key (If needed)

Congratulations!

Use any S3 compatible storage!

Specify your S3 bucket

Access key (If needed)

Congratulations!

Deci-AI / super-gradients connected to https://github.com/Deci-AI/super-gradients.git

Dataset_Setup_Instructions.md 15 KB Permalink History Raw

Computer Vision Datasets Setup

Classification Datasets

Detection Datasets

Segmentation Datasets

Pose Estimation Datasets

Oriented Box Detection Datasets

Comments

Use AWS S3 as storage!

Specify your S3 bucket

Access key (If needed)

Congratulations!

Use Google Cloud Storage!

Specify your Google Storage bucket

Service Account Key

Congratulations!

Use Azure Cloud Storage!

Specify your Azure Storage bucket

Access key (If needed)

Congratulations!

Use any S3 compatible storage!

Specify your S3 bucket

Access key (If needed)

Congratulations!

Deci-AI
/
super-gradients
connected to https://github.com/Deci-AI/super-gradients.git

Dataset_Setup_Instructions.md 15 KB

Permalink History Raw