Hi, I was trying to do inference with a previously trained Flux CNN that was saved with BSON with the following:
38 @load model_name model
39 model = model |> gpu
40 data = data |> gpu
41
42 println(typeof(model))
43 t = time()
44 predict = model(data)
This is the error message I got:
Chain{Tuple{Conv{2,2,typeof(elu),CuArray{Float32,4,Nothing},CuArray{Float32,1,Nothing}},Conv{2,2,typeof(elu),CuArray{Float32,4,Nothing},CuArray{Float32,1,Nothing}},Conv{2,2,typeof(elu),CuArray{Float32,4,Nothing},CuArray{Float32,1,Nothing}},BSON.__deserialized_types__.var"##913",Dense{typeof(identity),CuArray{Float32,2,Nothing},CuArray{Float32,1,Nothing}}}}
ERROR: MethodError: no method matching (::BSON.__deserialized_types__.var"##913")(::CuArray{Float32,4,Nothing})
The applicable method may be too new: running in world age 27198, while current world is 27199.
Closest candidates are:
#23(::Any) at D:\Work\Research\work\old\train_options.jl:78 (method too new to be called from this world context.)
Stacktrace:
[1] applychain(::Tuple{BSON.__deserialized_types__.var"##913",Dense{typeof(identity),CuArray{Float32,2,Nothing},CuArray{Float32,1,Nothing}}}, ::CuArray{Float32,4,Nothing}) at C:\Users\me\.julia\packages\Flux\2i5P1\src\layers\basic.jl:30 (repeats 4 times)
[2] (::Chain{Tuple{Conv{2,2,typeof(elu),CuArray{Float32,4,Nothing},CuArray{Float32,1,Nothing}},Conv{2,2,typeof(elu),CuArray{Float32,4,Nothing},CuArray{Float32,1,Nothing}},Conv{2,2,typeof(elu),CuArray{Float32,4,Nothing},CuArray{Float32,1,Nothing}},BSON.__deserialized_types__.var"##913",Dense{typeof(identity),CuArray{Float32,2,Nothing},CuArray{Float32,1,Nothing}}}})(::CuArray{Float32,4,Nothing}) at C:\Users\me\.julia\packages\Flux\2i5P1\src\layers\basic.jl:32
[3] inference(::Array{Float32,4}, ::String, ::Bool) at D:\Work\Research\work\train\inference.jl:44
[4] main(::Array{String,1}) at D:\Work\Research\work\train\inference.jl:96
[5] top-level scope at none:0
The error is from line 44. The first line of the output is what the println(typeof(model)) prints. What is “world age”? How do I solve this problem?
On another note, do I have to save Flux models with BSON? Or can I do it with other IO packages like JLD?