yax

projects/RbQ10/Project.toml

@@ -3,4 +3,8 @@ ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
 DimensionalData = "0703355e-b756-11e9-17c0-8b28908087d0"
 EasyHybrid = "61bb816a-e6af-4913-ab9e-91bff2e122e3"
 GLMakie = "e9467ef8-e4e7-5192-8a1a-b1aee30e663a"
+YAXArrays = "c21b50f5-aa40-41ea-b809-c0f5e47bfa5c"
 Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
+
+[sources]
+YAXArrays = {rev = "main", url = "https://github.com/JuliaDataCubes/YAXArrays.jl"}

projects/RbQ10/Q10_dd.jl

@@ -37,7 +37,7 @@ grads = backtrace(l)[1]
 # TODO: test DimArray inputs
 using DimensionalData, ChainRulesCore
 # mat = Matrix(df)'
-mat = Array(Matrix(df)')
+mat = Float32.(Array(Matrix(df)'))
 da = DimArray(mat, (Dim{:col}(Symbol.(names(df))), Dim{:row}(1:size(df,1))))
 
 ##! new dispatch
@@ -91,7 +91,7 @@ ar = rand(3,3)
 A = DimArray(ar, (Y([:a,:b,:c]), X(1:3)));
 grad = Zygote.gradient(x -> sum(x[Y=At(:b)]), A)
 
-xy = EasyHybrid.split_data((ds_p_f, ds_t), 0.8, shuffle=true, rng=Random.default_rng())
+# xy = EasyHybrid.split_data((ds_p_f, ds_t), 0.8, shuffle=true, rng=Random.default_rng())
 
 EasyHybrid.get_prediction_target_names(RbQ10)
 
@@ -100,3 +100,23 @@ xy1 = EasyHybrid.prepare_data(RbQ10, da)
 (x_train, y_train), (x_val, y_val) = EasyHybrid.split_data(da, RbQ10) # ; shuffleobs=false, split_data_at=0.8
 
 out = train(RbQ10, da, (:Q10, ); nepochs=200, batchsize=512, opt=Adam(0.01));
+
+using YAXArrays
+axDims = dims(da)
+
+ds_yax = YAXArray(axDims, da.data)
+
+ds_p_f = ds_yax[col=At(forcing_names ∪ predictor_names)]
+ds_t = ds_yax[col=At(target_names)]
+ds_t_nan = .!isnan.(ds_t) #  produces 1×35064 YAXArray{Float32, 2}, not a Bool
+ds_t_nan = map(x -> !isnan(x), ds_t) # 1×35064 YAXArray{Bool, 2}
+length(ds_t_nan)
+# is_no_nan = .!isnan.(y)
+
+
+ls = EasyHybrid.lossfn(RbQ10, ds_p_f, (ds_t, ds_t_nan), ps, st, LoggingLoss())
+ls_logs = EasyHybrid.lossfn(RbQ10, ds_p_f, (ds_t, ds_t_nan), ps, st, LoggingLoss(train_mode=false))
+acc_ = EasyHybrid.evaluate_acc(RbQ10, ds_p_f, ds_t, ds_t_nan, ps, st, [:mse, :r2], :mse, sum)
+
+# TODO how to proceed - would it work already for multiple targets?
+out_yax = train(RbQ10, ds_yax, (:Q10, ); nepochs=200, batchsize=512, opt=Adam(0.01));

src/train.jl

@@ -122,8 +122,10 @@ function train(hybridModel, data, save_ps;
     opt_state = Optimisers.setup(opt, ps)
 
     # ? initial losses
-    is_no_nan_t = .!isnan.(y_train)
-    is_no_nan_v = .!isnan.(y_val)
+    # is_no_nan_t = .!isnan.(y_train)
+    is_no_nan_t = map(x -> !isnan(x), y_train)
+    # is_no_nan_v = .!isnan.(y_val)
+    is_no_nan_v = map(x -> !isnan(x), y_val)
 
     l_init_train, _, init_ŷ_train =  evaluate_acc(hybridModel, x_train, y_train, is_no_nan_t, ps, st, loss_types, training_loss, agg)
     l_init_val, _, init_ŷ_val = evaluate_acc(hybridModel, x_val, y_val, is_no_nan_v, ps, st, loss_types, training_loss, agg)
@@ -194,7 +196,9 @@ function train(hybridModel, data, save_ps;
         for epoch in 1:nepochs
             for (x, y) in train_loader
                 # ? check NaN indices before going forward, and pass filtered `x, y`.
-                is_no_nan = .!isnan.(y)
+                # is_no_nan = .!isnan.(y)
+                is_no_nan = map(x -> !isnan(x), y) # doing this due to YAXArray Bool issue
+
                 if length(is_no_nan)>0 # ! be careful here, multivariate needs fine tuning
                     l, backtrace = Zygote.pullback((ps) -> lossfn(hybridModel, x, (y, is_no_nan), ps, st,
                         LoggingLoss(training_loss=training_loss, agg=agg)), ps)

src/utils/loss_fn.jl

@@ -57,7 +57,17 @@ function loss_fn(ŷ, y, y_nan, ::Val{:rmse})
     return sqrt(mean(abs2, (ŷ[y_nan] .- y[y_nan])))
 end
 function loss_fn(ŷ, y, y_nan, ::Val{:mse})
-    return mean(abs2, (ŷ[y_nan] .- y[y_nan]))
+    # Option 1: Convert to Array and compute MSE
+    #yh = Array(ŷ[y_nan])
+    #yt = Array(y[y_nan])
+    #return mean(abs2, yh .- yt)
+
+    # Option 2: Use YAXArray directly but map has to be used
+    return mean(x -> x, map((a,b)->(a-b)^2, ŷ[y_nan], y[y_nan]))
+
+    # Option 3 gives an error
+    #return mean(abs2, (ŷ[y_nan] .- y[y_nan])) # errors with ERROR: MethodError: no method matching to_yax(::Vector{Float32}) The function `to_yax` exists, but no method is defined for this combination of argument types.
+    # I guess our model output would need to yax and not Vector{Float32}
 end
 function loss_fn(ŷ, y, y_nan, ::Val{:mae})
     return mean(abs, (ŷ[y_nan] .- y[y_nan]))