Generative Flow Networks

Related tags

Deep Learninggflownet
Overview

Flow Network based Generative Models for Non-Iterative Diverse Candidate Generation

Implementation for our paper, submitted to NeurIPS 2021 (also check this high-level blog post).

This is a minimum working version of the code used for the paper, which is extracted from the internal repository of the Mila Molecule Discovery project. Original commits are lost here, but the credit for this code goes to @bengioe, @MJ10 and @MKorablyov (see paper).

Grid experiments

Requirements for base experiments:

  • torch numpy scipy tqdm

Additional requirements for active learning experiments:

  • botorch gpytorch

Molecule experiments

Additional requirements:

  • pandas rdkit torch_geometric h5py
  • a few biochemistry programs, see mols/Programs/README

For rdkit in particular we found it to be easier to install through (mini)conda. torch_geometric has non-trivial installation instructions.

We compress the 300k molecule dataset for size. To uncompress it, run cd mols/data/; gunzip docked_mols.h5.gz.

We omit docking routines since they are part of a separate contribution still to be submitted. These are available on demand, please do reach out to [email protected] or [email protected].

Comments
  • Error: Tensors used as indices must be long, byte or bool tensors

    Error: Tensors used as indices must be long, byte or bool tensors

    Dear authors, thanks for sharing the code for this wonderful work!

    I am currently trying to run the naive gflownet training code in molecular docking setting by running python gflownet.py under the mols directory. I have unzipped the datasets and have all requirements installed. And I have successfully run the model in the toy grid environment.

    However, I got this error when I run in the mols environment:

    Exception while sampling: tensors used as indices must be long, byte or bool tensors

    And when I further look up, it seems like the problem occurs around the line 70 in model_block.py. I tried to print out the stem_block_batch_idx but it doesn't seems like could be transfered to long type directly, which is required by an index:

    tensor([[-8.4156e-02, -4.2767e-02, -7.2483e-02, -3.3011e-02, -1.1865e-02, 2.0981e-03, 1.3293e-02, -7.3515e-03, -4.1853e-02, 2.1048e-02, 3.8597e-02, -1.5558e-02, 2.1581e-02, 4.9257e-03, 9.5167e-02, 4.0965e-02, 2.0146e-02, -5.5610e-02, -3.5318e-02, -3.1394e-02, 7.2078e-02, 1.8894e-02, -3.0249e-02, 2.9740e-02, 5.6950e-02, -3.8425e-02, 2.8620e-02, 9.2052e-02, -8.5357e-03, 1.6788e-02, 7.7801e-02, -4.2119e-02, 1.3606e-02, 7.5316e-02, 4.7131e-02, -4.3429e-03, 1.4157e-04, 2.0939e-02, -2.3499e-02, -6.5888e-02, -2.8960e-02, 3.1548e-02, -9.2680e-03, 5.4192e-02, -9.6579e-03, 2.0602e-02, 1.8935e-02, 4.1228e-03, -6.3467e-02, 3.6747e-02, 1.4168e-02, -6.1473e-03, -1.9472e-02, -3.3970e-02, -5.7308e-03, -4.6021e-02, -3.8956e-02, 4.7375e-02, -8.4562e-02, -1.0087e-02, 2.0478e-02, -6.8286e-02, 5.4663e-02, -5.1468e-02, 1.2617e-02, 2.4625e-02, 5.2167e-02, 5.7779e-02, -5.7788e-02, -1.3323e-02, 1.3913e-02, -7.4439e-02, -4.0981e-02, 5.0797e-02, -5.6230e-02, -5.0963e-02, -5.5488e-02, -2.7339e-02, 1.0469e-02, 3.4695e-02, -3.2623e-02, 7.6694e-03, -5.8748e-03, 7.0495e-02, -2.2805e-02, -5.4334e-03, -2.1636e-02, 1.9597e-02, 6.2370e-02, -2.4995e-02, 1.6165e-02, -4.6878e-03, 2.9743e-02, 1.2653e-02, -5.4271e-02, 1.1247e-02, -3.8340e-03, -4.7489e-02, 1.5719e-02, 3.2552e-02, 6.0665e-02, -1.2330e-02, 2.6115e-02, -2.7376e-02, 3.4152e-02, -1.0086e-02, -2.4257e-02, 3.2202e-02, -3.2659e-02, 8.6094e-02, -3.1996e-02, 7.8751e-02, 4.5367e-02, -3.8693e-02, -3.6531e-02, 6.7311e-03, 3.2884e-02, -3.2774e-02, -3.8855e-02, 2.8814e-02, 4.3942e-02, -1.3374e-02, 3.0905e-02, -7.0064e-02, -5.7230e-03, 4.5093e-02, 3.8167e-02, -3.0602e-02, -4.0387e-02, -1.5985e-02, -9.5962e-02, -1.1354e-02, 2.0879e-02, 1.4092e-02, -3.8405e-02, 1.4337e-02, -6.0682e-02, -9.0190e-03, -5.0898e-02, -4.7344e-02, 4.1045e-02, -6.7031e-02, 8.8112e-02, 3.2149e-02, 3.7748e-02, -4.0757e-02, 1.4378e-02, -1.0749e-01, 6.1679e-02, -6.7268e-03, -2.7889e-02, -5.9315e-02, -5.5883e-02, -2.6489e-02, 7.3640e-02, 1.8273e-02, -5.2330e-02, -7.7003e-05, 6.8413e-04, -1.4364e-01, -1.9389e-02, 4.5649e-02, -4.0468e-02, -4.2819e-02, 4.5874e-02, -1.6481e-02, 1.2627e-02, -8.4941e-02, -3.7458e-02, 2.1359e-02, -9.2863e-02, -3.4932e-03, 7.1990e-02, 6.2144e-02, 8.1462e-02, -2.0569e-02, 5.9194e-02, 1.6996e-03, 8.0618e-03, 6.1753e-02, 4.1602e-02, 1.0910e-02, 2.0523e-02, -9.9781e-04, 1.9131e-02, -1.0267e-02, -9.4474e-02, -3.5725e-02, 9.9953e-03, -4.3195e-02, -7.9051e-02, -3.1881e-02, 9.2158e-03, -9.6167e-04, -2.7508e-02, 7.1478e-02, -5.4107e-02, 8.0026e-02, -1.8887e-02, 4.6941e-02, 6.5166e-02, 1.2000e-02, 3.9906e-02, -2.8206e-02, 3.7483e-02, 3.5408e-02, -2.5863e-02, 2.3528e-02, 7.1814e-03, 8.0863e-02, -1.3736e-02, -8.5978e-02, -4.1238e-02, -1.2545e-02, 5.5479e-02, 7.3487e-03, 8.9125e-02, -3.4814e-02, -4.5358e-02, 4.9893e-02, 3.5286e-02, 3.2084e-02, 5.0868e-02, 2.3549e-02, -9.2907e-02, -6.9315e-03, -1.3088e-02, 8.7066e-02, 1.1554e-02, 1.3771e-02, -1.7489e-02, -5.2921e-02, 9.2110e-03, 1.6766e-02, 4.8030e-02, 1.4481e-02, 2.9254e-03, 3.5795e-02, 1.0397e-01, -2.0675e-03, -2.9916e-02, -5.3299e-02, -2.1396e-02, -5.3189e-02, 3.2805e-02, -2.6538e-03, -2.6352e-02, -1.2823e-02, 6.1972e-02, 5.4822e-02, 4.5579e-02, -3.6638e-02, 8.1013e-03, -5.6014e-02, 1.5187e-02, -6.5561e-02]], device='cuda:0', dtype=torch.float64, grad_fn=)

    I wonder if I am running the code in the correct way. Is this index correct and if so, do you know what's happening?

    opened by wenhao-gao 3
  • About Reproducibility Issues

    About Reproducibility Issues

    Hi there,

    Thank you very much for sharing the source codes.

    For reproducibility, I modified the codes as follows,

    https://github.com/GFNOrg/gflownet/blob/831a6989d1abd5c05123ec84654fb08629d9bc38/mols/gflownet.py#L84

    ---> self.train_rng = np.random.RandomState(142857)

    as well as to add

    torch.manual_seed(142857)
    torch.cuda.manual_seed(142857)
    torch.cuda.manual_seed_all(142857)
    

    However, I encountered an issue. I ran it more than 3 times with the same random seed, but the results are totally different (although they are close). I didn't modify other parts, except for addressing package compatibility issues.

    0 [1152.62, 112.939, 23.232] 100 [460.257, 44.253, 17.728] 200 [68.114, 6.007, 8.045]

    0 [1151.024, 112.603, 24.993] 100 [471.219, 45.525, 15.964] 200 [66.349, 6.174, 4.607]

    0 [1263.066, 124.094, 22.128] 100 [467.747, 44.899, 18.76] 200 [61.992, 5.715, 4.841]

    I am wondering whether you encountered such an issue before.

    Best,

    Dong

    opened by dongqian0206 2
  • Reward signal for grid environment?

    Reward signal for grid environment?

    Hello, I'm a bit confused where this reward function comes from: https://github.com/GFNOrg/gflownet/blob/831a6989d1abd5c05123ec84654fb08629d9bc38/grid/toy_grid_dag.py#L97

    My understanding is that the reward should be as defined in the paper (https://i.samkg.dev/2233/firefox_xGnEaZVBlN.png) - are these two equivalent in some way?

    opened by SamKG 1
  • Potential bug with `FlowNetAgent.sample_many`

    Potential bug with `FlowNetAgent.sample_many`

    Hi there!

    Thanks for sharing the code and just wanted to say I've enjoyed your paper. I was reading your code and noticed that there might be a subtle bug in the grid-env dag script. I might also have read it wrong...

    https://github.com/bengioe/gflownet/blob/dddfbc522255faa5d6a76249633c94a54962cbcb/grid/toy_grid_dag.py#L316-L320

    On line 316, we zip two things: zip([e for d, e in zip(done, self.envs) if not d], acts)

    Here done is a vector of bools of length batch-size, self.envs is a list of GridEnv of length n-envs or buffer-size, and acts is a vector of ints of length (n-envs or buffer-size,).

    By default, all the lengths of the above objects should be 16.

    I was reading through the code, and noticed that if any of the elements in done are True, then on line 316 we filter them out with if not d. If env[0] was "done", then we would have a list of 15 envs, basically self.envs[1:]. Then when you zip up the actions and the shorter list envs, the actions will be aligned incorrectly... We will basically end up with self.envs[1:] being aligned to actions act[:-1]. As a result, step is now length 15, and on the next line, we again line up the incorrect actions of length 16 with our step list of length 16.

    Perhaps we need to filter act based on the done vector? E.g act = act[done] after line 316?

    Maybe I've got this wrong, so apologies for the noise if that's the case, but thought I'd leave a note in case what I'm suggesting is the case.

    All the best!

    opened by fedden 1
  • Clarification regarding the number of molecular building blocks. Why they are different from JT-VAE?

    Clarification regarding the number of molecular building blocks. Why they are different from JT-VAE?

    Hello,

    First, I really enjoyed reading the paper. Amazing work!

    I have a question regarding the number of building blocks used for generating small molecules. Appendix A.3 of the paper states that there are a total of 105 unique building blocks (after accounting for different attachment points) and that they were obtained by the process suggested by the JT-VAE paper. (Jin et al. (2020)). However, in the JT-VAE paper, the total vocabulary size is $|\chi|=780$ obtained from the same ZINC dataset. My understanding is they are both the same. If that is correct, why are the number of building blocks different here? What am I missing? If they are not the same, can you please explain the difference?

    Thank you so much for your help

    opened by Srilok 1
Releases(paper_version)
Owner
Emmanuel Bengio
Emmanuel Bengio
Code and Data for the paper: Molecular Contrastive Learning with Chemical Element Knowledge Graph [AAAI 2022]

Knowledge-enhanced Contrastive Learning (KCL) Molecular Contrastive Learning with Chemical Element Knowledge Graph [ AAAI 2022 ]. We construct a Chemi

Fangyin 58 Dec 26, 2022
The official PyTorch implementation of the paper: *Xili Dai, Xiaojun Yuan, Haigang Gong, Yi Ma. "Fully Convolutional Line Parsing." *.

F-Clip — Fully Convolutional Line Parsing This repository contains the official PyTorch implementation of the paper: *Xili Dai, Xiaojun Yuan, Haigang

Xili Dai 115 Dec 28, 2022
A simple code to convert image format and channel as well as resizing and renaming multiple images.

Rename-Resize-and-convert-multiple-images A simple code to convert image format and channel as well as resizing and renaming multiple images. This cod

Happy N. Monday 3 Feb 15, 2022
ImageNet-CoG is a benchmark for concept generalization. It provides a full evaluation framework for pre-trained visual representations which measure how well they generalize to unseen concepts.

The ImageNet-CoG Benchmark Project Website Paper (arXiv) Code repository for the ImageNet-CoG Benchmark introduced in the paper "Concept Generalizatio

NAVER 23 Oct 09, 2022
PyTorch and Tensorflow functional model definitions

functional-zoo Model definitions and pretrained weights for PyTorch and Tensorflow PyTorch, unlike lua torch, has autograd in it's core, so using modu

Sergey Zagoruyko 590 Dec 22, 2022
basic tutorial on pytorch

Quick Tutorial on PyTorch PyTorch Basics Linear Regression Logistic Regression Artificial Neural Networks Convolutional Neural Networks Recurrent Neur

7 Sep 15, 2022
CLIP-GEN: Language-Free Training of a Text-to-Image Generator with CLIP

CLIP-GEN [简体中文][English] 本项目在萤火二号集群上用 PyTorch 实现了论文 《CLIP-GEN: Language-Free Training of a Text-to-Image Generator with CLIP》。 CLIP-GEN 是一个 Language-F

75 Dec 29, 2022
CNNs for Sentence Classification in PyTorch

Introduction This is the implementation of Kim's Convolutional Neural Networks for Sentence Classification paper in PyTorch. Kim's implementation of t

Shawn Ng 956 Dec 19, 2022
Gym for multi-agent reinforcement learning

PettingZoo is a Python library for conducting research in multi-agent reinforcement learning, akin to a multi-agent version of Gym. Our website, with

Farama Foundation 1.6k Jan 09, 2023
Official code for "Mean Shift for Self-Supervised Learning"

MSF Official code for "Mean Shift for Self-Supervised Learning" Requirements Python = 3.7.6 PyTorch = 1.4 torchvision = 0.5.0 faiss-gpu = 1.6.1 In

UMBC Vision 44 Nov 21, 2022
Deep-Learning-Book-Chapter-Summaries - Attempting to make the Deep Learning Book easier to understand.

Deep-Learning-Book-Chapter-Summaries This repository provides a summary for each chapter of the Deep Learning book by Ian Goodfellow, Yoshua Bengio an

Aman Dalmia 1k Dec 27, 2022
Image Super-Resolution by Neural Texture Transfer

SRNTT: Image Super-Resolution by Neural Texture Transfer Tensorflow implementation of the paper Image Super-Resolution by Neural Texture Transfer acce

Zhifei Zhang 413 Nov 30, 2022
The code from the paper Character Transformations for Non-Autoregressive GEC Tagging

Character Transformations for Non-Autoregressive GEC Tagging Milan Straka, Jakub Náplava, Jana Straková Charles University Faculty of Mathematics and

ÚFAL 5 Dec 10, 2022
Spatial Single-Cell Analysis Toolkit

Single-Cell Image Analysis Package Scimap is a scalable toolkit for analyzing spatial molecular data. The underlying framework is generalizable to spa

Laboratory of Systems Pharmacology @ Harvard 30 Nov 08, 2022
CATE: Computation-aware Neural Architecture Encoding with Transformers

CATE: Computation-aware Neural Architecture Encoding with Transformers Code for paper: CATE: Computation-aware Neural Architecture Encoding with Trans

16 Dec 27, 2022
Memory Defense: More Robust Classificationvia a Memory-Masking Autoencoder

Memory Defense: More Robust Classificationvia a Memory-Masking Autoencoder Authors: - Eashan Adhikarla - Dan Luo - Dr. Brian D. Davison Abstract Many

Eashan Adhikarla 4 Dec 25, 2022
Urban mobility simulations with Python3, RLlib (Deep Reinforcement Learning) and Mesa (Agent-based modeling)

Deep Reinforcement Learning for Smart Cities Documentation RLlib: https://docs.ray.io/en/master/rllib.html Mesa: https://mesa.readthedocs.io/en/stable

1 May 15, 2022
Model Zoo of BDD100K Dataset

Model Zoo of BDD100K Dataset

ETH VIS Group 200 Dec 27, 2022
Monocular Depth Estimation - Weighted-average prediction from multiple pre-trained depth estimation models

merged_depth runs (1) AdaBins, (2) DiverseDepth, (3) MiDaS, (4) SGDepth, and (5) Monodepth2, and calculates a weighted-average per-pixel absolute dept

Pranav 39 Nov 21, 2022
PyTorch implementation of the cross-modality generative model that synthesizes dance from music.

Dancing to Music PyTorch implementation of the cross-modality generative model that synthesizes dance from music. Paper Hsin-Ying Lee, Xiaodong Yang,

NVIDIA Research Projects 485 Dec 26, 2022