add untested manifold function

binaryva.lua

@@ -15,6 +15,9 @@ require 'LinearVA'
 --For loading data files
 require 'load'
 
+--For saving weights and biases
+require 'hdf5'
+
 require 'adagrad'
 
 data = load28('datasets/mnist.hdf5')
@@ -74,6 +77,7 @@ opfunc = function(batch)
     de_dw = KLD:backward(va:get(1).output, batch)
     encoder:backward(batch,de_dw)
 
+
     lowerbound = err  + KLDerr
     weights, grads = va:parameters()
 
@@ -107,4 +111,15 @@ while true do
     end
 
     print("\nEpoch: " .. epoch .. " Lowerbound: " .. lowerbound/data.train:size(1) .. " time: " .. sys.clock() - time)
+    if epoch % 2 == 0 then
+        local myFile = hdf5.open('params/epoch_' .. epoch .. '.hdf5', 'w')
+
+        myFile:write('weighttanh', va:get(3).weight)
+        myFile:write('biastanh', va:get(3).bias)
+        myFile:write('weightsigmoid', va:get(5).weight)
+        myFile:write('biassigmoid', va:get(5).bias)
+
+        myFile:close()
+    end
+
 end

plot.py

@@ -8,15 +8,38 @@
 """
 Code for creating manifold later!
 
-z = np.matrix([sp.norm.ppf(gridValues[i]),sp.norm.ppf(gridValues[j])]).T
-if continuous:
-    h_decoder = np.log(1 + np.exp(np.dot(W4,z) + b4))
-    y = 1 / (1 + np.exp(-(W5.dot(h_decoder) + b5)))
-else:
-    h_encoder = np.tanh(W4.dot(z) + b4)
-    y = 1 / (1 + np.exp(-(W5.dot(h_encoder) + b5)))
 """
 
+def manifold(gridSize, epoch):
+    f = h5py.File('params/epoch_' + epoch + '.hdf5','r')
+
+    wtanh = np.array(f["wtanh"])
+    btanh = np.array(f["btanh"])
+    wsig = np.array(f["wsig"])
+    bsig = np.array(f["bsig"])
+
+    wb = (wtanh,btanh,wsig,bsig)
+    gridValues = np.linspace(0.05,0.95,gridSize)
+
+    z = lambda i,j: np.matrix([sp.norm.ppf(gridValues[i]),sp.norm.ppf(gridValues[j])]).T
+
+    if continuous:
+        h_decoder = np.log(1 + np.exp(np.dot(W4,z) + b4))
+        y = 1 / (1 + np.exp(-(W5.dot(h_decoder) + b5)))
+
+    image = np.vstack([np.hstack([activation_binary(z(i+j),wb).reshape(shape) for j in xrange(numcols)]) for i in columns])
+
+    plt.imshow(image, interpolation='nearest', cmap='Greys')
+    plt.axis('off')
+    plt.show()
+
+def activation_binary(z, wb):
+    wtanh, btanh, wsig, bsig = wb
+    h = np.tanh(wtanh.dot(z) + btanh)
+    y = 1 / (1 + np.exp(-(wsig.dot(h) + bsig)))
+
+    return y
+
 def plotdigits(numcols):
     f = h5py.File('datasets/mnist.hdf5','r')
     data = np.array(f["x_train"])
@@ -31,5 +54,4 @@ def plotdigits(numcols):
     plt.axis('off')
     plt.show()
 
-plotdigits(5)
-
+manifold(10,5)