fsprojects / fsharp-ai-tools Goto Github PK

Hi How can I make Tensorflow.FSharp run?
I can't find these packages below:

In both .fsproj files, the following is found:

<Reference Include="..\lib\TensorFlow.Proto.dll" />

When I run the setup script, the file that is built into the lib folder is actually called

TensorFlow.FSharp.Proto.dll

Running df (v 3.0) |> DT.Eval (or similar) from dsl-live.fsx results in an error:

Cannot compute the partial derivative for node 'BroadcastTo' as it's unreachable from the output node(s).

As far as I can tell, this occurs only when evaluating scalar -> scalar derivatives. I've been able to narrow the error down to: https://github.com/fsprojects/TensorFlow.FSharp/blob/master/src/TensorFlow.FSharp/dsl.fs#L929

This tf { return (vec [1.0; 2.0]).[0..0] } no longer works as it returns a shape mismatch error. I guess maybe it's expecting a vector and is receiving a scalar?

Also, a question, it seems like tensor4 and tensor3 require that all dimensions are equal, is that correct?

Thanks

I get the following error when trying to build livecheck tooling using the supplied instructions:

git clone http://github.com/Microsoft/visualfsharp
cd visualfsharp
git fetch https://github.com/dsyme/visualfsharp livecheck
git checkout livecheck <--- error: pathspec 'livecheck' did not match any file(s) known to git..\build.cmd vs
cd ..

I'm trying to run the example code in README.md

let f (xs: DT<double>) = 
     sin (v 0.5 * sqr xs.[0] - v 0.25 * sqr xs.[1] + v 3.0) * -cos (v 2.0 * xs.[0] + v 1.0 - exp xs.[1])

and getting

error FS0001: The type 'DT<float>' does not support the operator 'Sin'

It would be nice if the F# bindings could automatically use the naming capabilities of TensorFlow for variables and other operations to when creating the graphs. That would allow Tensorboard to show the results directly without any human intervention.

Like this C# code:

graph.Const (Mean, "Mean")

Should just be "Mean" in F#

In the ImageClassifier.fsx example, the line below causes an error: Dimension -1 must be >= 0

let input_placeholder =
    graph.Placeholder
        (TFDataType.Float32, shape = TFShape(-1L, -1L, -1L, 3L),
         name = "new_input")

Removing the shape parameter allows the demo to run successfully. But something like this also works:

TFShape(Some ([|None; None; None; Some 3|] |> Array.map Dimension))

Should -1L be detected and mapped to None in the constructors, something like what happens in shape?