twitter/torch-autograd

Name: torch-autograd

Owner: Twitter, Inc.

Description: Autograd automatically differentiates native Torch code

Created: 2015-10-06 14:51:56.0

Updated: 2018-01-03 11:27:41.0

Pushed: 2017-08-29 17:35:10.0

Homepage:

Size: 2342

Language: Lua

GitHub Committers

Other Committers

README

Autograd

Autograd automatically differentiates native Torch code. Inspired by the original Python version.

Scope

Autograd has multiple goals:

• provide automatic differentiation of Torch expressions
• support arbitrary Torch types (e.g. transparent and full support for CUDA-backed computations)
• full integration with nn modules: mix and match auto-differentiation with user-provided gradients
• the ability to define any new nn compliant Module with automatic differentiation
• represent complex evaluation graphs, which is very useful to describe models with multiple loss functions and/or inputs
• graphs are dynamic, i.e. can be different at each function call: for loops, or conditional, can depend on intermediate results, or on input parameters
• enable gradients of gradients for transparent computation of Hessians

Updates

Jan 21, 2016: Two big new user-facing features:

• First, we now support direct assignment (so you can now do x[k] = v inside optimize=true autograd code, where k can be a number, table or LongTensor, and v can be a tensor or number, whichever is appropriate. Here's a few examples.
• Second, you can now take 2nd-order and higher gradients (supported in optimized mode. Either run autograd.optimize(true) or take the derivative of your function using df = autograd(f, {optimize = true}). Check out a simple example in our tests
• Plus, lots of misc bugfixes and new utilities to help with tensor manipulation (autograd.util.cat can work with numbers, or tensors of any time. autograd.util.cast can cast a nested table of tensors to any type you like).

Nov 16, 2015: Runtime performance was improved dramatically, as well as ease of use with better debugging tools. Performance is now within 30% of a statically described version of an equivalent model (nn and nngraph).

• a compute DAG is now generated and cached based on input tensors's dimensions
• the DAG is compiled into Lua code, with several optimizations
• all intermediate states (tensors) are saved and re-used in a tensor pool
• debugging facilities have been added: when debugging is enabled, a nan or inf will trigger a callback, that can be used to render a DOT representation of the graph (see debugging)
• now restricting user code to the functional API of Torch (a:add(b) forbidden, use res = torch.add(a,b) instead)
• additional control flags can be passed to d(f, {...}) to compute subparts of the graph (fprop or bprop), useful to generate a compiled fprop (see fine grained control)

Nov 6, 2015: initial release.

Install

• Install Torch (instructions here).
• Retrieve this repo
• Run: luarocks make

Examples

Autograd example

A simple neural network with a multinomial logistic loss:

ibraries:
require 'torch'
 = require 'autograd'

efine trainable parameters:
ms = {
 = {
  t.randn(100,50),
  t.randn(50,10),
,
 = {
  t.randn(50),
  t.randn(10),



efine model
alNet = function(params, x, y)
ocal h1 = t.tanh(x * params.W[1] + params.b[1])
ocal h2 = t.tanh(h1 * params.W[2] + params.b[2])
ocal yHat = h2 - t.log(t.sum(t.exp(h2)))
ocal loss = - t.sum(t.cmul(yHat, y))
eturn loss


radients:
ralNet = grad(neuralNet)

ome data:
t.randn(1,100)
t.Tensor(1,10):zero() y[1][3] = 1

ompute loss and gradients wrt all parameters in params:
ams, loss = dneuralNet(params, x, y)

n this case:
loss: is a scalar (Lua number)
dparams: is a table that mimics the structure of params; for
each Tensor in params, dparams provides the derivatives of the
loss wrt to that Tensor.

Important note: only variables packed in the first argument of the eval function will have their gradients computed. In the example above, if the gradients wrt x are needed, then x simply has to be moved into params. The params table can be arbitrarily nested.

See more complete examples in examples.

Assuming the model defined above, and a training set of {x,y} pairs, the model can easily be optimized using SGD:

i,sample in datasetIterator() do
- estimate gradients wrt params:
ocal grads, loss = dneuralNet(params, sample.x, sample.y)

- SGD step:
or i = 1,#params.W do
  -- update params with an arbitrary learning rate:
  params.W[i]:add(-.01, grads.W[i])
  params.b[i]:add(-.01, grads.b[i])
nd

Optimization

To enable the optimizer, which produces optimized representations of your loss and gradient functions (as generated lua code):

 = require 'autograd'
.optimize(true) -- global
l df = grad(f, { optimize = true }) -- for this function only
l grads = df(params)

Benefits:

• Intermediate tensors are re-used between invocations of df(), dramatically reducing the amount of garbage produced.
• Zero overhead from autograd itself, once the code for computing your gradients has been generated.
• On average, a 2-3x overall performance improvement.

Caveats:

• The generated code is cached based on the dimensions of the input tensors. If your problem is such that you have thousands of unique tensors configurations, you won't see any benefit.
• Each invocation of grad(f) produces a new context for caching, so be sure to only call this once.
• WARNING: Variables that you close over in an autograd function in optimize mode will never be updated – they are treated as static as soon as the function is defined.
• WARNING: If you make extensive use of control flow (any if-statements, for-loops or while-loops), you're better off using direct mode. In the best case, the variables used for control flow will be passed in as arguments, and trigger recompilation for as many possible branches as exist in your code. In the worst case, the variables used for control flow will be either computed internally, closed over, or not change in size or rank, and control flow changes will be completely ignored.

Wrapping nn modules

The nn library provides with all sorts of very optimized primitives, with gradient code written and optimized manually. Sometimes it's useful to rely on these for maximum performance.

Here we rewrite the neural net example from above, but this time relying on a mix of nn primitives and autograd-inferred gradients:

ibraries:
require 'torch'
 = require 'autograd'

efine trainable parameters:
ms = {
inear1 = {
  t.randn(50,100), -- note that parameters are transposed (nn convention for nn.Linear)
  t.randn(50),
,
inear2 = {
  t.randn(10,50),
  t.randn(10),



nstantiate nn primitives:
ote: we do this outside of the eval function, so that memory
s only allocated once; moving these calls to within the body
f neuralNet would work too, but would be quite slower.
ar1 = grad.nn.Linear(100, 50)
1 = grad.nn.Tanh()
ar2 = grad.nn.Linear(50, 10)
2 = grad.nn.Tanh()

efine model
alNet = function(params, x, y)
ocal h1 = acts1(linear1(params.linear1, x))
ocal h2 = acts2(linear2(params.linear2, h1))
ocal yHat = h2 - t.log(t.sum(t.exp(h2)))
ocal loss = - t.sum(t.cmul(yHat, y))
eturn loss


radients:
ralNet = grad(neuralNet)

ome data:
t.randn(1,100)
t.Tensor(1,10):zero() y[1][3] = 1

ompute loss and gradients wrt all parameters in params:
ams, loss = dneuralNet(params, x, y)

This code is stricly equivalent to the code above, but will be more efficient (this is especially true for more complex primitives like convolutions, …).

3rd party libraries that provide a similar API to nn can be registered like this:

l customnnfuncs = grad.functionalize('customnn')  -- requires 'customnn' and wraps it
le = customnnfuncs.MyNnxModule(...)

nder the hood, this is already done for nn:
.nn = grad.functionalize('nn')

On top of this functional API, existing nn modules and containers, with arbitarily nested parameters, can also be wrapped into functions. This is particularly handy when doing transfer learning from existing models:

efine a standard nn model:
l model = nn.Sequential()
l:add(nn.SpatialConvolutionMM(3, 16, 3, 3, 1, 1, 1, 1))
l:add(nn.Tanh())
l:add(nn.Reshape(16*8*8))
l:add(nn.Linear(16*8*8, 10))
l:add(nn.Tanh())
ote that this model could have been pre-trained, and reloaded from disk.

unctionalize the model:
l modelf, params = autograd.functionalize(model)

he model can now be used as part of a regular autograd function:
l loss = autograd.nn.MSECriterion()
alNet = function(params, x, y)
ocal h = modelf(params, x)
eturn loss(h, y)


ote: the parameters are always handled as an array, passed as the first
rgument to the model function (modelf). This API is similar to the other
odel primitives we provide (see below in "Model Primitives").

ote 2: if there are no parameters in the model, then you need to pass the input only, e.g.:
l model = nn.Sigmoid()
unctionalize :
l sigmoid = autograd.functionalize(model)

he sigmoid can now be used as part of a regular autograd function:
l loss = autograd.nn.MSECriterion()
alNet = function(params, x, y)
ocal h = sigmoid(x) -- please note the absence of params arg
eturn loss(h, y)

Creating auto-differentiated nn modules

For those who have a training pipeline that heavily relies on the torch/nn API, torch-autograd defines the autograd.nn.AutoModule and autograd.nn.AutoCriterion functions. When given a name, it will create a new class locally under autograd.auto.name. This class can be instantiated by providing a function, a weight, and a bias. They are also clonable, savable and loadable. Here we show an example of writing a 2-layer fully-connected module and an MSE criterion using AutoModule and AutoCriterion:

Here we rewrite the neural net example from above, but this time relying on a mix of nn primitives and autograd-inferred gradients:

efine functions for modules
inear
l linear  = function(input, weight, bias)
ocal y = weight * input + bias
eturn y


inear + ReLU
l linearReLU  = function(input, weight, bias)
ocal y = weight * input + bias
ocal output = torch.mul( torch.abs( y ) + y, 0.5)
eturn output


efine function for criterion
SE
l mse = function(input, target)
ocal buffer = input-target
eturn torch.sum( torch.cmul(buffer, buffer) ) / (input:dim() == 2 and input:size(1)*input:size(2) or input:size(1))


nput size, nb of hiddens
l inputSize, outputSize = 100, 1000

efine auto-modules and auto-criteria
nd instantiate them immediately
l autoModel = nn.Sequential()
l autoLinear1ReLU = autograd.nn.AutoModule('AutoLinearReLU')(linearReLU, linear1.weight:clone(), linear1.bias:clone())
l autoLinear2 = autograd.nn.AutoModule('AutoLinear')(linear, linear2.weight:clone(), linear2.bias:clone())
Model:add( autoLinear1ReLU )
Model:add( autoLinear2 )
l autoMseCriterion = autograd.nn.AutoCriterion('AutoMSE')(mse)
t this point, print(autograd.auto) should yield

 AutoLinearReLU : {...}
 AutoMSE : {...}
 AutoLinear : {...}


efine number of iterations and learning rate
l n = 100000
l lr = 0.001
l autoParams,autoGradParams = autoModel:parameters()
l unifomMultiplier = torch.Tensor(inputSize):uniform()

rain: this should learn how to approximate e^(\alpha * x)
ith an mlp aith both auto-modules and regular nn
i=1,n do
utoModel:zeroGradParameters()
ocal input = torch.Tensor(inputSize):uniform(-5,5):cmul(uniformMultiplier)
ocal target = input:clone():exp()
- Forward
ocal output = autoModel:forward(input)
ocal mseOut = autoMseCriterion:forward(output, target)
- Backward
ocal gradOutput = autoMseCriterion:backward(output, target)
ocal gradInput = autoModel:backward(input, gradOutput)
or i=1,#autoParams do
  autoParams[i]:add(-lr, autoGradParams[i])
nd

Gradient checks

For ease of mind (and to write proper tests), a simple grad checker is provided. See test.lua for complete examples. In short, it can be used like this:

arameters:
l W = t.Tensor(32,100):normal()
l x = t.Tensor(100):normal()

unction:
l func = function(inputs)
eturn t.sum(inputs.W * inputs.x)


heck grads wrt all inputs:
er:assert(gradcheck(func, {W=W, x=x}), 'incorrect gradients on W and x')

Model Primitives

To ease the construction of new models, we provide primitives to generate standard models.

Each constructor returns 2 things:

• f: the function, can be passed to grad(f) to get gradients
• params: the list of trainable parameters

Once instantiated, f and params can be used like this:

t = torch.randn(10)
 = f(params, input)
s = autograd(f)(params, input)

Current list of model primitives includes:

API:

rams = autograd.model.NeuralNetwork({
- number of input features:
nputFeatures = 10,

- number of hidden features, per layer, in this case
- 2 layers, each with 100 and 10 features respectively:
iddenFeatures = {100,10},

- activation functions:
ctivations = 'ReLU',

- if true, then no activation is used on the last layer;
- this is useful to feed a loss function (logistic, ...)
lassifier = false,

- dropouts:
ropoutProbs = {.5, .5},

API:

rams = autograd.model.SpatialNetwork({
- number of input features (maps):
nputFeatures = 3,

- number of hidden features, per layer:
iddenFeatures = {16, 32},

- poolings, for each layer:
oolings = {2, 2},

- activation functions:
ctivations = 'Sigmoid',

- kernel size:
ernelSize = 3,

- dropouts:
ropoutProbs = {.1, .1},

API:

rams = autograd.model.RecurrentNetwork({
- number of input features (maps):
nputFeatures = 100,

- number of output features:
iddenFeatures = 200,

- output is either the last h at step t,
- or the concatenation of all h states at all steps
utputType = 'last', -- or 'all'

API:

rams = autograd.model.RecurrentLSTMNetwork({
- number of input features (maps):
nputFeatures = 100,

- number of output features:
iddenFeatures = 200,

- output is either the last h at step t,
- or the concatenation of all h states at all steps
utputType = 'last', -- or 'all'

Loss Primitives

Similarly to model primitives, we provide common loss functions in autograd.loss:

ross entropy between 2 vectors:
for categorical problems, the target should be encoded as one-hot)
 = loss.crossEntropy(prediction, target)

inary cross entropy - same as above, but labels are considered independent bernoulli variables:
 = loss.binaryEntropy(prediction, target)

east squares - mean square error between 2 vectors:
 = loss.leastSquares(prediction, target)

Gradients of gradients

autograd can be called from within an autograd function, and the resulting gradients can used as part of your outer function:

l d = require 'autograd'
timize(true)
l innerFn = function(params)
- compute something...

l ddf = d(function(params)
ocal grads = d(innerFn)(params)
- do something with grads of innerFn...

l gradGrads = ddf(params) -- second order gradient of innerFn

(f, {
ebugHook = function(debugger, msg, gen)
  -- dump a dot representation of the graph:
  debugger.generateDot('result.dot')

  -- or show it (OSX only, uses Safari):
  debugger.showDot()

  -- print the generated source line that caused the inf/nan
  print(string.split(gen.source, "\n")[gen.line])
nd

l W = torch.Tensor(32,100):fill(.5)
l x = torch.Tensor(100):fill(.5)
l func = function(inputs)
eturn torch.sum(torch.div(inputs.W * inputs.x, 0))  -- DIV ZERO!

l dFunc = autograd(func, {
ebugHook = function(debugger, msg)
  debugger.showDot()
  print(msg)
  os.exit(0)
nd

c({W=W, x=x})

grad debugger detected a nan or inf value for locals[1]
: fn@path/to/code/example.lua:4

Finer-grain control over execution can also be achieved using these flags:

ll of these options default to true:
(f, {
ithForward = true | false,    -- compute the forward path
ithGradients = true | false,  -- compute the gradients (after forward)
artialGrad = true | false     -- partial grad means that d(f) expects grads wrt output


unning this:
 = grad(f, {withForward=true, withGradients=false})(inputs)
s equivalent to:
 = f(inputs)
.. but the function is compiled, and benefits from tensor re-use!

Licensed under the Apache License, Version 2.0. See LICENSE file.