Comments (4)
Good catch, this is because we don't have __lt
metamethod defined internally. I'll add this.
from torch-autograd.
An update —
We will not be able to directly support <,<=,>,>=,==
operations, because Lua requires that the variables on either side of the comparison have the same type. This will never be true when using autograd, because we wrap all values in objects to keep track of how they're used in the forward pass, to automatically infer the backward pass.
However, we can use our own utilities. So, I added util.le
, util.lt
, util.ge
, util.gt
, util.eq
so you can make comparisons safely in autograd. See the new LessThan test.
You're also keeping track of values in a dynamically-allocated array. This is the next thing we'll have to tackle, because autograd doesn't support assignment to arrays (autograd requires all method calls be pure functions). I'll make a little utility that could help with that.
from torch-autograd.
We added support for comparisons, as I mentioned above, and also now support concatenating numbers, using our own autograd.util.cat
function. It currently only supports catting variables of the same type (e.g. all FloatTensors or all numbers), and will probably stay that way for awhile, unless specific requests come up.
I modified your example only slightly, so please try this now:
t = require 'torch'
local grad = require 'autograd'
local util = require 'autograd.util'
params = {
a = t.randn(1,1)
}
f = function(params, x)
local result = t.sum(x * params.a)
-- sample from Bernoulli dist
local bernoulli = {}
if util.lt(torch.uniform(), t.sum(params.a)) then
bernoulli[1] = 1
end
return t.sum(util.cat(bernoulli) * result)
end
df = grad(f)
x = t.randn(1,1)
print(f(params, x))
print(df(params, x))
By the way, if you want to use optimized mode (autograd(f, {optimize=true})
), you will have to pre-generate the random numbers, and put them in the params
table. Otherwise, the if
statement will be optimized away entirely (we can't overload, and thus can't track, control flow in optimized mode).
from torch-autograd.
Closing. Reopen if something is wrong.
from torch-autograd.
Related Issues (20)
- discrepency between licenses HOT 1
- torch.add(tensor1, scalar, tensor2) not supported. HOT 1
- Support to torch.FloatTensor.resize
- "Failed to parse generated code" HOT 6
- Adding a torch-signal function
- "torch.DoubleTensor.t not currently supported". Workarounds? HOT 5
- Reusable AutoModule's HOT 1
- Question about implementing a pairwise L2 distances function HOT 1
- torch-cl?
- About tying the weights
- Wrong implemention for logsoftmax? HOT 1
- ABS still not implemented error but ABS code is actually there?
- Help with using autograd in training with wrapped NN modules HOT 4
- Autogradient of function cannot handle batched forward-prop HOT 1
- mailing list for autograd
- using `optimfn` with wrapped `nn` modules
- Feature: Autograd for tensor:split
- gradfuns.lua dimension problem on line 130
- FloatTensor.cdata not currently supported by autograd when implementing SSIM loss HOT 1
- Failed to parse generated code when optimize=true
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from torch-autograd.