深度學習聖經“花書”各章摘要與筆記整理

《深度學習》，又名“花書”。該書由三位大佬 Ian Goodfellow、Yoshua Bengio 和 Aaron Courville 撰寫，是深度學習領域奠基性的經典教材，被譽為深度學習的“聖經”。

原書內容非常充實，接近 800 頁。這本書內容很深很全面，但起點稍微高了一些，對數學理論基礎知識要求的比較多。因此，讀完之後，及時進行高度概括和經驗總結是十分有幫助的。石頭君最近在 GitHub 上發現一個關於花書各章摘要的專案，內容非常精煉，除了筆記的同時，部分章節還配備程式碼，值得推薦，我們一起來看一下。

該專案的名稱是：Deep-Learning-Book-Chapter-Summaries，作者是 Aman Dalmia 和 Ameya Godbole 兩位小哥。專案地址為：

https://github.com/dalmia/Deep-Learning-Book-Chapter-Summaries

主要內容

這份花書核心筆記主要涉及的章節包括：

• ch02 線性代數

• ch03 概率與資訊理論

• ch04 數值優化

• ch07 深度學習正則化

• ch08 深度模型中的優化

• ch09 卷積網路

• ch11 實踐方法論

• ch13 線性因子模型

筆記的形式是 .ipynb，便於在 Jupyter Notebook 上開啟和觀看。例如，我們來看一下第二章線性代數的筆記。

可見，Jupyter 筆記不僅包含了知識點的總結，也有相關程式碼。再來看第九章的卷積網路部分，配備了一些完整的圖片處理程式碼。

import numpy as np
from scipy import signal
from scipy import misc
import matplotlib.pyplot as plt

# %matplotlib inline

img = misc.ascent()
kernel = np.random.randn(5,5)
# kernel = np.array([[0,-10,0,10,0],[-10,-30,0,30,10],[0,-10,0,10,0]])

img = img.astype(np.float32)/255
orig_in = img

offsetx = offsety = 20
shift_in = np.zeros(orig_in.shape)
shift_in[offsetx:,offsety:] = img[:-offsetx,:-offsety]

rot_in = misc.imrotate(img, 90)
scale_in = misc.imresize(orig_in, 1.5)

output1 = signal.convolve2d(orig_in, kernel, mode='same')
output2 = signal.convolve2d(shift_in, kernel, mode='same')
output3 = signal.convolve2d(rot_in, kernel, mode='same')
output4 = signal.convolve2d(scale_in, kernel, mode='same')

fig, axes = plt.subplots(2, 4, figsize=(14, 7))
ax_orig = axes[0,0]
ax_shift = axes[0,1]
ax_rot = axes[0,2]
ax_scale = axes[0,3]

diff_orig = axes[1,0]
diff_shift = axes[1,1]
diff_rot = axes[1,2]
diff_scale = axes[1,3]

ax_orig.imshow(output1, cmap='gray')
ax_orig.set_title('Original')
ax_shift.imshow(output2, cmap='gray')
ax_shift.set_title('Shifted')
ax_rot.imshow(output3, cmap='gray')
ax_rot.set_title('Rotated')
ax_scale.imshow(output4, cmap='gray')
ax_scale.set_title('Scaled')

def shift(arr, offset):
output = np.zeros(arr.shape)
output[offset:, offset:] = arr[:-offset,:-offset]
return output

def rotate(arr, angle):
return misc.imrotate(arr, angle)

def resize(arr, scale):
return misc.imresize(arr, scale)

diff_orig.hist(np.ravel(output1),bins=100)
diff_orig.set_title('Output histogram')
diff_shift.hist(np.ravel(np.abs(output2-shift(output1, 20))),bins=100)
diff_shift.set_title('Shift histogram difference')
diff_rot.hist(np.ravel(np.abs(output3-rotate(output1, 10))),bins=100)
diff_rot.set_title('Rotate histogram difference')
diff_scale.hist(np.ravel(np.abs(output4-resize(output1, 1.5))),bins=100)
diff_scale.set_title('Scale histogram difference')

ax_orig.set_xticks([])
ax_shift.set_xticks([])
ax_rot.set_xticks([])
ax_scale.set_xticks([])

ax_orig.set_yticks([])
ax_shift.set_yticks([])
ax_rot.set_yticks([])
ax_scale.set_yticks([])

plt.tight_layout()
# plt.show()
plt.savefig('images/conv_equivariance.png')

對於池化層的程式碼示例：

import numpy as np
np.random.seed(101)

from scipy import signal
from scipy import misc
import matplotlib.pyplot as plt
%matplotlib inline

img = misc.ascent()
img = img.astype(np.float32)/255

# The image is more interesting here
orig_in = img[-200:,-300:-100]
offsetx = offsety = 15
shift_in = img[-200-offsetx:-offsetx,-300-offsety:-100-offsety]
kernel1 = np.random.randn(5,5)
kernel2 = np.random.randn(5,5)
kernel3 = np.random.randn(5,5)

def sigmoid(arr):
# Lazy implementation of sigmoid activation
return 1./(1 + np.exp(-arr))

def maxpool(arr, poolsize, stride):
# Lazy looping implementation of maxpool
output_shape = np.floor((np.array(arr.shape)-poolsize)/stride)+1
output_shape = output_shape.astype(np.int32)
output = np.zeros(output_shape)
for x in range(output_shape[0]):
for y in range(output_shape[1]):
output[x,y] = np.max(arr[x*stride:x*stride+poolsize,y*stride:y*stride+poolsize])
return output

output1_1 = signal.convolve2d(orig_in, kernel1, mode='valid')
pool1_1 = maxpool(output1_1, 2, 2)
actv1_1 = sigmoid(pool1_1)
output1_2 = signal.convolve2d(actv1_1, kernel2, mode='valid')
pool1_2 = maxpool(output1_2, 2, 2)
actv1_2 = sigmoid(pool1_2)
output1_3 = signal.convolve2d(actv1_2, kernel3, mode='valid')
pool1_3 = maxpool(output1_3, 2, 2)

output2_1 = signal.convolve2d(shift_in, kernel1, mode='valid')
pool2_1 = maxpool(output2_1, 2, 2)
actv2_1 = sigmoid(pool2_1)
output2_2 = signal.convolve2d(actv2_1, kernel2, mode='valid')
pool2_2 = maxpool(output2_2, 2, 2)
actv2_2 = sigmoid(pool2_2)
output2_3 = signal.convolve2d(actv2_2, kernel3, mode='valid')
pool2_3 = maxpool(output2_3, 2, 2)

fig, axes = plt.subplots(4, 3, figsize=(10, 10))

k1, k2, k3 = axes[0,:]
p1_1, p1_2, p1_3 = axes[1,:]
p2_1, p2_2, p2_3 = axes[2,:]
h1, h2, h3 = axes[3,:]

k1.imshow(kernel1, cmap='gray')
k1.set_title('kernel1')
k2.imshow(kernel2, cmap='gray')
k2.set_title('kernel2')
k3.imshow(kernel3, cmap='gray')
k3.set_title('kernel3')
k1.set_xticks([])
k2.set_xticks([])
k3.set_xticks([])
k1.set_yticks([])
k2.set_yticks([])
k3.set_yticks([])

p1_1.imshow(pool1_1, cmap='gray')
p1_1.set_title('pool1_1')
p1_2.imshow(pool1_2, cmap='gray')
p1_2.set_title('pool1_2')
p1_3.imshow(pool1_3, cmap='gray')
p1_3.set_title('pool1_3')
p1_1.set_xticks([])
p1_2.set_xticks([])
p1_3.set_xticks([])
p1_1.set_yticks([])
p1_2.set_yticks([])
p1_3.set_yticks([])

p2_1.imshow(pool2_1, cmap='gray')
p2_1.set_title('pool2_1')
p2_2.imshow(pool2_2, cmap='gray')
p2_2.set_title('pool2_2')
p2_3.imshow(pool2_3, cmap='gray')
p2_3.set_title('pool2_3')
p2_1.set_xticks([])
p2_2.set_xticks([])
p2_3.set_xticks([])
p2_1.set_yticks([])
p2_2.set_yticks([])
p2_3.set_yticks([])

h1.hist(np.ravel(np.abs(pool1_1-pool2_1)),bins=100)
h1.set_title('Pool 1 diff')
h2.hist(np.ravel(np.abs(pool1_2-pool2_2)),bins=100)
h2.set_title('Pool 2 diff')
h3.hist(np.ravel(np.abs(pool1_3-pool2_3)),bins=100)
h3.set_title('Pool 3 diff')

plt.tight_layout()
# plt.show()
plt.savefig('images/pool_invariance.png')

部落格筆記

該專案的作者還在自己的個人網站上釋出了花書的精煉筆記，地址為：

https://medium.com/inveterate-learner/tagged/deep-learning

附加資源

除了這份花書重點章節摘要之外，石頭君還推薦一份來自 Microsoft 計算機軟體工程師 Jeff Macaluso 總結的關於花書的經驗法則！

線上版閱讀地址：

https://jeffmacaluso.github.io/post/DeepLearningRulesOfThumb/

離線地址：

連結：https://pan.baidu.com/s/1eLlJy3xB6Hs0w_Q7bO536g

提取碼：7q1d

希望這份資源對你有所幫助~

更多 AI 乾貨，請關注公眾號：AI有道！