Storchastic is a PyTorch library for stochastic gradient estimation in Deep Learning

Last update: Dec 30, 2022

Related tags

Overview

Stochastic Deep Learning for Pytorch

Documentation on Read the Docs. Storchastic is a PyTorch library for stochastic gradient estimation in Deep Learning [1]. Many state of the art deep learning models use gradient estimation, in particular within the fields of Variational Inference and Reinforcement Learning. While PyTorch computes gradients of deterministic computation graphs automatically, it will not estimate gradients on stochastic computation graphs [2].

With Storchastic, you can easily define any stochastic deep learning model and let it estimate the gradients for you. Storchastic provides a large range of gradient estimation methods that you can plug and play, to figure out which one works best for your problem. Storchastic provides automatic broadcasting of sampled batch dimensions, which increases code readability and allows implementing complex models with ease.

When dealing with continuous random variables and differentiable functions, the popular reparameterization method [3] is usually very effective. However, this method is not applicable when dealing with discrete random variables or non-differentiable functions. This is why Storchastic has a focus on gradient estimators for discrete random variables, non-differentiable functions and sequence models.

Documentation on Read the Docs.

Example: Discrete Variational Auto-Encoder

Installation

pip install storchastic

Requires Pytorch 1.5 (older versions will not do!) and Pyro. The code is build on Python 3.7. The master branch works with PyTorch 1.7, but the version on pip is not compatible. Binaries will be updated soon.

Algorithms

Feel free to create an issue if an estimator is missing here.

Reparameterization [1, 3]
Score Function (REINFORCE) with Moving Average baseline [1, 4]
Score Function with Batch Average Baseline [5, 6]
Expected value for enumerable distributions
(Straight through) Gumbel Softmax [7, 8]
LAX, RELAX [9]
REBAR [10]
REINFORCE Without Replacement [6]
Unordered Set Estimator [13]

In development

Memory Augmented Policy Optimization [11]
Rao-Blackwellized REINFORCE [12]

Planned

Measure valued derivatives [1, 14]
ARM [15]
Automatic Credit Assignment [16]
...

References

[1] Monte Carlo Gradient Estimation in Machine Learning, Mohamed et al, 2019
[2] Gradient Estimation Using Stochastic Computation Graphs, Schulman et al, NeurIPS 2015
[3] Auto-Encoding Variational Bayes, Kingma and Welling, ICLR 2014
[4] Simple statistical gradient-following algorithms for connectionist reinforcement learning, Williams, Machine Learning 1992
[5] Variational inference for Monte Carlo objectives, Mnih and Rezende, ICML 2016
[6] Buy 4 REINFORCE Samples, Get a Baseline for Free!, Kool et al, ICLR Workshop dlStructPred 2019
[7] Categorical Reparameterization with Gumbel-Softmax, Jang et al, ICLR 2017
[8] The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables, Maddison et al, ICLR 2017
[9] Backpropagation through the Void: Optimizing control variates for black-box gradient estimation, Grathwohl et al, ICLR 2018
[10] REBAR: Low-variance, unbiased gradient estimates for discrete latent variable models, Tucker et al, NeurIPS 2017
[11] Memory Augmented Policy Optimization for Program Synthesis and Semantic Parsing, Liang et al, NeurIPS 2018
[12] Rao-Blackwellized Stochastic Gradients for Discrete Distributions, Liu et al, ICML 2019
[13] Estimating Gradients for Discrete Random Variables by Sampling without Replacement, Kool et al, ICLR 2020
[14] Measure-Valued Derivatives for Approximate Bayesian Inference, Rosca et al, Workshop on Bayesian Deep Learning (NeurIPS 2019)
[15] ARM: Augment-REINFORCE-Merge Gradient for Stochastic Binary Networks, Yin and Zhou, ICLR 2019
[16] Credit Assignment Techniques in Stochastic Computation Graphs, Weber et al, AISTATS 2019

Comments

ValueError: The value argument to log_prob must be a Tensor

I use python=3.8 and pytorch=1.8.1 and test the examplebernoulli_toy.py Error happens in the following line in expect.py log_probs = tensor.distribution.log_prob(tensor) error is ValueError: The value argument to log_prob must be a Tensor

Seems .log_prob() wants a tensor object as input. When I print the 'tensor', it outputs:

x: Stochastic tensor([[0., 0., 0., 0.],
        [0., 0., 0., 1.],
        [0., 0., 1., 0.],
        [0., 0., 1., 1.],
        [0., 1., 0., 0.],
        [0., 1., 0., 1.],
        [0., 1., 1., 0.],
        [0., 1., 1., 1.],
        [1., 0., 0., 0.],
        [1., 0., 0., 1.],
        [1., 0., 1., 0.],
        [1., 0., 1., 1.],
        [1., 1., 0., 0.],
        [1., 1., 0., 1.],
        [1., 1., 1., 0.],
        [1., 1., 1., 1.]]) Batch links: [('x', 16, tensor(0.0625))]

Could you offer some solutions? Thank you

opened by zhaoguangyuan123 8

Recursion limit reached with many cost nodes

Running backward() when there are a lot of cost nodes causes a recursion-limit-reached exception to be thrown.

I'm able to consistently generate this exception running the following:

import storch
import torch
from torch.distributions import Bernoulli
from storch.method import GumbelSoftmax

torch.manual_seed(0)

p = torch.tensor(0.5, requires_grad=True)

for i in range(1000):
    sample = GumbelSoftmax(f"sample_{i}")(Bernoulli(p))
    storch.add_cost(sample, f"cost_{i}")

storch.backward()

which produces the following trace:

Traceback (most recent call last):
  File "...", line 15, in <module>
    storch.backward()
  File ".../lib/python3.8/site-packages/storch/inference.py", line 217, in backward
    accum_loss._clean()
  File ".../lib/python3.8/site-packages/storch/tensor.py", line 401, in _clean
    node._clean()
  File ".../lib/python3.8/site-packages/storch/tensor.py", line 401, in _clean
    node._clean()
  File ".../lib/python3.8/site-packages/storch/tensor.py", line 401, in _clean
    node._clean()
  [Previous line repeated 994 more times]
  File ".../lib/python3.8/site-packages/storch/tensor.py", line 399, in _clean
    node._clean()
RecursionError: maximum recursion depth exceeded

bug

opened by csmith49 4

Compatibility with Conv networks

I'm currently working with CNNs, which allow input x with shape x.shape -> [N, C, H, W].

However, when I set N as the batch dim (use storch.denote_independent(x, 0, "data")) and perform RELAX("z", n_samples=1) sampling (it adds another plate), the sampled tensor becomes z.shape -> [Z, N, C, H, W] and is not compatible with Conv2D any more.

Is there any way to reshape z to be compatible with Conv2D?

Minimal reproduction code:

class RelaxSample(nn.Module):
    def __init__(self, k: int, c: int):
        super().__init__()
        self._k = k
        self._wv = nn.Parameter(torch.empty(k, c))
        self._method = RELAX("z", n_samples=1, in_dim=[k], rebar=True)

    def forward(self, latent):
        # [n, c, h, w]
        n, c, h, w = latent.shape
        # [n, h, w, c]
        q = latent.permute(0, 2, 3, 1)
        # [n, h, w, c], [k, c] -> [n, h, w, k]
        logit = torch.einsum("nhwc,kc->nhwk", q, self._wv)
        varPosterior = Categorical(logits=logit)
        # [z, n, h, w, k]
        z = self._method(varPosterior)
        # [z, n, c, h, w]
        z = torch.einsum("znhwk,kc->znhwc", z, self._wv).permute(0, 1, 4, 2, 3)
        # ********* z is not compatible with Conv2D *********
        return z

opened by xiaosu-zhu 3

RELAX/REBAR missing mc_sample

RELAX and REBAR gradient estimators don't appear to be usable on Bernoulli random variables - some operation expects them to have the mc_sample attribute, which is missing.

I can reproduce this result with the following:

import storch
import torch
from torch.distributions import Bernoulli
from storch.method import RELAX

torch.manual_seed(0)

p = torch.tensor(0.5, requires_grad=True)
d = Bernoulli(p)
sample = RELAX("sample")(d)
storch.add_cost(sample, "cost")
storch.backward()

which produces the following trace:

Traceback (most recent call last):
  File "...", line 13, in <module>
    sample = RELAX("sample")(d)
  File ".../lib/python3.8/site-packages/storch/method/relax.py", line 210, in __init__
    super().__init__(plate_name, sampling_method.set_mc_sample(self.mc_sample))
  File ".../lib/python3.8/site-packages/torch/nn/modules/module.py", line 771, in __getattr__
    raise ModuleAttributeError("'{}' object has no attribute '{}'".format(
torch.nn.modules.module.ModuleAttributeError: 'RELAX' object has no attribute 'mc_sample'

opened by csmith49 2

Plating to denote conditional independencies

Often, samples from a distribution are conditionally independent: for example in LDA. While it is possible to create a list of K conditionally independent variables by looping over sample steps, this is not vectorized and thus not efficient.

Plating could possibly be done by calling storch.plate() on a storch tensor.
enhancement

opened by HEmile 2

Examples with Bernoulli distributions failing

Issue

Running either of the examples with Bernoulli distributions (examples/bernoulli_toy.py and examples/bernoulli_grad_var.py) results in a ValueError, specifically:

ValueError: Input arguments must all be instances of numbers.Number or torch.tensor.

Environment

Running Python 3.8.5, with Pyro 1.5.0 and Torch 1.7.0.

Traceback

Running python3 examples/bernoulli_toy.py from the root of the repo produces (with better_exceptions enabled):

Traceback (most recent call last):
  File "examples/bernoulli_toy.py", line 29, in <module>
    experiment(Expect("x"))
    │          └ <class 'storch.method.method.Expect'>
    └ <function experiment at 0x7fbf47b3a0d0>
  File "examples/bernoulli_toy.py", line 19, in experiment
    x = method(b)
        │      └ Bernoulli(probs: torch.Size([4]), logits: torch.Size([4]))
        └ Expect(
  (sampling_method): Enumerate()
)
  File ".../python3.8/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
             │             │        └ {}
             │             └ (Bernoulli(probs: torch.Size([4]), logits: torch.Size([4])),)
             └ Expect(
  (sampling_method): Enumerate()
)
  File ".../python3.8/site-packages/storch/method/method.py", line 52, in forward
    return self.sample(distr)
           │           └ Bernoulli(probs: torch.Size([4]), logits: torch.Size([4]))
           └ Expect(
  (sampling_method): Enumerate()
)
  File ".../python3.8/site-packages/storch/method/method.py", line 114, in sample
    batch_weighting = self.sampling_method.plate_weighting(s_tensor, plate)
                      │                                    │         └ ('x', 16, tensor(0.0625))
                      │                                    └ tensor([[0., 0., 0., 0.],
        [0., 0., 0., 1.],
        [0., 0., 1., 0.],
        [0., 0., 1., 1.],
        [0., 1., 0., 0.]...
                      └ Expect(
  (sampling_method): Enumerate()
)
  File ".../python3.8/site-packages/storch/sampling/expect.py", line 82, in plate_weighting
    log_probs = tensor.distribution.log_prob(tensor)
                │                            └ tensor([[0., 0., 0., 0.],
        [0., 0., 0., 1.],
        [0., 0., 1., 0.],
        [0., 0., 1., 1.],
        [0., 1., 0., 0.]...
                └ tensor([[0., 0., 0., 0.],
        [0., 0., 0., 1.],
        [0., 0., 1., 0.],
        [0., 0., 1., 1.],
        [0., 1., 0., 0.]...
  File ".../python3.8/site-packages/torch/distributions/bernoulli.py", line 94, in log_prob
    logits, value = broadcast_all(self.logits, value)
            │       │             │            └ tensor([[0., 0., 0., 0.],
        [0., 0., 0., 1.],
        [0., 0., 1., 0.],
        [0., 0., 1., 1.],
        [0., 1., 0., 0.]...
            │       │             └ Bernoulli(probs: torch.Size([4]), logits: torch.Size([4]))
            │       └ <function broadcast_all at 0x7fbe57163430>
            └ tensor([[0., 0., 0., 0.],
        [0., 0., 0., 1.],
        [0., 0., 1., 0.],
        [0., 0., 1., 1.],
        [0., 1., 0., 0.]...
  File ".../python3.8/site-packages/torch/distributions/utils.py", line 24, in broadcast_all
    raise ValueError('Input arguments must all be instances of numbers.Number or torch.tensor.')
ValueError: Input arguments must all be instances of numbers.Number or torch.tensor.

bug

opened by csmith49 1

Abstract sampling procedures
We should abstract sampling procedures so that they are no longer part of a Method's inheritance tree, but rather as a separate object to plug and play. This would make it easier to combine various estimators with various sampling methods, whether biased or not.

Example applications:

https://openreview.net/forum?id=H1Xw62kRZ uses beam search with a biased REINFORCE method. Allowing to use any estimator with this beam search would help modularity and debugging.

Usually, normal MC sampling is enough, but eg when decoding text, we have to deal with eos's that complicate the sampling procedure.

Sampling without replacement is applicable to any unbiased estimator when correcting for the bias through the weighting. By making sampling modular, we can more easily test combinations of estimators.

enhancement
opened by HEmile 1
Add option to Method to ignore estimation

Methods like reparameterization do not require computing an estimator. We could skip some transposes by manually checking for this using a function in Method

opened by HEmile 1
Ignore deterministic tensor checking during sampling

Tensors resulting from sampling can not correspond with its plates. Solution: Create a "sampling" context in which the wrappers are temporarily disabled so that sampling is no longer a problem.
enhancement

opened by HEmile 1
Implement Expect

Implementing Expect is not as easy as I hoped. Unfortunately, in discrete VAEs, for example, you pass logits with shape [batch, latents, categories]. enumerate_support() does not recognize that batch is independent, but the different latents are dependent! It only enumerates the support over the categories dimension, but it should also enumerate over the latents dimension. To do this automatically however requires independency assumptions, which probably brings us back to Issue #4.

opened by HEmile 1
Implement the unordered set estimator

https://arxiv.org/abs/2002.06043

The unordered set estimator is mostly a sampling strategy. It probably also provides variance reduction for other estimators than the score function. Is it possible to make sampling modular from the rest of the estimators? In this case, reweighting the costs is probably enough...

It won't actually work with reparameterization: The reparameterization will not properly weight the different samples. Should rethink this.
enhancement help wanted

opened by HEmile 1
Bump certifi from 2020.12.5 to 2022.12.7
Bumps certifi from 2020.12.5 to 2022.12.7.

Commits

9e9e840 2022.12.07

b81bdb2 2022.09.24

939a28f 2022.09.14

aca828a 2022.06.15.2

de0eae1 Only use importlib.resources's new files() / Traversable API on Python ≥3.11 ...

b8eb5e9 2022.06.15.1

47fb7ab Fix deprecation warning on Python 3.11 (#199)

b0b48e0 fixes #198 -- update link in license

9d514b4 2022.06.15

4151e88 Add py.typed to MANIFEST.in to package in sdist (#196)

Additional commits viewable in compare view

Dependabot will resolve any conflicts with this PR as long as you don't alter it yourself. You can also trigger a rebase manually by commenting @dependabot rebase.

Dependabot commands and options

You can trigger Dependabot actions by commenting on this PR:

@dependabot rebase will rebase this PR

@dependabot recreate will recreate this PR, overwriting any edits that have been made to it

@dependabot merge will merge this PR after your CI passes on it

@dependabot squash and merge will squash and merge this PR after your CI passes on it

@dependabot cancel merge will cancel a previously requested merge and block automerging

@dependabot reopen will reopen this PR if it is closed

@dependabot close will close this PR and stop Dependabot recreating it. You can achieve the same result by closing it manually

@dependabot ignore this major version will close this PR and stop Dependabot creating any more for this major version (unless you reopen the PR or upgrade to it yourself)

@dependabot ignore this minor version will close this PR and stop Dependabot creating any more for this minor version (unless you reopen the PR or upgrade to it yourself)

@dependabot ignore this dependency will close this PR and stop Dependabot creating any more for this dependency (unless you reopen the PR or upgrade to it yourself)

@dependabot use these labels will set the current labels as the default for future PRs for this repo and language

@dependabot use these reviewers will set the current reviewers as the default for future PRs for this repo and language

@dependabot use these assignees will set the current assignees as the default for future PRs for this repo and language

@dependabot use this milestone will set the current milestone as the default for future PRs for this repo and language

You can disable automated security fix PRs for this repo from the Security Alerts page.

dependencies
opened by dependabot[bot] 0
Refactor: Use torchdim instead of Storchastics plating system
Storchastic uses a rather intricate system for batching over multiple dimensions, but it's rather buggy and hard to work with for end users. Recently PyTorch 1.12 introduced torchdim with first-class dimension objects that should serve the same purpose as plates in Storchastic. We will likely have much cleaner, easier to read, write and debug and faster code by adopting this new standard. See https://colab.research.google.com/drive/1BsVkddtVMX35aZAvo2GyI-wSFPVBCWuA#scrollTo=8511a637

It implements

Implicit batching: Two batch dimensions are joined together, just like in Storchastic.

Mixed named tensors: Only batch dimensions need to be named, event dimensions in Storchastic can just be numeric

enhancement
opened by HEmile 0
comparison between gumbel softmax, rebar, and relax

Hi, @HEmile , I test the example in examples/vae/discrete_vae.py, but find the gumbel softmax performs much better than rebar, relax and reinforce (testing loss after 10 epochs: 98 for gumbel, 165 for rebar, 208 for relax, 209 for reinforce). The performance is not that consistent with the results in your guide, where gumbel softmax is slightly worse than reinforce in terms of training loss. Is there anything wrong with the example? I don't change any parameters. Looking forward to your response.

opened by fnzhan 4

Concatenating Reparameterized and UnorderedSetEstimator samples into one tensor

Hi! I need to concatenate samples from these two methods into one tensor, so that I can sample from a continuous distribution using reparameterization, and sample from a discrete distribution using UnorderedSetEstimator. Is this functionality something that can be built in? It seems that this is not possible given the current iteration (see below for example error). Thank you!

import storch
import torch
import torch.distributions as td
method1 = storch.method.Reparameterization
method2 = storch.method.UnorderedSetEstimator

method1 = method1(plate_name="1",n_samples=25)
method2 = method2(plate_name="1",k=25)
p1 = td.Independent(td.Normal(loc=torch.zeros([1000,2]),scale=torch.ones([1000,2])),0)
p2 = td.Independent(td.OneHotCategorical(probs=torch.zeros([1000,3]).uniform_()),0)

samp1 = method1(p1)
samp2 = method2(p2)
samp1.shape # torch.Size([25, 1000, 2])
samp2.shape # torch.Size([25, 1000, 3])

storch.cat([samp1,samp2],2)
ValueError: Received a plate with name 1 that is not also an AncestralPlate.

enhancement

opened by DavidKLim 1

Implement GO/Generalized Stochastic Backpropagation

See https://hal.archives-ouvertes.fr/hal-02968975/document

Implement: For multivariate Bernoulli, Sample a single normal sample, then for each dimension, use this sample and flip the corresponding dimension.

Run the normal sample and the flipped samples. Weight the original sample with 1. The multiplicative is: the sum of the parameters for the normal sample. These are negated if the corresponding dimension is 1. Then for the flipped samples, use -parameter if it flipped to 0, otherwise use just the parameter (ie, not negated).

To make the zeroth-order correct, use importance sampling for the flipped samples. In the multiplicative estimator, divide again by this importance sampling to ensure correctness.

This doesn't use a baseline.

opened by HEmile 1

Storchastic is a PyTorch library for stochastic gradient estimation in Deep Learning

Related tags

Overview

Installation

Algorithms

In development

Planned

References

Comments

Issue

Environment

Traceback

Owner

Emile van Krieken

A PyTorch implementation of Learning to learn by gradient descent by gradient descent

The official implementation of You Only Compress Once: Towards Effective and Elastic BERT Compression via Exploit-Explore Stochastic Nature Gradient.

Trains an agent with stochastic policy gradient ascent to solve the Lunar Lander challenge from OpenAI

A Momentumized, Adaptive, Dual Averaged Gradient Method for Stochastic Optimization

Re-implementation of the Noise Contrastive Estimation algorithm for pyTorch, following "Noise-contrastive estimation: A new estimation principle for unnormalized statistical models." (Gutmann and Hyvarinen, AISTATS 2010)

Bayesian-Torch is a library of neural network layers and utilities extending the core of PyTorch to enable the user to perform stochastic variational inference in Bayesian deep neural networks

ESGD-M - A stochastic non-convex second order optimizer, suitable for training deep learning models, for PyTorch

Collision risk estimation using stochastic motion models

DeepLM: Large-scale Nonlinear Least Squares on Deep Learning Frameworks using Stochastic Domain Decomposition (CVPR 2021)

PyTorch implementation of SCAFFOLD (Stochastic Controlled Averaging for Federated Learning, ICML 2020).

Code for "Infinitely Deep Bayesian Neural Networks with Stochastic Differential Equations"

This project provides a stock market environment using OpenGym with Deep Q-learning and Policy Gradient.

PGPortfolio: Policy Gradient Portfolio, the source code of "A Deep Reinforcement Learning Framework for the Financial Portfolio Management Problem"(https://arxiv.org/pdf/1706.10059.pdf).

Binary Stochastic Neurons in PyTorch

PyTorch implementation for SDEdit: Image Synthesis and Editing with Stochastic Differential Equations

PyTorch implementation for Stochastic Fine-grained Labeling of Multi-state Sign Glosses for Continuous Sign Language Recognition.

PyTorch implementation for Score-Based Generative Modeling through Stochastic Differential Equations (ICLR 2021, Oral)

A fast, scalable, high performance Gradient Boosting on Decision Trees library, used for ranking, classification, regression and other machine learning tasks for Python, R, Java, C++. Supports computation on CPU and GPU.