reexport Flux, update CI

.github/workflows/CI.yml

@@ -16,18 +16,18 @@ jobs:
       fail-fast: false
       matrix:
         version:
-          - "1.6" # Latest
+          - "1.9" # Latest
         os:
           - ubuntu-latest
           - windows-latest
         arch:
           - x64
     steps:
-      # install CUDA
-      - uses: Jimver/[email protected]
-        id: cuda-toolkit
-        with:
-          cuda: '11.2.2'
+      # # install CUDA
+      # - uses: Jimver/[email protected]
+      #   id: cuda-toolkit
+      #   with:
+      #     cuda: '11.2.2'
 
       # check out the project and install Julia
       - uses: actions/checkout@v2

Project.toml

@@ -16,18 +16,20 @@ POMDPs = "a93abf59-7444-517b-a68a-c42f96afdd7d"
 Parameters = "d96e819e-fc66-5662-9728-84c9c7592b0a"
 Plots = "91a5bcdd-55d7-5caf-9e0b-520d859cae80"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
+Reexport = "189a3867-3050-52da-a836-e630ba90ab69"
 Statistics = "10745b16-79ce-11e8-11f9-7d13ad32a3b2"
 StatsBase = "2913bbd2-ae8a-5f71-8c99-4fb6c76f3a91"
 TensorBoardLogger = "899adc3e-224a-11e9-021f-63837185c80f"
 ValueHistories = "98cad3c8-aec3-5f06-8e41-884608649ab7"
 WeightsAndBiasLogger = "71805093-c1fc-4af5-8d0a-8dde53c6ac46"
 Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
+cuDNN = "02a925ec-e4fe-4b08-9a7e-0d78e3d38ccd"
 
 [compat]
-CUDA = "3, 4"
+CUDA = "3, 4, 5"
 ColorSchemes = "3"
 Distributions = "0.25"
-Flux = "0.13"
+Flux = "0.14"
 Images = "0.25"
 POMDPTools = "0.1"
 POMDPs = "0.9"

examples/adversarial/continuous_pendulum.jl

@@ -61,9 +61,6 @@ vline!([deg2rad(20), deg2rad(-20)])
 
 plot(-3:0.1:3, [exp.(logpdf(antagonist(𝒮_isarl.π).A, [0, 0, 0], [x]))[1] for x=-3:0.1:3], label="Px")
 
-antagonist(𝒮_isarl.π).A.logΣ([0, 0, 0])
-, [1])[1]
-
 # Solve with DQN
 𝒮_dqn = DQN(π=QS(as), S=S, N=N)
 π_dqn = solve(𝒮_dqn, mdp)
@@ -82,10 +79,6 @@ println("RARL Failure rate: ", pfail_rarl)
 pfail_isarl = Crux.failure(Sampler(mdp, protagonist(π_isarl), S=S, max_steps=100), Neps=Int(1e5), threshold=100)
 println("IS Failure rate: ", pfail_isarl)
 
-pol = AdvPol()
-
-𝒮_isarl.buffer
-
 
 pol = AdversarialPolicy(π_dqn, Pf(xs), ϵGreedyPolicy(Crux.LinearDecaySchedule(1., 0.1, floor(Int, N/2)), xs))
 𝒮_isarl = ISARL_Discrete(π=pol, S=S, N=N, xlogprobs=xlogprobs)

src/Crux.jl

@@ -3,13 +3,14 @@ module Crux
 
     set_crux_warnings(val::Bool) = global CRUX_WARNINGS = val
 
+    using Reexport
     using Random
     using Distributions
     using POMDPs
     using POMDPTools:render
     using Parameters
     using TensorBoardLogger
-    using Flux
+    @reexport using Flux
     using Zygote
     import Zygote: ignore_derivatives
     using Flux.Optimise: train!
@@ -24,6 +25,7 @@ module Crux
     using Base.Iterators: partition
     using WeightsAndBiasLogger
     using Dates
+
 
     extra_functions = Dict()
     function set_function(key, val)

src/policies.jl

@@ -210,6 +210,7 @@ function exploration(π::MixtureNetwork, s; kwargs...)
 		end
 		indices
 	end
+    println("weights: ", αs, "indices: ", indices)
     a = hcat([exploration(d, s[:, i])[1] for (d, i) in zip(π.networks, indices)]...)
 
     return a, logpdf(π, s, a)

test/policy_tests.jl

@@ -151,7 +151,7 @@ USE_CUDA && @test all([Crux.device(n)== gpu for n in p_gpu.networks])
 # @test all(Crux.value(p, s) .≈ Crux.value(p_gpu, s))
 # @test Crux.valueall(p, s) == [Crux.value(p.networks[1], s), Crux.value(p.networks[2], s)]
 # @test Crux.valueall(p, s, a) == [Crux.value(p.networks[1], s, a), Crux.value(p.networks[2], s, a)]
-@test_broken try
+@test_broken try # I think this is because the mixture network isn't designed for batch actions?
 	action(p, s)
 	true
 catch