Coco detection¶
This module allows to read multiple datasets in COCO JSON format from a relative path stored in the ~/.aloception/alodataset_config.json file. By default, if the configuration file does not exist or the database directory is not found, the module will perform a command line user prompt to store/overwrite alodataset_config.json file and the new database directory.
See also
CocoDetection pytorch module
Base dataset module
How to setup your data? tutorial.
Basic use¶
In order to use the module, img_folder and ann_file must be given as a minimum requirements:
from alodataset import CocoBaseDataset
coco_dataset = CocoBaseDataset(
img_folder = "val2017",
ann_file = "annotations/instances_val2017.json"
)
If the database does not exist or is wrong, a command line user prompt was executed:
1[WARNING] "coco" does not exist in config file. Please write the coco root directory: "user prompt"
2[WARNING] "old_path_directory" path does not exists for dataset: "coco". Please write the new directory: "user prompt"
After its correct initialization, the module allows the reading of individual images in two ways:
# Get a frame by index
frame0 = coco_dataset.getitem(0)
# Get a random frame in batch
framer = next(iter(coco_dataset.stream_loader()))
Also, a iterable object can be generated using train_loader() or stream_loader() functions:
# Frame iteration by individual images
for frame in coco_dataset.stream_loader():
# Render each image
frame.get_view().render()
# Frames iteration by images in a decided batches
for frame in coco_dataset.train_loader(batch_size = 2):
# Transform each list in a batch, and then render it
frames = Frame.batch_list(frames)
frames.get_view().render()
Note
The dataset and paths of COCO detection 2017 dataset were taken as example.
Change of default database¶
name attribute is equal to coco as default. Another dataset can be configurated
using the relevant paths and a decided name. An example of how use
Cottontail-Rabbits Dataset is showing below:
from alodataset import CocoBaseDataset
coco_dataset = CocoBaseDataset(
img_folder = "valid",
ann_file = "valid/_annotations.coco.json"
)
See also
For a custom dataset, see How setup your data tutorial.
Class filtering¶
Coco detection allows to select a desired set of classes, automatically managing the amount, labeling and ID change
corresponding to the case. The classes to be filtered are exposed by classes parameter which by default does not filter
the dataset to any specific class.
For finetune applications, it is usual to manipulate databases in order to get a single class (for example, people):
from alodataset import CocoBaseDataset
coco_dataset = CocoBaseDataset(
classes = ["person"],
img_folder = "val2017",
ann_file = "annotations/instances_val2017.json"
)
Now, coco_dataset should handled images with boxes that have person as label class.
Warning
Each element in the desired list of classes must be one element in CATEGORIES attribute.
Download a minimal dataset sample¶
For quick configuration issues, coco_detection_dataset has the option to download and use a sample of 8 examples
of the original dataset, without the need to download the entire dataset. To download and/or access these samples,
sample parameter has been declared:
from alodataset import CocoBaseDataset
coco_dataset = CocoBaseDataset(sample = True)
for frame in coco_dataset.stream_loader():
# Render each image
frame.get_view().render()
Warning
The purpose of the sample is to verify the correct functioning of the module. It is not advisable to use such small samples for training.
Coco detection API¶
Read a dataset in COCO JSON format and transform each element
in a Frame object. Ideal for object detection applications.
- class alodataset.coco_detection_dataset.CocoBaseDataset(img_folder=None, ann_file=None, name='coco', return_masks=False, classes=None, fix_classes_len=None, **kwargs)¶
Bases:
Generic[torch.utils.data.dataset.T_co]