bug fixes

Author: fwaris

GCN/GCNModel.fs

@@ -10,6 +10,7 @@ let gcnLayer in_features out_features hasBias (adj:TorchTensor) =
     let bias = if hasBias then Parameter(randName(),Float32Tensor.empty([|out_features|],requiresGrad=true)) |> Some else None
     let parms = [| yield weight; if hasBias then yield bias.Value|]
     Init.kaiming_uniform(weight.Tensor) |> ignore
+    if hasBias then Init.uniform(bias.Value.Tensor,0.,1.0) |> ignore
 
     Model.create(parms,fun wts t -> 
         use support = t.mm(wts.[0])
@@ -22,9 +23,7 @@ let gcnLayer in_features out_features hasBias (adj:TorchTensor) =
 let create nfeat nhid nclass dropout adj =
     let gc1 = gcnLayer nfeat nhid true adj
     let gc2 = gcnLayer nhid nclass true adj        
-    // let relu = ReLU()
-    // let logm = LogSoftmax(1L)
-    let drp = if dropout then Dropout() |> M else Model.nop
+    let drp = if dropout > 0.0 then Dropout(dropout) |> M else Model.nop
 
     fwd3 gc1 gc2 drp (fun t g1 g2 drp -> 
         use t = gc1.forward(t)

GCN/TorchSharp.Fun.fs

@@ -162,14 +162,10 @@ let inline (->>) m1 m2 = compose (M m1) (None,M m2)
 let inline (=>>) m1 (n,m2) = compose (M m1) (Some n, M m2)
 
 module Tensor = 
+    //Note: ensure 't matches tensor datatype otherwise ToArray might crash the app (i.e. exception cannot be caught)
     let private _getData<'t> (t:TorchTensor) =
         let s = t.Data<'t>()
-        let xs = Array.zeroCreate s.Length
-        for i in 0 .. s.Length-1 do
-            xs.[i] <- s.[i]
-        
-        //s.ToArray()
-        xs
+        s.ToArray()
 
     let getData<'t> (t:TorchTensor) =
         if t.device_type <> DeviceType.CPU then 

GCN/Train.fs

@@ -18,7 +18,7 @@ let run (datafolder,no_cuda,fastmode,epochs,dropout,lr,hidden,seed,weight_decay)
 
     let nclass = labels.max().ToInt64() + 1L
 
-    let model = GCNModel.create features.shape.[1] (int64 hidden) nclass true adj
+    let model = GCNModel.create features.shape.[1] (int64 hidden) nclass dropout adj
     let loss = NN.Functions.nll_loss()
 
     if cuda then
@@ -29,10 +29,13 @@ let run (datafolder,no_cuda,fastmode,epochs,dropout,lr,hidden,seed,weight_decay)
     let train epoch =
         let t = DateTime.Now
         model.Module.Train()
+        let parms = model.Module.parameters()
         optimizer.zero_grad()
         let output = model.forward(features)
         let loss_train = loss.Invoke(output.[ idx_train], labels.[idx_train])
+        let ls = float loss_train
         let acc_train = Utils.accuracy(output.[idx_train], labels.[idx_train])
+        printfn $"training - loss: {ls}, acc: {acc_train}"
         loss_train.backward()
         optimizer.step()
 

GCN/Utils.fs

@@ -35,10 +35,12 @@ let sparse_mx_to_torch_sparse_tensor (m:Matrix<float32>) =
     let idxs = Seq.append rows cols |> Seq.toArray
     let idx1 = idxs |> Int64Tensor.from |> fun x -> x.view(2L,-1L)
     let vals = coo |> Seq.map(fun (r,c,v) -> v) |> Seq.toArray |> Float32Tensor.from     
-    Float32Tensor.sparse(idx1,vals,[|int64 m.RowCount; int64 m.ColumnCount|])
+    let t = Float32Tensor.sparse(idx1,vals,[|int64 m.RowCount; int64 m.ColumnCount|])
+    let dt = TorchSharp.Fun.Tensor.getData<float32>(t.to_dense())
+    t
 
 let accuracy(output:TorchTensor, labels:TorchTensor) = 
-    let predsData = TorchSharp.Fun.Tensor.getData<int64>(output)
+    let predsData = TorchSharp.Fun.Tensor.getData<float32>(output)
     let preds  = predsData |> Array.chunkBySize (int output.shape.[1]) |> Array.map maxIdx
     let lbls = TorchSharp.Fun.Tensor.getData<int64>(labels)
     let correct = Array.zip preds lbls |> Array.filter (fun (a,b) -> a = b) |> Array.length |> float