Datasets

Each dataset in this competition comes from one of following 5 domains: image, video, speech, text or tabular. Every dataset is formatted in TFRecords and split into a train set (with true labels) and a test set (without true labels). The data loading process is done in the ingestion program (thus common to all participants), which parses these TFRecords to a `tf.data.Dataset` object. Each of its examples is of the form

(example, labels)

where `example` is a dense 4-D Tensor of dtype tf.float32 and of shape

(sequence_size, row_count, col_count, num_channels)

and `labels` is a 1-D Tensor of shape

(output_dim,).

Here `output_dim` represents number of classes of the multilabel classification task.

Metadata

The metadata of each dataset contains info such as the shape of examples, number of examples, number of classes, etc. These info can be accessed by calling different functions found at here.

Although the domain information is not given directly in the metadata, it can be inferred from metadata by a function similar to this one.

Specification for text datasets

Although it is straight-forward to interpret the 4-D Tensor representation of each example for most domains, we need to make some manual choices to encode text datasets. The choices we made are: 

• For English, split the original document by space to tokenize; For Chinese, consider each character as a token;
• Construct a vocabulary and map each of these tokens to an integer index;
• Replace each token by the index (cast as tf.float32);
• Each example (document) is then a sequence of integers.

The mapping from token to integer index can be accessed by calling

token_to_index = metadata.get_channel_to_index_map()

Embedding weights:

In the Docker image running by the platform (evariste/autodl:gpu-latest),  a built-in embedding model is provided for Chinese and English respectively, and the path of the embedding models is "/app/embedding". Both the embedding models are from fastText (Chinese, English). In addition, pre-trained weights using BERT can also be found in the same folder.

Alternative way to download the Docker image:

Since the Docker image (evariste/autodl:gpu-latest) is larger than 11GB, the usual way of downloading it using docker pull could be difficult. Thus we provide an alternative way for downloading:

1. Download the 3 files via: https://pan.baidu.com/s/1cxDDSZRSyGT6fH82cNkGRQ

2. Merge them using command:

   cat autodl-gpu-latest.tar.part* > autodl-gpu-latest.tar

3. Load the image to Docker using:
   

   docker load < autodl-gpu-latest.tar

Public datasets

We provide a list of public datasets. You will have access to the data (training set and test set) AND the true labels for these datasets. Notice that the video datasets do not include a sound track.

• num_train/num_test: number of training/test examples
• sequence_size/row_count/col_count/num_channels: shape of the examples. -1 means the value varies from one example to another.
• output_dim: number of classes

These data were re-formatted from original public datasets. If you use them, please make sure to acknowledge the original data donnors (see "Source" in the data table) and check the tems of use.

To download all public datasets at once:

cd autodl_starting_kit_stable
python download_public_datasets.py

Format and use your own datasets

We provide toolkit to participants to format their own datasets to the same format of this challenge. If you want to practice designing algorithms with your own datasets, follow these steps. 

Competition protocol

This challenge has two phases. This is the feedback phase: when you submit your code, you get immediate feedback on 5 feedback datasets. In the final test phase, you will be evaluated on several new datasets. Eligible participants to the final phase will be notified when and where to submit their code for a final blind test. The ranking in the final phase will count towards determining the winners.

Code submitted is trained and tested automatically, without any human intervention. Code submitted on "All datasets" is run on all five feedback datasets in parallel on separate compute workers, each one with its own time budget. 

The identities of the datasets used for testing on the platform are concealed. The data are provided in a raw form (no feature extraction) to encourage researchers to use Deep Learning methods performing automatic feature learning, although this is NOT a requirement. All problems are multi-label classification problems. The tasks are constrained by the time budget (20 minutes/dataset). 

Here is some pseudo-code of the evaluation protocol:

# For each dataset, our evaluation program calls the model constructor
# IMPORTANT: this initilization step doesn't consume time in the total time budget
#            so one should carry out meta-learning or loading pre-trained weights in this step.
#            This step should not exceed 20min. Otherwise the submission will fail.
M = Model(metadata=dataset_metadata)

# Initialize
remaining_time budget = overall_time_budget
start_time = time()

# Ingestion program calls multiple times train and test:
repeat until M.done_training or remaining_time_budget < 0
{
  M.train(training_data, remaining_time_budget)
  remaining_time_budget = start_time + overall_time_budget - time.time()

  results = M.test(test_data, remaining_time_budget)
  remaining_time_budget = start_time + overall_time_budget - time.time()

  # Results made available to scoring program (run in separate container)
  save(results)
}

It is the responsibility of the participants to make sure that neither the "train" nor the "test" methods exceed the “remaining_time_budget”. The method “train” can choose to manage its time budget such that it trains in varying time increments. There is pressure that it does not use all "overall_time_budget" at the first iteration because we use the area under the learning curve as metric.

Metrics

The participants can train in batches of pre-defined duration to incrementally improve their performance, until the time limit is attained. In this way we can plot learning curves: "performance" as a function of time. Each time the "train" method terminates, the "test" method is called and the results are saved, so the scoring program can use them, together with their timestamp.

We treat both multi-class and multi-label problems alike. Each label/class is considered a separate binary classification problem, and we compute the normalized AUC (NAUC or Gini coefficient)

    2 * AUC - 1

as score for each prediction, here AUC is the usual area under ROC curve (ROC AUC).

For each dataset, we compute the Area under Learning Curve (ALC). The learning curve is drawn as follows:

• at each timestamp t, we compute s(t), the normalized AUC (see above) of the most recent prediction. In this way, s(t) is a step function w.r.t time t;
• in order to normalize time to the [0, 1] interval, we perform a time transformation by
  
  where T is the time budget (of default value 1200 seconds = 20 minutes) and t0 is a reference time amount (of default value 60 seconds).
• then compute the area under learning curve using the formula
  
  we see that s(t) is weighted by 1/(t + t0)), giving a stronger importance to predictions made at the beginning of th learning curve.

After we compute the ALC for all 5 datasets, the overall ranking is used as the final score for evaluation and will be used in the learderboard. It is computed by averaging the ranks (among all participants) of ALC obtained on the 5 datasets.

Examples of learning curves:

Challenge Rules

• General Terms: This challenge is governed by the General ChaLearn Contest Rule Terms, the Codalab Terms and Conditions, and the specific rules set forth.
• Announcements: To receive announcements and be informed of any change in rules, the participants must provide a valid email.
• Conditions of participation: Participation requires complying with the rules of the challenge. Prize eligibility is restricted by US government export regulations, see the General ChaLearn Contest Rule Terms. The organizers, sponsors, their students, close family members (parents, sibling, spouse or children) and household members, as well as any person having had access to the truth values or to any information about the data or the challenge design giving him (or her) an unfair advantage, are excluded from participation. A disqualified person may submit one or several entries in the challenge and request to have them evaluated, provided that they notify the organizers of their conflict of interest. If a disqualified person submits an entry, this entry will not be part of the final ranking and does not qualify for prizes. The participants should be aware that ChaLearn and the organizers reserve the right to evaluate for scientific purposes any entry made in the challenge, whether or not it qualifies for prizes.
• Dissemination: The challenge is part of the official selection of the NeurIPS 2019 conference. Top ranking participants will be invited to submit a paper to a special issue on Automated Machine learning of the IEEE transactions PAMI.
• Registration: The participants must register to Codalab and provide a valid email address. Teams must register only once and provide a group email, which is forwarded to all team members. Teams or solo participants registering multiple times to gain an advantage in the competition may be disqualified.
• Anonymity: The participants who do not present their results at the workshop can elect to remain anonymous by using a pseudonym. Their results will be published on the leaderboard under that pseudonym, and their real name will remain confidential. However, the participants must disclose their real identity to the organizers to claim any prize they might win. See our privacy policy for details.
• Submission method: The results must be submitted through this CodaLab competition site. The number of submissions per day and maximum total computational time are restrained and subject to change, according to the number of participants. Using multiple accounts to increase the number of submissions is NOT permitted. In case of problem, send email to autodl@chalearn.org. The entries must be formatted as specified on the Instructions page.
• Prizes: The three top ranking participants in the Final phase blind testing may qualify for prizes. To compete for prizes, the participants must make a valid submission on the Feed-back phase; their last submission in the Feed-back phase will be forwarded to the Final phase for ranking using five or more new datasets. Only submissions ranked better than one of the baseline3 (code in Quick Start page) performances (i.e. among the performances of the accounts: baseline3, baseline3_a, baseline3_b, baseline3_c, baseline3_d) in the Feed-back phase will be eligible for the final ranking. And only participants whose performances exceed that of the best Baseline3 entry will qualify for winning prizes in the Final phase. To be ranked, the participants must also fill out a fact sheet (link to be given after the end of competition) briefly describing their methods. There is no other publication requirement. The winners will be required to make their code publicly available under an OSI-approved license such as, for instance, Apache 2.0, MIT or BSD-like license, if they accept their prize, within a week of the deadline for submitting the final results. Entries exceeding the time budget will not qualify for prizes. In case of a tie, the prize will go to the participant who submitted his/her entry first. Non winners or entrants who decline their prize retain all their rights on their entries and are not obliged to publicly release their code.

Frequently Asked Questions

Can organizers compete in the challenge?

No, they can make entries that show on the leaderboard for test purposes and to stimulate participation, but they are excluded from winning prizes. Excluded entrants include: baseline0, baseline1, baseline2, baseline3, baseline3_a, baseline3_b, baseline3_c, baseline3_d, baiyu, eric, hugo.jair, juliojj, Lukasz, madclam, Pavao, shangeth, thomas, tthomas, Zhen, Zhengying.

Are there prerequisites to enter the challenge?

No, except accepting the TERMS AND CONDITIONS.

Can I enter any time?

Yes, until the challenge deadline.

Where can I download the data?

You can download "public data" only from the Data page. The data on which your code is evaluated cannot be downloaded, it will be visible to your code only, on the Codalab platform.

How do I make submissions?

To make a valid challenge entry, make sure to click first the orange button "All datasets", then click the blue button on the upper right side "Upload a Submission". This will ensure that you submit on all 5 datasets of the challenge simultaneously. You may also make a submission on a single dataset for debug purposes, but it will not count towards the final ranking.

Do you provide tips on how to get started?

We provide a Starting Kit in Python with step-by-step instructions in a Jupyter notebook called "tutorial.ipynb", which can be found in the github repository https://github.com/zhengying-liu/autodl_starting_kit_stable. You can also have a well rendered preview here.

Are there publication opportunities?

Yes. Top ranking participants will be invited to submit papers to a special issue of the IEEE transaction journal PAMI on Automated Machine Learning and will be entered in a contest for the best paper. Deadline to be define.

There will be 2 best paper awards of $1000 ("best paper" and "best student paper").

Are there prizes?

Yes, a 4000 USD prize pool.

Do I need to submit code to participate?

Yes, participation is by code submission.

When I submit code, do I surrender all rights to that code to the SPONSORS or ORGANIZERS?

No. You just grant to the ORGANIZERS a license to use your code for evaluation purposes during the challenge. You retain all other rights.

If I win, I must submit a fact sheet, do you have a template?

Yes, please download it [HERE].

What is your CPU/GPU computational configuration?

We are running your submissions on Google Cloud NVIDIA Tesla P100 GPUs. In non peak times we are planning to use 10 workers, each of which will have one NVIDIA Tesla P100 GPU (running CUDA 10 with drivers cuDNN 7.5) and 4 vCPUs, with 26 GB of memory, 100 GB disk.

The PARTICIPANTS will be informed if the computational resources increase. They will NOT decrease.

Can I pre-train a model on my local machine and submit it?

This is not explicitly forbidden, but it is discouraged. We prefer if all calculations are performed on the server. If you submit a pre-trained model, you will have to disclose it in the fact sheets. 

Will there be a final test round on separate datasets?

YES. The ranking of participants will be made from a final blind test made by evaluating a SINGLE SUBMISSION made on the final test submission site. The submission will be evaluated on five new datasets in a completely "blind testing" manner. The final test ranking will determine the winners.

What is my time budget?

Each execution must run in less than 20 minutes (1200 seconds) for each dataset.

Does the time budget correspond to wall time or CPU/GPU time?

Wall time.

My submission seems stuck, how long will it run?

In principle no more than its time budget. We kill the process if the time budget is exceeded. Submissions are queued and run on a first time first serve basis. We are using several identical servers. Contact us if your submission is stuck more than 24 hours. Check on the leaderboard the execution time.

How many submissions can I make?

Five per day (and up to a total of 100), but up to a total computational time of 5 hours (submissions taking longer will be aborted). This may be subjet to change, according to the number of participants. Please respect other users. It is forbidden to register under multiple user IDs to gain an advantage and make more submissions. Violators will be DISQUALIFIED FROM THE CONTEST.

Do my failed submissions count towards my number of submissions per day?

No. Please contact us if you think the failure is due to the platform rather than to your code and we will try to resolve the problem promptly.

What happens if I exceed my time budget?

This should be avoided. In the case where a submission exceeds 20 minutes of time budget for a particular task (dataset), the submission handling process (ingestion program in particular) will be killed when time budget is used up and predictions made so far (with their corresponding timestamps) will be used for evaluation. In the other case where a submission exceeds the total compute time per day, all running tasks will be killed by CodaLab and the status will be marked 'Failed' and a score of -1.0 will be produced.

The time budget is too small, can you increase it?

No sorry, not for this challenge.

What metric are you using?

All problems are multi-label problems and we treat them as multiple 2-class classification problems. For a given dataset, all binary classification problems are scored with the ROC AUC and results are averaged (over all classes/binary problems). For each time step at which you save results, this gives you one point on the learning curve. The final score for one dataset is the area under the learning curve. The overall score on all 5 datasets is the average rank on the 5 datasets. For more details, go to 'Get Started' -> 'Instructions' -> 'Metrics' section.

Which version of Python are you using?

The code was tested under Python 3.5. We are running Python 3.5 on the server and the same libraries are available.

Can I use something else than Python code?

Yes. Any Linux executable can run on the system, provided that it fulfills our Python interface and you bundle all necessary libraries with your submission.

Do I have to use TensorFlow?

No. We use TFRecords to format the datasets in a uniform manner, but you can use other software to process the data, including PyTorch (included in the Docker, see the following question).

Which docker are you running on Codalab?

evariste/autodl:gpu-latest, see the Dockerfile and some instructions on dockerhub.

How do I test my code in the same environment that you are using before submitting?

When you submit code to Codalab, your code is executed inside a Docker container. This environment can be exactly reproduced on your local machine by downloading the corresponding docker image. The docker environment of the challenge contains Anaconda libraries, TensorFlow, and PyTorch (among other things). 
Non GPU users, if you are new to Docker, follow these instructions to install docker. You may then use the docker evariste/autodl:cpu-latest. See details in the Starting Kit that can be downloaded from the Instructions page. GPU users, follow these more detailed instructions.

What is meant by "Leaderboard modifying disallowed"?

Your last submission is shown automatically on the leaderboard. You cannot choose which submission to select. If you want another submission than the last one you submitted to "count" and be displayed on the leaderboard, you need to re-submit it.

Can I register multiple times?

No. If you accidentally register multiple times or have multiple accounts from members of the same team, please notify the ORGANIZERS. Teams or solo PARTICIPANTS with multiple accounts will be disqualified.

How can I create a team?

We have disabled Codalab team registration. To join as a team, just share one account with your team. The team leader is responsible for making submissions and observing the rules.

How can I destroy a team?

You cannot. If you need to destroy your team, contact us.

Can I join or leave a team?

It is up to you and the team leader to make arrangements. However, you cannot participate in multiple teams.

Can cheat by trying to get a hold of the evaluation data and/or future frames while my code is running?

No. If we discover that you are trying to cheat in this way you will be disqualified. All your actions are logged and your code will be examined if you win.

Can I give an arbitrary hard time to the ORGANIZERS?

ALL INFORMATION, SOFTWARE, DOCUMENTATION, AND DATA ARE PROVIDED "AS-IS". UPSUD, CHALEARN, IDF, AND/OR OTHER ORGANIZERS AND SPONSORS DISCLAIM ANY EXPRESSED OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR ANY PARTICULAR PURPOSE, AND THE WARRANTY OF NON-INFRIGEMENT OF ANY THIRD PARTY'S INTELLECTUAL PROPERTY RIGHTS. IN NO EVENT SHALL ISABELLE GUYON AND/OR OTHER ORGANIZERS BE LIABLE FOR ANY SPECIAL, INDIRECT OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER ARISING OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE OF SOFTWARE, DOCUMENTS, MATERIALS, PUBLICATIONS, OR INFORMATION MADE AVAILABLE FOR THE CHALLENGE. In case of dispute or possible exclusion/disqualification from the competition, the PARTICIPANTS agree not to take immediate legal action against the ORGANIZERS or SPONSORS. Decisions can be appealed by submitting a letter to the CHALEARN president, and disputes will be resolved by the CHALEARN board of directors. See contact information.

Where can I get additional help?

For questions of general interest, THE PARTICIPANTS should post their questions to the forum.

Other questions should be directed to the organizers.

Credits

This challenge would not have been possible without the help of many people.

Main organizers:

• Olivier Bousquet (Google, Switzerland)
• André Elisseef (Google, Switzerland)
• Isabelle Guyon (U. Paris-Saclay; UPSud/INRIA, France and ChaLearn, USA)
• Zhengying Liu (U. Paris-Saclay; UPSud, France)
• Wei-Wei Tu (4paradigm, China)

Other contributors to the organization, starting kit, and datasets, include:

• Stephane Ayache (AMU, France)
• Hubert Jacob Banville (INRIA, France)
• Mahsa Behzadi (Google, Switzerland)
• Kristin Bennett (RPI, New York, USA)
• Hugo Jair Escalante (IANOE, Mexico and ChaLearn, USA)
• Sergio Escalera (U. Barcelona, Spain and ChaLearn, USA)
• Gavin Cawley (U. East Anglia, UK)
• Baiyu Chen (UC Berkeley, USA)
• Albert Clapes i Sintes (U. Barcelona, Spain)
• Bram van Ginneken (Radboud U. Nijmegen, The Netherlands)
• Alexandre Gramfort (U. Paris-Saclay; INRIA, France)
• Yi-Qi Hu (4paradigm, China)
• Julio Jacques Jr. (U. Barcelona, Spain)
• Meysam Madani (U. Barcelona, Spain)
• Tatiana Merkulova (Google, Switzerland)
• Adrien Pavao (U. Paris-Saclay; INRIA, France and ChaLearn, USA)
• Shangeth Rajaa (BITS Pilani, India)
• Herilalaina Rakotoarison (U. Paris-Saclay, INRIA, France)
• Mehreen Saeed (FAST Nat. U. Lahore, Pakistan)
• Marc Schoenauer (U. Paris-Saclay, INRIA, France)
• Michele Sebag (U. Paris-Saclay; CNRS, France)
• Danny Silver (Acadia University, Canada)
• Lisheng Sun (U. Paris-Saclay; UPSud, France)
• Sebastien Treger (La Pallaisse, France)
• Fengfu Li (4paradigm, China)
• Lichuan Xiang (4paradigm, China)
• Jun Wan (Chinese Academy of Sciences, China)
• Mengshuo Wang (4paradigm, China)
• Jingsong Wang (4paradigm, China)
• Ju Xu (4paradigm, China)
• Zhen Xu (Ecole Polytechnique and U. Paris-Saclay; INRIA, France)

The challenge is running on the Codalab platform, administered by Université Paris-Saclay and maintained by CKCollab LLC, with primary developers:

• Eric Carmichael (CKCollab, USA)
• Tyler Thomas (CKCollab, USA)

ChaLearn is the challenge organization coordinator. Google is the primary sponsor of the challenge and helped defining the tasks, protocol, and data formats. 4Paradigm donated prizes, datasets, and contributed to the protocol, baselines methods and beta-testing. Other institutions of the co-organizers provided in-kind contributions, including datasets, data formatting, baseline methods, and beta-testing.

Contact the organizers.