活动地址：CSDN21天学习挑战赛
学习的最大理由是想摆脱平庸，早一天就多一份人生的精彩；迟一天就多一天平庸的困扰。
宫颈癌是发生在子宫颈的癌症。如果宫颈癌在早期发现很容易治愈。然而，由于该领域缺乏专业知识，宫颈癌筛查和治疗计划面临的最大挑战之一是确定合适的治疗方法，而如果能够实时确定宫颈癌的类型，治疗工作将得到极大提高。
该案例的数据可在https://www.kaggle.com/c/intel-mobileodt-cervical-cancer-screening中找到。该数据集是开发基于图像准确识别女性子宫颈类型的算法挑战的一部分。这样可以防止无效治疗，并允许医者为需要更高级治疗的病例提供适当的转诊。
首先，在https://github.com/ml-resources/deeplearning-keras/tree/ed1/ch05中下载GitHub库。将训练集和测试集下载并保存到data文件夹。
·train.7z：训练集。图像按其标记的类别进行组织，分为Type_1、Type_2和Type_3。
·test.7z：测试集。
将训练图像存储在train文件夹下的子文件夹Type_1、Type_2和Type_3中。存储格式为.jpg文件，并根据不同类别存储在对应的文件夹中。通过字符串读取图像路径，函数用图像的文件夹名称来获取相应标签，并将图像路径和标签作为两个并行的数组返回。接下来介绍数据预处理步骤。
数据预处理
宫颈图像具有不同的大小，且具有高分辨率。对于CNN，输入数据不仅需要统一的大小，并且还需要足够的分辨率以便区分分类中的主要特征，而较低的分辨率可以避免计算能力上的限制。
def processCervicalData():
# image resizing
imgPaths = []
labels = []
trainingDirs = ['/deeplearning-keras/ch05/data/train']
for dir in trainingDirs:
newFilePaths, newLabels, numLabels = readFilePaths(dir)
if len(newFilePaths) > 0:
imgPaths += newFilePaths
labels += newLabels
imgPaths, labels = shuffle(imgPaths, labels)
labelCount = labelsCount(labels)
type1Count = labelCount[0]
type2Count = labelCount[1]
type3Count = labelCount[2]
print("Count of type1 : ", type1Count)
print("Count of type2 : ", type2Count)
print("Count of type3 : ", type3Count)
print("Total Number of data samples: " + str(len(imgPaths)))
print("Number of Classes: " + str(numLabels))
newShape = [(256,256,3)]
destDir = ['/deeplearning-keras/ch05/data/resized_imgs']
for newImgShape, destFolder in zip(newShape,destDir):
for i, path,label in zip(count(),imgPaths,labels):
split_path = path.split('/')
newPath = 'size'+str(newImgShape[0])+'_'+split_path[-1]
newPath = '/'.join([destFolder]+split_path[8:-1]+[newPath])
add_flip = True
if label == 1:
add_flip = False
# Used to exclude corrupt data
try:
resizeImage(path, maxSize=newImgShape, savePath=newPath, addFlip=add_flip)
except OSError:
print("Error at path " + path)
经网络一方面需要图像的大小恒定，另一方面需要图像尽量小以留出足够的内存空间进行模型训练，但图像需要足够的像素以根据重要特征进行分类。我们半自动地选择了256×256像素的图像大小。
利用以下代码对图像大小进行等比例调整：
def resizeImage(imgPath, maxSize=(256,256,3), savePath=None, addFlip=False):
ImageFile.LOAD_TRUNCATED_IMAGES = True
img = Image.open(imgPath)
# set aspect ratio
if type(img) == type(np.array([])):
img = Image.fromarray(img)
img.thumbnail(maxSize, Image.ANTIALIAS)
tmpImage = (np.random.random(maxSize)*255).astype(np.uint8)
resizedImg = Image.fromarray(tmpImage)
resizedImg.paste(img,((maxSize[0]-img.size[0])//2, (maxSize[1]-img.size[1])//2))
if savePath:
resizedImg.save(savePath)
if addFlip:
flip = resizedImg.transpose(Image.FLIP_LEFT_RIGHT)
if savePath:
splitPath = savePath.split('/')
flip_path = '/'.join(splitPath[:-1] + ['flipped_'+splitPath[-1]])
flip.save(flip_path)
return np.array(resizedImg, dtype=np.float32), np.array(flip,dtype=np.float32)
return np.array(resizedImg, dtype=np.float32)
利用训练集对CNN进行训练。作为宫颈癌模型训练的一部分，读入调整后的图像路径，用于该项目的其余部分：
def processCervicalData():
# image resizing
imgPaths = []
labels = []
trainingDirs = ['/deeplearning-keras/ch05/data/train']
for dir in trainingDirs:
newFilePaths, newLabels, numLabels = readFilePaths(dir)
if len(newFilePaths) > 0:
imgPaths += newFilePaths
labels += newLabels
imgPaths, labels = shuffle(imgPaths, labels)
labelCount = labelsCount(labels)
将数据分成训练集和验证集。80%的数据作为训练集用于拟合模型；20%的数据组成验证集，用于模型超参数调整的同时提供训练集拟合模型的无偏评估。拆分数据后，将路径和标签保存为CSV文件，并注明它们所属的数据集，以确保两个数据集中的数据不会混淆：
trainCSV = '/Users/manpreet.singh/git/deeplearning/deeplearning-keras/ch05/csvs/train_set.csv'
validCSV = '/Users/manpreet.singh/git/deeplearning/deeplearning-keras/ch05/csvs/valid_set.csv'
training_portion = .8
split_index = int(training_portion * len(imagePaths))
X_train_paths, y_train = imagePaths[:split_index], labels[:split_index]
X_valid_paths, y_valid = imagePaths[split_index:], labels[split_index:]
print("Train size: ")
print(len(X_train_paths))
print("Valid size: ")
print(len(X_valid_paths))
savePaths(trainCSV, X_train_paths, y_train)
savePaths(validCSV, X_valid_paths, y_valid)
train_csv = 'csvs/train_set.csv'
valid_csv = 'csvs/valid_set.csv'
X_train_paths, y_train = getSplitData(train_csv)
X_valid_paths, y_valid = getSplitData(valid_csv)
n_classes = max(y_train) + 1
数据已被分为训练集和验证集，相应的路径储存在X_train_paths和X_valid_paths中，如上述代码的输出所示。
利用独热编码将分类变量表示为二进制向量。独热编码将数据样本的标签表示为具有单个1值的零向量，1值的索引对应于表示有效的真值标签。独热编码是真实标签的一种有用的表示方式，且易用于神经网络，有利于计算模型损失并反向传播。
y_train = oneHotEncode(y_train, n_classes)
y_valid = oneHotEncode(y_valid, n_classes)
由于图像太多而导致无法将所有图像作为NumPy数组读入内存，因此创建生成器来分批读取图像到内存中，随机添加的样板能有效地增加数据集
我们创建具有两个全连接层的模型，且第二个全连接层比第一个小，然后是输出层。正如convModel所描述的那样，我们在第一层并行运行3×3、4×4和5×5的过滤器，而在接下来的卷积层中，由于内存限制而减小深度，只运行3×3和5×5的过滤器。
在Keras可以通过逐个添加层帮助我们实现层堆叠，而且在convModel中，每层的深度会依次递减。具体是将输入层、批量归一化层、卷积层，最大池化层、丢弃层、指数线性单位（ELU）层和softmax层堆叠在一起。其中在最大池化层后添加丢弃层是很有必要的，一方面是因为丢弃会对结果产生很大的影响，另一方面是因为如果在最大池化层之前添加，丢弃层的效果可能会减弱。
具体的步骤如下：
1.首先，创建CNN模型：
2.使用前面的convModel拟合训练集。为了使目标函数最小化，我们将自适应矩估计（Adam）优化器作为优化算法。Adam优化器是RMSProp和矩估计的结合，不仅对内存的要求比较低，而且几乎不需要对超参数进行调整。
3.保持学习率为0.001。如果将学习率设置得很低，也就意味着对权重的更新非常慢，训练的进展也将会非常缓慢。反之，当学习率维持在很高的水平，将会导致损失函数过早收敛于局部值。
4.使用categorical_crossentropy损失函数，测量分类模型的性能和准确率，并将结果作为衡量模型性能的指标。
整个工程代码如下：
import numpy as np
import os
from sklearn.utils import shuffle
from PIL import Image
from PIL import ImageFile
import matplotlib.image as mpimg
from itertools import count
import scipy.misc as sci
import scipy.ndimage.interpolation as scizoom
import random
import time
from multiprocessing.pool import ThreadPool
from keras.models import Model
import math
from keras import optimizers
from keras.layers import Conv2D, MaxPooling2D, Dense, Input, Flatten, Dropout, concatenate
from keras.layers.normalization import BatchNormalization
ImageFile.LOAD_TRUNCATED_IMAGES = True
def labelsCount(labels):
labelCountDic = dict()
for label in labels:
if label in labelCountDic:
labelCountDic[label] += 1
else:
labelCountDic[label] = 1
return labelCountDic
def readFilePaths(dir, no_labels=False, label_type=None):
filePaths = []
labels = []
labels2Nums = dict()
numLabels = None
for dirName, subDirs, Files in os.walk(dir):
if len(subDirs) > 0:
numLabels = len(subDirs)
for i, subDir in enumerate(subDirs):
labels2Nums[subDir] = i
else:
cervicalType = dirName.split('/')[-1]
for img in Files:
if '.jpg' in img.lower():
filePaths.append(os.path.join(dirName, img))
if no_labels:
labels.append(img)
elif type(label_type) is int:
labels.append(label_type)
else:
labels.append(labels2Nums[cervicalType])
if type(numLabels) is int:
return filePaths, labels, numLabels
else:
return filePaths, labels, None
def resizeImage(imgPath, maxSize=(256,256,3), savePath=None, addFlip=False):
ImageFile.LOAD_TRUNCATED_IMAGES = True
img = Image.open(imgPath)
if type(img) == type(np.array([])):
img = Image.fromarray(img)
img.thumbnail(maxSize, Image.ANTIALIAS)
tmpImage = (np.random.random(maxSize)*255).astype(np.uint8)
resizedImg = Image.fromarray(tmpImage)
resizedImg.paste(img,((maxSize[0]-img.size[0])//2, (maxSize[1]-img.size[1])//2))
if savePath:
resizedImg.save(savePath)
if addFlip:
flip = resizedImg.transpose(Image.FLIP_LEFT_RIGHT)
if savePath:
splitPath = savePath.split('/')
flip_path = '/'.join(splitPath[:-1] + ['flipped_'+splitPath[-1]])
flip.save(flip_path)
return np.array(resizedImg, dtype=np.float32), np.array(flip,dtype=np.float32)
return np.array(resizedImg, dtype=np.float32)
def processCervicalData():
# image resizing
imgPaths = []
labels = []
trainingDirs = ['/deeplearning-keras/ch05/data/train']
for dir in trainingDirs:
newFilePaths, newLabels, numLabels = readFilePaths(dir)
if len(newFilePaths) > 0:
imgPaths += newFilePaths
labels += newLabels
imgPaths, labels = shuffle(imgPaths, labels)
labelCount = labelsCount(labels)
type1Count = labelCount[0]
type2Count = labelCount[1]
type3Count = labelCount[2]
print("Count of type1 : ", type1Count)
print("Count of type2 : ", type2Count)
print("Count of type3 : ", type3Count)
print("Total Number of data samples: " + str(len(imgPaths)))
print("Number of Classes: " + str(numLabels))
newShape = [(256,256,3)]
destDir = ['/deeplearning-keras/ch05/data/resized_imgs']
for newImgShape, destFolder in zip(newShape,destDir):
for i, path,label in zip(count(),imgPaths,labels):
split_path = path.split('/')
newPath = 'size'+str(newImgShape[0])+'_'+split_path[-1]
newPath = '/'.join([destFolder]+split_path[8:-1]+[newPath])
add_flip = True
if label == 1:
add_flip = False
# Used to exclude corrupt data
try:
resizeImage(path, maxSize=newImgShape, savePath=newPath, addFlip=add_flip)
except OSError:
print("Error at path " + path)
'''
************************************************Cervical Training*************************************************
'''
def getSteps(n_samples,batch_size,n_augs=1):
n_samples = n_samples*(n_augs+1)
steps_per_epoch = n_samples//batch_size + 1
if n_samples % batch_size == 0:
train_steps_per_epoch = n_samples//batch_size
return steps_per_epoch
# 独热代码
def oneHotEncode(labels, n_classes):
one_hots = []
for label in labels:
one_hot = [0]*n_classes
if label >= len(one_hot):
print("Labels out of bounds\nCheck your n_classes parameter")
return
one_hot[label] = 1
one_hots.append(one_hot)
return np.array(one_hots,dtype=np.float32)
def getSplitData(csv_file_path):
paths = []
labels = []
with open(csv_file_path, 'r') as f:
for line in f:
split_line = line.strip().split(',')
paths.append(split_line[0])
labels.append(int(split_line[1]))
return paths,labels
def savePaths(csv_file_path, paths, labels):
with open(csv_file_path, 'w') as csv_file:
for path,label in zip(paths,labels):
csv_file.write(path + ',' + str(label) + '\n')
def image_generator(file_paths, labels, batch_size, resize_dims=None, randomly_augment=False,rand_order=True):
if randomly_augment:
batch_size = int(batch_size/2) # maintains batch size despite image additions
aug_paths = file_paths
aug_labels = labels
else:
aug_paths, aug_labels = [], []
while True:
if rand_order:
file_paths,labels = shuffle(file_paths,labels)
aug_paths, aug_labels = shuffle(aug_paths, aug_labels)
for batch in range(0, len(file_paths), batch_size):
rpaths = []
rlabels = []
if randomly_augment:
rpaths = aug_paths[batch:batch+batch_size]
rlabels = aug_labels[batch:batch+batch_size]
images, batch_labels = convertImages(file_paths[batch:batch+batch_size],
labels[batch:batch+batch_size],
resize_dims=resize_dims,
rpaths=rpaths,
rlabels=rlabels)
yield images, batch_labels
def rotate(image, angle, ones=None, random_fill=True, color_range=255):
if not random_fill:
return sci.imrotate(image, angle).astype(np.float32)
elif ones == None:
ones = sci.imrotate(np.ones_like(image),angle)
rot_image = sci.imrotate(image, angle).astype(np.float32)
edge_filler = np.random.random(rot_image.shape).astype(np.float32)*color_range
rot_image[ones[:,:,:]!=1] = edge_filler[ones[:,:,:]!=1]
return rot_image
def translate(img, row_amt, col_amt, color_range=255):
translation = np.random.random(img.shape).astype(np.float32)*color_range
if row_amt > 0:
if col_amt > 0:
translation[row_amt:,col_amt:] = img[:-row_amt,:-col_amt]
elif col_amt < 0:
translation[row_amt:,:col_amt] = img[:-row_amt,-col_amt:]
else:
translation[row_amt:,:] = img[:-row_amt,:]
elif row_amt < 0:
if col_amt > 0:
translation[:row_amt,col_amt:] = img[-row_amt:,:-col_amt]
elif col_amt < 0:
translation[:row_amt,:col_amt] = img[-row_amt:,-col_amt:]
else:
translation[:row_amt,:] = img[-row_amt:,:]
else:
if col_amt > 0:
translation[:,col_amt:] = img[:,:-col_amt]
elif col_amt < 0:
translation[:,:col_amt] = img[:,-col_amt:]
else:
return img.copy()
return translation.astype(img.dtype)
def random_zoom(image, max_zoom=1/6., allow_out_zooms=False):
color_range = 255
if allow_out_zooms:
zoom_factor = 1 + (random.random()-.5)*max_zoom*2
else:
zoom_factor = 1 + random.random()*max_zoom
while zoom_factor == 1:
zoom_factor = 1 + (random.random()-0.5)*max_zoom
img_height, img_width = image.shape[:2]
zoomed_h = round(img_height*zoom_factor)
zoomed_w = round(img_width*zoom_factor)
diff_h = abs(zoomed_h-img_height)
diff_w = abs(zoomed_w-img_width)
start_row = round(diff_h/2)
start_col = round(diff_w/2)
if zoom_factor > 1:
end_row = start_row+img_height
end_col = start_col+img_width
zoom_img = scizoom.zoom(image,(zoom_factor,zoom_factor,1),output=np.uint8)[start_row:end_row,
start_col:end_col]
# Zoom out on image
elif zoom_factor < 1:
temp = scizoom.zoom(image,(zoom_factor,zoom_factor,1),output=np.uint8)
temp_height, temp_width = temp.shape[:2]
zoom_img = np.random.random(image.shape)*color_range # Random pixels instead of black space for out zoom
zoom_img[start_row:start_row+temp_height,
start_col:start_col+temp_width] = temp[:,:]
return zoom_img.astype(np.float32)
# random augment - returns a randomly rotated, translated, or scaled copy of an image
def randomAugment(image, rotation_limit=45, shift_limit=10,
zoom_limit=1/3., random_fill=True):
augmentation_type = random.randint(0,4)
# Rotation
if augmentation_type >= 0 and augmentation_type <= 1:
random_angle = random.randint(-rotation_limit,rotation_limit)
while random_angle == 0:
random_angle = random.randint(-rotation_limit,rotation_limit)
aug_image = rotate(image,random_angle,random_fill=random_fill)
# Translation
elif augmentation_type >= 2 and augmentation_type <=3:
row_shift = random.randint(-shift_limit, shift_limit)
col_shift = random.randint(-shift_limit, shift_limit)
aug_image = translate(image,row_shift,col_shift)
# Scale
else:
aug_image = random_zoom(image,max_zoom=zoom_limit)
return aug_image
# convert randoms
def convertRandoms(paths, labels, resize_dims=None, warp_ok=False):
images = []
for i,path in enumerate(paths):
try:
if resize_dims and not warp_ok:
img = resizeImage(path, maxsizes=resize_dims)
else:
img = mpimg.imread(path)
if resize_dims:
img = sci.imresize(img, resize_dims)
img = randomAugment(img)
except OSError:
# Uses augmented version of next image in list
if i == 0:
if resize_dims and not warp_ok:
img = resizeImage(paths[i+1],maxsizes=resize_dims)
else:
img = mpimg.imread(paths[i+1])
if resize_dims:
img = sci.imresize(img, resize_dims)
img = randomAugment(img)
labels[i] = labels[i+1]
# Uses most recent original image
elif i > 0:
img = randomAugment(images[-1])
labels[i] = labels[i-1]
images.append(img)
return images, labels
# convert images
def convertImages(paths, labels, resize_dims=None, warp_ok=False, rpaths=[], randomly_augment=True, rlabels=[]):
if len(rpaths) > 0 and len(rlabels) > 0:
# pool = Pool(processes=1)
# result = pool.apply_async(convert_randoms, (rpaths,rlabels,resize_dims))
rand_imgs, rand_labels = convertRandoms(rpaths,rlabels,resize_dims)
images = []
for i,path in enumerate(paths):
try:
if resize_dims and not warp_ok:
img = resizeImage(path)
else:
img = mpimg.imread(path)
if resize_dims:
img = sci.imresize(img, resize_dims)
except OSError:
# Uses augmented version of next image in list
if i == 0:
if resize_dims and not warp_ok:
img = resizeImage(paths[i+1])
else:
img = mpimg.imread(paths[i+1])
if resize_dims:
img = sci.imresize(img, resize_dims)
img = randomAugment(img)
labels[i] = labels[i+1]
# Uses most recent original image
elif i > 0:
sub_index = -1
if randomly_augment:
sub_index = -2
img = randomAugment(images[sub_index])
labels[i] = labels[i-sub_index]
images.append(img)
if len(rpaths) > 0 and len(rlabels) > 0:
# result.wait()
# rand_imgs, rand_labels = result.get()
images = images+rand_imgs
labels = np.concatenate([labels,rand_labels],axis=0)
return np.array(images,dtype=np.float32), labels
return np.array(images,dtype=np.float32), labels
# cnn model
def convModel(first_conv_shapes=[(4,4),(3,3),(5,5)], conv_shapes=[(3,3),(5,5)], conv_depths=[12,12,11,8,8], dense_shapes=[100,50,3], image_shape=(256,256,3), n_labels=3):
stacks = []
pooling_filter = (2,2)
pooling_stride = (2,2)
inputs = Input(shape=image_shape)
zen_layer = BatchNormalization()(inputs)
for shape in first_conv_shapes:
stacks.append(Conv2D(conv_depths[0], shape, padding='same', activation='elu')(zen_layer))
layer = concatenate(stacks,axis=-1)
layer = BatchNormalization()(layer)
layer = MaxPooling2D(pooling_filter,strides=pooling_stride,padding='same')(layer)
layer = Dropout(0.05)(layer)
for i in range(1,len(conv_depths)):
stacks = []
for shape in conv_shapes:
stacks.append(Conv2D(conv_depths[i],shape,padding='same',activation='elu')(layer))
layer = concatenate(stacks,axis=-1)
layer = BatchNormalization()(layer)
layer = Dropout(i*10**-2+.05)(layer)
layer = MaxPooling2D(pooling_filter,strides=pooling_stride, padding='same')(layer)
layer = Flatten()(layer)
fclayer = Dropout(0.1)(layer)
for i in range(len(dense_shapes)-1):
fclayer = Dense(dense_shapes[i], activation='elu')(fclayer)
fclayer = BatchNormalization()(fclayer)
outs = Dense(dense_shapes[-1], activation='softmax')(fclayer)
return inputs, outs
def max_index(array):
# ** Returns index of maximum value in an array **
max_i = 0
for j in range(1,len(array)):
if array[j] > array[max_i]: max_i = j
return max_i
# get confidence for the predicted class
def confid(predictions,conf):
for i,prediction in enumerate(predictions):
max_i = max_index(prediction)
predictions[i][max_i] = conf
for j in range(len(prediction)):
if j != max_i:
predictions[i][j] = (1-conf)/(len(prediction)-1)
return predictions
# save predictions into csv
def save(csv_file_path, names, predictions, header=None):
with open(csv_file_path, 'w') as f:
f.write(header+'\n')
for name,logit in zip(names,predictions):
f.write(name+',')
for i,element in enumerate(logit):
if i == logit.shape[0]-1: f.write(str(element)+'\n')
else: f.write(str(element)+',')
# train a cnn model
def cervicalTraining():
resizedImageDir = ['/Users/manpreet.singh/git/deeplearning/deeplearning-keras/ch05/data/resized_imgs/train']
imagePaths = []
labels = []
for i, resizedPath in enumerate(resizedImageDir):
new_paths, new_labels, n_classes = readFilePaths(resizedPath)
if len(new_paths) > 0:
imagePaths += new_paths
labels += new_labels
imagePaths, labels = shuffle(imagePaths, labels)
trainCSV = '/Users/manpreet.singh/git/deeplearning/deeplearning-keras/ch05/csvs/train_set.csv'
validCSV = '/Users/manpreet.singh/git/deeplearning/deeplearning-keras/ch05/csvs/valid_set.csv'
training_portion = .8
split_index = int(training_portion * len(imagePaths))
X_train_paths, y_train = imagePaths[:split_index], labels[:split_index]
X_valid_paths, y_valid = imagePaths[split_index:], labels[split_index:]
print("Train size: ")
print(len(X_train_paths))
print("Valid size: ")
print(len(X_valid_paths))
savePaths(trainCSV, X_train_paths, y_train)
savePaths(validCSV, X_valid_paths, y_valid)
train_csv = 'csvs/train_set.csv'
valid_csv = 'csvs/valid_set.csv'
X_train_paths, y_train = getSplitData(train_csv)
X_valid_paths, y_valid = getSplitData(valid_csv)
n_classes = max(y_train) + 1
y_train = oneHotEncode(y_train, n_classes)
y_valid = oneHotEncode(y_valid, n_classes)
batch_size = 110
add_random_augmentations = False
resize_dims = None
n_train_samples = len(X_train_paths)
train_steps_per_epoch = getSteps(n_train_samples, batch_size, n_augs=1)
n_valid_samples = len(X_valid_paths)
valid_steps_per_epoch = getSteps(n_valid_samples, batch_size, n_augs=0)
train_generator = image_generator(X_train_paths, y_train, batch_size,
resize_dims=resize_dims,
randomly_augment=add_random_augmentations)
valid_generator = image_generator(X_valid_paths, y_valid, batch_size,
resize_dims=resize_dims, rand_order=False)
'''
modeling
'''
n_classes = 3
image_shape = (256, 256, 3)
first_conv_shapes = [(4, 4), (3, 3), (5, 5)]
conv_shapes = [(3, 3), (5, 5)]
conv_depths = [12, 12, 11, 8, 8]
dense_shapes = [100, 50, n_classes]
inputs, outs = convModel(first_conv_shapes, conv_shapes, conv_depths, dense_shapes, image_shape, n_classes)
model = Model(inputs=inputs, outputs=outs)
learning_rate = .0001
for i in range(20):
if i > 4:
learning_rate = .00001 # Anneals the learning rate
adam_opt = optimizers.Adam(lr=learning_rate)
model.compile(loss='categorical_crossentropy', optimizer=adam_opt, metrics=['accuracy'])
history = model.fit_generator(train_generator, train_steps_per_epoch, epochs=1,
validation_data=valid_generator, validation_steps=valid_steps_per_epoch,
max_queue_size=1)
model.save('/Users/manpreet.singh/git/deeplearning/deeplearning-keras/ch05/models/model.h5')
'''
get predictions
'''
data_path = '/Users/manpreet.singh/git/deeplearning/deeplearning-keras/ch05/data/test'
model_path = '/Users/manpreet.singh/git/deeplearning/deeplearning-keras/ch05/models/model.h5'
resize_dims = (256, 256, 3)
test_divisions = 20 # Used for segmenting image evaluation in threading
batch_size = 100 # Batch size used for keras predict function
ins, outs = convModel()
model = Model(inputs=ins, outputs=outs)
model.load_weights(model_path)
test_paths, test_labels, _ = readFilePaths(data_path, no_labels=True)
print(str(len(test_paths)) + ' testing images')
pool = ThreadPool(processes=1)
portion = len(test_paths) // test_divisions + 1 # Number of images to read in per pool
async_result = pool.apply_async(convertImages, (test_paths[0 * portion:portion * (0 + 1)],
test_labels[0 * portion:portion * (0 + 1)], resize_dims))
total_base_time = time.time()
test_imgs = []
predictions = []
for i in range(1, test_divisions + 1):
base_time = time.time()
print("Begin set " + str(i))
while len(test_imgs) == 0:
test_imgs, _ = async_result.get()
img_holder = test_imgs
test_imgs = []
if i < test_divisions:
async_result = pool.apply_async(convertImages, (test_paths[i * portion:portion * (i + 1)],
test_labels[0 * portion:portion * (0 + 1)],
resize_dims))
predictions.append(model.predict(img_holder, batch_size=batch_size, verbose=0))
print("Execution Time: " + str((time.time() - base_time) / 60) + 'min\n')
predictions = np.concatenate(predictions, axis=0)
print("Total Execution Time: " + str((time.time() - total_base_time) / 60) + 'mins')
conf = .95 # Prediction confidence
savePredictions = '/Users/manpreet.singh/git/deeplearning/deeplearning-keras/ch05/predictions.csv'
predictions = confid(predictions, conf)
header = 'image_name,Type_1,Type_2,Type_3'
save(savePredictions, test_labels, predictions, header)
if __name__ == '__main__':
processCervicalData()
cervicalTraining()
2092
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
