def image_triple_generator(lst_files,
input_shape,
batch_size,
crop_shape=None):
pos_ratio, neg_ratio = 1, 1
pos_limit, neg_limit = 1, 4
pos_factor, neg_factor = 1, 1.01
img_cache = {}
datagen_args = dict(horizontal_flip=True)
datagen_left = IDG(**datagen_args)
datagen_right = IDG(**datagen_args)
f = np.load(lst_files)
lst, y = f['lst'], f['label']
num_batches = len(y) // batch_size + 1
clss = np.unique(y)
num_clss = clss.shape[0]
kmap = {v: k for k, v in enumerate(clss)}
label_set = [np.where(y == c)[0] for c in clss]
step = 0
while True:
step += 1
#loop per epoch
for bid in range(num_batches):
id_left, id_right, y_diff = utils.gen_pairs(
y, kmap, label_set, batch_size, pos_ratio, neg_ratio)
Xleft = utils.process_images([lst[i] for i in id_left],
datagen_left, img_cache, input_shape,
crop_shape)
Xright = utils.process_images([lst[i] for i in id_right],
datagen_right, img_cache,
input_shape, crop_shape)
Y_diff = np.array(y_diff)
yield [Xleft, Xright], Y_diff
if step % 10 is 0:
pos_ratio = min(pos_ratio * pos_factor, pos_limit)
neg_ratio = min(neg_ratio * neg_factor, neg_limit)
on_windows = False
if on_windows:
data_directory = 'D:\\Data\\Sketches\\png'
path_delim = '\\'
else:
data_directory = '../data/places_200'
path_delim = '/'
# From https://stackoverflow.com/questions/46717742/split-data-directory-into-training-and-test-directory-with-sub-directory-structu
#image dimensions
img_height = 299
img_width = 299
train_datagen = IDG(samplewise_std_normalization=True,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True,
validation_split=0.2)
batch_size = 100
class_mode = 'categorical'
color_mode = 'rgb'
shuffle = True
train_generator = train_datagen.flow_from_directory(data_directory,
target_size=(img_width,
img_height),
batch_size=batch_size,
class_mode=class_mode,
color_mode=color_mode,
shuffle=shuffle,
def data_generator(X, y, num_classes, batch_size=32):
return IDG(rescale=1. / 225, shear_range=0.2, zoom_range=0.2, horizontal_flip=True) \
.flow(X, to_categorical(y, num_classes), batch_size=batch_size)
if on_windows:
data_directory = 'D:\\Data\\Sketches\\png'
path_delim = '\\'
else:
data_directory = '../data/Sketches/png'
path_delim = '/'
# From https://stackoverflow.com/questions/46717742/split-data-directory-into-training-and-test-directory-with-sub-directory-structu
#image dimensions
img_height = 224
img_width = 224
train_datagen = IDG(featurewise_std_normalization=True,
shear_range=0.25,
zoom_range=0.25,
rotation_range=45,
horizontal_flip=True,
vertical_flip=True,
validation_split=0.2)
batch_size = 100
class_mode = 'categorical'
color_mode = 'grayscale'
shuffle = True
train_generator = train_datagen.flow_from_directory(data_directory,
target_size=(img_width,
img_height),
batch_size=batch_size,
class_mode=class_mode,
color_mode=color_mode,
mdl_clssify.add(Conv2D(64,3,3, input_shape=(64,64,3),activation='relu'))
mdl_clssify.add(MaxPooling2D(pool_size=(2,2)))
mdl_clssify.add(Conv2D(64,3,3,activation='relu'))
mdl_clssify.add(MaxPooling2D(pool_size=(2,2)))
#mdl.clssify.add(Dropout(0.5))
mdl_clssify.add(Flatten())
mdl_clssify.add(Dense(output_dim=128,activation='relu'))
mdl_clssify.add(Dense(output_dim=1 ,activation='sigmoid'))
mdl_clssify.compile(optimizer='adam',loss='binary_crossentropy',metrics=['accuracy'])
trainData_generat = IDG(
rescale=1./255,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True)
testData_generat = IDG(rescale=1./255)
set_train = trainData_generat.flow_from_directory('D:/final year project/AMD/data/train',
target_size=(64, 64), batch_size=9,
class_mode='binary')
set_test = testData_generat.flow_from_directory('D:/final year project/AMD/data/test',
target_size=(64, 64), batch_size=9,
class_mode='binary')
#pdb.set_trace()
import cv2 #for reading the images into grayscale vectors
from PIL import Image #used to manipulate the images into the appropriate sizes
import matplotlib.pyplot as plt
# %matplotlib inline
from keras.preprocessing.image import ImageDataGenerator as IDG #keras's built in image augmentor, which is perfect due to our limited data set.
size = (160, 160) #the new size of each image after resizing. Dimensions are still up in the air
#using IDG, make an augmentor with the parameters indicating what augmentations can take place.
aug_generator = IDG(rescale=1./255, #this is to normalize each pixel's numerical values to between 0 and 1, since RGB values go from 0 to 255.
rotation_range=45,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True,
fill_mode='nearest',
validation_split=0.2) #some of our data will be used for validation, so split it off
normal_generator = IDG(rescale=1./255) #no need to have augmentation parameters since the model isn't being trained to fit data generated by this.
# Just normalize the data appropriately.
#we'll use our augmented data generator instead of just extracting our training data. IDG has something perfect for this,
#considering we have the directories already.
#needed data generators
training_generator = aug_generator.flow_from_directory(train_dir,
target_size=size, #force resizes all input images, super nice.
labels[i * batch_size:(i + 1) * batch_size] = labels_batch
i += 1
if i * batch_size >= sample_count:
break
return features, labels
if __name__ == "__main__":
##using VGG16 to pre-extract feature used to train the FC layer
conv_base = VGG16(weights='imagenet',
include_top=False,
input_shape=(150, 150, 3))
#print conv_base.summary()
train_dir = '/Users/han/keras_use/examples/CNN/dogsvscats/train'
validation_dir = '/Users/han/keras_use/examples/CNN/dogsvscats/validation'
datagen = IDG(rescale=1. / 255)
batch_size = 20
train_features, train_labels = extract_features(train_dir, 2000)
validation_features, validation_labels = extract_features(
validation_dir, 1000)
train_features = np.reshape(train_features, (2000, 4 * 4 * 512))
validation_features = np.reshape(validation_features, (1000, 4 * 4 * 512))
##FC layer
model = models.Sequential()
model.add(layers.Dense(256, activation='relu', input_dim=4 * 4 * 512))
model.add(layers.Dropout(0.5))
model.add(layers.Dense(1, activation='sigmoid'))
model.compile(optimizer=optimizers.RMSprop(lr=1e-4),
loss='binary_crossentropy',
metrics=['accuracy'])
def showIm(n):
plt.figure()
plt.imshow(np.array(xTrain[n]).reshape((28, 28)), cmap='binary_r')
plt.xlabel(f"Value: {yTrain[n]}", fontsize=18)
plt.show()
showIm(randint(0, 10000))
xTrain = np.reshape(xTrain, (-1, 28, 28, 1))
print(xTrain.shape)
dataGeneration = IDG(rescale=1. / 255,
rotation_range=20,
zoom_range=0.2,
width_shift_range=0.1,
height_shift_range=0.1,
validation_split=0.2)
trainDataGen = dataGeneration.flow(xTrain,
yTrain,
batch_size=64,
subset='training')
valDataGen = dataGeneration.flow(xTrain,
yTrain,
batch_size=64,
subset='validation')
model = Sequential()
model.add(
C, W, H = 3, 50, 50
TRAIN_DIR = '/home/main/programming/kaggle/fisheries/data/train/'
TEST_DIR = '/home/main/programming/kaggle/fisheries/data/test/'
ORDER = ['ALB', 'BET', 'DOL', 'LAG', 'NoF', 'OTHER', 'SHARK', 'YFT']
NDIM = len(ORDER)
idg = IDG(
# featurewise_center=True,
# featurewise_std_normalization=True,
rotation_range=5.,
width_shift_range=0.05,
height_shift_range=0.05,
zoom_range=(0.95, 1 / 0.95),
channel_shift_range=0.5,
horizontal_flip=True,
vertical_flip=True,
fill_mode='constant',
cval=0,
)
os.makedirs('/tmp/fisheries', exist_ok=True)
mem = Memory(cachedir='/tmp/fisheries', verbose=0)
@mem.cache
def read_imagefile(fn, c, w, h):
img = load_img(fn,
keep_aspect_ratio=True,
train_cats_dir = '/Users/han/keras_use/examples/CNN/dogsvscats/train/cats'
fnames = [
os.path.join(train_cats_dir, fname)
for fname in os.listdir(train_cats_dir)
]
# print fnames
img_path = fnames[3]
print img_path
img = image.load_img(img_path, target_size=(150, 150))
# print type(img)
x = image.img_to_array(img)
x = x.reshape((1, ) + x.shape)
# print type(x),x.shape
##Image data augmentation
i = 0
datagen = IDG(rotation_range=60,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True,
fill_mode='nearest')
for batch in datagen.flow(x, batch_size=1):
plt.figure(i)
imgplot = plt.imshow(image.array_to_img(batch[0]))
plt.savefig(str(i) + "cat.png")
i += 1
if i % 4 == 0: break
plt.show()
def train(self,
loss = 'categorical_crossentropy',
optimizer = keras.optimizers.rmsprop(lr=0.0001, decay=1e-6),
metrics = ['accuracy'],
batch_size = 32,
epochs = 10,
verbose = 1,
generator = False,
use_saved_model = False):
if self.trained:
if len(os.listdir('./photos')) <= self.cls_num:
print('\nAlready using the newest model!\nOffer more data to re-train.')
return
if os.path.exists('./models') and use_saved_model:
self.model = keras.models.load_model(os.listdir('./models')[-1])
self.trained = True
self.result = 'unused'
return
if self._read_data():
self.data, self.label = self._preprocessing(self.data)
if verbose:
print('\nSuccessfully load data: ', end = '')
print(self.data.shape)
print('\nNow start build model... ')
self.do_nothing(2)
self._build_model(optimizer, loss, metrics, self.cls_num)
if verbose:
print('\nSuccessfully built model.')
print('\nModel summary:')
print(self.model.summary())
rst = raw_input('Start training? [Y/n] ')
if rst in ['n','N']:
print('Abort')
return
print('\ntraining...')
self.do_nothing(2)
if generator:
data_generator = IDG(featurewise_center=False,
samplewise_center=False,
featurewise_std_normalization=False,
samplewise_std_normalization=False,
zca_whitening=False,
rotation_range=15,
width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
vertical_flip=True)
# Required for featurewise_center,
# featurewise_std_normalization
# and zca_whitening.
#data_generator.fit(self.data)
self.model.fit_generator(data_generator.flow(self.data,
self.label,
batch_size),
steps_per_epoch = self.data.shape[0] / batch_size,
epochs = epochs)
else:
self.model.fit(x = self.data,
y = self.label,
batch_size = batch_size,
epochs = epochs,
verbose = verbose)
self.trained = True
self.result = 'unused'
self.model.save_weights(os.path.join(os.getcwd(),
'models',
time.strftime('%Y%m%d%H%M')+'.h5'))
print('\33[1mSuccessfully trained the model!\33[0m')
return
print('\nError occurs when trying to train the model. Abort')
x = Flatten()(xception.output)
prediction = Dense(len(folders), activation='relu')(x)
model = Model(inputs=xception.input,outputs=prediction)
model.summary()
model.compile(
loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
train_datagen = IDG(rescale = 1./255,
shear_range = 0.2,
zoom_range = 0.2,
horizontal_flip = True)
test_datagen = IDG(rescale = 1./255)
val_datagen = IDG(rescale = 1./255)
training_set = train_datagen.flow_from_directory(train_path,
target_size = (299, 299),
batch_size = 32,
class_mode = 'categorical')
test_set = test_datagen.flow_from_directory(test_path,
target_size = (299, 299),
batch_size = 32,
class_mode = 'categorical')
def train(model: keras.models.Model,
optimizer: dict,
save_path: str,
train_dir: str,
valid_dir: str,
batch_size: int = 32,
epochs: int = 10,
samples_per_epoch=1000,
pretrained=None,
augment: bool = True,
weight_mode=None,
verbose=0,
**kwargs):
""" Trains the model with the given configurations. """
shape = model.input_shape[1:3]
optimizer_cpy = optimizer.copy()
shared_gen_args = {
'rescale': 1. / 255, # to preserve the rgb palette
}
train_gen_args = {}
if augment:
train_gen_args = {
"fill_mode": 'reflect',
'horizontal_flip': True,
'vertical_flip': True,
'width_shift_range': .15,
'height_shift_range': .15,
'shear_range': .5,
'rotation_range': 45,
'zoom_range': .2,
}
gen = IDG(**{**shared_gen_args, **train_gen_args})
gen = gen.flow_from_directory(train_dir,
target_size=shape,
batch_size=batch_size,
seed=SEED)
val_count = len(
glob(os.path.join(valid_dir, '**', '*.jpg'), recursive=True))
valid_gen = IDG(**shared_gen_args)
optim = getattr(keras.optimizers, optimizer['name'])
if optimizer.pop('name') != 'sgd':
optimizer.pop('nesterov')
schedule = optimizer.pop('schedule')
if schedule == 'decay' and 'lr' in optimizer.keys():
initial_lr = optimizer.pop('lr')
else:
initial_lr = 0.01
optim = optim(**optimizer)
callbacks = [
utils.checkpoint(save_path),
utils.csv_logger(save_path),
]
if pretrained is not None:
if not os.path.exists(pretrained):
raise FileNotFoundError()
model.load_weights(pretrained, by_name=False)
if verbose == 1:
print("Loaded weights from {}".format(pretrained))
if optimizer_cpy['name'] == 'sgd':
if schedule == 'decay':
callbacks.append(utils.step_decay(epochs, initial_lr=initial_lr))
elif schedule == 'big_drop':
callbacks.append(utils.constant_schedule())
model.compile(optim,
loss='categorical_crossentropy',
metrics=['accuracy', top3_acc])
create_xml_description(save=os.path.join(save_path, 'model_config.xml'),
title=model.name,
epochs=epochs,
batch_size=batch_size,
samples_per_epoch=samples_per_epoch,
augmentations=augment,
schedule=schedule,
optimizer=optimizer_cpy,
**kwargs)
if weight_mode:
class_weights = [[key, value] for key, value in weight_mode.items()]
filen = os.path.join(save_path, 'class_weights.npy')
np.save(filen, class_weights)
h = None # has to be initialized here, so we can reference it later
try:
with warnings.catch_warnings():
warnings.simplefilter("ignore")
h = model.fit_generator(
gen,
steps_per_epoch=samples_per_epoch / batch_size,
epochs=epochs,
validation_data=valid_gen.flow_from_directory(
valid_dir,
target_size=shape,
batch_size=batch_size,
seed=SEED),
validation_steps=val_count / batch_size,
callbacks=callbacks,
class_weight=weight_mode,
verbose=2)
except KeyboardInterrupt:
save_results(verbose=1, save_path=save_path, model=model, hist=h)
return
save_results(verbose=1, save_path=save_path, model=model, hist=h)
"""
Created on Sun Mar 11 23:52:26 2018
@author: Ibo Turk
"""
from keras.preprocessing.image import ImageDataGenerator as IDG
from IboTurk_part0 import data as x_train
from IboTurk_part0 import data_values as y_train
from matplotlib import pyplot as plt
from keras.utils import np_utils
from preprocessing import ZiscFunction
ITdatagen = IDG(samplewise_center=False,
rotation_range=360,
shear_range=360,
data_format="channels_last",
fill_mode="constant",
preprocessing_function=ZiscFunction)
#y_train = np_utils.to_categorical(y_train, 2)
#print(y_train)
#ITdatagen.fit(x_train)
#batches = 0
#for x_batch, y_batch in ITdatagen.flow(x_train, y_train, batch_size=32):
# batches += 1
# if batches >= 1:
# we need to break the loop by hand because
# the generator loops indefinitely
# break
#plt.subplot(121)
from keras.preprocessing.image import ImageDataGenerator as IDG
import numpy as np
import os
from scipy.misc import imread
datagen = IDG(horizontal_flip=True, vertical_flip=True)
# def load_images_from_folder(folder):
# images = []
# for filename in os.listdir('1'):
# if any([filename.endswith(x) for x in ['.jpeg', '.jpg','.JPG']]):
# img = Image.open(os.path.join(folder, filename))
# img=np.array(img)
# datagen.fit(img.reshape(1,1696,2544,3))
# i=0
# for batch in datagen.flow(img.reshape(1,1696,2544,3), batch_size=1,save_to_dir='preview',
# save_prefix='lol', save_format='jpeg'):
# i += 1
# if i == 4:
# break
# if img is not None:
# images.append(img)
# return images
for root, dirs, files in os.walk('3'):
for file in files:
if file.endswith(".jpg") or file.endswith(".JPG"):
img = imread(os.path.join(root, file))
k = img.shape
i = 0
for batch in datagen.flow(img.reshape(1, k[0], k[1], k[2]),
batch_size=1,
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(128, (3, 3), activation="relu"))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(128, (3, 3), activation="relu"))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Flatten())
model.add(layers.Dense(512, activation='relu'))
model.add(layers.Dense(1, activation='sigmoid'))
model.compile(loss='binary_crossentropy',
optimizer=optimizers.RMSprop(lr=1e-4))
train_datagen = IDG(rescale=1. / 255)
test_datagen = IDG(rescale=1. / 255)
train_generator = train_datagen.flow_from_directory(train_dir,
target_size=(150, 150),
batch_size=20,
class_mode='binary')
validation_generator = test_datagen.flow_from_directory(validation_dir,
target_size=(150, 150),
batch_size=20,
class_mode='binary')
history = model.fit_generator(train_generator,
steps_per_epoch=100,
epochs=30,
print("Network Parameters:\n",opt.get_config())
model = MohlerNet4.build(width=32,height=32,depth=3,classes=3)
model.compile(loss="categorical_crossentropy", optimizer=opt,
metrics=["accuracy"])
# train the network
print("[INFO] training network...")
numEpochs = 50
batch_size = 16
if dataAugmentation == True:
train_datagen = IDG(
rotation_range=20,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range = 0.2,
zoom_range = 0.2,
horizontal_flip=True)
train_datagen.fit(trainX)
H = model.fit_generator(train_datagen.flow(trainX,trainY,batch_size=batch_size),
steps_per_epoch=(2250//batch_size)
,epochs=numEpochs,
validation_data=(testX,testY))
else:
H = model.fit(trainX, trainY, validation_data=(testX, testY),
batch_size=batch_size, epochs=numEpochs, verbose=1)
#evaluate the network
print("[INFO] evaluating network...")
