def aug_method():
# 增强方法1: 图片变形
# 1) 水平翻转: 最常用!! 对卷积神经网络而言, 左右互换(水平翻转)差异很大
aug = image.HorizontalFlipAug(0.5)
apply(img, aug)
plt.gcf().suptitle('HorizontialFlipAug')
plt.show()
# 2) 随机裁剪: 卷积神经网络对位置敏感, 虽然Pool池化可以减小位置敏感, 但仍不不够
# 2.1) 裁剪指定(width x height) 的图片
aug = image.RandomCropAug((200, 200))
apply(img, aug)
plt.show()
# 2.2) 裁剪指定'长宽比'范围大小且不小于最小值的随机图片, 并resize为指定(width x height)大小
aug = image.RandomSizedCropAug(size=(100, 100),
min_area=.1,
ratio=(0.5, 2.0))
apply(img, aug)
plt.show()
# 增强方法2: 颜色变化
# 1) 亮度变化:
aug = image.BrightnessJitterAug(
brightness=0.5
) # brightness : The brightness jitter ratio range, [0, 1]
apply(img, aug)
plt.show()
# 2) 色调变化:
aug = image.HueJitterAug(hue=0.5)
apply(img, aug)
plt.show()
def get_augs(shape):
train_augs = [
image.HorizontalFlipAug(.5),
image.RandomCropAug((shape, shape))
]
valid_augs = train_augs
tests_augs = [image.CenterCropAug((shape, shape))]
return train_augs, valid_augs, tests_augs
def transform_train(data, label, data_shape=(3, 363, 363), resize=363):
# data [height, width, channel]
# im = data.astype('float32') / 255,
ctx = data.context
im = resize_longer(data.asnumpy(), resize, data_shape)
im = rotate_image(im)
im = nd.array(im, ctx=ctx)
im = im.astype('float32') / 255
auglist = [
image.RandomCropAug((331, 331)),
image.ColorJitterAug(0.3, 0.3, 0.3),
image.RandomGrayAug(0.5),
# image.ColorNormalizeAug(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
image.ColorNormalizeAug(mean=(0.417, 0.402, 0.366),
std=(0.081, 0.079, 0.080)),
]
for aug in auglist:
im = aug(im)
# im = im.clip(0, 1)
im = nd.transpose(im, (2, 0, 1))
return im, nd.array([label], ctx=ctx).asscalar()
def augment_data(imags, label):
datas, labels = [], []
n = imags.shape[0]
for k in range(n):
imag = imags[k].reshape(shape=(1, imags[k].shape[0], imags[k].shape[1],
imags[k].shape[2]))
for i in range(4):
trans_band = transform(
imag, image.HorizontalFlipAug(.5)).astype('float32')
datas.append(img_norm(trans_band))
labels.append(label[k].asscalar())
for i in range(9):
trans_band = transform(
imag, image.RandomSizedCropAug((75, 75), .75,
(.8, 1.2))).astype('float32')
datas.append(img_norm(trans_band))
labels.append(label[k].asscalar())
# brightness augmenter
for i in range(9):
trans_band = transform(
imag, image.BrightnessJitterAug(.1)).astype('float32')
datas.append(img_norm(trans_band))
labels.append(label[k].asscalar())
# random crop augmenter
for i in range(9):
trans_band = resize(transform(imag, image.RandomCropAug((50, 50))),
75).astype('float32')
datas.append(img_norm(trans_band))
labels.append(label[k].asscalar())
# center crop augmenter
trans_band = resize(transform(imag, image.CenterCropAug((50, 50))),
75).astype('float32')
datas.append(img_norm(trans_band))
labels.append(label[k].asscalar())
ds = nd.concat(*datas, dim=0)
return ds, labels
from mxnet import nd, gluon, image
batch_size = 64
train_augs = [image.HorizontalFlipAug(.5), image.RandomCropAug((224, 224))]
test_augs = [image.CenterCropAug((224, 224))]
data_dir = '/home/zhbli/temp/dog_dataset'
train_imgs = gluon.data.vision.ImageFolderDataset(
data_dir + '/train', transform=lambda X, y: transform(X, y, train_augs))
test_imgs = gluon.data.vision.ImageFolderDataset(
data_dir + '/test', transform=lambda X, y: transform(X, y, test_augs))
def transform(data, label, augs):
data = data.astype('float32')
for aug in augs:
data = aug(data)
data = nd.transpose(data, (2, 0, 1))
return data, nd.array([label]).asscalar().astype('float32')
def getDataset():
train_data = gluon.data.DataLoader(train_imgs, batch_size, shuffle=True)
test_data = gluon.data.DataLoader(test_imgs, batch_size)
return train_data, test_data
for file_name in files:
if not file_name.endswith('.png'):
continue
img_dir = os.path.join(path_crocodile, file_name)
img_arr = mx.image.imread(img_dir)
img_arr = nd.transpose(img_arr, (2, 0, 1))
img_arr = (img_arr.astype(np.float32) / 255)
img_arr = nd.array(img_arr.reshape((1, ) + img_arr.shape))
img_list.append(img_arr)
img_list_crocodile.append(img_arr)
size = 28
train_augs = [
image.HorizontalFlipAug(.5),
image.ColorJitterAug(.5, .5, .5),
image.RandomCropAug((size, size)),
]
test_augs = [image.CenterCropAug((size, size))]
def apply_aug_list(img, augs):
for f in augs:
img = f(img)
return img
def try_gpu():
try:
ctx = mx.gpu()
_ = nd.array([0], ctx=ctx)
except:
from mxnet import nd
import utils
def apply(img, aug, n=3):
X = [aug(img.astype('float32')) for _ in range(n * n)]
Y = nd.stack(*X).clip(0, 255) / 255
utils.show_images(Y, n, n, figsize=(256, 256))
# 水平翻转
aug = image.HorizontalFlipAug(.5)
apply(img, aug)
# 随机裁剪一个块, 256 * 256
aug = image.RandomCropAug([256, 256])
apply(img, aug)
# 颜色变化
# 随机将亮度增加或减少一个量
aug = image.BrightnessJitterAug(.5)
apply(img, aug)
# 随机色调变化
aug = image.HueJitterAug(.5)
apply(img, aug)
def apply_aug_list(img, augs):
for aug in augs:
img = aug(img)
def apply_aug_list(img, augs):
for f in augs:
img = f(img)
return img
def train(train_augs, test_augs, learning_rate=.1):
batch_size = 128
num_epochs = 10
ctx = utils.try_all_gpus()
loss = gluon.loss.SoftmaxCrossEntropyLoss()
train_data, test_data = get_data(
batch_size, train_augs, test_augs)
net = utils.resnet18(10)
net.initialize(ctx=ctx, init=init.Xavier())
net.hybridize()
trainer = gluon.Trainer(net.collect_params(),
'sgd', {'learning_rate': learning_rate})
utils.train(
train_data, test_data, net, loss, trainer, ctx, num_epochs)
train_augs = [
image.HorizontalFlipAug(.5),
image.RandomCropAug((32,32))
]
test_augs = [
image.CenterCropAug((32,32))
]
train(train_augs, test_augs)
"""*********************************************************************************************************************
对CIFAR10数据集进行图片增强.
"""
def apply_aug_list(img, augs):
"""对输入图片, 按照augs列表中aug方法逐个增强"""
for aug in augs:
img = aug(img)
return img
# 训练集增强方法
train_augs = [image.HorizontalFlipAug(0.5), image.RandomCropAug(size=(28, 28))]
# 测试集增强方法
test_augs = [image.CenterCropAug(size=(28, 28))]
def get_transform(augs):
""" """
def transform(data, label):
""" """
data = data.astype('float32')
if augs is not None:
data = apply_aug_list(data, augs)
data = nd.transpose(data, axes=(2, 0, 1)) / 255
return data, label.astype('float32')
from src import utils
import mxnet as mx
import base64
from mxnet import image
from mxnet import nd
from collections import namedtuple
train_augs = [
image.HorizontalFlipAug(1.),
image.RandomCropAug((224, 224))
]
test_augs = [
image.CenterCropAug((224, 224))
]
def transform(data, label, augs):
data = data.astype('float32')
for aug in augs:
data = aug(data)
data = nd.transpose(data, (2, 0, 1))
return data, nd.array([label]).asscalar().astype('float32')
net = mx.mod.Module.load('../model/food', 0, True)
net.bind(data_shapes=[('data', (64, 3, 244, 244))])
batch = namedtuple('batch', ['data'])
def score_image(image_base64_string):
with open('target_image.jpg', 'wb') as f:
f.write(base64.b64decode(image_base64_string))
f.close()
import mxnet as mx
import numpy as np
import os
import Inception3_modify
from mxnet import gluon, image, nd, ndarray, init
from mxnet.gluon.model_zoo import vision as models
import sys
sys.path.append('..')
import utils
train_augs = [image.HorizontalFlipAug(.5), image.RandomCropAug((299, 299))]
test_augs = [image.CenterCropAug((299, 299))]
def transform(data, label, augs):
data = data.astype('float32')
for aug in augs:
data = aug(data)
data = nd.transpose(data, (2, 0, 1))
return data, nd.array([label]).asscalar().astype('float32')
ctx = mx.gpu()
data_dir = './data'
batch_size = 16
epochs = 1
###################################### true train data
train_imgs = gluon.data.vision.ImageFolderDataset(
data_dir + '/train_valid_test/Images',
for i in range(nrows):
for j in range(ncols):
figs[i][j].imshow(imgs[i * ncols + j].asnumpy())
figs[i][j].axes.get_xaxis().set_visible(False)
figs[i][j].axes.get_yaxis().set_visible(False)
plt.show()
self.__X = [
aug(self.__img.astype("float32"))
for _ in range(self.__n * self.__n)
]
self.__Y = nd.stack(*self.__X).clip(0, 255) / 255
show_images(self.__Y, self.__n, self.__n, figsize=(8, 8))
if __name__ == "__main__":
ia = IamgeAugment(input_path="cat.jpg", n=3)
# 变形
# 50% 几率水平翻转图片
ia.apply(aug=image.HorizontalFlipAug(.5))
# 剪裁
# 随机从原图片剪裁出 200×200 的一部分 , 可以图片先变大在剪小
ia.apply(aug=image.RandomCropAug([200, 200]))
# 随机从图片中剪裁一个区域要求是剪裁面积占原图片 >= 0.1 且剪裁的区域长/宽在 0.5~2 之间后将剪裁后的图片放缩到 200×200
ia.apply(aug=image.RandomSizedCropAug((200, 200), .1, (.5, 2)))
# 色调
# 随机将亮度增加或者减小在 0-50% 间的一个量
ia.apply(aug=image.BrightnessJitterAug(.5))
# 随机色调变化
ia.apply(aug=image.HueJitterAug(.5))
# -*- coding: utf-8 -*-
# @Time : 18-1-30 下午1:25
# @Author : lyyang
from mxnet.gluon.data import vision
from mxnet import image
from mxnet import nd
train_test_augs = [
image.RandomCropAug((image_size,image_size))
]
def transform(data, label, augs=None):
data = data.astype('float32')
if augs:
for aug in augs:
data = aug(data)
data = nd.transpose(data, (2,0,1))
return data, nd.array([label]).asscalar().astype('float32')
# 读取原始图像文件。flag=1说明输入图像有三个通道（彩色）。
train_ds = vision.ImageFolderDataset(input_str + 'train', flag=1,transform=lambda X, y: transform(X, y, train_test_augs))
test_ds = vision.ImageFolderDataset(input_str + 'test', flag=1,transform=lambda X, y: transform(X, y, train_test_augs))
loader = gluon.data.DataLoader
train_data = loader(train_ds, batch_size, shuffle=True, last_batch='keep')
test_data = loader(train_ds, batch_size, shuffle=True, last_batch='keep')
from mxnet import gluon
from mxnet import image
from mxnet import init
from mxnet.gluon.data import vision
from mxnet import nd
import net as net_collection
import utils
import numpy as np
input_str = '/data/datasets/cifar-10/'
batch_size = 32
image_size = 32
n_class = 10
train_test_augs = [image.RandomCropAug((image_size, image_size))]
# 设定训练数据扩充形式
def transform_train(data, label):
im = data.astype('float32') / 255
auglist = image.CreateAugmenter(data_shape=(3, 32, 32),
resize=0,
rand_crop=False,
rand_resize=False,
rand_mirror=True,
mean=np.array([0.4914, 0.4822, 0.4465]),
std=np.array([0.2023, 0.1994, 0.2010]),
brightness=0,
contrast=0,
saturation=0,