Flux.jl: tensorflow error: you must feed a value for placeholder tensor 'placeholder_8' with dtype float

[ExpandingMan]

I'm probably getting a bit ahead of things here since I'm using master of TensorFlow.jl as I'm on 0.6 and the latest release doesn't work. TensorFlow itself seems to be working just fine on master.

Anyway, I get the following error when attempting to train the MNIST example:

ERROR: LoadError: Tensorflow error: Status: You must feed a value for placeholder tensor 'placeholder_8' with dtype float
         [[Node: placeholder_8 = Placeholder[_class=[], dtype=DT_FLOAT, shape=[], _device="/job:localhost/replica:0/task:0/gpu:0"]()]]
         [[Node: placeholder_8/_17 = _Recv[client_terminated=false, recv_device="/job:localhost/replica:0/task:0/cpu:0", send_device="/job:localhost/replica:0/task:0/gpu:0", send_device_incarnation=1, tensor_name="edge_3_placeholder_8", tensor_type=DT_FLOAT, _device="/job:localhost/replica:0/task:0/cpu:0"]()]]

Stacktrace:
 [1] check_status at /home/expandingman/.julia/v0.6/TensorFlow/src/core.jl:402 [inlined]
 [2] run(::TensorFlow.Session, ::Array{TensorFlow.Port,1}, ::Array{Any,1}, ::Array{TensorFlow.Port,1}, ::Array{Ptr{Void},1}) at /home/expandingman/.julia/v0.6/TensorFlow/src/run.jl:100
 [3] run(::TensorFlow.Session, ::Array{TensorFlow.Tensor{Float32},1}, ::Dict{Any,Any}) at /home/expandingman/.julia/v0.6/TensorFlow/src/run.jl:169
 [4] run(::TensorFlow.Session, ::TensorFlow.Tensor{Float32}, ::Dict{Any,Any}) at /home/expandingman/.julia/v0.6/TensorFlow/src/run.jl:187
 [5] back!(::Flux.TF.Exec, ::Array{Float32,2}, ::Array{Float64,2}) at /home/expandingman/.julia/v0.6/Flux/src/backend/tensorflow/model.jl:44
 [6] macro expansion at /home/expandingman/.julia/v0.6/Flux/src/training.jl:36 [inlined]
 [7] macro expansion at /home/expandingman/.julia/v0.6/Juno/src/progress.jl:128 [inlined]
 [8] macro expansion at /home/expandingman/.julia/v0.6/Flux/src/training.jl:15 [inlined]
 [9] macro expansion at /home/expandingman/.julia/v0.6/Juno/src/progress.jl:128 [inlined]
 [10] #train!#119(::Array{##3#4,1}, ::Int64, ::Float64, ::Function, ::Function, ::Flux.TF.Model, ::Array{Tuple{Array{Float64,1},Array{Int64,1}},1}) at /home/expandingman/.julia/v0.6/Flux/src/training.jl:29
 [11] (::Flux.#kw##train!)(::Array{Any,1}, ::Flux.#train!, ::Flux.TF.Model, ::Array{Tuple{Array{Float64,1},Array{Int64,1}},1}) at ./<missing>:0
 [12] include_from_node1(::String) at ./loading.jl:552
 [13] include(::String) at ./sysimg.jl:14
while loading /home/expandingman/src/test_flux.jl, in expression starting on line 22

As I suspect this is mostly the fault of my insisting on using TensorFlow master as a result of using 0.6, I'll probably look into patching this myself until the 0.6 ecosystem matures.

Update: It seems that placeholder_8 is a placeholder that is getting added to the TensorFlow graph object, but is disconnected from the actual graph (i.e. not used by any functions). Again, this is in the basic MNIST example. Apparently, it is the placeholder for the gradients.

[MikeInnes]

Yes, this issue is a valid one with Flux. We changed our approach to gradients to make things more consistent but were waiting on https://github.com/malmaud/TensorFlow.jl/pull/215 to complete it. Should be fixed soon.

[ExpandingMan]

Thanks for the update. Indeed it seems that TensorFlow merged their pull request already. In that case I'll wait rather than trying to get this going through some hack.

By the way, the warnings about the dot operators seem very hard to fix. Will this require some way of passing multiple function arguments to graph, as in graph(::typeof(broadcast), ::typeof(+), args...) or will there be some simpler solution?

[MikeInnes]

That's one option. Right now I'm trying to implement #31 in DataFlow.jl, which is a slightly different design. With that you'd dispatch on graph(::Broadcast{typeof(+)}, args...)

[MikeInnes]

So now DataFlow just interprets f.(x) as broadcast(f, x). This is pretty easy to handle in Flux; the right graph method gets called in Flux already, so we just need to add the overloads.

[baggepinnen]

I am still having this issue, has it been resolved on some branch other than the latest tagged version perhaps?

[MikeInnes]

This should be fixed in the latest release.