autodl-gpu

English | 简体中文

1. NeurIPS AutoDL Challenge 1'st Solution

1st solution for AutoDL Challenge@NeurIPS, competition rules can be found at AutoDL Competition.

1.0.1. Motivation

There exists a series of common and tough problems in the real world, such as limited resources (CPU/ memory), skewed data, hand-craft features, model selection, network architecture details tuning, sensitivity of pre-trained models, sensitivity of hyperparameters and so on. How to solve them wholly and efficiently?

1.0.2. Solution

AutoDL concentrates on developing generic algorithms for multi-label classification problems in ANY modalities: image, video, speech, text and tabular data without ANY human intervention. Ten seconds at the soonest, our solution achieved SOTA performances on all the 24 offline datasets and 15 online datasets, beating a number of top players in the world.

1.1. Table of Contents

• 1. NeurIPS AutoDL Challenge 1'st Solution
  • 1.0.1. Motivation
  • 1.0.2. Solution
  • 1.1. Table of Contents
  • 1.2. Features
  • 1.3. Evaluation
  • 1.4. Installation
    • 1.4.1. With pip
  • 1.5. Quick Tour
    • 1.5.1. Run local test tour
    • 1.5.2. Tour of Image Classification
    • 1.5.3. Tour of Video Classification
    • 1.5.4. Tour of Speech Classification
    • 1.5.5. Tour of Text Classification
    • 1.5.6. Tour of Tabular Classification
  • 1.6. Public Datasets
    • 1.6.1. Optional: Download public datasets
    • 1.6.2. Public datasets sample info
  • 1.7. Usage for AutoDL local development and testing
  • 1.8. Contributing
  • 1.9. Contact us
  • 1.10. Join the Community
  • 1.11. License

1.2. Features

• Full-AutoML/AutoDL: Fully automated Deep Learning without ANY human intervention covering the whole pipelines.
• Generic & Universal: Supporting ANY modality(image, video, speech, text, tabular) data, and ANY classification problems including binary-class, multi-class and multi-label problems.
• SOTA: Winner solution of AutoDL challenge, involving both tranditional machine learning models and deep learning model backbones.
• Out-of-the-Box: You can use the solution out-of-the-box.
• Fast: You can train your model in ten seconds at the soonest to get highly competitive performance.
• Real-time: You can get the performance feedback(AUC score) in real time.

1.3. Evaluation

• Feedback-phase leaderboard: DeepWisdom Top 1, average rank 1.2, won 4 out of 5 datasets.

• Final-phase leaderboard visualization: DeepWisdom Top 1, average rank 1.2, won 7 out of 10 datasets.

1.4. Installation

This repo is tested on Python 3.6+, PyTorch 1.0.0+ and TensorFlow 2.0.

You should install AutoDL in a virtual environment. If you're unfamiliar with Python virtual environments, check out the user guide.

Create a virtual environment with the version of Python you're going to use and activate it.

Now, if you want to use AutoDL, you can install it with pip.

1.4.1. With pip

AutoDL can be installed using pip as follows:

pip install autodl-gpu 
pip install autodl-gpu=1.0.0 

1.5. Quick Tour

1.5.1. Run local test tour

see Quick Tour - Run local test tour.

1.5.2. Tour of Image Classification

see Quick Tour - Image Classification Demo.

1.5.3. Tour of Video Classification

see Quick Tour - Video Classification Demo.

1.5.4. Tour of Speech Classification

see Quick Tour - Speech Classification Demo.

1.5.5. Tour of Text Classification

see Quick Tour - Text Classification Demo.

1.5.6. Tour of Tabular Classification

see Quick Tour - Tabular Classification Demo.

1.6. Public Datasets

1.6.1. Optional: Download public datasets

python download_public_datasets.py

1.6.2. Public datasets sample info

#	Name	Type	Domain	Size	Source	Data (w/o test labels)	Test labels
1	Munster	Image	HWR	18 MB	MNIST	munster.data	munster.solution
2	City	Image	Objects	128 MB	Cifar-10	city.data	city.solution
3	Chucky	Image	Objects	128 MB	Cifar-100	chucky.data	chucky.solution
4	Pedro	Image	People	377 MB	PA-100K	pedro.data	pedro.solution
5	Decal	Image	Aerial	73 MB	NWPU VHR-10	decal.data	decal.solution
6	Hammer	Image	Medical	111 MB	Ham10000	hammer.data	hammer.solution
7	Kreatur	Video	Action	469 MB	KTH	kreatur.data	kreatur.solution
8	Kreatur3	Video	Action	588 MB	KTH	kreatur3.data	kreatur3.solution
9	Kraut	Video	Action	1.9 GB	KTH	kraut.data	kraut.solution
10	Katze	Video	Action	1.9 GB	KTH	katze.data	katze.solution
11	data01	Speech	Speaker	1.8 GB	--	data01.data	data01.solution
12	data02	Speech	Emotion	53 MB	--	data02.data	dat02.solution
13	data03	Speech	Accent	1.8 GB	--	data03.data	data03.solution
14	data04	Speech	Genre	469 MB	--	data04.data	data04.solution
15	data05	Speech	Language	208 MB	--	data05.data	data05.solution
16	O1	Text	Comments	828 KB	--	O1.data	O1.solution
17	O2	Text	Emotion	25 MB	--	O2.data	O2.solution
18	O3	Text	News	88 MB	--	O3.data	O3.solution
19	O4	Text	Spam	87 MB	--	O4.data	O4.solution
20	O5	Text	News	14 MB	--	O5.data	O5.solution
21	Adult	Tabular	Census	2 MB	Adult	adult.data	adult.solution
22	Dilbert	Tabular	--	162 MB	--	dilbert.data	dilbert.solution
23	Digits	Tabular	HWR	137 MB	MNIST	digits.data	digits.solution
24	Madeline	Tabular	--	2.6 MB	--	madeline.data	madeline.solution

1.7. Usage for AutoDL local development and testing

w1. Git clone the repo

cd <path_to_your_directory>
git clone https://github.com/DeepWisdom/AutoDL.git

1. Prepare pretrained models. Download model speech_model.h5 and put it to AutoDL_sample_code_submission/at_speech/pretrained_models/ directory.

2. Optional: run in the exact same environment as on the challenge platform with docker.

   • CPU

   cd path/to/autodl/
   docker run -it -v "$(pwd):/app/codalab" -p 8888:8888 evariste/autodl:cpu-latest
   

   • GPU

   nvidia-docker run -it -v "$(pwd):/app/codalab" -p 8888:8888 evariste/autodl:gpu-latest
   

3. Prepare sample datasets, using the toy data in AutoDL_sample_data or download new datasets.

4. Run local test

python run_local_test.py

The full usage is

python run_local_test.py -dataset_dir='AutoDL_sample_data/miniciao' -code_dir='AutoDL_sample_code_submission'

Then you can view the real-time feedback with a learning curve by opening the HTML page in AutoDL_scoring_output/.

Details can be seen in AutoDL Challenge official starting_kit.

1.8. Contributing

Feel free to dive in! Open an issue or submit PRs.

1.9. Contact us

1.10. Join the Community

Scan QR code and join AutoDL community!

1.11. License

Apache License 2.0