added on kaggle with 96.8

Assignment2/main.py

@@ -48,15 +48,16 @@ def preprocessMnist(X):
 def taskMnist():
 	XTrain, YTrain, XVal, YVal, XTest, _ = loadMnist()
 	# Create a NeuralNetwork object 'nn1' as follows with optimal parameters. For parameter definition, refer to py file.
-	lr,batchSize,epochs = 0.01,256,50
+	lr,batchSize,epochs = 0.005,256,10000
 	nn1 = nn.NeuralNetwork(lr, batchSize, epochs)
 	# Add layers to neural network corresponding to inputs and outputs of given data
 	input_layer = XTrain.shape[1]
-	hidden_layer = 256
+	hidden_layer1 = 128
+	# hidden_layer2 = 64
 	output_layer = 10
-# 	activation_fn = 'softmax' # 'relu'
-	nn1.addLayer(nn.FullyConnectedLayer(input_layer,hidden_layer,'relu'))
-	nn1.addLayer(nn.FullyConnectedLayer(hidden_layer,output_layer,'softmax'))
+	nn1.addLayer(nn.FullyConnectedLayer(input_layer,hidden_layer1,'relu'))
+	# nn1.addLayer(FullyConnectedLayer(hidden_layer1,hidden_layer2,'relu'))
+	nn1.addLayer(nn.FullyConnectedLayer(hidden_layer1,output_layer,'softmax'))
 	###############################################
 	# TASK 3b (Marks 13) - YOUR CODE HERE
 	# raise NotImplementedError

Assignment2/nn.py

@@ -26,23 +26,24 @@ class NeuralNetwork:
 		# validY - Corresponding list of validation data labels
 		for i in range(self.epochs):
 
-			dataSize = trainX.shape[0]
-			rndIndex = np.random.permutation(dataSize)
-			trainX,trainY = trainX[rndIndex],trainY[rndIndex]
-			print('Epoch',i+1)
-			batchDataX = trainX[:self.batchSize]
-			batchDataY = trainY[:self.batchSize]
-#       prediction = predict(batchDataX)
+			# dataSize = trainX.shape[0]
+			# rndIndex = np.random.permutation(dataSize)
+			# trainX,trainY = trainX[rndIndex],trainY[rndIndex]
+			numRand = np.random.randint(0,trainX.shape[0]-self.batchSize-1)
+			batchDataX = trainX[numRand:numRand+self.batchSize]
+			batchDataY = trainY[numRand:numRand+self.batchSize]
 			activations = []
 			activations.append(batchDataX)
 			for l in self.layers:
 				activations.append(l.forwardpass(activations[-1]))
 			L = self.crossEntropyLoss(batchDataY,activations[-1])
+			# prediction = predict(batchDataX)
+			# print(L)
 			delta = self.crossEntropyDelta(batchDataY,activations[-1])
 			for l in reversed(self.layers):
 				prev_activation = activations[self.layers.index(l)]
 				delta = l.backwardpass(prev_activation,delta)
-				l.updateWeights(0.01)  
+				l.updateWeights(self.lr)
 		# The methods trains the weights and baises using the training data(trainX, trainY)
 		# Feel free to print accuracy at different points using the validate() or computerAccuracy() functions of this class
 		###############################################
@@ -130,8 +131,8 @@ class FullyConnectedLayer:
 		###############################################
 		# TASK 1a (Marks 0) - YOUR CODE HERE
 		# raise NotImplementedError
-		self.weights = np.random.randn(in_nodes,out_nodes)/np.sqrt(in_nodes)
-		self.biases = np.zeros(out_nodes)
+		self.weights = np.random.randn(in_nodes,out_nodes)*1e-6
+		self.biases = np.zeros(out_nodes)*1e-6
 		###############################################
 		# NOTE: You must NOT change the above code but you can add extra variables if necessary
 		
@@ -173,7 +174,7 @@ class FullyConnectedLayer:
 		###############################################
 		# TASK 1c (Marks 3) - YOUR CODE HERE
 		exps = np.exp(X)
-		return exps / np.sum(exps)
+		return exps / np.sum(exps,axis=1,keepdims=True)
 		raise NotImplementedError
 		###############################################
 
@@ -186,11 +187,8 @@ class FullyConnectedLayer:
 		# Hint: You might need to compute Jacobian first
 		###############################################
 		# TASK 1f (Marks 7) - YOUR CODE HERE
-		nonZero = delta[delta!=0]
-		indexList = []
-		for i in range(len(delta)):
-			temp = list(delta[i])
-			indexList.append(temp.index(nonZero[i]))
+
+		indexList = np.nonzero(delta)[1].reshape(-1)
 		for i in range(len(indexList)):
 			temp = X[i][indexList[i]]
 			X[i] = -temp*X[i]
@@ -251,8 +249,8 @@ class FullyConnectedLayer:
 		# This function should actually update the weights using the gradients computed in the backwardpass
 		###############################################
 		# TASK 1h (Marks 2) - YOUR CODE HERE
-		self.weights = self.weights - lr*(self.weightsGrad)
-		self.biases = self.biases - lr*(self.biasesGrad)
+		self.weights += lr*(self.weightsGrad)
+		self.biases += lr*(self.biasesGrad)
     
 # 		raise NotImplementedError
 		###############################################