## Introduction

• The goal of this tutorial is to explain how to structure a deep learning project.
• Splitting your data into training, dev and test sets can be disastrous if not done correctly. In this tutorial, we will walk you through some best practices when splitting your dataset.

## Choosing train, train-dev, dev and test sets

• Guideline: Choose a dev set and test set to reflect data you expect to get in the future.
• It is important to choose the dev and test sets from the same distribution and it must be taken randomly from all the data.
• If the training set and dev sets have to have different distributions (due to the lack of sufficient data), it is good practice to introduce a train-dev set that has the same distribution as the training set. This train-dev set will be used to measure how much the model is overfitting.
• Guideline: The dev and test sets should be just big enough to represent accurately the performance of the model.
• The size of the dev and test set should be big enough for the dev and test results to be representative of the performance of the model. If the dev set has 100 examples, the dev accuracy can vary a lot depending on the chosen dev set. For bigger datasets ((\\rangle 1M\) examples), the dev and test set can have around 10,000 examples each for instance (only 1% of the total data).

## Objectives

• These guidelines translate into best practices for code:
• The split between train/dev/test should always be the same across experiments.
• Otherwise, different models are not evaluated in the same conditions.
• We should have a reproducible script to create the train/dev/test split.
• We need to test if the dev and test sets should come from the same distribution.

## Dataset structure

• The best and most secure way to split the data into these three sets is to have one directory for train, one for dev and one for test.
• For instance if you have a dataset of images, you could have a structure like this with $$80\%$$ in the training set, $$10\%$$ in the dev set and $$10\%$$ in the test set.

  data/
train/
img_000.jpg
...
img_799.jpg
dev/
img_800.jpg
...
img_899.jpg
test/
img_900.jpg
...
img_999.jpg


## Reproducibility is important

• Often a dataset will come either in one big set that you will split into train, dev and test. Academic datasets often come already with a train/test split (to be able to compare different models on a common test set). You will therefore have to build yourself the train/dev split before beginning your project.
• A good practice that is true for every software, but especially in machine learning, is to make every step of your project reproducible. It should be possible to start the project again from scratch and create the same exact split between train, dev and test sets.
• The cleanest way to do it is to have a build_dataset.py file that will be called once at the start of the project and will create the split into train, dev and test. Optionally, calling build_dataset.py can also download the dataset. We need to make sure that any randomness involved in build_dataset.py uses a fixed seed so that every call to python build_dataset.py will result in the same output.
• Never split data manually (by moving files into different folders one by one), because you wouldn’t be able to reproduce it.

## Implementation details

• Let’s illustrate the good practices with a simple example. We have filenames of images that we want to split into train, dev and test. Here is a way to split the data into three sets: 80% train, $$10\%) dev and \(10\%$$ test.

  filenames = ['img_000.jpg', 'img_001.jpg', ...]

split_1 = int(0.8 * len(filenames))
split_2 = int(0.9 * len(filenames))
train_filenames = filenames[:split_1]
dev_filenames = filenames[split_1:split_2]
test_filenames = filenames[split_2:]


### Ensure that train/dev/test belong to the same distribution (if possible)

• Often we have a big dataset and want to split it into train, dev and test set. In most cases, each split will have the same distribution as the others.
• What could go wrong?
• Suppose that we’ve divided our dataset into 10 groups. Let’s assume that the first 100 images (img_000.jpg to img_099.jpg) have label 0, the next 100 have label 1, … and the last 100 images have label 9.
• In that case, the above code will make the dev set only have label 8, and the test set only label 9.
• We therefore need to ensure that the filenames are correctly shuffled before splitting the data.

  filenames = ['img_000.jpg', 'img_001.jpg', ...]
random.shuffle(filenames)  # randomly shuffles the ordering of filenames

split_1 = int(0.8 * len(filenames))
split_2 = int(0.9 * len(filenames))
train_filenames = filenames[:split_1]
dev_filenames = filenames[split_1:split_2]
test_filenames = filenames[split_2:]

• This should give approximately the same distribution for train, dev and test sets. If necessary, it is also possible to split each class into $$80\%/10\%/10\%$$ so that the distribution is the same in each set.

### No randomness makes things reproducible!

• We talked earlier about making our dataset building script reproducible. Here we need to make sure that the train/dev/test split stays the same across every run of python build_dataset.py.
• The code in the above section doesn’t ensure reproducibility, since each time you run it you will have a different split.
• To make sure to have the same split each time this code is run, we need to fix the random seed before shuffling the filenames.
• Here is a good way to remove any randomness in the process:

  filenames = ['img_000.jpg', 'img_001.jpg', ...]
filenames.sort()  # make sure that the filenames have a fixed order before shuffling
random.seed(230)
random.shuffle(filenames) # shuffles the ordering of filenames (deterministic given the chosen seed)

split_1 = int(0.8 * len(filenames))
split_2 = int(0.9 * len(filenames))
train_filenames = filenames[:split_1]
dev_filenames = filenames[split_1:split_2]
test_filenames = filenames[split_2:]

• The call to filenames.sort() makes sure that if you build filenames in a different way, the output is still the same.

## Citation

If you found our work useful, please cite it as:

@article{Chadha2020DistilledSplittingDatasets,
title   = {Splitting Datasets},