keras 二分类

文件结构:

• classify 
  • train.py
  • test.py
  • train
    • cat
      • xxx.jpg xxx.jpg
    • dog
      • xxx.jpg xxx.jpg
  • test
    • xxx.jpg xxx.jpg

train.py:

import glob
from PIL import Image, ImageOps
import numpy as np


from sklearn.model_selection import train_test_split
from sklearn.preprocessing import LabelBinarizer

import keras
from keras.utils import to_categorical

z = glob.glob('./train/*/*.jpg')
ori_label = []
ori_imgs = []
assert type(z) == list

img_size_x = 112
img_size_y = 112

for fn in z:
    if fn[-3:] != 'jpg':
        continue
    ori_label.append(fn.split('/')[-2])
    new_img = Image.open(fn)
    ori_imgs.append(ImageOps.fit(new_img, (img_size_x, img_size_y), Image.ANTIALIAS).convert('RGB'))

imgs = np.array([np.array(im) for im in ori_imgs])
imgs = imgs.reshape(imgs.shape[0], img_size_x, img_size_y, 3) / 255

lb = LabelBinarizer()
label = lb.fit_transform(ori_label)
label = to_categorical(label)

trainX, validX, trainY, validY = train_test_split(imgs, label, test_size=0.2, random_state=42)

from keras.layers import Input, Conv2D, BatchNormalization, Activation, MaxPooling2D, GlobalMaxPooling2D, Dense, Dropout
from keras.models import Model, load_model
from keras.optimizers import SGD

IM_imput = Input((img_size_x, img_size_y, 3))
IM = Conv2D(32,(5,5))(IM_imput)
IM = BatchNormalization(axis = 3)(IM)
IM = Activation('relu')(IM)
IM = MaxPooling2D(strides=(2,2))(IM)
IM = Conv2D(64,(5,5))(IM_imput)
IM = BatchNormalization(axis = 3)(IM)
IM = Activation('relu')(IM)
IM = GlobalMaxPooling2D()(IM)
IM = Dense(64,activation='relu')(IM)
IM = Dropout(0.5)(IM)
IM = Dense(2,activation='relu')(IM)

model = Model(inputs=IM_input, outputs=IM)
model.summary()

sgd = SGD(lr=0.01, momentum=0.9, decay=0.002, nesterov=False)
model.compile(loss='binary_crossentropy',optimizer=sgd,metrics=['acc'])

from keras.callbacks import LearningRateScheduler, EarlyStopping, ModelCheckpoint

batch_size = 32

annealer = LearningRateScheduler(lambda x: 1e-3*0.9**x)
early_stop = EarlyStopping(patience=8)
modelsave = ModelCheckpoint(filepath='model.h5',save_best_only=True,verbose=1)

model.fit(
    trainX, trainY, batch_size = batch_size,
    epochs=20,
    validation_data = (validX, validY),
    callbacks = [annealer,early_stop,modelsave]
)

test.py:

import os
from keras.models import load_model
import numpy as np
from PIL import Image
import cv2

model = load_model('./model.h5')

#single_img:

input = cv2.imread('0_1.jpg',cv2.IMREAD_COLOR)
input = cv2.resize(input, (112, 112),interpolation = cv2.INTER_CUBIC)
input= np.array(input).reshape(1,112,112,3)/255
pre_y = model.predict(input)
label = np.argmax(pre_y)
print(label)

#file_img:
'''
predict_dir = 'train/wu'

def get_inputs(src=[]):
  pre_x = []
  for s in src:
    input = cv2.imread('0_1.jpg',cv2.IMREAD_COLOR)
    input = cv2.resize(input, (112, 112),interpolation = cv2.INTER_CUBIC)
    #input= np.array(input).reshape(1,112,112,3)/255
    pre_x.append(input)
  pre_x = np.array(pre_x) / 255.0
  return pre_x

test = os.listdir(predict_dir)
images = []

for testpath in test:
  if testpath.endswith('jpg'):
      fd = os.path.join(predict_dir,testpath)
      images.append(fd)

pre_x = get_inputs(images)
pre_y = model.predict(pre_x)
print(pre_y)
'''