Andrew Ng - Deep Learning - course1 - week 4, deep neural network

1.Code

#-*- utf-8 -*-
"""
date: 2018/12/3
cost: a day
"""

"""
log
今日发现的cost下降十分慢，与答案有出入的原因：
update写错
（另一个注意点：initialize parameters ------/np.sqrt() )
"""

import matplotlib.pyplot as plt
import numpy as np
import lr_utils
import testCases
from dnn_utils import relu, sigmoid


def get_data_set(layer_dims):

	train_x_orig, train_y_orig, test_x_orig, test_y_orig, classes= lr_utils.load_dataset()
	train_x = train_x_orig.reshape(train_x_orig.shape[0], -1).T / 255
	train_y = train_y_orig
	test_x = test_x_orig.reshape(test_x_orig.shape[0], -1).T / 255
	test_y = test_y_orig

	return train_x, train_y, test_x, test_y, classes


def initialize_parameters(layer_dims):

	np.random.seed(3)
	parameters = {}
	L = len(layer_dims)

	for x in range(1, L):
		parameters["W" + str(x)] = np.random.randn(layer_dims[x], layer_dims[x - 1]) / np.sqrt(layer_dims[x-1])
		parameters["b" + str(x)] = np.zeros((layer_dims[x],1))

		assert(parameters["W" + str(x)].shape == (layer_dims[x], layer_dims[x - 1]) )
		assert(parameters["b" + str(x)].shape == (layer_dims[x],1))

	return parameters


def sigmoid(Z):

	return 1 / (1 + np.exp(- Z))


def relu(Z):

	return np.maximum(0, Z)

def forward(X, parameters):
	A_caches = []
	A_caches.append(X)

	L = len(parameters) // 2
	A_prev = X
	for l in range(1, L+1):
		Z = np.dot(parameters["W" + str(l)], A_prev) + parameters["b" + str(l)]
		if l != L:
			A = relu(Z)			
		else:
			A = sigmoid(Z)
		A_prev = A
		A_caches.append(A)

	return A, A_caches



def relu_backward(dAL, A):

	dZ = np.array(dAL, copy=True)
	#dZ = dAL
	dZ[A == 0] = 0

	return dZ



def backward(AL, Y, parameters, A_caches):
	grads = {}	
	L = len(parameters) // 2
	dAL = - np.divide(Y, AL) + np.divide(1 -Y, 1- AL)
	#dAL = (1 - Y) / (1 - AL) - Y / AL
	grads["dA" + str(L)] = dAL
	
	for l in reversed(range(1, L+1)):
		if l == L:
			dZ = dAL *A_caches[l] * (1 - A_caches[l])
		else:
			dZ = relu_backward(dA, A_caches[l])
		dW = np.dot(dZ, A_caches[l - 1].T) / Y.shape[1]
		db = np.sum(dZ, axis = 1, keepdims = True) / Y.shape[1]
		grads["dW" + str(l)] = dW
		grads["db" + str(l)] = db
		if l != 1:
			dA = dA_prev = np.dot(parameters["W" + str(l)].T, dZ)
			grads["dA" + str(l-1)] = dA

	return grads


def update(parameters, grads, learning_rate):

	L = len(parameters) // 2
	for l in range(1,L+1):
		parameters["W" + str(l)] -= grads["dW" + str(l)] * learning_rate
		parameters["b" + str(l)] -= grads["db" + str(l)] * learning_rate

	return parameters



def dnn_model(train_x, train_y, parameters, num_iterations = 3000, learning_rate = 0.0075, print_cost = False, isPlot = False):

	costs = []
	for i in range(num_iterations):
		AL, A_caches = forward(train_x,  parameters)
		grads = backward(AL, train_y, parameters, A_caches)
		parameters = update(parameters,grads,learning_rate)

		if i % 100 == 0:
			cost = - np.mean(train_y * np.log(AL) + (1 - train_y) * np.log(1 - AL))
			costs.append(cost)

			if print_cost:
				print("第" + str(i+1) + "次迭代 cost = " + str(cost))
	if isPlot:
		plt.plot(costs)
		plt.xlabel("迭代次数(每百次)")
		plt.ylabel("cost")
		plt.show()

	return parameters


def predict(X, parameters):

	Y_forward,cache = forward(X, parameters)
	Y = np.array([[1 if x > 0.5 else 0 for x in Y_forward.squeeze()]])


	return Y
 


if __name__ == "__main__":

	layer_dims = [12288, 20, 7, 5, 1]
	train_x, train_y, test_x, test_y, classes = get_data_set(layer_dims)
	parameters = initialize_parameters(layer_dims)

	parameters = dnn_model(train_x, train_y, parameters, num_iterations = 3000, learning_rate = 0.0075, print_cost = True, isPlot = False)

	predict_train_y = predict(train_x, parameters)
	predict_test_y = predict(test_x, parameters)

	train_correct_rate = float(np.sum(train_y == predict_train_y) / train_y.shape[1])
	test_correct_rate = float(np.sum(test_y == predict_test_y) / test_y.shape[1])

	print("correct rate of train: " + str(train_correct_rate * 100) + "%") 
	print("correct rate of test: " + str(test_correct_rate * 100) + "%")

2.Outcome

3.Learning Reflection
3.1 pay much attention to unit test
3.2 make it easy