Neural Turing Machine (NTM) & Differentiable Neural Computer (DNC) with pytorch & visdom

Overview

Neural Turing Machine (NTM) &

Differentiable Neural Computer (DNC) with

pytorch & visdom


  • Sample on-line plotting while training(avg loss)/testing(write/read weights & memory) NTM on the copy task (top 2 rows, 1st row converges to sequentially write to lower locations, 2nd row converges to sequentially write to upper locations) and DNC on the repeat-copy task (3rd row) (the write/read weights here are after location focus so are no longer necessarily normalized within each head by design):

  • Sample loggings while training DNC on the repeat-copy task (we use WARNING as the logging level currently to get rid of the INFO printouts from visdom):
[WARNING ] (MainProcess) <===================================>
[WARNING ] (MainProcess) bash$: python -m visdom.server
[WARNING ] (MainProcess) http://localhost:8097/env/daim_17051000
[WARNING ] (MainProcess) <===================================> Agent:
[WARNING ] (MainProcess) <-----------------------------======> Env:
[WARNING ] (MainProcess) Creating {repeat-copy | } w/ Seed: 123
[WARNING ] (MainProcess) Word     {length}:   {4}
[WARNING ] (MainProcess) Words #  {min, max}: {1, 2}
[WARNING ] (MainProcess) Repeats  {min, max}: {1, 2}
[WARNING ] (MainProcess) <-----------------------------======> Circuit:    {Controller, Accessor}
[WARNING ] (MainProcess) <--------------------------------===> Controller:
[WARNING ] (MainProcess) LSTMController (
  (in_2_hid): LSTMCell(70, 64, bias=1)
)
[WARNING ] (MainProcess) <--------------------------------===> Accessor:   {WriteHead, ReadHead, Memory}
[WARNING ] (MainProcess) <-----------------------------------> WriteHeads: {1 heads}
[WARNING ] (MainProcess) DynamicWriteHead (
  (hid_2_key): Linear (64 -> 16)
  (hid_2_beta): Linear (64 -> 1)
  (hid_2_alloc_gate): Linear (64 -> 1)
  (hid_2_write_gate): Linear (64 -> 1)
  (hid_2_erase): Linear (64 -> 16)
  (hid_2_add): Linear (64 -> 16)
)
[WARNING ] (MainProcess) <-----------------------------------> ReadHeads:  {4 heads}
[WARNING ] (MainProcess) DynamicReadHead (
  (hid_2_key): Linear (64 -> 64)
  (hid_2_beta): Linear (64 -> 4)
  (hid_2_free_gate): Linear (64 -> 4)
  (hid_2_read_mode): Linear (64 -> 12)
)
[WARNING ] (MainProcess) <-----------------------------------> Memory:     {16(batch_size) x 16(mem_hei) x 16(mem_wid)}
[WARNING ] (MainProcess) <-----------------------------======> Circuit:    {Overall Architecture}
[WARNING ] (MainProcess) DNCCircuit (
  (controller): LSTMController (
    (in_2_hid): LSTMCell(70, 64, bias=1)
  )
  (accessor): DynamicAccessor (
    (write_heads): DynamicWriteHead (
      (hid_2_key): Linear (64 -> 16)
      (hid_2_beta): Linear (64 -> 1)
      (hid_2_alloc_gate): Linear (64 -> 1)
      (hid_2_write_gate): Linear (64 -> 1)
      (hid_2_erase): Linear (64 -> 16)
      (hid_2_add): Linear (64 -> 16)
    )
    (read_heads): DynamicReadHead (
      (hid_2_key): Linear (64 -> 64)
      (hid_2_beta): Linear (64 -> 4)
      (hid_2_free_gate): Linear (64 -> 4)
      (hid_2_read_mode): Linear (64 -> 12)
    )
  )
  (hid_to_out): Linear (128 -> 5)
)
[WARNING ] (MainProcess) No Pretrained Model. Will Train From Scratch.
[WARNING ] (MainProcess) <===================================> Training ...
[WARNING ] (MainProcess) Reporting       @ Step: 500 | Elapsed Time: 30.609361887
[WARNING ] (MainProcess) Training Stats:   avg_loss:         0.014866309287
[WARNING ] (MainProcess) Evaluating      @ Step: 500
[WARNING ] (MainProcess) Evaluation        Took: 1.6457400322
[WARNING ] (MainProcess) Iteration: 500; loss_avg: 0.0140423600748
[WARNING ] (MainProcess) Saving Model    @ Step: 500: /home/zhang/ws/17_ws/pytorch-dnc/models/daim_17051000.pth ...
[WARNING ] (MainProcess) Saved  Model    @ Step: 500: /home/zhang/ws/17_ws/pytorch-dnc/models/daim_17051000.pth.
[WARNING ] (MainProcess) Resume Training @ Step: 500
...

What is included?

This repo currently contains the following algorithms:

  • Neural Turing Machines (NTM) [1]
  • Differentiable Neural Computers (DNC) [2]

Tasks:

  • copy
  • repeat-copy

Code structure & Naming conventions

NOTE: we follow the exact code structure as pytorch-rl so as to make the code easily transplantable.

  • ./utils/factory.py

We suggest the users refer to ./utils/factory.py, where we list all the integrated Env, Circuit, Agent into Dict's. All of the core classes are implemented in ./core/. The factory pattern in ./utils/factory.py makes the code super clean, as no matter what type of Circuit you want to train, or which type of Env you want to train on, all you need to do is to simply modify some parameters in ./utils/options.py, then the ./main.py will do it all (NOTE: this ./main.py file never needs to be modified).

  • namings

To make the code more clean and readable, we name the variables using the following pattern:

  • *_vb: torch.autograd.Variable's or a list of such objects
  • *_ts: torch.Tensor's or a list of such objects
  • otherwise: normal python datatypes

Dependencies


How to run:

You only need to modify some parameters in ./utils/options.py to train a new configuration.

  • Configure your training in ./utils/options.py:
  • line 12: add an entry into CONFIGS to define your training (agent_type, env_type, game, circuit_type)
  • line 28: choose the entry you just added
  • line 24-25: fill in your machine/cluster ID (MACHINE) and timestamp (TIMESTAMP) to define your training signature (MACHINE_TIMESTAMP), the corresponding model file and the log file of this training will be saved under this signature (./models/MACHINE_TIMESTAMP.pth & ./logs/MACHINE_TIMESTAMP.log respectively). Also the visdom visualization will be displayed under this signature (first activate the visdom server by type in bash: python -m visdom.server &, then open this address in your browser: http://localhost:8097/env/MACHINE_TIMESTAMP)
  • line 28: to train a model, set mode=1 (training visualization will be under http://localhost:8097/env/MACHINE_TIMESTAMP); to test the model of this current training, all you need to do is to set mode=2 (testing visualization will be under http://localhost:8097/env/MACHINE_TIMESTAMP_test).
  • Run:

python main.py


Implementation Notes:

The difference between NTM & DNC is stated as follows in the DNC[2] paper:

Comparison with the neural Turing machine. The neural Turing machine (NTM) was the predecessor to the DNC described in this work. It used a similar architecture of neural network controller with read–write access to a memory matrix, but differed in the access mechanism used to interface with the memory. In the NTM, content-based addressing was combined with location-based addressing to allow the network to iterate through memory locations in order of their indices (for example, location n followed by n+1 and so on). This allowed the network to store and retrieve temporal sequences in contiguous blocks of memory. However, there were several drawbacks. First, the NTM has no mechanism to ensure that blocks of allocated memory do not overlap and interfere—a basic problem of computer memory management. Interference is not an issue for the dynamic memory allocation used by DNCs, which provides single free locations at a time, irrespective of index, and therefore does not require contiguous blocks. Second, the NTM has no way of freeing locations that have already been written to and, hence, no way of reusing memory when processing long sequences. This problem is addressed in DNCs by the free gates used for de-allocation. Third, sequential information is preserved only as long as the NTM continues to iterate through consecutive locations; as soon as the write head jumps to a different part of the memory (using content-based addressing) the order of writes before and after the jump cannot be recovered by the read head. The temporal link matrix used by DNCs does not suffer from this problem because it tracks the order in which writes were made.

We thus make some effort to put those two together in a combined codebase. The classes implemented have the following hierarchy:

  • Agent
    • Env
    • Circuit
      • Controller
      • Accessor
        • WriteHead
        • ReadHead
        • Memory

The part where NTM & DNC differs is the Accessor, where in the code NTM uses the StaticAccessor(may not be an appropriate name but we use this to make the code more consistent) and DNC uses the DynamicAccessor. Both Accessor classes use _content_focus() and _location_focus()(may not be an appropriate name for DNC but we use this to make the code more consistent). The _content_focus() is the same for both classes, but the _location_focus() for DNC is much more complicated as it uses dynamic allocation additionally for write and temporal link additionally for read. Those focus (or attention) mechanisms are implemented in Head classes, and those focuses output a weight vector for each head (write/read). Those weight vectors are then used in _access() to interact with the external memory.

A side note:

The sturcture for Env might look strange as this class was originally designed for reinforcement learning settings as in pytorch-rl; here we use it for providing datasets for supervised learning, so the reward, action and terminal are always left blank in this repo.


Repos we referred to during the development of this repo:


The following paper might be interesting to take a look:)

Neural SLAM: We present an approach for agents to learn representations of a global map from sensor data, to aid their exploration in new environments. To achieve this, we embed procedures mimicking that of traditional Simultaneous Localization and Mapping (SLAM) into the soft attention based addressing of external memory architectures, in which the external memory acts as an internal representation of the environment. This structure encourages the evolution of SLAM-like behaviors inside a completely differentiable deep neural network. We show that this approach can help reinforcement learning agents to successfully explore new environments where long-term memory is essential. We validate our approach in both challenging grid-world environments and preliminary Gazebo experiments. A video of our experiments can be found at: \url{https://goo.gl/RfiSxo}.

@article{zhang2017neural,
  title={Neural SLAM},
  author={Zhang, Jingwei and Tai, Lei and Boedecker, Joschka and Burgard, Wolfram and Liu, Ming},
  journal={arXiv preprint arXiv:1706.09520},
  year={2017}
}


Citation

If you find this library useful and would like to cite it, the following would be appropriate:

@misc{pytorch-dnc,
  author = {Zhang, Jingwei},
  title = {jingweiz/pytorch-dnc},
  url = {https://github.com/jingweiz/pytorch-dnc},
  year = {2017}
}
Owner
Jingwei Zhang
Jingwei Zhang
A new benchmark for Icon Question Answering (IconQA) and a large-scale icon dataset Icon645.

IconQA About IconQA is a new diverse abstract visual question answering dataset that highlights the importance of abstract diagram understanding and c

Pan Lu 24 Dec 30, 2022
Using modified BiSeNet for face parsing in PyTorch

face-parsing.PyTorch Contents Training Demo References Training Prepare training data: -- download CelebAMask-HQ dataset -- change file path in the pr

zll 1.6k Jan 08, 2023
A Python training and inference implementation of Yolov5 helmet detection in Jetson Xavier nx and Jetson nano

yolov5-helmet-detection-python A Python implementation of Yolov5 to detect head or helmet in the wild in Jetson Xavier nx and Jetson nano. In Jetson X

12 Dec 05, 2022
PyTorch implementation of the supervised learning experiments from the paper Model-Agnostic Meta-Learning (MAML)

pytorch-maml This is a PyTorch implementation of the supervised learning experiments from the paper Model-Agnostic Meta-Learning (MAML): https://arxiv

Kate Rakelly 516 Jan 05, 2023
Flexible-Modal Face Anti-Spoofing: A Benchmark

Flexible-Modal FAS This is the official repository of "Flexible-Modal Face Anti-

Zitong Yu 22 Nov 10, 2022
An image base contains 490 images for learning (400 cars and 90 boats), and another 21 images for testingAn image base contains 490 images for learning (400 cars and 90 boats), and another 21 images for testing

SVM Données Une base d’images contient 490 images pour l’apprentissage (400 voitures et 90 bateaux), et encore 21 images pour fait des tests. Prétrait

Achraf Rahouti 3 Nov 30, 2021
BuildingNet: Learning to Label 3D Buildings

BuildingNet This is the implementation of the BuildingNet architecture described in this paper: Paper: BuildingNet: Learning to Label 3D Buildings Arx

16 Nov 07, 2022
Code for the paper "Learning-Augmented Algorithms for Online Steiner Tree"

Learning-Augmented Algorithms for Online Steiner Tree This is the code for the paper "Learning-Augmented Algorithms for Online Steiner Tree". Requirem

0 Dec 09, 2021
Implementation of the Point Transformer layer, in Pytorch

Point Transformer - Pytorch Implementation of the Point Transformer self-attention layer, in Pytorch. The simple circuit above seemed to have allowed

Phil Wang 501 Jan 03, 2023
Official implementation of the paper DeFlow: Learning Complex Image Degradations from Unpaired Data with Conditional Flows

DeFlow: Learning Complex Image Degradations from Unpaired Data with Conditional Flows Official implementation of the paper DeFlow: Learning Complex Im

Valentin Wolf 86 Nov 16, 2022
Application of the L2HMC algorithm to simulations in lattice QCD.

l2hmc-qcd 📊 Slides Recent talk on Training Topological Samplers for Lattice Gauge Theory from the Machine Learning for High Energy Physics, on and of

Sam Foreman 37 Dec 14, 2022
2D Human Pose estimation using transformers. Implementation in Pytorch

PE-former: Pose Estimation Transformer Vision transformer architectures perform very well for image classification tasks. Efforts to solve more challe

Panteleris Paschalis 23 Oct 17, 2022
A Python 3 package for state-of-the-art statistical dimension reduction methods

direpack: a Python 3 library for state-of-the-art statistical dimension reduction techniques This package delivers a scikit-learn compatible Python 3

Sven Serneels 32 Dec 14, 2022
a reimplementation of UnFlow in PyTorch that matches the official TensorFlow version

pytorch-unflow This is a personal reimplementation of UnFlow [1] using PyTorch. Should you be making use of this work, please cite the paper according

Simon Niklaus 134 Nov 20, 2022
This repo holds the code of TransFuse: Fusing Transformers and CNNs for Medical Image Segmentation

TransFuse This repo holds the code of TransFuse: Fusing Transformers and CNNs for Medical Image Segmentation Requirements Pytorch=1.6.0, 1.9.0 (=1.

Rayicer 93 Dec 19, 2022
Implementation of GGB color space

GGB Color Space This package is implementation of GGB color space from Development of a Robust Algorithm for Detection of Nuclei and Classification of

Resha Dwika Hefni Al-Fahsi 2 Oct 06, 2021
Recognize Handwritten Digits using Deep Learning on the browser itself.

MNIST on the Web An attempt to predict MNIST handwritten digits from my PyTorch model from the browser (client-side) and not from the server, with the

Harjyot Bagga 7 May 28, 2022
Official implementation of "SinIR: Efficient General Image Manipulation with Single Image Reconstruction" (ICML 2021)

SinIR (Official Implementation) Requirements To install requirements: pip install -r requirements.txt We used Python 3.7.4 and f-strings which are in

47 Oct 11, 2022
Deep Learning as a Cloud API Service.

Deep API Deep Learning as Cloud APIs. This project provides pre-trained deep learning models as a cloud API service. A web interface is available as w

Wu Han 4 Jan 06, 2023
PClean: A Domain-Specific Probabilistic Programming Language for Bayesian Data Cleaning

PClean: A Domain-Specific Probabilistic Programming Language for Bayesian Data Cleaning Warning: This is a rapidly evolving research prototype.

MIT Probabilistic Computing Project 190 Dec 27, 2022