The environmental impact of present ML models and how can we improve it DSOC R&D 李 星 2020/07/21

※ 掲載されている内容等は発表時点の情報です。 ※ 公開に当たり、資料の⼀部を変更・削除している場合があります。

Data Strategy and Operation Center We also have materials in Japanese. https://speakerdeck.com/sansandsoc

Data Strategy and Operation Center self-introduction I joined Sansan DSOC from OCT/2019 as a new graduate student. I am in charge of exploiting Sansan’s data through various machine learning methods in an efficient way. Now I am currently focusing on recommendation system and some customized small NLP tasks. 李 星 XING LI

Data Strategy and Operation Center The size of DL models is getting larger and larger…(NLP) [16]

Data Strategy and Operation Center The size of DL models is getting larger and larger…(CV) [4]

Data Strategy and Operation Center The training time of DL models is also increasing [17]

Data Strategy and Operation Center Method p_c: the average power draw (in watts) from all CPU sockets during training p_r: the average power draw from all DRAM (main memory) sockets p_g: the average power draw of a GPU during training g: the number of GPUs used to train t: the total training time 1.58: Power Usage Effectiveness(PUE). [13] 0.954: Average CO2 produced for power consumed in the U.S. [14] [2] [2]

Data Strategy and Operation Center A vivid comparison [2]

Data Strategy and Operation Center Common Model’s carbon emission table [2]

Data Strategy and Operation Center Where does the cost actually come from? [20]

Data Strategy and Operation Center Two aspects to make models more energy efficient without unacceptably losing performance

Data Strategy and Operation Center Algorithm 1. Mixed Precision ( FP16 & FP32 ) 2. Model Distillation 3. Model Pruning 4. Weight Quantization & Sharing 5. Others

Data Strategy and Operation Center Mixed precision training iteration for a layer. [8] Algorithm ─ Mixed Precision ─ Where is the “mixed” coming from?

Data Strategy and Operation Center Algorithm ─ Mixed Precision ─ Performance remain [8]

Data Strategy and Operation Center Algorithm ─ Mixed Precision ─ Mainstream Library Support PyTorch Mixed Precision Tutorial: https://pytorch.org/docs/stable/notes/amp_examples.html TensorFlow Mixed Precision Guide: https://www.tensorflow.org/guide/mixed_precision

Data Strategy and Operation Center Algorithm ─ Model Distillation ─ Core Idea [11]

Data Strategy and Operation Center Algorithm ─ Model Distillation ─ An example [9] [9]

Data Strategy and Operation Center Algorithm ─ Model Distillation ─ Useful Distilled Models’ Implementation Github: https://github.com/dkozlov/awesome-knowledge-distillation

Data Strategy and Operation Center Algorithm ─ Model Pruning ─ Basic Framework & Concepts [15] Training Pruning Fine-tuning

Data Strategy and Operation Center Algorithm ─ Model Pruning ─ Original Training Result [15]

Data Strategy and Operation Center Algorithm ─ Model Pruning ─ After only pruning [15]

Data Strategy and Operation Center Algorithm ─ Model Pruning ─ Prune+Retrain [15]

Data Strategy and Operation Center Algorithm ─ Model Pruning ─ Iteratively [15]

Data Strategy and Operation Center Algorithm ─ Model Pruning ─ Mainstream Library Support PyTorch Pruning Tutorial: https://pytorch.org/tutorials/intermediate/pruning_tutorial.html TensorFlow Pruning Tutorial: https://www.tensorflow.org/model_optimization/guide/pruning

Data Strategy and Operation Center Algorithm ─ Weight Quantization & Sharing ─ Core Idea [5]

Data Strategy and Operation Center [5] Algorithm ─ Weight Quantization & Sharing ─ Initialization of K-means Three different methods for centroids initialization. Distribution of weights (◼blue) and distribution of codebook before (×green cross) and after fine-tuning (●red dot)

Data Strategy and Operation Center Algorithm ─ Weight Quantization & Sharing ─ Further compression trick(Huffman Coding) [5]

Data Strategy and Operation Center Algorithm ─ Others (Most of them involve the redesign to original network architectures) • Special designed network architectures: • ShuffleNet, MobileNet, BottleNet, SqueezeNet[6] and etc…. • Winograd Transformation • Low Rank Approximation • Binary/Ternary Net • …

Data Strategy and Operation Center Hardware All in one word: to minimize the memory access! Actually, I am neither going to nor able to discuss how to design the chips~(TT) But we could know how good our models run on a specific hardware platform so that we can decide to continue optimising our algorithm or buy a better x(C/G/T)PU.

Data Strategy and Operation Center Hardware ─ Roofline Model ─ Definition [18] [18]

Data Strategy and Operation Center Hardware ─ Roofline Model ─ How to use it [18]

Data Strategy and Operation Center Hardware ─ Choose more environmental friendly hardware ─ xPU matters [1]

Data Strategy and Operation Center Hardware ─ Choose more environmental friendly hardware ─ Location matters [1]

Data Strategy and Operation Center Hardware ─ Choose more environmental friendly hardware ─ Platform matters [2]

Data Strategy and Operation Center Hardware ─ Choose more environmental friendly hardware ─ Same platform in different locations [1] Amazon Web Services

Data Strategy and Operation Center Summary If you don’t want to touch the network architecture: ● Algorithm - Mixed Precision(FP16&FP32) ● Algorithm - Model Distillation ● Algorithm - Model Pruning ● Algorithm - Weight Quantization & Sharing ● Hardware - Use roofline to help you increase energy efficiency ● Hardware - Carefully decide the device, location and platform. If you can be able to design new model: ● Special designed network architectures: ● Winograd Transformation ● Low Rank Approximation ● Binary/Ternary Net If you can start from hardware level: ● My unknown area...

Data Strategy and Operation Center Others ─ What about planting a tree? 6 Trees for life ~= 1 tonne of CO2 [19]

Data Strategy and Operation Center Others ─ Sansan cares about it !

No content

Data Strategy and Operation Center References 1. Quantifying the Carbon Emissions of Machine Learning (https://arxiv.org/pdf/1910.09700.pdf) 2. Energy and Policy Considerations for Deep Learning in NLP (https://arxiv.org/pdf/1906.02243.pdf) 3. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter (https://arxiv.org/pdf/1910.01108.pdf) 4. Neural Network Architectures(https://towardsdatascience.com/neural-network-architectures-156e5bad51ba) 5. DEEP COMPRESSION: COMPRESSING DEEP NEURAL NETWORKS WITH PRUNING, TRAINED QUANTIZATION AND HUFFMAN CODING(https://arxiv.org/pdf/1510.00149.pdf) 6. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size(https://arxiv.org/pdf/1602.07360.pdf) 7. Deep Learning Performance Documentation Nvidia (https://docs.nvidia.com/deeplearning/performance/mixed-precision- training/index.html#mptrain__fig1) 8. MIXED PRECISION TRAINING (https://arxiv.org/pdf/1710.03740.pdf) 9. Distilling the Knowledge in a Neural Network(https://arxiv.org/pdf/1503.02531.pdf) 10. Distilling Task-Specific Knowledge from BERT into Simple Neural Networks(https://arxiv.org/pdf/1903.12136.pdf) 11. Knowledge Distillation: Simplified (https://towardsdatascience.com/knowledge-distillation-simplified-dd4973dbc764) 12. ML CO2 IMPACT: https://mlco2.github.io/impact/#home 13. Rhonda Ascierto. 2018. Uptime Institute Global Data Center Survey. Technical report, Uptime Institute. 14. EPA. 2018. Emissions & Generation Resource Integrated Database (eGRID). Technical report, U.S. Environmental Protection Agency. 15. Learning both Weights and Connections for Efficient Neural Networks(https://papers.nips.cc/paper/5784-learning-both-weights- and-connections-for-efficient-neural-network.pdf) 16. GPT-3: The New Mighty Language Model from OpenAI(https://mc.ai/gpt-3-the-new-mighty-language-model-from-openai-2/) 17. AI and Compute(https://openai.com/blog/ai-and-compute/) 18. Performance Analysis(HPC Course, University of Bristol) 19. Reduce your carbon footprint by Planting a tree(https://co2living.com/reduce-your-carbon-footprint-by-planting-a-tree/) 20. EIE: Efficient Inference Engine on Compressed Deep Neural Network(https://arxiv.org/pdf/1602.01528.pdf)