def predict_dog(model, predict_img=test_jinmao):
img = Image.open(predict_img)
test = np.asarray(img, dtype="float32")
sample = np.empty((1, DOG_PIC_WIDTH, DOG_PIC_HEIGHT, 3))
sample[0, :, :, :] = test.reshape(DOG_PIC_WIDTH, DOG_PIC_HEIGHT, 3)
predictions = model.predict(sample)
write_log(str(predictions), file=log_file)
def load_model_and_evaluate(model_name):
model = keras.models.load_model(os.path.join(test_6_MODEL, model_name))
test_inputs, test_labels = create_test_samples()
if model:
test_loss, test_acc = model.evaluate(test_inputs, test_labels)
write_log("loss:%s \n acc:%s" % (str(test_loss), str(test_acc)))
return model
def predict_dog(dog_pic_path: str, model_name=model_name):
img = Image.open(os.path.join(test_dir, dog_pic_path))
arr = np.asarray(img, dtype="float32").reshape(585, 460, 3)
data = np.empty((1, 585, 460, 3), dtype="float32")
data[0, :, :, :] = arr
model = keras.models.load_model(os.path.join(test_6_MODEL, model_name))
results = model.predict(data)
write_log(str(results), file="_test_7.log")
def test_model():
dog_dir = [DOG_SUMU_DIR, DOG_JINMAO_DIR]
pic_width = DOG_PIC_WITDH
pic_height = DOG_PIC_HEIGHT
trains, labels = generate_trans_data(dog_dir, pic_width, pic_height)
print(trains.shape, labels.shape)
write_log(str(labels), file="test_8.log")
model = DogsKindModl(pic_width, pic_height)
model.compile_model()
model.fit_model(trains, labels)
model.save_model()
def opencv_CascadeClassify():
"""
使用OpenCV的人脸检测器进行人脸的初步检测，
使用框架训练CNN网络进行人脸的二分类判定，
将两部分合在一起完成人脸检测。
此环节需注意根据应用场景调整参数，
做到性能与召回率的平衡。
args:
returns:
(x,y,w,h) type: tuple
x,y 返回人脸矩形的左上角坐标
w,y 返回矩形的宽和高
"""
for xml in os.listdir(haar_cascade_dir):
xml = os.path.join(haar_cascade_dir, xml)
# write_log(str(xml),file=lg)
face_cascade = cv.CascadeClassifier(xml)
# eye_cascade = cv.CascadeClassifier(harr_cascade_eye_xml)
img = cv.imread(man_test_jpg)
print(type(img))
print(str(img.shape))
gray = cv.cvtColor(img, cv.COLOR_BGR2GRAY)
print(type(gray))
write_log(str(type(img)), file=lg)
write_log(str(img), file=lg)
write_log(str(type(gray)), file=lg)
write_log(str(gray), file=lg)
cv.waitKey(0)
faces = face_cascade.detectMultiScale(
img,
scaleFactor=1.1,
minNeighbors=3,
minSize=(20, 30),
# flags=cv.cv.CV_HAAR_SCALE_IMAGE
)
print("length of faces is %d" % (len(faces)))
if len(faces) == 0:
# write_log("failed this file %s"%(xml),file=lg)
continue
else:
# write_log("success this file %s"%(xml),file=lg)
for x, y, w, h in faces:
cv.rectangle(img, (x, y), (x + w, y + h), (0, 0, 0), 2)
write_log(str(w) + str(h), file=lg)
return faces
# cv.imshow("myself.com",img)
# cv.waitKey(0)
# write_log(str(faces),file=lg)
break
def generate_model(model_name: str):
train_size = len(os.listdir(DOG_SUMU_DIR)) + len(
os.listdir(DOG_JINMAO_DIR))
train_fi = len(os.listdir(DOG_JINMAO_DIR))
# (72,585,460,3)
trains = np.empty((train_size, DOG_PIC_WIDTH, DOG_PIC_HEIGHT, 3),
dtype="float32")
labels = np.empty((train_size), dtype="int32")
for ind in range(len(os.listdir(DOG_JINMAO_DIR))):
img = Image.open(
os.path.join(DOG_JINMAO_DIR,
os.listdir(DOG_JINMAO_DIR)[ind]))
arr = np.asarray(img, dtype="float32")
trains[ind, :, :, :] = arr.reshape(DOG_PIC_WIDTH, DOG_PIC_HEIGHT,
3) # (460,585,3)
labels[ind] = 0
for ind in range(len(os.listdir(DOG_SUMU_DIR))):
img = Image.open(
os.path.join(DOG_SUMU_DIR,
os.listdir(DOG_SUMU_DIR)
[ind])) # tf.image.decode_jpeg(img,channels=3)
arr = np.asarray(img, dtype="float32")
trains[ind + train_fi, :, :, :] = arr.reshape(DOG_PIC_WIDTH,
DOG_PIC_HEIGHT, 3)
labels[ind + train_fi] = 1
# (74,460,585,3)
write_log(str(labels), file=log_file)
model = keras.Sequential([
keras.layers.Flatten(input_shape=(DOG_PIC_WIDTH, DOG_PIC_HEIGHT, 3)),
keras.layers.Dense(2,
activation=tf.nn.relu,
input_shape=(DOG_PIC_WIDTH, DOG_PIC_HEIGHT, 3)),
keras.layers.Dense(2, activation=tf.nn.relu),
keras.layers.Dense(2, activation=tf.nn.softmax),
])
model.compile(
optimizer=tf.train.AdamOptimizer(),
loss="sparse_categorical_crossentropy",
metrics=["accuracy"],
)
# expect (585,460,3)
model.fit(trains, labels, epochs=1, steps_per_epoch=20)
model.save(test_6_MODEL + "/" + model_name)
def predict_apicture(pic_path, model_path, dog_width, dog_height):
"""
ARGS:
pic_path:预测图片路径
model_path:模型存储路径
RETURNS:
result: 分类标签名
"""
img = Image.open(pic_path)
arr = np.asarray(img, dtype="float32")
data = np.empty((1, dog_width, dog_height, 3), dtype="float32")
data[0, :, :, :] = arr.reshape(dog_width, dog_height, 3)
from keras.models import load_model
model = load_model(model_path)
results = model.predict(data)
write_log(str(results), file="test_8.log")
pass
def generate_model_then_predict_model(model_name: str, epochs=2):
train_size = len(os.listdir(DOG_SUMU_DIR)) + len(
os.listdir(DOG_JINMAO_DIR))
train_fi = len(os.listdir(DOG_JINMAO_DIR))
trains = np.empty((train_size, DOG_PIC_WIDTH, DOG_PIC_HEIGHT, 3),
dtype="float32")
labels = np.empty((train_size), dtype="int32")
for ind in range(len(os.listdir(DOG_JINMAO_DIR))):
img = Image.open(
os.path.join(DOG_JINMAO_DIR,
os.listdir(DOG_JINMAO_DIR)[ind]))
arr = np.asarray(img, dtype="float32")
trains[ind, :, :, :] = arr.reshape(DOG_PIC_WIDTH, DOG_PIC_HEIGHT,
3) # (460,585,3)
labels[ind] = 0
for ind in range(len(os.listdir(DOG_SUMU_DIR))):
img = Image.open(
os.path.join(DOG_SUMU_DIR,
os.listdir(DOG_SUMU_DIR)
[ind])) # tf.image.decode_jpeg(img,channels=3)
arr = np.asarray(img, dtype="float32")
trains[ind + train_fi, :, :, :] = arr.reshape(DOG_PIC_WIDTH,
DOG_PIC_HEIGHT, 3)
labels[ind + train_fi] = 1
write_log(str(trains.shape), file=log_file)
model.compile(
optimizer=tf.train.AdamOptimizer(),
loss="sparse_categorical_crossentropy",
metrics=["accuracy"],
)
# model.fit(trains,epochs=epochs,steps_per_epoch=20)
model.fit(trains)
model.save(test_6_MODEL + "/" + model_name)
(2 0)
k(x, y) = x ( ) y.T
(0 1)
"""
M = np.array([[2, 0], [0, 1.0]])
return np.dot(np.dot(x, M), y.T)
h = 0.02 # step size in the mesh
# we create an instance of SVM and fit out data.
clf = svm.SVC(kernel=my_kernel)
result = clf.fit(x, y)
write_log(str(type(clf)), file=log_file)
write_log(str(result), file=log_file)
x_min, x_max = x[:, 0].min() - 1, x[:, 0].max() + 1
y_min, y_max = x[:, 1].min() - 1, x[:, 1].max() + 1
xx, yy = np.meshgrid(np.arange(x_min, x_max, h), np.arange(y_min, y_max, h))
z = clf.predict(np.c_[xx.ravel(), yy.ravel()])
# Put the result into a color plot
z = z.reshape(xx.shape)
plt.pcolormesh(xx, yy, z, cmap=plt.cm.Paired)
# Plot also the training points
plt.scatter(x[:, 0], x[:, 1], c=y, cmap=plt.cm.Paired, edgecolors="k")
对于有颜色的图片，会在三个channel上都计算梯度。 那么相应的幅值就是三个channel上最大的幅值， 角度(方向)是最大幅值所对应的角。 方向是像素强度变化方向 https://blog.csdn.net/passball/article/details/82254256 https://blog.csdn.net/wjb820728252/article/details/78395092 """ image = man_test_jpg im = cv2.imread(image) rgb = np.asarray(im) hog = cv2.HOGDescriptor() h = hog.compute(im) gx = cv2.Sobel(im, cv2.CV_32F, 1, 0, ksize=1) # im 在 x 方向上的梯度 gy = cv2.Sobel(im, cv2.CV_32F, 0, 1, ksize=1) mag, angle = cv2.cartToPolar(gx, gy, angleInDegrees=True) # 梯度值 和方向(角度) write_log(str(mag), file=lg) write_log(str(angle), file=lg) write_log(str(mag.shape) + "size:" + str(mag.size), file=lg) write_log(str(angle.shape) + "size:" + str(angle.size), file=lg) # write_log(str(rgb.shape)) # write_log(str(h.shape)) # write_log(str(rgb.size)) write_log(str(h.size), file=lg)
# linear: （x,x')
# polynomial: is specified by keyword degree, by coef0.
# rbf: . is specified by keyword gamma, must be greater than 0.
# sigmoid (), where is specified by coef0.
# write_log(str(datas),file=lf)
# write_log(str(labels),file=lf)
# write_log(str(datas.shape),file=lf)
# write_log(str(labels.shape),file=lf)
cf = SVR(kernel="rbf")
print(datas.size, datas.shape)
datas = datas.reshape((4, 220 * 220 * 3))
write_log(str(datas[0]), file=lf)
write_log(str(datas[0].shape), file=lf)
labels = ["1", "2", "3", "4"]
cf.fit(datas, labels)
result = cf.predict(datas[0:])
write_log(str(result), file=lf)
# avr = avr/4
# avr = avr.astype("uint8")
# avr_im = Image.fromarray(avr)
# avr_im.show()
# print(avr.shape)
# print(avr.size)
import cv2 as cv from __init__ import write_log, you_jpg lg = "_test_cnn_9_9.log" img = cv.imread(you_jpg) hog = cv.HOGDescriptor() foundLocations, foundWeights = hog.detectMultiScale(img) foundLocations_d, weights = hog.detect(img) descriptors = hog.compute(img) write_log(str(foundLocations), file=lg) write_log(str(foundLocations_d), file=lg) write_log(str(descriptors), file=lg)
from sklearn import datasets from __init__ import write_log log_file = "_test_cnn_9_4.log" diabetes = datasets.load_diabetes() write_log(str(len(diabetes.data[0])), file=log_file)
"Pullover",
"Dress",
"Coat",
"Sandal",
"Shirt",
"Sneaker",
"Bag",
"Ankle boot",
]
fashion_mnist = keras.datasets.fashion_mnist
(x, y), (x_t, y_t) = fashion_mnist.load_data()
x = x / 255.0
x_t = x_t / 255.0
write_log(str(x), file="_test_5.log")
write_log(str(y), file="_test_5.log")
write_log(str(x_t), file="_test_5.log")
model = keras.Sequential([
keras.layers.Flatten(input_shape=(28, 28)),
keras.layers.Dense(128, activation=tf.nn.relu),
keras.layers.Dense(10, activation=tf.nn.softmax),
])
model.compile(
optimizer=tf.train.AdamOptimizer(),
loss="sparse_categorical_crossentropy",
metrics=["accuracy"],
)
在scikit-learn库中，有AdaBoostRegression（回归）和AdaBoostClassifier（分类）两个
在对和AdaBoostClassifier进行调参时，主要是对两部分进行调参：1）AdaBoost框架调参；2）弱分类器调参
"""
from sklearn.datasets import load_iris
from sklearn.ensemble import AdaBoostClassifier
# 导包
from sklearn.model_selection import cross_val_score
from __init__ import write_log
log_file = "test_cnn_9_1.log"
# 载入数据，sklearn中自带的iris数据集
iris = load_iris()
write_log(str(iris), file=log_file)
"""
AdaBoostClassifier参数解释
base_estimator:弱分类器，默认是CART分类树：DecisionTressClassifier
algorithm：在scikit-learn实现了两种AdaBoost分类算法，即SAMME和SAMME.R，
SAMME就是原理篇介绍到的AdaBoost算法，指Discrete AdaBoost
SAMME.R指Real AdaBoost，返回值不再是离散的类型，而是一个表示概率的实数值，算法流程见后文
两者的主要区别是弱分类器权重的度量，SAMME使用了分类效果作为弱分类器权重，SAMME.R使用了预测概率作为弱分类器权重。
SAMME.R的迭代一般比SAMME快，默认算法是SAMME.R。因此，base_estimator必须使用支持概率预测的分类器。
loss：这个只在回归中用到，不解释了
n_estimator:最大迭代次数，默认50。在实际调参过程中，常常将n_estimator和学习率learning_rate一起考虑
learning_rate:每个弱分类器的权重缩减系数v。f_k(x)=f_{k-1}*a_k*G_k(x)。较小的v意味着更多的迭代次数，默认是1，也就是v不发挥作用。
另外的弱分类器的调参，弱分类器不同则参数不同，这里不详细叙述
"""
# 构建模型
clf = AdaBoostClassifier(n_estimators=100) # 弱分类器个数设为100