Several things seem to be happening here: - TensorFlow.jl does not export `tanh` (but does export `tanh_`, note underscore) - There is ambiguity among several `sigmoid` method signatures, which may cause the import to fail (try calling `TF.API.sigmoid` directly to see the warning). - Julia does not raise an import error because the name `tanh` already exists in every module namespace by default, due to implicit imports.
I would suggest to file an issue (or pull-request) at the TensorFlow.jl repository for the first two. On Wed, May 18, 2016 at 4:38 AM, Jon Norberg <[email protected]> wrote: > I am following the examples in tensor flow.jl I get all the examples to > work nicely. however, when I want to change an activation function to tan > or sigmoid I get an error. I suspect I am doing something wrong rather than > a problem with the package so I ask here... > > If I do this: > > using TensorFlow > using TensorFlow.Train > using TensorFlow.InputData > import TensorFlow: DT_FLOAT32 > import TensorFlow.API: relu, tanh, > sigmoid,softmax_cross_entropy_with_logits, AdamOptimizer, arg_max, equal, > cast > methods(relu) > > I get: > > 2 methods for generic function relu: > relu(features::Union{TensorFlow.CoreTypes.AbstractTensor,Void}) at > /Users/Raukhur/.julia/v0.4/TensorFlow/src/API/TfNn.jl:1006 > relu(features::Union{TensorFlow.CoreTypes.AbstractTensor,Void}, > name::Union{AbstractString,Void}) at > /Users/Raukhur/.julia/v0.4/TensorFlow/src/API/TfNn.jl:1006 > > but if I do this > > methods(tanh) > > I get 13 methods for tanh but none for tensor flow related types and for > > methods(sigmoid) I get > > LoadError: UndefVarError: sigmoid not defined > while loading In[42], in expression starting on line 6 > > So it seems relu gets imported but not tanh or sigmoid. also > > methods(TensorFlow.API.tanh) nor > methods(TensorFlow.tanh) works > > Its probably something simple I have overlooked....
