brunjlar / neural Goto Github PK

View Code? Open in Web Editor NEW

122.0 122.0 17.0 13.41 MB

Neural Nets in native Haskell

License: MIT License

Haskell 99.77% Nix 0.23%

neural's People

Contributors

Stargazers

Watchers

Forkers

cutsea110 alexanderkjeldaas ahumay snowmantw tshimanga santolucito capn-freako kevinniechen hardentoo fragamus almays andwilley mariszo admiraloom

neural's Issues

Compile error when building from Hackage

Hi,
I'm trying to install neural 0.3.0.1 from Hackage, and I'm getting build errors.
I know that github says build is green, however Hackage says build is failing, here: https://hackage.haskell.org/package/neural-0.3.0.1/reports/

Any chance this could be fixed?

This is my build log:

[13 of 22] Compiling Data.FixedSize.Vector ( src/Data/FixedSize/Vector.hs, dist/dist-sandbox-50b32bd7/build/Data/FixedSize/Vector.o )

src/Data/FixedSize/Vector.hs:104:31: error:
    • Couldn't match type ‘1 + n’ with ‘n + 1’
      Expected type: VS.Vector (n + 1) a
        Actual type: VS.Vector (1 + n) a
      NB: ‘+’ is a type function, and may not be injective
    • In the second argument of ‘($)’, namely ‘VS.cons x xs’
      In the expression: Vector $ VS.cons x xs
      In an equation for ‘cons’:
          cons x (Vector xs) = Vector $ VS.cons x xs
    • Relevant bindings include
        xs :: VS.Vector n a (bound at src/Data/FixedSize/Vector.hs:104:16)
        cons :: a -> Vector n a -> Vector (n + 1) a
          (bound at src/Data/FixedSize/Vector.hs:104:1)

src/Data/FixedSize/Vector.hs:112:28: error:
    • Couldn't match type ‘n + 1’ with ‘1 + n0’
      Expected type: VS.Vector (1 + n0) a
        Actual type: VS.Vector (n + 1) a
      NB: ‘+’ is a type function, and may not be injective
      The type variable ‘n0’ is ambiguous
    • In the first argument of ‘VS.head’, namely ‘v’
      In the expression: VS.head v
      In an equation for ‘vhead’: vhead (Vector v) = VS.head v
    • Relevant bindings include
        v :: VS.Vector (n + 1) a
          (bound at src/Data/FixedSize/Vector.hs:112:15)
        vhead :: Vector (n + 1) a -> a
          (bound at src/Data/FixedSize/Vector.hs:112:1)

src/Data/FixedSize/Vector.hs:120:37: error:
    • Couldn't match type ‘n + 1’ with ‘1 + n’
      Expected type: VS.Vector (1 + n) a
        Actual type: VS.Vector (n + 1) a
      NB: ‘+’ is a type function, and may not be injective
    • In the first argument of ‘VS.tail’, namely ‘v’
      In the second argument of ‘($)’, namely ‘VS.tail v’
      In the expression: Vector $ VS.tail v
    • Relevant bindings include
        v :: VS.Vector (n + 1) a
          (bound at src/Data/FixedSize/Vector.hs:120:15)
        vtail :: Vector (n + 1) a -> Vector n a
          (bound at src/Data/FixedSize/Vector.hs:120:1)
cabal: Leaving directory '/tmp/cabal-tmp-5629/neural-0.3.0.1'
cabal: Error: some packages failed to install:
neural-0.3.0.1 failed during the building phase. The exception was:
ExitFailure 1

Adding convolutional version of MNIST example?

Hi Lars,

Thanks for this library!

After studying the code in your MNIST.hs example, I assume that you are not, yet, using convolutional layers in your MNIST digit classification example. Is that correct? If so, I'd like to contribute to this project, by adding that (i.e. - modifying the MNIST.hs example to use convolutional layers).

Did you already have an idea in mind, as to how you'd like to proceed with this?

And, finally, I was hoping you might have time to consider these two questions:

In MNIST.hs, why is the MNISTModel type defined as:
```
type MNISTModel = Classifier (Matrix 28 28) 10 Img Digit
```
when the learning appears to work on the flattened 784-element vector, instead?
This choice seems to force mnistModel.f to come inside the learning loop and, therefore, be defined as a Diff, whereas otherwise it might have been defined as a simple function (just flattening the 2D image into a 1D vector, and remaining outside of the learning loop).
Referring to this comment:

A Classifier f n b c is a Model that classifies items of type b into categories of type c, using a component with input shape f and output shape Vector n.

Why is the apparent type redundancy required?
In other words, why not just Classifier f n (or, Classifier b c)?

Thanks!
-db

OpenCL support

I know its too much to ask, but can i expect OpenCL support for AMD/nVidia gpu cards?

I think haskell is best for pretty much anything, so I wish to have a haskell library that supports gpu computations and can ease the level of experimentation/research that I can do with neural nets!

neural-0.3.0.1 will not build because of fixed vector

I have no idea how to fix these.

The library will not build, here is my log:

Configuring neural-0.3.0.1...
Building neural-0.3.0.1...
Preprocessing library neural-0.3.0.1...
[ 1 of 22] Compiling Data.Utils.Statistics ( src/Data/Utils/Statistics.hs, dist/build/Data/Utils/Statistics.o )
[ 2 of 22] Compiling Data.Utils.Arrow ( src/Data/Utils/Arrow.hs, dist/build/Data/Utils/Arrow.o )
[ 3 of 22] Compiling Data.MyPrelude   ( src/Data/MyPrelude.hs, dist/build/Data/MyPrelude.o )
[ 4 of 22] Compiling Data.Utils.Analytic ( src/Data/Utils/Analytic.hs, dist/build/Data/Utils/Analytic.o )
[ 5 of 22] Compiling Data.Utils.Cache ( src/Data/Utils/Cache.hs, dist/build/Data/Utils/Cache.o )
[ 6 of 22] Compiling Data.Utils.Stack ( src/Data/Utils/Stack.hs, dist/build/Data/Utils/Stack.o )
[ 7 of 22] Compiling Data.Utils.Traversable ( src/Data/Utils/Traversable.hs, dist/build/Data/Utils/Traversable.o )
[ 8 of 22] Compiling Data.Utils.List  ( src/Data/Utils/List.hs, dist/build/Data/Utils/List.o )
[ 9 of 22] Compiling Data.Utils.Pipes ( src/Data/Utils/Pipes.hs, dist/build/Data/Utils/Pipes.o )
[10 of 22] Compiling Data.Utils.Random ( src/Data/Utils/Random.hs, dist/build/Data/Utils/Random.o )
[11 of 22] Compiling Numeric.Neural.Model ( src/Numeric/Neural/Model.hs, dist/build/Numeric/Neural/Model.o )
[12 of 22] Compiling Data.FixedSize.Class ( src/Data/FixedSize/Class.hs, dist/build/Data/FixedSize/Class.o )
[13 of 22] Compiling Data.FixedSize.Vector ( src/Data/FixedSize/Vector.hs, dist/build/Data/FixedSize/Vector.o )

src/Data/FixedSize/Vector.hs:74:16: error:
    * Couldn't match type `Int'
                     with `finite-typelits-0.1.3.0:Data.Finite.Internal.Finite n'
      Expected type: (Index (Vector n) -> a) -> Vector n a
        Actual type: (finite-typelits-0.1.3.0:Data.Finite.Internal.Finite
                        n
                      -> a)
                     -> Vector n a
    * In the expression: Vector . VS.generate
      In an equation for `generate': generate = Vector . VS.generate
      In the instance declaration for `FixedSize (Vector n)'
    * Relevant bindings include
        generate :: (Index (Vector n) -> a) -> Vector n a
          (bound at src/Data/FixedSize/Vector.hs:74:5)

src/Data/FixedSize/Vector.hs:104:31: error:
    * Couldn't match type `1 + n' with `n + 1'
      Expected type: VS.Vector (n + 1) a
        Actual type: VS.Vector (1 + n) a
      NB: `+' is a type function, and may not be injective
    * In the second argument of `($)', namely `VS.cons x xs'
      In the expression: Vector $ VS.cons x xs
      In an equation for `cons':
          cons x (Vector xs) = Vector $ VS.cons x xs
    * Relevant bindings include
        xs :: VS.Vector n a (bound at src/Data/FixedSize/Vector.hs:104:16)
        cons :: a -> Vector n a -> Vector (n + 1) a
          (bound at src/Data/FixedSize/Vector.hs:104:1)

src/Data/FixedSize/Vector.hs:112:28: error:
    * Couldn't match type `n + 1' with `1 + n0'
      Expected type: VS.Vector (1 + n0) a
        Actual type: VS.Vector (n + 1) a
      NB: `+' is a type function, and may not be injective
      The type variable `n0' is ambiguous
    * In the first argument of `VS.head', namely `v'
      In the expression: VS.head v
      In an equation for `vhead': vhead (Vector v) = VS.head v
    * Relevant bindings include
        v :: VS.Vector (n + 1) a
          (bound at src/Data/FixedSize/Vector.hs:112:15)
        vhead :: Vector (n + 1) a -> a
          (bound at src/Data/FixedSize/Vector.hs:112:1)

src/Data/FixedSize/Vector.hs:120:37: error:
    * Couldn't match type `n + 1' with `1 + n'
      Expected type: VS.Vector (1 + n) a
        Actual type: VS.Vector (n + 1) a
      NB: `+' is a type function, and may not be injective
    * In the first argument of `VS.tail', namely `v'
      In the second argument of `($)', namely `VS.tail v'
      In the expression: Vector $ VS.tail v
    * Relevant bindings include
        v :: VS.Vector (n + 1) a
          (bound at src/Data/FixedSize/Vector.hs:120:15)
        vtail :: Vector (n + 1) a -> Vector n a
          (bound at src/Data/FixedSize/Vector.hs:120:1)
cabal: Leaving directory '/tmp/cabal-tmp-8105/neural-0.3.0.1'

What does "generation" mean?

Does the word "generation", as used in the reporting from this library, mean:

a full epoch, or
just a mini-batch iteration?

How to isolate a particular hidden layer result, in a CNN?

Referring to Lars' MNIST.hs example, and given this layout of a CNN:

    c = reLULayer
      . cArr (Diff toVector)
      . (convolution (Proxy :: DP.Proxy 7) 3 reLULayer :: Component (Volume 28 28 1) (Volume 8 8 8))
      . cArr (Diff fromMatrix)

what's the best way to make one of the 8 8x8 convolution results available to the reporting pipe?

I'd like to have the reporting pipe dump this "image" every so often, so I can see the convolution kernels "tuning" themselves to certain image features.

Also, am I correct in assuming that I'll fail if I try to brute force this by picking apart the model component, via pattern matching, due to the existentially hidden shape of the model parameter set?

Please fill out `description:` field

As you can see on https://hackage.haskell.org/package/neural-0.1.0.0, the description says "Please see README.md". If you upload packages to Hackage, you're expected to write a short abstact in the description field. An optional README does not replace the description field, but rather expands on it, since most tooling doesn't have access to the README, and in that context referring to an inaccessible README is rather pointless.

brunjlar / neural Goto Github PK

neural's People

Contributors

Stargazers

Watchers

Forkers

neural's Issues

Compile error when building from Hackage

Adding convolutional version of MNIST example?

OpenCL support

neural-0.3.0.1 will not build because of fixed vector

What does "generation" mean?

How to isolate a particular hidden layer result, in a CNN?

Please fill out `description:` field

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent