Update README and upload neuralnetwork.py

Author: rahuldesai1

hw4/README.md

@@ -1,18 +1,19 @@
 # Introduction to Tensorflow
-Please submit each notebook by the below deadline to gradescope.
+Please submit each notebook by the below deadline to gradescope. This is just to keep you on pace to complete the homeworks in time.
 
-**DEADLINE:** Saturday (10/19)
+NOTE: Deadlines have been shifted around.
+
+**DEADLINE:** Sunday (10/20)
+1. Start with `neural_network.py`. In this file you will be implementing a neural network from scratch using numpy! This is a tough one! Feel free to consult Google or anyone in the club if you need help!
+
+**DEADLINE:** Tuesday (10/22)
 
 2. Do the Tenorflow Tutorial notebook. You should finish this notebook by Saturday.
 
-**DEADLINE:** Monday (10/21) --> For both of these
+**DEADLINE:** Thursday (10/24) --> For both of these
 
 3. Next, you will be implementing a Denoising Autoencoder in tensorflow! In this homework you will gain more tf experience and train a model on AWS! 
 
 4. Finish with the Keras is Cool notebook. This will introduce you to Keras. It should not take too long and is not nearly as important as the other homeworks.
 
-**DEADLINE:** Thursday (10/24)
-NOTE: This is not been pushed to github yet.... will be release in the near future. 
-1. Start with `neural_network.py`. In this file you will be implementing a neural network from scratch using numpy! Don't worry tho...there is plenty of guidance and if you have any questions feel free to reach out.
-
 

hw4/neuralnetwork.py

@@ -0,0 +1,53 @@
+import numpy as np
+
+class NeuralNet:
+
+    #constructor, we will hardcode this to a 1 hidden layer network, for simplicity
+    #the problem we will grade on is differentiating 0 and 1s
+    #Some things/structuure may need to be changed. What needs to stay consistant is us being able to call
+    #forward with 2 arguments: a data point and a label. Strange architecture, but should be good for learning
+    def __init__(self, input_size=784, hidden_size=100, output_size=1):
+        self.input_size = input_size
+        self.hidden_size = hidden_size
+        self.output_size = output_size
+        #YOUR CODE HERE, initialize appropriately sized weights/biases with random paramters
+        self.weight1 = 
+        self.bias1 = 
+        self.weight2 = 
+        self.bias2 = 
+
+    #Potentially helpful, np.dot(a, b), also @ is the matrix product in numpy (a @ b)
+
+    #loss function, implement L1 loss
+    #YOUR CODE HERE
+    def loss(self, y0, y1):
+
+    #relu and sigmoid, nonlinear activations
+    #YOUR CODE HERE
+    def relu(self, x):
+
+    #You also may want the derivative of Relu and sigmoid
+
+    def sigmoid(self, x):
+
+    #forward function, you may assume x is correct input size
+    #have the activation from the input to hidden layer be relu, and from hidden to output be sigmoid
+    #have your forward function call backprop: we won't be doing batch training, so for EVERY SINGLE input,
+    #we will update our weights. This is not always (maybe not even here) possible or practical, why?
+    #Also, normally forward doesn't take in labels. Since we'll have forward call backprop, it'll take in labels
+    #YOUR CODE HERE
+    def forward(self, x, label):
+
+    #implement backprop, might help to have a helper function update weights
+    #Recommend you check out the youtube channel 3Blue1Brown and their video on backprop
+    #YOUR CODE HERE
+    def backprop(self, x, label): #What else might we need to take in as arguments? Modify as necessary
+
+        #Compute the gradients first
+        #First will have to do with combining derivative of sigmoid, output layer, and what else?
+        #np.sum(x, axis, keepdims) may be useful
+
+        #Update your weights and biases. Use a learning rate of 0.1, and update on every call to backprop
+        lr = .1
+
+