|
import numpy as np
from os import makedirs, path, listdir
from shutil import copyfile
np.random.seed(42)
if not path.isdir('dogs-vs-cats/train/dog'):
testRatio = 0.25
subdirs = ['train/', 'test/']
for subdir in subdirs:
labeldirs = ['dogs/', 'cats/']
for labldir in labeldirs:
newdir = 'dogs-vs-cats/' + subdir + labldir
makedirs(newdir, exist_ok=True)
for file in listdir('dogs-vs-cats/train'):
src = 'dogs-vs-cats/train/' + file
if path.isfile(src):
if np.random.rand() < testRatio: dst_dir = 'test/'
else: dst_dir = 'train/'
if file.startswith('cat'):
dst = 'dogs-vs-cats/' + dst_dir + 'cats/' + file
copyfile(src, dst)
elif file.startswith('dog'):
dst = 'dogs-vs-cats/' + dst_dir + 'dogs/' + file
copyfile(src, dst)
from tensorflow.keras.models import Sequential, load_model
from tensorflow.keras.layers import Dense, Conv2D, MaxPool2D, Flatten, BatchNormalization
from tensorflow.keras.preprocessing.image import ImageDataGenerator
reducePicDim = 256
try:
CNN = load_model("dogVScat.h5")
except:
trainDatagen = ImageDataGenerator(rotation_range=30, rescale=1./255, horizontal_flip=0.1)
trainGenerator = trainDatagen.flow_from_directory(
directory=r"./dogs-vs-cats/train/",
target_size=(reducePicDim, reducePicDim),
color_mode="rgb", batch_size=64,
class_mode="categorical", shuffle=True, seed=42)
CNN = Sequential()
CNN.add(Conv2D(32,(5,5),activation='relu',input_shape=(reducePicDim,reducePicDim,3)))
CNN.add(MaxPool2D(pool_size=(3, 3)))
CNN.add(BatchNormalization())
CNN.add(Conv2D(32,(5,5),activation='relu'))
CNN.add(MaxPool2D(pool_size=(3, 3)))
CNN.add(BatchNormalization())
CNN.add(Conv2D(64,(3,3),activation='relu'))
CNN.add(MaxPool2D(pool_size=(2, 2)))
CNN.add(BatchNormalization())
CNN.add(Conv2D(64,(3,3),activation='relu'))
CNN.add(Flatten())
CNN.add(Dense(100,activation='relu'))
CNN.add(Dense(50,activation='relu'))
CNN.add(Dense(2,activation='softmax'))
CNN.summary()
CNN.compile(optimizer='adam',loss='categorical_crossentropy',metrics=['accuracy'])
CNN.fit_generator(generator=trainGenerator, epochs=10, verbose=True)
CNN.save("dogVScat.h5")
testDatagen = ImageDataGenerator(rescale=1./255)
testGenerator = testDatagen.flow_from_directory(
directory=r"./dogs-vs-cats/test/",
target_size=(reducePicDim, reducePicDim),
color_mode="rgb", batch_size=8,
class_mode="categorical", shuffle=False)
myloss, acc = CNN.evaluate_generator(testGenerator, steps=len(testGenerator), verbose=True)
print('Acc: %.3f' % (acc * 100.0))
testGenerator.reset()
yP = CNN.predict_generator(testGenerator, steps=len(testGenerator), verbose=True)
yPClass = np.argmax(yP,axis=1)
cats = np.sum(testGenerator.classes == 0)
dogs = np.sum(testGenerator.classes == 1)
catsAsDogs = np.sum( np.abs(yPClass[testGenerator.classes == 0] -0) )
dogsAsCats = np.sum( np.abs(yPClass[testGenerator.classes == 1] -1) )
confMatrix = np.array([[(cats-catsAsDogs)/cats, catsAsDogs/cats],
[dogsAsCats/dogs, (dogs-dogsAsCats)/dogs]])
print(confMatrix)
import tensorflow as tf
from tensorflow.keras.models import Model
def heatmap(img, model):
for l in model.layers: #*\label{code:catsdogs:1}
if isinstance(l, Conv2D): lastConvLayer = l #*\label{code:catsdogs:2}
calcFeaturesAndPred = Model([model.input], [lastConvLayer.output, model.output]) #*\label{code:catsdogs:3}
img = tf.Variable(img); #*\label{code:catsdogs:7}
with tf.GradientTape() as tape:
featureMaps, predictions = calcFeaturesAndPred(img[np.newaxis,...]) #*\label{code:catsdogs:4}
maxActivation = np.argmax(predictions[0]) #*\label{code:catsdogs:5}
predictedClassEntry = predictions[:, maxActivation] #*\label{code:catsdogs:6}
grads = tape.gradient(predictedClassEntry, featureMaps) #*\label{code:catsdogs:8}
pooledGrads = np.mean(grads, axis=(0, 1, 2)) #*\label{code:catsdogs:9}
weightedFeatures = featureMaps * pooledGrads #*\label{code:catsdogs:10}
heatmapImg = np.sum(weightedFeatures, axis=-1).squeeze() #*\label{code:catsdogs:17}
heatmapImg = np.maximum(heatmapImg, 0)
if np.max(heatmapImg) > 0 : heatmapImg /= np.max(heatmapImg)
return heatmapImg, predictions.numpy()
from matplotlib import cm
from PIL import Image as PILImage
from tensorflow.keras.preprocessing import image
def overlayHeatmap(img, heatmapImg):
heatmapImg = cm.Reds(heatmapImg)[..., :3] #*\label{code:catsdogs:11}
heatmapImg = image.array_to_img(heatmapImg) #*\label{code:catsdogs:12}
heatmapImg = heatmapImg.resize(img.shape[:-1], resample=PILImage.BICUBIC) #*\label{code:catsdogs:13}
heatmapImg = image.img_to_array(heatmapImg) #*\label{code:catsdogs:14}
imGray = 0.2989*img[:,:,0] + 0.5870*img[:,:,1] + 0.1140*img[:,:,2] #*\label{code:catsdogs:15}
imGray = cm.gray(imGray)[..., :3] #*\label{code:catsdogs:16}
if np.max(imGray) <= 1: imGray = 255*imGray
if np.max(heatmapImg) <= 1 : heatmapImg = 255*heatmapImg
superimposedImg = np.minimum(heatmapImg * 0.6 + 0.2*imGray, 255).astype(np.uint8)
return superimposedImg
import matplotlib.pyplot as plt
plt.rcParams.update({'figure.max_open_warning': 0})
imageList = ['hund1.jpg', 'hund2.jpg','hund3.jpg', 'hund4.jpg',
'katze1.jpg', 'katze2.jpg', 'katze3.jpg', 'katze4.jpg',
'ratte.jpg', 'luchs.jpg', 'wolf.jpg', 'katze1freigestellt.jpg']
for imageFile in imageList:
imgSize = CNN.input_shape[1:-1]
img = image.load_img(imageFile, target_size=imgSize)
img = image.img_to_array(img)/255.0
plt.figure(); plt.imshow(img)
hm, predictions = heatmap(img, CNN)
plt.figure(); plt.imshow(hm, cmap=cm.Reds)
if np.argmax(predictions[0]) == 0: classstring = 'cat'
else: classstring = 'dog'
print(imageFile,' predicted as ',classstring, ' with ', predictions)
fusion = overlayHeatmap(img, hm)
plt.figure(); plt.imshow(fusion)
name = imageFile.split('.')[0]
plt.title(name+str(predictions))
name = 'heat'+name+'.png'
image.array_to_img(fusion).save(name, 'PNG')
¤ Dauer der Verarbeitung: 0.1 Sekunden
(vorverarbeitet)
¤
|
Haftungshinweis
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung ist noch experimentell.
|
