Getting started

In this chapter you’ll be introduced to the basic usage of dbcollection and its most commonly used features. More advanced methods and features will be addressed in the next chapters, but for now lets take a quick look at how we can start using this package.

Before you start using this package, there are a few things you should be aware in order to improve your experience when managing datasets. First, setting up the paths to store data/metada files is important in order to keep things organized and in the same place. Next, you should also take a look at the available datasets data file’s size in disk in order to check if you have enough disk space to store them.

After this is done, take a look at the contents of the package before proceeding to learning how to use the API of this module in the basic usage section.

Setting up paths

dbcollection uses a json cache file stored in your home directory (~/dbcollection.json) to log what datasets have been loaded, what tasks have been used and where data is stored. When first installing this package, the default paths where downloaded data files and .h5 metadata files are stored are defined in the cache file to store data in ~\dbcollection\.

You can chose to store your data files or metadata files into separate dirs on a disk or on separate disks by modifying the paths of these two fields: default_cache_dir and default_download_dir. You can either manually change the paths in the ~\dbcollection.json file under the info key or you can do it by using dbcollection. To use the latter, you can do the following:

>>> # import the package
>>> import dbcollection as dbc
>>> # directly access the fields and assign new paths
>>> dbc.cache.cache_dir = 'new/cache/path'
>>> dbc.cache.download_dir = 'new/download_data/path'

This will change the paths where the metadata files (cache_dir) and the downloaded files (download_dir) are stored in disk. In case you want to reset the paths to the original defaults you can simply do:

>>> # reset the metadata cache dir to the default path
>>> dbc.cache.reset_cache_dir()
>>> # reset the data download dir to the default path
>>> dbc.cache.reset_download_dir()

With this you should be able to easily locate where data/metadata files are being stored in disk without having to manually specify a path everytime you setup a new dataset.

Contents of the package

The dbcollection package comes with several API methods and other features to help managing datasets with very little overhead. These include:

  • API methods for managing datasets and fetching data;
  • API for managing/querying the cache file;
  • A list of available datasets;
  • utility methods for parsing loaded data, load different types of data files, downloading urls, string convertions, or constructing a tree of files in a dir.

The following sections will describe in more detail these features one by one and how to use them. Next, comes a brief tutorial on how to use this package to quickly start hacking new stuff with it.

Basic usage

To use this package you first need to import it.

>>> import dbcollection as dbc

Then, to load a dataset, all you need to do is call the load() method with the name of the dataset you want to load. For example, lets load the mnist dataset.

>>> mnist = dbc.load('mnist')

This returns a DataLoader object which contains all necessary methods to fetch data for this dataset.

Notice the name of the dataset is all lower case. The name of a dataset must be the exact one, and to have the right one you should check the list of available datasets to see the correct name. To do this you can use the info_datasets() method to list all available datasets names and tasks.

>>> dbc.info_datasets()
Available datasets in cache for load
- mnist  ['classification']

Available datasets for download
- caltech_pedestrian  ['detection', 'detection_10x', 'detection_30x']
- cifar10  ['classification']
- cifar100  ['classification']
- coco  ['caption_2015', 'caption_2016', 'detection_2015', 'detection_2016', 'keypoints_2016']
- flic  ['keypoints']
- ilsvrc2012  ['classification', 'raw256']
- inria_pedestrian  ['detection']
- leeds_sports_pose  ['keypoints', 'keypoints_original']
- leeds_sports_pose_extended  ['keypoints']
- mnist  ['classification']
- mpii_pose  ['keypoints', 'keypoints_full']
- pascal_voc_2007  ['detection']
- pascal_voc_2012  ['detection']
- ucf_101  ['recognition']
- ucf_sports  ['recognition']

This returns two lists, one for used datasets on your system, and the other is a list of all available datasets to download and their respective tasks for processing. Please notice that the mnist dataset we’ve just loaded has the classification task setup. This is due to this task being the default task that is selected if no task is specified at loading time. Also, a list of all available tasks is displayed in the Available datasets for download list.


For more information about the available datasets and tasks see here.

Returning to the previous example about loading the mnist dataset, the resulting data loading object contains several methods to fetch data from the metadata file, as well as other information like the task name, the set splits, where the data files are located, etc.

>>> mnist.
mnist.data_dir          mnist.hdf5_filepath     mnist.object_field_id(  mnist.size(
mnist.db_name              mnist.object_fields     mnist.task
mnist.get(              mnist.list(             mnist.root_path         mnist.test
mnist.hdf5_file         mnist.object(           mnist.sets              mnist.train

The API methods for fetching and querying the metadata file are quite handy. For example, to see how the metadata file is structured and what data fields it contains, you simply have to use the info() method in order to have an idea of how data is organized.


> Set: test
- classes,        shape = (10, 2),          dtype = uint8
- images,         shape = (10000, 28, 28),  dtype = uint8,  (in 'object_ids', position = 0)
- labels,         shape = (10000,),         dtype = uint8,  (in 'object_ids', position = 1)
- object_fields,  shape = (2, 7),           dtype = uint8
- object_ids,     shape = (10000, 2),       dtype = uint8

(Pre-ordered lists)
- list_images_per_class,  shape = (10, 1135),  dtype = int32

> Set: train
- classes,        shape = (10, 2),          dtype = uint8
- images,         shape = (60000, 28, 28),  dtype = uint8,  (in 'object_ids', position = 0)
- labels,         shape = (60000,),         dtype = uint8,  (in 'object_ids', position = 1)
- object_fields,  shape = (2, 7),           dtype = uint8
- object_ids,     shape = (60000, 2),       dtype = uint8

(Pre-ordered lists)
- list_images_per_class,  shape = (10, 6742),  dtype = int32

This way, you get a general idea of how the dataset’s data is split and what fields compose each set, and also their type or shape. This method and its output are described in more detail in the fetching data section.

To fetch data, you can use two methods to retrieve a chunk of data by using the get() and object() methods. These two are complementary to one another, but when you need to fetch data from a single field you use the get() method, and when you need to retrieve data from a set of fields you’ll use the object() method. For example, lets retrieve the first 10 images from the training set of mnist:

>>> imgs = mnist.get('train', 'images', range(10))
>>> imgs.shape
(10, 28, 28)

Fetching data is simple! If can retrieve this same data in two other ways. The first way is to grab the train set data altogether and then using the same method:

>>> train = mnist.train  # get a data loader object of the train set
>>> train
SetLoader: set<train>, len<60000>
>>> imgs = train.get('images', range(10))
>>> imgs.shape
(10, 28, 28)

The difference here is that you can grab the train set as a separate object and do all your operations with it. Also, here you don’t have to explicitly define the set to fetch data from, but you still have to define the field name.

The second way you can fetch data is by targeting the actual field you want to retrieve data from. Just like the previous examples, we can grab the first 10 images from the mnist train set in the following ways:

>>> # First way
>>> images = mnist.train.images  # get a data loader object of the images field
>>> images
FieldLoader: <HDF5 dataset "images": shape (60000, 28, 28), type "|u1">
>>> images.get(range(10))
>>> imgs.shape
(10, 28, 28)

>>> # Second way
>>> imgs = images[0:10]
>>> imgs.shape
(10, 28, 28)

For single fields you can do array slicing operations likewith numpy arrays. All of these operations convey the same results, and it is up to the user to decide which one fits his/hers needs best.

We’ve see so far how fetching data from single fields is done, but most cases you want to grab sets of related data fields like, for example, the image and label. This information is conveyd by two key fields in the metadata files that relate different fields ids with each other: the object_fields and object_ids fields. The object_ids field is a list of indexes of fields defined in the object_fields field. So, to get the right label for a given image you just need to collect the ids of each field and then fetch their data. To do this, we’ll use the object() method to grab the ids of the fields for the 100th item:

>>> # Grab the ids of the image and label fields of the 100th element
>>> ids = mnist.object('train', 99)
>>> ids
array([99,  1], dtype=uint8)
>>> # Fetch the image data
>>> img = mnist.get('train', 'images', ids[0])
>>> img.shape
(28, 28)
>>> # Fetch the label data
>>> label = mnist.get('train', 'labels', ids[1])
>>> label

Another way you can do this to get the same data, without having to manually fetch data from every field, is to use the convert_to_value argument in object() and set it to True. This will automatically fetch the data of all fields and return them in a list.

>>> # Grab the ids of the image and label fields of the 100th element
>>> (img, label) = mnist.object('train', 99, convert_to_value=True)
>>> img.shape
(28, 28)
>>> label

As you can see, this can be quite handy when multiple fields compose an object element. You’ll mostly use a combination of get() and object() to fetch data from a dataset in your code, and this is all you’ll probably need.

The two last methods I would like to point to are list() and size(). The list() method lists all data fields available for each set and the size() method returns the size of a field. The purpose of these methods is to mearly serve as information source for the user.

With this information, you should be able to have a sufficient understanding of how dbcollection works and how to take advantage of its features. In the tutorial we’ll dive deeper on more advanced features and use cases that can help you get more out of this module.