vor rewrite2: vielleicht besser nicht Tensor fitten

1 files changed, 38 insertions, 35 deletions

diff --git a/boston.py b/boston.py
index 5996d59..ecf35c2 100644
--- a/boston.py
+++ b/boston.py
@@ -19,34 +19,39 @@ x_train, x_test, y_train, y_test = train_test_split(
 with tf.device('/cpu:0'):
     sess = tf.InteractiveSession()
 
-    vecUnfittedInput = tf.placeholder(tf.float32, [13], name="input")
-    numTarget = tf.placeholder(tf.float32, [1], name="target")
+    vecInput = tf.placeholder(tf.float32, [13], name="input")
+    # 'unfitted' means not squeezed between 0 and 1, but real values
+    numUnfittedTarget = tf.placeholder(tf.float32, [1], name="target")
 
     # numbers to scale the output between 0-1 and back to prices
     numFactor = tf.placeholder(tf.float32, [1], name="factor")
-    # ^ max(input_matrix) - min(input_matrix)
     numOffset = tf.placeholder(tf.float32, [1], name="offset")
-    # ^ min(input_matrix)
 
-    # scale the unscaled input
-    vecInput = tf.div(tf.sub(vecUnfittedInput, numOffset), numFactor)
-
-    vecBias = tf.Variable(tf.truncated_normal([13], stddev=0.1), name="bias")
-    vecWeights = tf.Variable(tf.truncated_normal([1, 13], stddev=0.1), name="weight")
+    vecBias = tf.Variable(tf.random_uniform([13], minval=0, maxval=1), name="bias")
+    vecBias = tf.Print(vecBias, [vecBias], "vecbias")
+    # vecWeights is not a true vector, it is turned by 90 degrees
+    vecWeights = tf.Variable(tf.random_uniform([1, 13], minval=0, maxval=1), name="weight")
+    vecWeights = tf.Print(vecWeights, [vecWeights], "vecweights")
 
+    # sum((i + b) * w)
     matWeighted = tf.mul(tf.add(vecInput, vecBias), vecWeights)
-    vecNetinput = tf.reduce_sum(matWeighted)
+    matWeighted = tf.Print(matWeighted, [matWeighted], "matweighted: ")  # DEBUG
+    vecNetinput = tf.reduce_sum(matWeighted, 1)
     vecNetinput = tf.Print(vecNetinput, [vecNetinput], "netinput: ")  # DEBUG
+    vecLayerOutput = tf.sigmoid(vecNetinput)
+    vecLayerOutput = tf.Print(vecLayerOutput, [vecLayerOutput], "vecLayerOutput: ")  # DEBUG
 
-    vecUnfittedOutput = tf.sigmoid(vecNetinput)
-
-    # unscale the output
-    vecOutput = tf.add(tf.mul(vecUnfittedOutput, numFactor), numOffset)
     # since this is the last layer, vecOutput is a numOutput
-    numOutput = vecOutput
+    numOutput = tf.reduce_sum(vecLayerOutput)
+    # unscale the output
+    numUnfittedOutput = tf.add(tf.mul(numOutput, numFactor), numOffset)
+    numUnfittedOutput = tf.Print(numUnfittedOutput, [numUnfittedOutput], "unfitted output: ")  # DEBUG
 
-    numDifference = tf.sub(numOutput, numTarget)
-    numMSE = tf.reduce_mean(tf.square(numDifference))
+    numUnfittedTarget = tf.Print(numUnfittedTarget, [numUnfittedTarget], "numUnfittedTarget: ")  # DEBUG
+    numDifference = tf.sub(numUnfittedOutput, numUnfittedTarget)
+    numDifference = tf.Print(numDifference, [numDifference], "numDifference: ")
+    numMSE = tf.reduce_sum(tf.square(numDifference))
+    numMSE = tf.Print(numMSE, [numMSE], "MSE: ")
 
     summaryMSE = tf.scalar_summary("MSE", numMSE)
     summaryDifference = tf.scalar_summary("Mean difference", numDifference)
@@ -55,35 +60,33 @@ with tf.device('/cpu:0'):
 
     # all variables have to be specified here
     sess.run(tf.initialize_all_variables())
-
-    # mse = tf.Print(mse, [mse], "mse: ")
     train_step = tf.train.GradientDescentOptimizer(0.01).minimize(numMSE)
 
 
-factor = 0  #max(target) - min(target)
-offset = 0  #min(target)
+factor = max(target) - min(target)
+offset = min(target)
 
 for count in range(0, len(x_train)):
     trainsteps = 100
+    print("count " + str(count))
     for i in range(0, trainsteps):  # 100 epochs
-        #if i % 10 == 9:
-            #feed = {
-                #"input": x_test[testcount],  # x_test,
-                #"target": y_test[testcount],  # [y_test],
-                #"factor": factor,
-                #"offset": offset
-            #}
-            #result = sess.run([summaries, numMSE], feed_dict=feed)
+        if i % 10 == 9:
+            feed = {
+                vecInput: x_test[0],
+                numUnfittedTarget: [y_test[0]],
+                numFactor: [factor],
+                numOffset: [offset]
+            }
+            result = sess.run([summaries, numMSE], feed_dict=feed)
+            #print("test " + result[0])
             #writer.add_summary(result[0], count * trainsteps + i)
 
-        y_t = 1 #float(y_train[count].tolist())
-        print(y_t)
         sess.run(train_step,
                  feed_dict={
-                     "input": 1,  #x_train[count].tolist(),
-                     "target": 1,  #y_t,
-                     "factor": 1,  #[[factor]],
-                     "offset": 1,  #[[offset]]
+                     vecInput: x_train[count],
+                     numUnfittedTarget: [y_train[count]],
+                     numFactor: [factor],
+                     numOffset: [offset]
                  })
 
 print("finished training")