Faces autoencoder with PCA
This is an experiment that will be used to create real time face generator using autoencoders and principal component analysis.
Final solution will be a website with sliders that user can move around to generate faces in real time in the browser using TensorflowJS.
This work is inspired by carykh. This version is rewritten in jupyter notebook using Tensorflow’s Keras for model training as it is easier to export keras model to TensorflowJS format.
The purpose of this experiment is to create keras model of autoencoder that will be converted to TensorflowJS version which can be run in the browser.
Github url: https://github.com/sikora507/faces-autoencoder
To run this experiment, you need to have docker and docker-compose installed.
Then you can run it by typing docker-compose up in the terminal.
import os
import tensorflow as tf
import numpy as np
from scipy import misc
import random
import math
import face_recognition
from skimage import transform
import os.path
import urllib.request
import time
import requests
import matplotlib.pyplot as plt
import json
from sklearn.decomposition import PCA
from tensorflow.keras.layers import Input, Dense, Conv2D, MaxPooling2D, UpSampling2D, Reshape, Flatten
from tensorflow.keras.models import Model, Sequential
from tensorflow.keras import backend as K
%matplotlib inline
Set global variables
IMAGE_WIDTH = 64
IMAGE_HEIGHT = 64
DENSE_SIZE = 200
Set up variables and download helpers
START_MONTH = 11
START_DATE = 17
START_SLOT = 6
DESIRED_X = 32
DESIRED_Y = 21
DESIRED_SIZE = 24
NAMES_PER_FILE = 100
TEMP_FILENAME = "temp.png"
monthOn = 0
dayOn = 0
daysInMonth = [31,29,31,30,31,30,31,31,30,31,30,31]
monthNames = ['january','february','march','april','may','june','july','august','september','october','november','december']
def getAvg(face, landmark):
cum = np.zeros((2))
for point in face[landmark]:
cum[0] += point[0]
cum[1] += point[1]
return cum/len(face[landmark])
def getNorm(a):
return (a-np.mean(a))/np.std(a)
Download all the data
imageCounter = 0
for monthOn in range(START_MONTH,12):
thisStartDate = 1
if monthOn == START_MONTH:
thisStartDate = START_DATE
for dayOn in range(thisStartDate,daysInMonth[monthOn]+1):
response = urllib.request.urlopen("https://www.famousbirthdays.com/"+monthNames[monthOn]+str(dayOn)+".html")
pageSource = response.read().splitlines()
lineOn = 0
while str(pageSource[lineOn]) != "b'<div class=\"container people-list\">'":
lineOn += 1
thisStartSlot = 0
if monthOn == START_MONTH and dayOn == START_DATE:
thisStartSlot = START_SLOT
for slotOn in range(0,48):
while "class=\"face person-item\"" not in str(pageSource[lineOn]):
lineOn += 1
iul = str(pageSource[lineOn])
pnl = str(pageSource[lineOn+4])
imageURL = iul[iul.index("background: url(")+16:iul.index(") no-repeat center center")]
personName = ""
age = ""
pnl_s = 2
if pnl[pnl_s] == ' ':
pnl_s += 1
if "," in pnl:
personName = pnl[pnl_s:pnl.index(",")]
age = pnl[pnl.index(",")+2:-1]
else:
personName = pnl[pnl_s:pnl.index("(")-1]
age = pnl[pnl.index("(")+1:pnl.index(")")]
if slotOn < thisStartSlot or imageURL == 'https://www.famousbirthdays.com/faces/large-default.jpg' or personName == "Ronan Domingo" or personName == "Glam And Gore" or personName == "Edith Piaf" or personName == "Lexi Marie":
print(personName+" SKIPPED!")
else:
img_data = requests.get(imageURL).content
with open("temp.png", 'wb') as handler:
handler.write(img_data)
image_face_info = face_recognition.load_image_file(TEMP_FILENAME)
face_landmarks = face_recognition.face_landmarks(image_face_info)
image_numpy = misc.imread(TEMP_FILENAME)
colorAmount = 0
if len(image_numpy.shape) == 3:
nR = getNorm(image_numpy[:,:,0])
nG = getNorm(image_numpy[:,:,1])
nB = getNorm(image_numpy[:,:,2])
colorAmount = np.mean(np.square(nR-nG))+np.mean(np.square(nR-nB))+np.mean(np.square(nG-nB))
if len(face_landmarks) == 1 and colorAmount >= 0.04: # We need there to only be one face in the image, AND we need it to be a colored image.
leftEyePosition = getAvg(face_landmarks[0],'left_eye')
rightEyePosition = getAvg(face_landmarks[0],'right_eye')
nosePosition = getAvg(face_landmarks[0],'nose_tip')
mouthPosition = getAvg(face_landmarks[0],'bottom_lip')
centralPosition = (leftEyePosition+rightEyePosition)/2
faceWidth = np.linalg.norm(leftEyePosition-rightEyePosition)
faceHeight = np.linalg.norm(centralPosition-mouthPosition)
if faceWidth >= faceHeight*0.7 and faceWidth <= faceHeight*1.5:
faceSize = (faceWidth+faceHeight)/2
toScaleFactor = faceSize/DESIRED_SIZE
toXShift = (centralPosition[0])
toYShift = (centralPosition[1])
toRotateFactor = np.arctan2(rightEyePosition[1]-leftEyePosition[1],rightEyePosition[0]-leftEyePosition[0])
rotateT = transform.SimilarityTransform(scale=toScaleFactor,rotation=toRotateFactor,translation=(toXShift,toYShift))
moveT = transform.SimilarityTransform(scale=1,rotation=0,translation=(-DESIRED_X,-DESIRED_Y))
outputArr = transform.warp(image=image_numpy,inverse_map=(moveT+rotateT))[0:IMAGE_HEIGHT,0:IMAGE_WIDTH]
misc.imsave("../data/dataFace"+str(imageCounter)+".png",outputArr)
if imageCounter%NAMES_PER_FILE == 0:
f = open("../names/name"+str(imageCounter)+".txt","w+")
fame = str(slotOn)
if monthOn == 1 and dayOn == 29:
fame *= 4
f.write(personName+","+fame+","+age+"\n")
if imageCounter%NAMES_PER_FILE == (NAMES_PER_FILE-1):
f.close()
print("DAY "+monthNames[monthOn]+" "+str(dayOn)+": I just used person "+personName+" to create image number "+str(imageCounter))
imageCounter += 1
time.sleep(0.5)
lineOn += 1
Store images count
images_count = len(os.listdir('../data/'))
Setup helper functions
def CreateAutoEncoder():
input_img = Input(shape=(IMAGE_HEIGHT, IMAGE_WIDTH, 3)) # adapt this if using `channels_first` image data format
x = Conv2D(40, (3, 3), activation='relu', padding='same')(input_img)
x = MaxPooling2D((2, 2), padding='same')(x)
x = Conv2D(80, (3, 3), activation='relu', padding='same')(x)
x = MaxPooling2D((2, 2), padding='same')(x)
x = Conv2D(150, (3, 3), activation='relu', padding='same')(x)
x = MaxPooling2D((2, 2), padding='same')(x)
x = Conv2D(200, (3, 3), activation='relu', padding='same')(x)
x = MaxPooling2D((2, 2), padding='same')(x)
x = Flatten()(x)
encoded = Dense(DENSE_SIZE, activation='relu')(x)
y = Dense(3200, activation='relu')(encoded)
y = Reshape([4,4,200])(y)
y = UpSampling2D((2, 2))(y)
y = Conv2D(150, (3, 3), activation='relu', padding='same')(y)
y = UpSampling2D((2, 2))(y)
y = Conv2D(80, (3, 3), activation='relu', padding='same')(y)
y = UpSampling2D((2, 2))(y)
y = Conv2D(40, (3, 3), activation='relu', padding='same')(y)
y = UpSampling2D((2, 2))(y)
y = Conv2D(15, (3, 3), activation='relu', padding='same')(y)
decoded = Conv2D(3, (3, 3), activation='sigmoid', padding='same')(y)
autoencoder = Model(input_img, decoded)
#optimizer = tf.keras.optimizers.Adam(lr=0.005)
autoencoder.compile(optimizer='adagrad', loss='binary_crossentropy')
return autoencoder
def WriteOutputFileResult(epoch):
if epoch%10!=0:
return
imageIndex = np.random.randint(images_count)
output = autoencoder.predict(batch_data[imageIndex].reshape(1,64,64,3))
exampleImage = np.empty([IMAGE_HEIGHT*2,IMAGE_WIDTH,3])
exampleImage[0:IMAGE_HEIGHT,0:IMAGE_WIDTH,0:3] = batch_data[imageIndex,0:IMAGE_HEIGHT,0:IMAGE_WIDTH,0:3]
exampleImage[IMAGE_HEIGHT:IMAGE_HEIGHT*2,0:IMAGE_WIDTH,0:3] = output[0,0:IMAGE_HEIGHT,0:IMAGE_WIDTH,0:3]
exampleImage = np.clip(exampleImage, 0, 1)
misc.imsave('../modelExamples/encoder'+str(epoch)+'.png',exampleImage)
save_path='../models/keras.h5'
autoencoder.save(save_path)
print("MODEL SAVED: "+str(save_path))
Train autoencoder
from tensorflow.keras.models import load_model
modelPath = '../models/keras.h5'
if os.path.isfile(modelPath):
print('Loading existing model')
autoencoder = load_model('../models/keras.h5', custom_objects={"GlorotUniform": tf.keras.initializers.glorot_uniform})
else:
print('Creating new model')
autoencoder = CreateAutoEncoder()
batch_data = np.empty([images_count,IMAGE_HEIGHT,IMAGE_WIDTH,3])
for example in range(images_count):
imageIndex = example# + int(math.floor(random.randrange(images_count)))
imagio = misc.imread('../data/dataFace'+str(imageIndex)+'.png')
batch_data[example] = imagio[:,:,0:3]/255.0
autoencoder.fit(batch_data, batch_data,
epochs=10000,
batch_size=100,
shuffle=True,
#validation_data=(batch_data, batch_data),
callbacks=[
#tf.keras.callbacks.ModelCheckpoint('../models/keras.h5', monitor='val_loss', verbose=0, save_best_only=True, save_weights_only=False, mode='auto', period=1),
tf.keras.callbacks.LambdaCallback(on_epoch_end=lambda epoch, logs: WriteOutputFileResult(epoch))
])
Tear apart encoder and decoder
[layer.name for layer in autoencoder.layers]
['input_3',
'conv2d_18',
'max_pooling2d_8',
'conv2d_19',
'max_pooling2d_9',
'conv2d_20',
'max_pooling2d_10',
'conv2d_21',
'max_pooling2d_11',
'flatten_2',
'dense_4',
'dense_5',
'reshape_2',
'up_sampling2d_8',
'conv2d_22',
'up_sampling2d_9',
'conv2d_23',
'up_sampling2d_10',
'conv2d_24',
'up_sampling2d_11',
'conv2d_25',
'conv2d_26']
encoder = Sequential()
for i in range(0,11):
encoder.add(autoencoder.layers[i])
encoder.build()
decoder = Sequential()
for i in range(11,len(autoencoder.layers)):
decoder.add(autoencoder.layers[i])
decoder.build(input_shape=(None, 200))
def getEncodedImage(index):
image = misc.imread('../data/dataFace'+str(index)+'.png')
image = image[:,:,0:3]/255.0
return image
Preview random image to be encoded
randIndex = np.random.randint(0, images_count)
image = getEncodedImage(randIndex)
plt.imshow(image)
/usr/local/lib/python3.5/dist-packages/ipykernel_launcher.py:2: DeprecationWarning: `imread` is deprecated!
`imread` is deprecated in SciPy 1.0.0, and will be removed in 1.2.0.
Use ``imageio.imread`` instead.
<matplotlib.image.AxesImage at 0x7faf480ad128>
encoded = encoder.predict(image.reshape([-1, image.shape[0], image.shape[1], image.shape[2]]))
print('Encoded image:\n' ,encoded[0])
Encoded image:
[0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 1.6740656 0.
0. 1.2446964 0. 0. 4.6464124 2.7161005
4.744044 0. 0. 0. 1.0253751 0.
0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0.
0.516001 0. 0. 0. 0. 0.
0. 0. 0. 0. 2.0416033 0.
1.2738434 0. 0. 0. 4.646125 0.
2.3047757 0. 0. 0.6224123 4.424849 0.
0. 1.1240004 0. 0. 0. 1.548488
0. 0.4614426 0. 0. 0. 0.
0. 1.1851753 1.2601664 0. 0. 0.
0. 0. 0. 0. 0. 0.
0.5733514 0. 0. 0. 0. 0.18939778
1.6195832 4.0283566 0. 0. 0. 1.4311539
0. 1.9477886 1.0686957 0. 0. 1.4714222
0. 0. 0. 0. 0. 0.
0. 1.5717533 0. 0. 0. 1.5706655
0. 0. 4.989835 0. 0. 0.
0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0.
0.87802136 0. 0. 0. 0. 0.
0. 0. 0. 1.3519242 0. 0.
2.0465598 3.8467922 0. 0. 0.17653233 0.
2.6257434 1.0397722 0. 0. 1.2316633 0.
0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 1.5195677 0.
0. 0. 0. 1.062976 0. 0.99829024
2.1957307 0. 3.938364 0. 0.5726944 0.
0. 0. 0. 0.5094159 0. 0.
0. 0. 0. 0. 4.072344 0.
0. 0. 0. 0. 3.8419755 0.
0. 0. ]
decoded = decoder.predict(encoded)
Recreate image from encoded state
plt.imshow(decoded[0])
<matplotlib.image.AxesImage at 0x7faefc7a5cc0>
Create dataframe to inspect all encoded images together
import pandas as pd
all_rows = []
for i in range(images_count):
image = getEncodedImage(i)
encoded_internal = encoder.predict(image.reshape([-1, image.shape[0], image.shape[1], image.shape[2]]))[0]
all_rows.append(encoded_internal)
df = pd.DataFrame(all_rows)
df.describe()
/usr/local/lib/python3.5/dist-packages/ipykernel_launcher.py:2: DeprecationWarning: `imread` is deprecated!
`imread` is deprecated in SciPy 1.0.0, and will be removed in 1.2.0.
Use ``imageio.imread`` instead.
| 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | ... | 190 | 191 | 192 | 193 | 194 | 195 | 196 | 197 | 198 | 199 | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| count | 527.0 | 527.000000 | 527.0 | 527.000000 | 527.0 | 527.0 | 527.0 | 527.000000 | 527.0 | 527.0 | ... | 527.000000 | 527.0 | 527.0 | 527.000000 | 527.000000 | 527.0 | 527.000000 | 527.0 | 527.000000 | 527.000000 |
| mean | 0.0 | 0.383350 | 0.0 | 0.497579 | 0.0 | 0.0 | 0.0 | 1.563971 | 0.0 | 0.0 | ... | 0.588612 | 0.0 | 0.0 | 0.292458 | 0.928342 | 0.0 | 0.658084 | 0.0 | 0.906207 | 0.520338 |
| std | 0.0 | 1.190903 | 0.0 | 1.101257 | 0.0 | 0.0 | 0.0 | 1.455407 | 0.0 | 0.0 | ... | 1.390174 | 0.0 | 0.0 | 1.263822 | 1.070300 | 0.0 | 1.130944 | 0.0 | 1.259891 | 1.474352 |
| min | 0.0 | 0.000000 | 0.0 | 0.000000 | 0.0 | 0.0 | 0.0 | 0.000000 | 0.0 | 0.0 | ... | 0.000000 | 0.0 | 0.0 | 0.000000 | 0.000000 | 0.0 | 0.000000 | 0.0 | 0.000000 | 0.000000 |
| 25% | 0.0 | 0.000000 | 0.0 | 0.000000 | 0.0 | 0.0 | 0.0 | 0.294710 | 0.0 | 0.0 | ... | 0.000000 | 0.0 | 0.0 | 0.000000 | 0.000000 | 0.0 | 0.000000 | 0.0 | 0.000000 | 0.000000 |
| 50% | 0.0 | 0.000000 | 0.0 | 0.000000 | 0.0 | 0.0 | 0.0 | 1.333912 | 0.0 | 0.0 | ... | 0.000000 | 0.0 | 0.0 | 0.000000 | 0.655429 | 0.0 | 0.000000 | 0.0 | 0.324040 | 0.000000 |
| 75% | 0.0 | 0.000000 | 0.0 | 0.328242 | 0.0 | 0.0 | 0.0 | 2.321734 | 0.0 | 0.0 | ... | 0.510635 | 0.0 | 0.0 | 0.000000 | 1.580515 | 0.0 | 0.958642 | 0.0 | 1.432640 | 0.000000 |
| max | 0.0 | 7.928043 | 0.0 | 8.070414 | 0.0 | 0.0 | 0.0 | 8.701365 | 0.0 | 0.0 | ... | 12.323253 | 0.0 | 0.0 | 14.501927 | 5.239333 | 0.0 | 6.052303 | 0.0 | 6.644118 | 9.616698 |
8 rows × 200 columns
Save mean from encodings to file
mean = df.mean().values.tolist()
text_file = open("../models/mean.json", "w")
text_file.write(json.dumps(mean))
text_file.close()
Filter out columns all containing zeros
m2 = df.loc[:, (df != 0).any(axis=0)]
m2.describe()
| 1 | 3 | 7 | 10 | 13 | 16 | 17 | 18 | 19 | 21 | ... | 179 | 180 | 183 | 184 | 190 | 193 | 194 | 196 | 198 | 199 | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| count | 527.000000 | 527.000000 | 527.000000 | 527.000000 | 527.000000 | 527.000000 | 527.000000 | 527.000000 | 527.000000 | 527.000000 | ... | 527.000000 | 527.000000 | 527.000000 | 527.000000 | 527.000000 | 527.000000 | 527.000000 | 527.000000 | 527.000000 | 527.000000 |
| mean | 0.383350 | 0.497579 | 1.563971 | 1.032203 | 0.645219 | 1.415085 | 0.818090 | 0.795896 | 1.025534 | 0.855816 | ... | 0.830482 | 0.720698 | 0.921920 | 1.092613 | 0.588612 | 0.292458 | 0.928342 | 0.658084 | 0.906207 | 0.520338 |
| std | 1.190903 | 1.101257 | 1.455407 | 1.125508 | 1.383625 | 1.554425 | 1.545218 | 1.385235 | 1.114971 | 1.103555 | ... | 1.060839 | 1.136248 | 1.201186 | 1.640923 | 1.390174 | 1.263822 | 1.070300 | 1.130944 | 1.259891 | 1.474352 |
| min | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | ... | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 |
| 25% | 0.000000 | 0.000000 | 0.294710 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | ... | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 |
| 50% | 0.000000 | 0.000000 | 1.333912 | 0.756519 | 0.000000 | 1.057168 | 0.000000 | 0.000000 | 0.758654 | 0.457324 | ... | 0.382845 | 0.000000 | 0.473265 | 0.104616 | 0.000000 | 0.000000 | 0.655429 | 0.000000 | 0.324040 | 0.000000 |
| 75% | 0.000000 | 0.328242 | 2.321734 | 1.556919 | 0.611775 | 2.378852 | 1.168464 | 1.147271 | 1.744949 | 1.374141 | ... | 1.381343 | 1.199598 | 1.443289 | 1.651781 | 0.510635 | 0.000000 | 1.580515 | 0.958642 | 1.432640 | 0.000000 |
| max | 7.928043 | 8.070414 | 8.701365 | 5.879054 | 10.312868 | 6.995349 | 9.534487 | 8.131895 | 6.139172 | 6.481678 | ... | 5.561953 | 6.141798 | 6.244396 | 9.122747 | 12.323253 | 14.501927 | 5.239333 | 6.052303 | 6.644118 | 9.616698 |
8 rows × 107 columns
pca_columns = len(m2.columns)
This number shows how many principal components are needed
We can take only non-zero columns from our encodings table
This will allow to use fewer components
pca_columns
107
Manual PCA
np_all_rows = np.asarray(all_rows)
pca = PCA(n_components=pca_columns)
pca.fit(np_all_rows)
values = np.sqrt(pca.explained_variance_)
vectors = pca.components_
json_vectors = json.dumps(vectors.tolist())
json_values = json.dumps(values.tolist())
all_rows_pca = pca.transform(np_all_rows)
sliders_min = all_rows_pca.min(axis=0)
sliders_max = all_rows_pca.max(axis=0)
json_sliders_min = json.dumps(sliders_min.tolist())
json_sliders_max = json.dumps(sliders_max.tolist())
text_file = open("../models/eigenVectors.json", "w")
text_file.write(json_vectors)
text_file.close()
text_file = open("../models/eigenValues.json", "w")
text_file.write(json_values)
text_file.close()
text_file = open("../models/slidersMin.json", "w")
text_file.write(json_sliders_min)
text_file.close()
text_file = open("../models/slidersMax.json", "w")
text_file.write(json_sliders_max)
text_file.close()
Load data as it will be loaded by tensorflowJS in the browser
with open('../models/eigenVectors.json') as f:
eigenVectors = np.asarray(json.load(f))
with open('../models/eigenValues.json') as f:
eigenValues = np.asarray(json.load(f))
with open('../models/mean.json') as f:
meanData = np.asarray(json.load(f))
eigenvectorInverses = np.linalg.pinv(eigenVectors)
traits = encoded[0] - meanData
image_pca = np.matmul(traits, eigenvectorInverses) / eigenValues
image_pca
array([-1.21494064e+00, -4.95205738e-01, 1.58773161e+00, 1.03317212e+00,
1.64704590e-02, 8.42720705e-02, 6.85427567e-02, -2.79387655e+00,
1.87637418e+00, 7.37616443e-02, 2.54778020e-01, -1.77222121e-01,
1.13375160e+00, -1.30607305e-01, -5.21181651e-01, 1.04267917e+00,
-5.30473454e-01, -9.19137177e-01, -6.33939878e-01, 1.90321028e+00,
1.06104141e+00, -7.84540597e-01, 6.10745502e-01, 7.75544652e-01,
-3.78812174e-01, -8.48271968e-01, -9.15274639e-01, -5.84614425e-02,
4.95471393e-01, -8.18511609e-01, 1.17733230e-01, 7.96533237e-01,
3.81170546e-01, -7.34867326e-01, -1.80153036e-01, 2.69993923e-01,
-4.40699800e-01, 4.58954842e-01, 1.20221894e+00, 1.05707949e+00,
-9.26709735e-01, -2.62151517e-01, 4.12873764e-01, 9.51895180e-01,
-1.82035008e-01, 2.05909696e-01, -3.62205891e-02, 1.64909150e-01,
3.50870550e-02, -6.51322007e-01, -1.10071462e+00, 1.10239015e-01,
1.00940874e+00, -3.22872511e-01, -6.86396836e-01, -3.72447165e-01,
1.11109849e+00, 6.57288975e-01, -1.00576561e-01, 1.89648021e+00,
3.43962315e-01, 2.34161901e-01, 3.68600332e-01, 8.69722112e-01,
8.62145163e-01, 9.93404808e-01, -1.97469710e-01, 7.98476353e-01,
-1.10964469e+00, -4.45596697e-01, -5.13701822e-01, 7.46462772e-01,
-1.43335627e+00, -5.69662544e-01, -5.31171657e-01, -6.39782067e-01,
-2.70193727e-01, -9.27538014e-01, 6.65862011e-02, -3.60068660e-01,
1.08171159e+00, -2.99478694e-01, 8.84023699e-01, 4.59543898e-01,
9.50298384e-01, -7.86344394e-02, -1.28323107e+00, 1.15862300e+00,
7.38062574e-01, -8.21171997e-01, 1.04588902e+00, -1.54296072e+00,
9.66833325e-01, -4.98282965e-01, 8.00057643e-01, -3.95012717e-01,
9.15517991e-01, -1.49782181e+00, 5.22660921e-01, 5.27705578e-01,
1.26221635e-03, 1.14477946e+00, 1.60914136e+00, 9.85090220e-01,
-7.20066725e-01, -7.78477121e-01, -6.47627714e-01])
recovered_from_pca = meanData.copy()
for i in range(pca_columns):
recovered_from_pca += image_pca[i] * eigenValues[i] * eigenVectors[i]
recovered_from_pca
array([-8.30572673e-07, -8.26042793e-15, 7.28689154e-07, -1.38158148e-13,
-4.69476584e-07, -1.24504785e-07, -4.88584873e-07, 1.88553600e-13,
1.75781359e-07, 8.27562621e-07, 1.67406559e+00, 2.58648698e-07,
-1.61953981e-07, 1.24469638e+00, -1.84599036e-07, 3.55158112e-07,
4.64641237e+00, 2.71610045e+00, 4.74404383e+00, 1.98366497e-13,
-1.74709958e-07, -7.35730921e-14, 1.02537513e+00, 5.83238056e-07,
-1.10047388e-13, 3.03989825e-07, -1.23132841e-13, -3.87376862e-07,
1.58832694e-07, -2.72547013e-07, 3.58801530e-14, 1.76143822e-14,
-4.30501988e-14, -2.59269578e-07, -2.51543659e-07, 9.44383460e-14,
5.16000986e-01, 2.46116744e-07, 3.67596578e-14, 6.68791436e-07,
-3.52233793e-08, 4.35242120e-14, -3.67191394e-08, 1.43867990e-13,
2.98658325e-07, 3.94759415e-08, 2.04160333e+00, -1.50137861e-07,
1.27384341e+00, -1.00947029e-13, -1.36277723e-07, -1.51687690e-13,
4.64612484e+00, -1.96762745e-13, 2.30477571e+00, -3.33317077e-08,
-2.60209557e-07, 6.22412324e-01, 4.42484903e+00, -3.28825352e-08,
-3.63440664e-08, 1.12400043e+00, -6.86406919e-08, -2.03170814e-14,
1.06426262e-07, 1.54848802e+00, 1.47350922e-07, 4.61442590e-01,
1.71426241e-13, -8.23370235e-14, -1.34194914e-07, 6.22532487e-08,
-1.05264673e-07, 1.18517530e+00, 1.26016641e+00, 3.41220108e-14,
5.01983140e-08, -1.56444773e-07, 8.23085090e-14, 1.12135995e-13,
-2.42597609e-13, 1.87335794e-08, 2.74907211e-08, -1.98532163e-13,
5.73351383e-01, -2.23147896e-08, 1.45389737e-07, 1.52400662e-13,
2.31082514e-13, 1.89397782e-01, 1.61958325e+00, 4.02835655e+00,
-3.84828551e-08, -2.53931234e-08, -5.17085792e-08, 1.43115389e+00,
-7.45952721e-08, 1.94778860e+00, 1.06869566e+00, -9.69328784e-14,
3.07462389e-14, 1.47142220e+00, 1.67131066e-13, 1.36094738e-07,
1.29731295e-13, -6.22669381e-09, -3.90126941e-08, 3.06171007e-09,
1.82071362e-08, 1.57175326e+00, 2.47293505e-13, -2.96403527e-14,
0.00000000e+00, 1.57066548e+00, 0.00000000e+00, 0.00000000e+00,
4.98983479e+00, 0.00000000e+00, 4.21225554e-14, 9.77273817e-14,
1.91644877e-13, 0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00, 1.33545952e-13, 0.00000000e+00,
0.00000000e+00, -1.43735718e-13, -2.28697269e-14, 0.00000000e+00,
8.78021359e-01, 0.00000000e+00, 0.00000000e+00, -7.04210995e-15,
0.00000000e+00, 0.00000000e+00, -5.41233725e-15, -1.47889947e-13,
0.00000000e+00, 1.35192418e+00, 0.00000000e+00, -3.06933298e-14,
2.04655981e+00, 3.84679222e+00, 0.00000000e+00, 0.00000000e+00,
1.76532328e-01, 0.00000000e+00, 2.62574339e+00, 1.03977215e+00,
2.33606537e-13, -4.60990024e-14, 1.23166335e+00, 0.00000000e+00,
0.00000000e+00, -1.22700461e-13, 0.00000000e+00, 7.23830718e-14,
0.00000000e+00, 1.28057287e-14, 6.03857242e-14, 0.00000000e+00,
-3.15511506e-14, 0.00000000e+00, 1.51956773e+00, 2.93515212e-14,
0.00000000e+00, 1.56125113e-14, -2.00616569e-14, 1.06297600e+00,
0.00000000e+00, 9.98290241e-01, 2.19573069e+00, 5.76066972e-14,
3.93836403e+00, 0.00000000e+00, 5.72694421e-01, 3.06946742e-13,
2.72847717e-13, 0.00000000e+00, 0.00000000e+00, 5.09415925e-01,
-7.86610360e-15, 0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
0.00000000e+00, 0.00000000e+00, 4.07234383e+00, 0.00000000e+00,
0.00000000e+00, -2.00222976e-13, -5.84567117e-14, 0.00000000e+00,
3.84197545e+00, 0.00000000e+00, 3.00013486e-13, -6.78797296e-15])
decoded = decoder.predict(recovered_from_pca.reshape(1,DENSE_SIZE))
plt.imshow(decoded[0])
<matplotlib.image.AxesImage at 0x7faefc6df1d0>
Save all images as encodings in json file
json_all_rows_pca = json.dumps(all_rows_pca.tolist())
text_file = open("../models/facesPca.json", "w")
text_file.write(json_all_rows_pca)
text_file.close()
This experiment was made after reading Hands-On Machine Learning with Scikit-Learn and TensorFlow. It has principal component analysis well covered.
