Merge pull request #1081 from ChrisRackauckas-Claude/fix-nan-gradient-handling

Skip gradient updates when gradients contain NaN or Inf

Author: ChrisRackauckas

lib/OptimizationOptimisers/Project.toml

@@ -4,33 +4,34 @@ authors = ["Vaibhav Dixit <[email protected]> and contributors"]
 version = "0.3.13"
 
 [deps]
-OptimizationBase = "bca83a33-5cc9-4baa-983d-23429ab6bcbb"
-SciMLBase = "0bca4576-84f4-4d90-8ffe-ffa030f20462"
+Logging = "56ddb016-857b-54e1-b83d-db4d58db5568"
 Optimisers = "3bd65402-5787-11e9-1adc-39752487f4e2"
+Optimization = "7f7a1694-90dd-40f0-9382-eb1efda571ba"
+OptimizationBase = "bca83a33-5cc9-4baa-983d-23429ab6bcbb"
 Reexport = "189a3867-3050-52da-a836-e630ba90ab69"
-Logging = "56ddb016-857b-54e1-b83d-db4d58db5568"
+SciMLBase = "0bca4576-84f4-4d90-8ffe-ffa030f20462"
 
-[extras]
-ForwardDiff = "f6369f11-7733-5829-9624-2563aa707210"
-Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
-MLUtils = "f1d291b0-491e-4a28-83b9-f70985020b54"
-MLDataDevices = "7e8f7934-dd98-4c1a-8fe8-92b47a384d40"
-Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
-Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
-ComponentArrays = "b0b7db55-cfe3-40fc-9ded-d10e2dbeff66"
-Lux = "b2108857-7c20-44ae-9111-449ecde12c47"
-Printf = "de0858da-6303-5e67-8744-51eddeeeb8d7"
+[sources]
+OptimizationBase = {path = "../OptimizationBase"}
 
 [compat]
-julia = "1.10"
-OptimizationBase = "4"
-SciMLBase = "2.122.1"
+Logging = "1.10"
 Optimisers = "0.2, 0.3, 0.4"
+OptimizationBase = "4"
 Reexport = "1.2"
-Logging = "1.10"
+SciMLBase = "2.122.1"
+julia = "1.10"
 
-[sources]
-OptimizationBase = {path = "../OptimizationBase"}
+[extras]
+ComponentArrays = "b0b7db55-cfe3-40fc-9ded-d10e2dbeff66"
+ForwardDiff = "f6369f11-7733-5829-9624-2563aa707210"
+Lux = "b2108857-7c20-44ae-9111-449ecde12c47"
+MLDataDevices = "7e8f7934-dd98-4c1a-8fe8-92b47a384d40"
+MLUtils = "f1d291b0-491e-4a28-83b9-f70985020b54"
+Printf = "de0858da-6303-5e67-8744-51eddeeeb8d7"
+Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
+Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
+Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 
 [targets]
 test = ["ComponentArrays", "ForwardDiff", "Lux", "MLDataDevices", "MLUtils", "Random", "Test", "Zygote", "Printf"]

lib/OptimizationOptimisers/src/OptimizationOptimisers.jl

@@ -67,6 +67,7 @@ function SciMLBase.__solve(cache::OptimizationCache{O}) where {O <: AbstractRule
     breakall = false
     progress_id = :OptimizationOptimizersJL
     for epoch in 1:epochs, d in data
+
         if cache.f.fg !== nothing && dataiterate
             x = cache.f.fg(G, θ, d)
             iterations += 1
@@ -106,7 +107,7 @@ function SciMLBase.__solve(cache::OptimizationCache{O}) where {O <: AbstractRule
         if cache.progress
             message = "Loss: $(round(first(first(x)); digits = 3))"
             @logmsg(LogLevel(-1), "Optimization", _id=progress_id,
-                message=message, progress=iterations / maxiters)
+                message=message, progress=iterations/maxiters)
         end
         if cache.solver_args.save_best
             if first(x)[1] < first(min_err)[1]  #found a better solution
@@ -129,7 +130,12 @@ function SciMLBase.__solve(cache::OptimizationCache{O}) where {O <: AbstractRule
                 break
             end
         end
-        state, θ = Optimisers.update(state, θ, G)
+        # Skip update if gradient contains NaN or Inf values
+        if all(isfinite, G)
+            state, θ = Optimisers.update(state, θ, G)
+        elseif cache.progress
+            @warn "Skipping parameter update due to NaN or Inf in gradients at iteration $iterations" maxlog=10
+        end
     end
     cache.progress && @logmsg(LogLevel(-1), "Optimization",
         _id=progress_id, message="Done", progress=1.0)

lib/OptimizationOptimisers/test/runtests.jl

@@ -134,3 +134,47 @@ end
 
     @test res.objective < 1e-4
 end
+
+@testset "NaN/Inf gradient handling" begin
+    # Test that optimizer skips updates when gradients contain NaN or Inf
+    # Function that can produce NaN due to sqrt of negative number
+    function weird_nan_function(x, p)
+        val = (p[1] - x[1])^2 + p[2] * (x[2] - x[1]^2)^2
+        # sqrt of a value that can become negative produces NaN
+        val += sqrt(max(x[1], 0.0)) * 0.01
+        return val
+    end
+
+    x0 = [-0.5, 0.1]  # Start with negative x[1] to trigger sqrt of negative
+    _p = [1.0, 100.0]
+
+    optprob = OptimizationFunction(weird_nan_function, OptimizationBase.AutoZygote())
+    prob = OptimizationProblem(optprob, x0, _p)
+
+    # Should not throw error and should complete all iterations
+    sol = solve(prob, Optimisers.Adam(0.01), maxiters = 50, progress = false)
+
+    # Verify solution completed all iterations
+    @test sol.stats.iterations == 50
+
+    # Verify parameters are not NaN (would be NaN if updates were applied with NaN gradients)
+    @test all(!isnan, sol.u)
+    @test all(isfinite, sol.u)
+
+    # Function with 1/x that can produce Inf gradient when x is very small
+    function weird_inf_function(x, p)
+        val = (p[1] - x[1])^2 + p[2] * (x[2] - x[1]^2)^2
+        # 1/(x[1] + 0.01) can have very large gradient near x[1] = -0.01
+        val += 0.01 / (abs(x[1] - 0.1) + 1e-8)
+        return val
+    end
+
+    optprob_inf = OptimizationFunction(weird_inf_function, OptimizationBase.AutoZygote())
+    prob_inf = OptimizationProblem(optprob_inf, x0, _p)
+
+    sol_inf = solve(prob_inf, Optimisers.Adam(0.01), maxiters = 50, progress = false)
+
+    @test sol_inf.stats.iterations == 50
+    @test all(!isnan, sol_inf.u)
+    @test all(isfinite, sol_inf.u)
+end