Skip to content

Dataset Importation

Zero provides comprehensive support for various custom dataset formats, making it simple to import them into the Web App. Additionally, some Kompress Algorithms can execute without requiring any custom datasets.

Dataset Preparation Guidelines

Image Classification

For image classification tasks, Zero is compatible with the standard ImageNet dataset format. This format necessitates organizing images into folders representing respective categories, with validation splits already performed. The prescribed structure is as follows:

ImageNetDataset/
| -- train
|   | -- Class_A
|   |   | -- image1.jpg
|   |   | -- image2.jpg
|   |   | -- ...
|   | -- Class_B
|   |   | -- image1.jpg
|   |   | -- image2.jpg
|   |   | -- ...
|   | -- ...
| -- val
|   | -- Class_A
|   |   | -- image1.jpg
|   |   | -- image2.jpg
|   |   | -- ...
|   | -- Class_B
|   |   | -- image1.jpg
|   |   | -- image2.jpg
|   |   | -- ...
|   | -- ...

Object Detection

Zero extends support to both COCO and VOC dataset formats for object detection tasks. The formats are detailed below:

COCO Format

CocoDataset/
| -- train2017
|   | -- (Contains training images)
| -- val2017
|   | -- (Contains validation images)
| -- annotations
|   | -- instances_train2017.json  (Train Annotations)
|   | -- instances_val2017.json    (Val Annotations)

VOC Format

VOCdevkit/
| -- VOC2012
|   | -- Annotations
|   |   | -- (Contains annotation files)
|   | -- JPEGImages
|   |   | -- (Contains JPEG images)

Segmentation

For image segmentation tasks, NYUN-KOMPRESS exclusively supports the VOC Format. The layout is structured as follows:

VOCdevkit/
| -- VOC2012
|   | -- Annotations
|   |   | -- (Contains annotation files)
|   | -- JPEGImages
|   |   | -- (Contains JPEG images)
|   | -- ImageSets
|   |   | -- (Contains ImageSets files)
|   | -- SegmentationClass
|   |   | -- (Contains SegmentationClass files)
|   | -- SegmentationObject
|   |   | -- (Contains SegmentationObject files)

NYUN-Adapt requires the dataset to be in the following format : 

custom_data/
| -- dataset
|   | -- images (images in jpg format)
|   | -- labels (segmentation maps in png format)
|   | -- splits 
|   |   | -- train.txt (names of the images for training (without the extension))
|   |   | -- val.txt (names of the images for validation (without the extension))

Pose Detection

For pose detection tasks, Zero exclusively supports the COCO-Pose Format. The layout is structured as follows:

coco-pose/
| -- images
|   | -- train2017
|   |   | -- Contains training images(JPEG format)
|   | -- val2017
|   |   | -- Contains testing images (JPEG images)
| -- annotations
|   |   | -- person_keypoints_train2017.json (Contains annotation file in coco-pose format)
|   |   | -- person_keypoints_val2017.json (Contains annotation file in coco-pose format)

LLM Text Generation

Zero supports loading any text dataset compatible with the Hugging Face datasets.load_dataset for HF datasets or datasets.load_from_disk for custom datasets. It supports two main dataset formats:

LLM - Single Column

This format is suitable for use cases where the dataset is already formatted with a single text column. For example, "wikitext" dataset with "text" as the TEXT_COLUMN.

LLM - Multi Columns

This format is suitable for datasets with multiple columns. The multi-column dataset can also handle simple formatting of the dataset for instructional use cases. (See example usage below)

Please note that there are no limitations regarding the loading of datasets with either single-column or multi-column structures. However, it is advisable to opt for loading multi-column datasets using option 2, especially when straightforward formatting is desired.

Parameters for Dataset Loading:

Parameter Default Value Description
DATASET_SUBNAME null Subname of the dataset if applicable.
TEXT_COLUMN text Specifies the text columns to be used.
If multiple columns are present, they should be
separated by commas.
SPLIT train Specifies the split of the dataset to load,
such as 'train', 'validation', or 'test'.
FORMAT_STRING null If provided, this string is
used to format the dataset.
It allows for customization
of the dataset's text representation.

Dataset Formatting:

The process responsible for formatting the dataset based on the provided parameters follows these steps:

  1. If no format string is provided, the TEXT_COLUMN and the dataset are used as is.
  2. If a format string is provided, it is applied to format the dataset. The format string should contain placeholders for the columns specified in TEXT_COLUMN.
  3. Once the dataset is formatted, a log message is generated to indicate that the format string was found and used. It also displays a sample of the formatted dataset.
  4. Finally, the dataset is mapped to replace the original "text" columns with the newly formatted "text" column (or) a new "text" column is created.

Example usage for Multi-column dataset:

Params for Alpaca Dataset (yahma/alpaca-cleaned)

DATASET_SUBNAME - null
TEXT_COLUMN - input,output,instruction
SPLIT - train
FORMAT_STRING - Instruction:\n{instruction}\n\nInput:\n{input}\n\nOutput:\n{output}

Alpaca dataset before Formatting

input instruction output
How to bake a cake Step 1: Preheat the oven to... A delicious homemade cake...
Introduction to Python Python is a high-level... Learn Python programming...
History of the Roman Empire The Roman Empire was... Explore the rise and fall...
... ... ...

Alpaca dataset after formatting with the input params

input instruction output text
How to bake a cake Step 1: Preheat the oven to... A delicious homemade cake... Instruction:\nStep 1: Preheat...\n\nInput:\nHow to bake a cake\n\nOutput:\nA delicious homemade cake...
Introduction to Python Python is a high-level... Learn Python programming... Instruction:\nPython is a hi...\n\nInput:\nIntroduction to Python\n\nOutput:\nLearn Python programming...
History of the Roman Empire The Roman Empire was... Explore the rise and fall... Instruction:\nThe Roman Emp...\n\nInput:\nHistory of the Roman Empire\n\nOutput:\nExplore the rise and fall...
... ... ... ...

Text Classification

Zero supports token classification and text clasification. Users can load any text dataset compatible with the Hugging Face datasets.load_dataset for HF datasets or datasets.load_from_disk for custom datasets. Make sure the uploaded dataset has the following format:

Parameter Data Type Default Value Description
input_column str "text" The name of the input text column : "tokens" - for token classification (new, pos, chunk) and "text" - for text classification
target_column str "label" the target column in the dataset : "ner_tags" : NER , "pos_tags": POS tagging, "chunk_tags": Chunking, "label" : text classification

Example dataset formats :

TEXT CLASSIFICATION DATASET

text label
I rented I AM CURIOUS-YELLOW from my video store because of all the controversy that surrounded it when it was first released in 1967. I also... 0
"I Am Curious: Yellow" is a risible and pretentious steaming pile. It doesn't matter what one's political views are because this film can hardly be taken... 0
If only to avoid making this type of film in the future. This film is interesting as an experiment but tells no cogent story.

One might...		 0
This film was probably inspired by Godard's Masculin, féminin and I urge you to see that film instead.

The film has two strong elements and...		 0
Oh, brother...after hearing about this ridiculous film for umpteen years all I can think of is that old Peggy Lee song..

"Is that all there is?"...		 0

TOKEN CLASSIFICATION DATASET

id tokens pos_tags
0 ["EU", "rejects", "German", "call", "to", ...] [22, 42, 16, 21, 35, 37, 16, 21, ...]
1 ["Peter", "Blackburn"] [22, 22]
2 ["BRUSSELS", "1996-08-22"] [22, 11]
3 ["The", "European", "Commission", "said", "on", ...] [12, 22, 22, 38, 15, 22, 28, 38, ...]
4 ["Germany", "'s", "representative", "to", ...] [22, 27, 21, 35, 12, 22, 22, 27, ...]

Summarization

Zero supports loading any text dataset compatible with the Hugging Face datasets.load_dataset for HF datasets or datasets.load_from_disk for custom datasets. Make sure the uploaded dataset has the following format:

document summary
The full cost of damage in Newton Stewart, one of the areas worst affected, is still... Clean-up operations are continuing across the Scottish Borders and Dumfries and...
A fire alarm went off at the Holiday Inn in Hope Street at about 04:20 BST on Saturday... Two tourist buses have been destroyed by fire in a suspected arson attack in Belfas...
Ferrari appeared in a position to challenge until the final laps, when the Mercedes... Lewis Hamilton stormed to pole position at the Bahrain Grand Prix ahead of Mercedes...
John Edward Bates, formerly of Spalding, Lincolnshire, but now living in London,... A former Lincolnshire Police officer carried out a series of sex attacks on...
Patients and staff were evacuated from Cerahpasa hospital on Wednesday after a man... An armed man who locked himself into a room at a psychiatric hospital in Istanbul has...

The dataset params for summarization tasks in Adapt are as follows :

Key Value Description
DATASET_SUBNAME null Subname of the dataset if applicable.
input_column 'document' Name of the input column containing the text corpus
target_column 'summary' Name of the target column containing the summarized text

Question Answering

The dataset for question answering must follow the general Adapt dataset. Adapt currently support extractive question-answering and hence requires :

  • CONTEXT - Text column that contains the context
  • QUESTION - Contains the Question
  • ANSWER - A column containing dictionary entries with the answer and index of context from which the answer starts.
    example : { "text": [ "...answer text..." ], "answer_start": [ 276 ] }

The column names can be different from this, but they should be mentioned in

  • input_column : 'context'
  • input_question_column : 'question'
  • target_column : 'answer'

Dataset arguments in question answering:

Key Value Description
DATASET_SUBNAME null Subname of the dataset if applicable.
input_column 'context' Name of the input column
input_question_column 'question' Name of the input question column
target_column 'answer' Name of the target column
squad_v2_format False #True Whether the data follows SQUAD-V2 Format (True/False)

Example of a default dataset for Question Answering :

context question answers
Beyoncé Giselle Knowles-Carter ... When did Beyonce start becoming... { "text": [ "in the late 1990s" ],...}
Beyoncé Giselle Knowles-Carter What areas did Beyonce compete in... { "text": [ "singing and dancing" ],... }
Beyoncé Giselle Knowles-Carter When did Beyonce leave Destiny's Chil... { "text": [ "2003" ], "answer_start": [ 526 ...] }
Beyoncé Giselle Knowles-Carter In what city and state did Beyonce... { "text": [ "Houston, Texas" ],... }

Translation

Zero supports loading any text dataset compatible with the Hugging Face datasets.load_dataset for HF datasets or datasets.load_from_disk for custom datasets. Make sure the uploaded dataset has the following format:

id translation
0 { "ca": "Source: Project GutenbergTranslation: Josep Carner", "de": "Source: Project Gutenberg" }
1 { "ca": "Les Aventures De Tom Sawyer", "de": "Die Abenteuer Tom Sawyers" }
2 { "ca": "Mark Twain", "de": "Mark Twain" }
3 { "ca": "PREFACI.", "de": "Vorwort des Autors." }
4 { "ca": "La major part de les aventures consignades en aquest llibre succeïren de bo de bo; una o dues són experiments de la meva collita; la resta pertanye...", "de": "" }

Dataset Arguments in Translation:

Key Value Description
source_lang '' The key of source language as per the given dataset
target_lang '' The key of target language as per the given dataset
DATASET_SUBNAME null Subset for the dataset (for multilingual datasets, the subname generally represents the pair of language used for translation)

Importing Your Dataset

There are two different ways to import your dataset into Zero:

Using Data File Upload in the Web-App

Utilize the Import Functionality in the Models and Datasets Tab.

Using Dataset Relative Folder Path

Users can keep the dataset uploaded into the connected infrastructure and provide the "Dataset Relative Folder Path" - relative location with respect to "User Data/Model Folder Path" as provided while connecting the infrastructure.

An example folder structure for relative datasets can be:

/home/nyunuser/prayog/custom_data/
| -- datasets
|   | -- ImageNetDataset
|   |   | -- (... as above)
|   | -- CocoDataset
|   |   | -- (... as above)
|   | -- Samvaad
|   |   | -- ...

And hence the "Dataset Relative Folder Path" can be:

  1. datasets/ImageNetDataset
  2. datasets/CocoDataset
  3. datasets/Samvaad

Note: For LLM tasks, the data folder must be loadable by datasets.load_from_disk and should return a datasets.DatasetDict object and not a datasets.Dataset object.