Two SimpleChains overwrite each other's predictions?

[jmsull]

The predictions of two SCs which should have nothing to do with each other seem to be overwriting each other? This seems like a bug, but let me know if I am doing something obviously wrong?

Here is a ~MWE of some code I have been running:

using SimpleChains
sc1 = SimpleChain(
    static(1),
    TurboDense{true}(tanh, 128), 
    TurboDense{true}(tanh, 128), 
    TurboDense{false}(identity, 1), 
    )
p1 = SimpleChains.init_params(sc1);

sc2 = SimpleChain(
    static(1),
    TurboDense{true}(tanh, 128),  
    TurboDense{true}(tanh, 128), 
    TurboDense{false}(identity, 1), 
    )
p2 = SimpleChains.init_params(sc2); 

@assert p1 != p2

X = randn(Float32,1,10); #some inputs
sc1(X, p1)[1,1] ≈ sc2(X, p2)[1,1] #false, not the same

y1 = sc1(X, p1) 
y2 = sc2(X, p2)
y1[1,1] == y2[1,1] #true, the same

[mohdibntarek]

I can reproduce this.

[chriselrod]

The predictions of two SCs which should have nothing to do with each other seem to be overwriting each other?

Note

julia> y1 === y2
true

julia> sc1 === sc2
true

The two scs are actually indistinguishable.

A SimpleChain overwrites its previous answer to avoid allocating:

julia> @benchmark $sc2($X, $p2)
BenchmarkTools.Trial: 10000 samples with 7 evaluations.
 Range (min … max):  4.615 μs …  8.601 μs  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     4.643 μs              ┊ GC (median):    0.00%
 Time  (mean ± σ):   4.649 μs ± 61.947 ns  ┊ GC (mean ± σ):  0.00% ± 0.00%

       ▁▄▇██▄                                                 
  ▂▂▃▄▆███████▆▄▃▂▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▂▂▂▂▂▂▂▂▂▂▂ ▃
  4.62 μs        Histogram: frequency by time         4.8 μs <

 Memory estimate: 0 bytes, allocs estimate: 0.

So evaluating the same SimpleChain twice is expected to do this.

But there should at least be a ! function that lets you pass your own arrays to update in place (and then perhaps some other convenient wrapper that allocates an unaliased array).

Another hacky workaround:

julia> @inline mytanh(x) = Base.FastMath.tanh_fast(x)
mytanh (generic function with 1 method)

julia> sc3 = SimpleChain(
           static(1),
           TurboDense{true}(tanh, 128),  
           TurboDense{true}(mytanh, 128), 
           TurboDense{false}(identity, 1), 
           )
SimpleChain with the following layers:
TurboDense static(128) with bias.
Activation layer applying: tanh
TurboDense static(128) with bias.
Activation layer applying: mytanh
TurboDense static(1) without bias.

julia> sc1 === sc3 # not the same chain
false

julia> y3 = sc3(X, p2)
1×10 StrideArray{Float32, 2, (1, 2), Tuple{StaticInt{1}, Int64}, Tuple{Nothing, Nothing}, Tuple{StaticInt{1}, StaticInt{1}}, Vector{UInt8}} with indices static(1):static(1)×Base.OneTo(10):
 0.337447  0.488095  -0.00920447  -0.2206  -0.242737  0.150495  -0.502708  0.291047  -0.895863  -0.337516

julia> y1 = sc1(X, p1)
1×10 StrideArray{Float32, 2, (1, 2), Tuple{StaticInt{1}, Int64}, Tuple{Nothing, Nothing}, Tuple{StaticInt{1}, StaticInt{1}}, Vector{UInt8}} with indices static(1):static(1)×Base.OneTo(10):
 -0.0441075  -0.0641206  0.00119711  0.0287531  0.0316527  …  0.0660765  -0.037995  0.118903  0.0441165

julia> y1 === y3 # different chains, different answers
false

julia> y3
1×10 StrideArray{Float32, 2, (1, 2), Tuple{StaticInt{1}, Int64}, Tuple{Nothing, Nothing}, Tuple{StaticInt{1}, StaticInt{1}}, Vector{UInt8}} with indices static(1):static(1)×Base.OneTo(10):
 0.337447  0.488095  -0.00920447  -0.2206  -0.242737  0.150495  -0.502708  0.291047  -0.895863  -0.337516

[jmsull]

Thanks for the reply!

Ok I didn't realize the scs were indistinguishable (and see that now) - but to confirm, this is the expected behavior? Is this written in the docs or source comments anywhere?

Yes, something like a wrapper or ! function would be great if that's possible!