Esempio n. 1
0
import os
import numpy as np

from keras.preprocessing.image import img_to_array, load_img
from data_utils import prepare_input_image_batch, postprocess_output, resize
from model_utils import generate_mobilenet_model


IMAGE_FOLDER_PATH = r"D:\Yue\Documents\Datasets\MSCOCO\val\valset\\"
batch_size = 10
image_size = 256

model = generate_mobilenet_model(img_size=image_size)
model.load_weights('weights/model.h5')

X = []
files = os.listdir(IMAGE_FOLDER_PATH)

files = files[:100]
for i, filename in enumerate(files):
    img = img_to_array(load_img(os.path.join(IMAGE_FOLDER_PATH, filename))) / 255.
    img = resize(img, (image_size, image_size, 3)) * 255.  # resize needs floats to be in 0-1 range, preprocess needs in 0-255 range
    X.append(img)

    if i % (len(files) // 20) == 0:
        print("Loaded %0.2f percentage of images from directory" % (i / float(len(files)) * 100))

X = np.array(X, dtype='float32')
print("Images loaded. Shape = ", X.shape)

X_lab, X_features = prepare_input_image_batch(X, batchsize=batch_size)
	sepia_filter = np.array([[.393, .769, .189], [.349, .686, .168], [.272, .534, .131]])
	scale = 0.01
	sepia_img = img.dot(sepia_filter.T * scale)
	sepia_img /= sepia_img.max()
	sepia_img = skimage.img_as_ubyte(sepia_img)
	return sepia_img 
	#variance_generator = lambda i,j,h: 0.25*(i+j)/1022. + 0.001
	#variances = np.fromfunction(variance_generator,(224,224,3))
	#sepia_img = skimage.util.random_noise(sepia_img, mode='gaussian', seed=1)
	#sepia_img = skimage.util.random_noise(sepia_img, mode='localvar', local_vars=variances)
	
	#return skimage.img_as_ubyte(skimage.color.rgb2gray(sepia_img))
	
	return sepia_img
	
model = generate_mobilenet_model()
model.load_weights(_MODEL_WEIGHTS_NAME)

def get_images_names(path):
    images = os.listdir(path)
    final = []
    for image in images:
        final.append(image.rstrip())
    return images

images_names = get_images_names(_IMAGE_PATH)

for img in images_names:
	x_b = img_to_array(load_img(_IMAGE_PATH+img, target_size= (224,224)))
	x_b = np.expand_dims(x_b, axis=0)
	x_b = x_b/255 #conversion to a range of -1 to 1. Explanation saved.
import os

from data_utils import train_generator, val_batch_generator
from model_utils import generate_mobilenet_model
from train_utils import TensorBoardBatch

from keras.callbacks import ModelCheckpoint

nb_train_images = 60000  # there are 82783 images in MS-COCO, set this to how many samples you want to train on.
batch_size = 125

model = generate_mobilenet_model(lr=1e-3)
model.summary()

# continue training if weights are available
#if os.path.exists('weights/mobilenet_model.h5'):
#    model.load_weights('weights/mobilenet_model.h5')

# use Batchwise TensorBoard callback
tensorboard = TensorBoardBatch(batch_size=batch_size)
checkpoint = ModelCheckpoint('weights/mobilenet_model_v2.h5',
                             monitor='loss',
                             verbose=1,
                             save_best_only=True,
                             save_weights_only=True)
callbacks = [checkpoint, tensorboard]

model.fit_generator(generator=train_generator(batch_size),
                    steps_per_epoch=nb_train_images // batch_size,
                    epochs=100,
                    verbose=1,