Merge pull request #4 from aicenter/dev_tpm_2023

Fix minor bugs

examples/base_spsn/set_classification.jl

@@ -29,14 +29,14 @@ end
 function knownsetmixture(μs, Σs, λs, prior)
     components = map(zip(μs, Σs, λs)) do ps
         pc = Poisson(log(ps[3]))
-        pf = MvNormalParams(ps[1], ps[2])
+        pf = MvNormalParams(Float32.(ps[1]), Float32.(ps[2]))
         SetNode(pf, pc)
     end
     SumNode(components, prior)    
 end
 
 ### Creating and sampling known model
-m1 = knownsetmixture(μ, Σ, λ, [1., 1., 1.])
+m1 = knownsetmixture(μ, Σ, λ, [1f0, 1, 1])
 nbags = 100
 bags, baglabels = randwithlabel(m1, nbags)
 

examples/base_spsn/set_clustering.jl

@@ -31,13 +31,13 @@ end
 function knownsetmixture(μs, Σs, λs, prior)
     components = map(zip(μs, Σs, λs)) do ps
         pc = Poisson(log(ps[3]))
-        pf = MvNormalParams(ps[1], ps[2])
+        pf = MvNormalParams(Float32.(ps[1]), Float32.(ps[2]))
         SetNode(pf, pc)
     end
     SumNode(components, prior)    
 end
 ### Creating and sampling known model
-m1 = knownsetmixture(μ, Σ, λ, [1., 1., 1.])
+m1 = knownsetmixture(μ, Σ, λ, [1f0, 1, 1])
 nbags = 300
 bags, baglabels = randwithlabel(m1, nbags)
 

examples/mutagenesis/base_model.ipynb

@@ -5,7 +5,15 @@
    "execution_count": 1,
    "id": "bbb741d7",
    "metadata": {},
-   "outputs": [],
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "\u001b[36m\u001b[1m[ \u001b[22m\u001b[39m\u001b[36m\u001b[1mInfo: \u001b[22m\u001b[39mPrecompiling SumProductSet [d0366596-3556-49ae-b3ef-851ab4ad1106]\n"
+     ]
+    }
+   ],
    "source": [
     "using DrWatson\n",
     "@quickactivate \n",
@@ -354,7 +362,7 @@
       "Epoch 98 - acc: | 0.890  0.773 | \n",
       "Epoch 99 - acc: | 0.890  0.773 | \n",
       "Epoch 100 - acc: | 0.890  0.773 | \n",
-      "108.356103 seconds (247.62 M allocations: 59.257 GiB, 10.02% gc time, 60.86% compilation time)\n"
+      "159.992875 seconds (247.45 M allocations: 59.246 GiB, 8.70% gc time, 72.44% compilation time)\n"
      ]
     }
    ],
@@ -372,6 +380,28 @@
   {
    "cell_type": "code",
    "execution_count": 10,
+   "id": "e18eaadc",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "1×100 Matrix{Float64}:\n",
+       " -189.62  -208.523  -248.18  -266.854  …  -259.407  -218.131  -210.526"
+      ]
+     },
+     "execution_count": 10,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "logpdf(m, ds_train)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 11,
    "id": "d5fa4303",
    "metadata": {},
    "outputs": [
@@ -381,7 +411,7 @@
        "516"
       ]
      },
-     "execution_count": 10,
+     "execution_count": 11,
      "metadata": {},
      "output_type": "execute_result"
     }
@@ -392,7 +422,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 11,
+   "execution_count": 12,
    "id": "f23baf1c",
    "metadata": {},
    "outputs": [

src/SumProductSet.jl

@@ -21,18 +21,17 @@ include("modelnodes/modelnode.jl")
 include("distributions/distributions.jl")
 include("modelbuilders.jl")
 include("util.jl")
+include("loss.jl")
 include("reflectinmodel.jl")
 
-export ZIPoisson, Poisson, Geometric, Categorical, MvNormal, MvNormalParams
+export Poisson, Geometric, Categorical, MvNormal, MvNormalParams
 export logpdf, logjnt
 export SumNode, ProductNode, SetNode
 export rand, randwithlabel
-export setmixture, gmm, sharedsetmixture, spn
+export gmm, setmixture, sharedsetmixture, spn
 
 export reflectinmodel
-export em_loss, ce_loss, ssl_loss
-
-export VAE, Encoder, Decoder, SplitLayer, elbo, reconstruct_loss
+export em_loss, ce_loss
 
 Base.show(io::IO, ::MIME"text/plain", n::AbstractModelNode) = HierarchicalUtils.printtree(io, n)
 

src/loss.jl

@@ -0,0 +1,20 @@
+
+"""
+    em_loss(m::SumNode, xu)
+Expetation maximization algorithm objective (multiplied by the minus sign) `for the top SumNode` layer for unlabeled data `xu``.  
+"""
+function em_loss(m::SumNode, xu)
+    # E-step for unlabeled data
+    p = []
+    Flux.Zygote.ignore() do 
+        p = softmax(logjnt(m, xu); dims=1)
+    end
+    # minus M-step objective
+    -mean(p .* logjnt(m, xu))
+end 
+
+"""
+    ce_loss(m::SumNode, xl, yl::Vector{Int})
+Cross entropy loss / negative log-likelihood loss for labeled data `xl` with corresponding labels `yl`. 
+"""
+ce_loss(m::SumNode, xl, yl::Vector{Int}) = -mean(logjnt(m, xl)[CartesianIndex.(yl, 1:length(yl))])