Download as PDF, PPTX
Uploaded byTaehoon Kim
Random Thoughts on Paper Implementations [KAIST 2018]
This document provides random thoughts on implementing machine learning papers. It discusses what types of papers to implement, including computer vision, NLP, reinforcement learning and more. It recommends specific papers and code repositories. It also discusses whether to use TensorFlow or PyTorch and mentions grants and competitions for implementing papers.
More Related Content
![Deep learning for fun and profit [pyconde 2018]](https://cdn.slidesharecdn.com/ss_thumbnails/deeplearningforfunandprofitpyconde2018-181029150851-thumbnail.jpg?width=640&height=640&fit=bounds)
Similar to Random Thoughts on Paper Implementations [KAIST 2018]
Random Thoughts on Paper Implementations [KAIST 2018]
- 1. Random thoughts on PaperImplementation Taehoon Kim / carpedm20
- 2. Taehoon Kim /carpedm20 /
- 3.
- 4.
- 5. Sukhbaatar, Sainbayar, JasonWeston, and Rob Fergus. "End-to-end memory networks." Advances in neural information processing systems. 2015.
- 6. Sukhbaatar, Sainbayar, JasonWeston, and Rob Fergus. "End-to-end memory networks." Advances in neural information processing systems. 2015.
- 7. Sukhbaatar, Sainbayar, JasonWeston, and Rob Fergus. "End-to-end memory networks." Advances in neural information processing systems. 2015.
- 8.
- 9.
- 10.
- 11. Lua Torch → SimpleTranslation
- 12.
- 13.
- 14.
- 15. Received questions aboutimplementation
- 16.
- 17. 1. What toimplement 2. What to read 3. TensorFlow or PyTorch? 4. Grants and Competitions
- 18. 1. What toimplement
- 19. Computer Vision Natural LanguageProcessing Reinforcement Learning Semi/Un-supervised Learning …
- 20. Computer Vision Natural LanguageProcessing Reinforcement Learning Semi/Un-supervised Learning …
- 21.
- 22. Too many GANs.. Relativelyeasier to implement many but not much of a difference
- 23. Reasoning Question Answering Controllable TextGeneration Unsupervised Embedding Exploration Without Reward Neural Programmers Neural Architecture Search Hierarchical Representation Learning Video Generation Program Induction Self-Play Learning Adversarial Attack Multi-agent RL Model-based RL Meta Learning Learned Data Augmentation Speech Synthesis Music Generation
- 24.
- 25. https://blog.openai.com/evolved-policy-gradients/ Adaptive learning fromprior experience with Evolution Strategies (ES)
- 26. https://blog.openai.com/evolved-policy-gradients/ Easiest way :learning from pretrained network But how can we do better?
- 27. https://blog.openai.com/evolved-policy-gradients/ 𝜃 Loss : 𝑓𝑥; 𝜃 − 𝑦 ( Neural Network : 𝑓 𝑥; 𝜃
- 28. https://blog.openai.com/evolved-policy-gradients/ 𝜃𝜙 Loss : 𝐿𝑥, 𝑦; 𝜙 Neural Network : 𝑓 𝑥; 𝜃
- 29.
- 30.
- 31.
- 32. Ganin, Yaroslav, etal. "Synthesizing Programs for Images using Reinforced Adversarial Learning." arXiv preprint arXiv:1804.01118 (2018). Distributed RL + GAN
- 33. Ganin, Yaroslav, etal. "Synthesizing Programs for Images using Reinforced Adversarial Learning." arXiv preprint arXiv:1804.01118 (2018).
- 34. Ganin, Yaroslav, etal. "Synthesizing Programs for Images using Reinforced Adversarial Learning." arXiv preprint arXiv:1804.01118 (2018).
- 35.
- 36.
- 37.
- 38. 2. What toRead
- 39. “How to implementa paper FASTER?”
- 40. Read others codea lot
- 41. Recommendations • NLP: End-to-EndMemory Network • https://github.com/carpedm20/MemN2N-tensorflow • Vision: (Fast) Style Transfer • https://github.com/lengstrom/fast-style-transfer • RL: Asynchronous Methods for Deep Reinforcement Learning • https://github.com/openai/universe-starter-agent (TBH code is dirty but there are lots of things to learn) • https://github.com/ikostrikov/pytorch-a3c • Etc: Neural Turing Machine • https://github.com/loudinthecloud/pytorch-ntm
- 42. More (nerdy) recommendations •from tensorflow.contrib.seq2seq import Helper: • https://github.com/keithito/tacotron/blob/master/models/helpers.py • tf.while_loop: • https://github.com/melodyguan/enas/blob/master/src/ptb/ptb_enas_child.py • import tensorflow.contrib.graph_editor as ge: • https://github.com/openai/gradient-checkpointing/blob/master/memory_saving_gradients.py • Google’s production level code example: • https://github.com/tensorflow/tensor2tensor • The Annotated Transformer • http://nlp.seas.harvard.edu/2018/04/03/attention.html • Relatively safe-to-read codes: • https://github.com/tensorflow/models
- 43. Two useful codesnippets
- 44.
- 45.
- 46.
- 48.
- 49.
- 50. https://github.com/carpedm20/SPIRAL-tensorflow/blob/master/utils/image.py 64 batch image→ 8 × 8 image grid summary
- 51. https://github.com/carpedm20/SPIRAL-tensorflow/blob/master/utils/image.py 64 batch image→ 8 × 8 image grid summary
- 52. Read more recentand correct code Never read carpedm20’s code (no joke. BAD examples)
- 53. 3. TensorFlow orPyTorch?
- 54.
- 55.
- 56. • Dirty longcode • More stressful debugging • Faster development (1.5.0: Jan. 18, 1.6.0-1.7.0: Mar. 18, 1.8.0: Apr. 18, possibly more reliable) • Easier (partial) save and load models (tf.Supervisor) • Harder dynamic computation (tf.fold) • XLA, TensorBoard, TPU, tf.eager, Multi-node distribution, Documentation
- 57. … compiling partsof the computational graph with XLA (a TensorFlow Just-In-Time compiler) and … Espeholt, Lasse, et al. "IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures." arXiv preprint arXiv:1802.01561 (2018).
- 58. tensorboard --port 6006--debugger_port 6064 https://github.com/tensorflow/tensorboard/tree/master/tensorboard/plugins/debugger
- 59.
- 60.
- 61.
- 62. • Clean shortcode • Less stressful debugging • Slower development (0.2.0: Aug. 17, 0.3.0: Dec. 17) • Dirty (partial) save and load models • Easier dynamic computation
- 63. 4. Grants andCompetitions
- 64. This is notabout
- 66.
- 67.
- 68.
- 70.
- 71.
- 72.
- 73.
- 74.
- 75.
- 76.
- 78.
- 79.
- 80.
- 81. Seize the day:) Great opportunities await you
- 82.
- 84.
- 85.
- 87. Teaching Machines toUnderstand Visual Manuals via Attention Supervision for Object Assembly
- 88. Teaching Machines toUnderstand Visual Manuals via Attention Supervision for Object Assembly (work in progress)
- 89.
- 90.
- 91.
- 92.
- 93.
- 94.
- 95.
- 96.
- 97.