reorg

net.py

@@ -10,6 +10,10 @@ def unpickle(file):
 
 classes = ["airplane","automobile","bird","cat","deer","dog","frog","horse","ship","truck"];
 
+classcount = 10
+
+p = 0
+
 class NeuralNet:
 
 	def __init__(self):
@@ -18,33 +22,94 @@ def __init__(self):
 		np.random.ranf((10,1)) * np.random.ranf((1,10)),
 		np.random.ranf((10,1)) * np.random.ranf((1,10))]	
 
+	def predictThis(self,image,c,correct):
+		out0 = np.matmul(self.layer[0] , image)
+		maxout0 = np.amax(out0)
+		out0 = [x/(maxout0+10) for x in out0]
+
+		pout1 = np.matmul(self.layer[1] , out0)
+		maxout1 = np.amax(pout1)
+		pout1 = [x/(maxout1+10) for x in pout1]
+
+		out2 = np.matmul(self.layer[2] , pout1)
+		maxout2 = np.amax(out2)
+		out2 = [x/(maxout2+10) for x in out2]
+
+		out2 = np.array(out2).ravel()
+
+
+		out3 = np.matmul(self.layer[3] , out2)
+		maxout3 = np.amax(out3)
+		out3 = [x/(maxout3+10) for x in out3]
+
+		# print(" predicted = ",np.argmax(out3)," class = ",c)
+
+		if(np.argmax(out3) == c):
+			correct+=1
+
+		return correct
+
 	def trainThis(self,image,c):
-		out1 = np.matmul(self.layer[0] , image)
+		out0 = np.matmul(self.layer[0] , image)
+		maxout0 = np.amax(out0)
+		out0 = [x/(maxout0+10) for x in out0]
+
+		out1 = np.matmul(self.layer[1] , out0)
+		maxout1 = np.amax(out1)
+		out1 = [x/(maxout1+10) for x in out1]
+
+		out2 = np.matmul(self.layer[2] , out1)
+		maxout2 = np.amax(out2)
+		out2 = [x/(maxout2+10) for x in out2]
+
+		out2 = np.array(out2).ravel()
+		out3 = np.matmul(self.layer[3] ,out2)
+		maxout3 = np.amax(out3)
+		out3 = [x/(maxout3+10) for x in out3]
+
+
+		ideal = np.zeros(10)
+		ideal.fill(0.1)
+		ideal[c] = 0.9
+
+		# print(" ideal: ",ideal)
+		# print(" error: ",error)
+		# print("Layer 4","\t\t\t","Error")
+		# [ print(out3[i],"\t\t",error[i]) for i in range(10) ]
+
+		# backprop for layer 3
+		layer3ideal = np.matmul(np.asmatrix(ideal).transpose(),np.asmatrix(out2))
+		l3diff =  layer3ideal - self.layer[3]
+		self.layer[3] = self.layer[3] + (l3diff/5)
 
-		out1 = [x/10000 for x in out1]
-		out2 = np.matmul(self.layer[1] , out1)
-		out3 = np.matmul(self.layer[2] , out2)
-		out4 = np.matmul(self.layer[3] , out3)
-		out4 = [x/100 for x in out4]
 
-		# print(out1)
-		# print(out2)
-		# print(out3)
-		# print(out4)
-		print(" predicted = ",np.argmax(out4)," class = ",c)
-		print()
-		# ideal = 
+		global p
+		p = net.predictThis(image,c,p)
 
-		return out4;
+		return out3;
 
 	def printLayer(self,n):
 		print(self.layer[n])
 
 
 net = NeuralNet();
 
-[net.trainThis(imgset[b'data'][i],imgset[b'labels'][i]) for i in range(10)]
+trainCount = 1800
+
+[net.trainThis(imgset[b'data'][i],imgset[b'labels'][i]) for i in range(trainCount)]
+
+print("--------------------------------------------")
+print("Now for the tests: ")
 
+print(p)
+
+correct = 0
+total = 50
+
+for i in range(2001,2001+total):
+	correct = net.predictThis(imgset[b'data'][i],imgset[b'labels'][i],correct)
+
+print("\n\nAccuracy: ",correct,"/",total)
 
 
 # print(flayer)
@@ -58,14 +123,17 @@ def printLayer(self,n):
 # tp print image data
 # print(imgset[b'data'])
 
-[3072 x 3072][3072 x 1] = [3072 x 1]   -- layer 1
+[3072 x 3072][3072 x 1] = [3072 x 1]   -- layer 0
 
-[10 x 3072][3072 x 1] = [10 x 1]       -- layer 2
+[10 x 3072][3072 x 1] = [10 x 1]       -- layer 1
 
-[10 x 10][10 x 1] = [10 x 1]           -- layer 3
+[3072 x 10][10 x 1] = [3072 x 1]
 
-[10 x 10][10 x 1] = [10 x 1]           -- layer 4
+[10 x 10][10 x 1] = [10 x 1]           -- layer 2
+[10 x 10] = [10 x 1] [10 x 1].T
 
+[10 x 10][10 x 1] = [10 x 1]           -- layer 3
+[10 x 10] = [10 x 1] [1 x 10]t 
 
 '''
 

tf.py

@@ -0,0 +1,4 @@
+import tensorflow as tf
+hello = tf.constant('Hello, TensorFlow!')
+sess = tf.Session()
+print(sess.run(hello))