阿里天池2018广东工业智造图像比赛
该笔记是我们小队，关于阿里天池2018广东工业智造图像比赛的解决方案。首次参加天池比赛，这里记录一下自己的处理过程。模型是VGG19 + group normalization 这里是比赛链接。目前开源到github：
https://github.com/Qinxianshen/tianchi_2018_guangdong_image
- 分析题意，确定目标
- 数据预处理：包括图片读取，数据增强，划分训练集、测试集
- 选择模型训练
- 调参
1.分析题意，确定目标
将大赛给的四百多张铝型材图片，根据所损坏的程度分成12类。确定为图形分类问题。
| 瑕疵名 | 提交结果 |
|---|---|
| 正常 | norm |
| 不导电 | defect1 |
| 擦花 | defect2 |
| 横条压凹 | defect3 |
| 桔皮 | defect4 |
| 漏底 | defect5 |
| 碰伤 | defect6 |
| 起坑 | defect7 |
| 凸粉 | defect8 |
| 涂层开裂 | defect9 |
| 脏点 | defect10 |
| 其他 | defect11 |
label_warp = {'正常': 0,
'不导电': 1,
'擦花': 2,
'横条压凹': 3,
'桔皮': 4,
'漏底': 5,
'碰伤': 6,
'起坑': 7,
'凸粉': 8,
'涂层开裂': 9,
'脏点': 10,
'其他': 11,
}
2.数据预处理：包括图片读取，数据增强，划分训练集、测试集
（1）图片读取
测试集
def read_img(path):
map_path, map_relative = [path +x for x in os.listdir(path) if os.path.isfile(path + x) ], [y for y in os.listdir(path)]
return map_path, map_relative
path_root_test = 'Q:/Github/GuanDong-AL-train/data/guangdong_round1_test_a_20180916/'
# test data
map_path_test, map_relative_test = read_img(path_root_test)
test_file = pd.DataFrame({'img_path': map_path_test})
test_file.to_csv('data/test.csv', index=False)
训练集
#train data
img_path, label = [], []
for first_dir in map_relative_train:
first_path = path_root_train + first_dir + '/'
if '无瑕疵样本' in first_path:
map_path_train_norm, map_relative_train_norm = read_img(first_path)
for temp in map_path_train_norm:
img_path.append(temp)
label.append('正常')
else:
map_path_train_other, map_relative_train_other = read_img(first_path)
for second_dir in map_relative_train_other:
second_path = first_path + second_dir + '/'
if '其他' in second_path:
map_path_train_other_11, map_relative_train_other_11 = read_img(second_path)
for third_dir in map_relative_train_other_11:
if '.DS_Store' == third_dir:
continue
third_path = second_path + third_dir + '/'
map_path_train_other_11_son_img_path, map_relative_train_other_11_son = read_img(third_path)
for temp in map_path_train_other_11_son_img_path:
if temp == 'Q:/Github/GuanDong-AL-train/data/guangdong_round1_train2_20180916/瑕疵样本/其他/粘接/.DS_Store':
continue
img_path.append(temp)
label.append('其他')
else:
map_path_train_other_son_imgs_path, map_relative_train_other_son = read_img(second_path)
for temp in map_path_train_other_son_imgs_path:
img_path.append(temp)
label.append(second_dir)
label_file = pd.DataFrame({'img_path': img_path, 'label': label})
label_file['label'] = label_file['label'].map(label_warp)
label_file.to_csv('data/label.csv', index=False)
（2）数据增强
#对train 数据增强
import random
from PIL import Image
from PIL import ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True
angles = [0,90,180,270]
angles = list(angles)
angles_num = len(angles)
for i in img_path:
index = random.randint(0, angles_num - 1)
im = Image.open(i)
im.rotate(angles[index])
im.save(i)
print(i +" ----- is ok")
（3）划分训练集、测试集
#读取的图片，顺序打乱，划分测试和训练
def read_new_img(img_path,label):
imgs=[]
labels = label
for idx in img_path:
img = io.imread(idx)
img = transform.resize(img, (image_size, image_size))
imgs.append(img)
print(idx + '------------is ok ')
x_data, x_label = np.array(imgs), np.array(def_one_hot(np.array(labels)))
data = []
sigle_data = []
for i in range(len(x_data)):
sigle_data.append(x_data[i])
sigle_data.append(x_label[i])
data.append(sigle_data)
sigle_data = []
# 打乱顺序
data = np.array(data)
num_example = data.shape[0]
np.random.shuffle(data)
# 将所有数据分为训练集和验证集
ratio = 0.8
img_train = []
label_train = []
img_test = []
label_test = []
for i in range(int(ratio*num_example)):
img_train.append(data[i][0])
label_train.append(data[i][1])
img_train = np.array(img_train)
label_train = np.array(label_train)
for i in range(int(ratio*num_example),num_example):
img_test.append(data[i][0])
label_test.append(data[i][1])
img_test = np.array(img_test)
label_test = np.array(label_test)
print(img_train)
return img_train,img_test,label_train,label_test
3.选择模型训练
我们队选择了VGG19模型 唯一不同的是，加入了2018年刚出的group normalization.下面是论文地址：
以下是代码实现：group normalization
def norm(x, norm_type, is_train,i, G=32, esp=1e-5):
with tf.variable_scope('{}_norm_{}'.format(norm_type,i)):
if norm_type == 'none':
output = x
elif norm_type == 'batch':
output = tf.contrib.layers.batch_norm(
x, center=True, scale=True, decay=0.999,
is_training=is_train, updates_collections=None
)
elif norm_type == 'group':
# normalize
# tranpose: [bs, h, w, c] to [bs, c, h, w] following the paper
x = tf.transpose(x, [0, 3, 1, 2])
N, C, H, W = x.get_shape().as_list()
G = min(G, C)
x = tf.reshape(x, [-1, G, C // G, H, W]) # <------------------------------这里源码错了 需要改成这样 https://github.com/shaohua0116/Group-Normalization-Tensorflow/issues/1
mean, var = tf.nn.moments(x, [2, 3, 4], keep_dims=True)
x = (x - mean) / tf.sqrt(var + esp)
# per channel gamma and beta
gamma = tf.get_variable('gamma', [C],
initializer=tf.constant_initializer(1.0))
beta = tf.get_variable('beta', [C],
initializer=tf.constant_initializer(0.0))
gamma = tf.reshape(gamma, [1, C, 1, 1])
beta = tf.reshape(beta, [1, C, 1, 1])
output = tf.reshape(x, [-1, C, H, W]) * gamma + beta # 《------------------------ 这里源码错了 需要改成这样 https://github.com/shaohua0116/Group-Normalization-Tensorflow/issues/1
# tranpose: [bs, c, h, w, c] to [bs, h, w, c] following the paper
output = tf.transpose(output, [0, 2, 3, 1])
else:
raise NotImplementedError
return output
VGG19网络结构
# build_network
W_conv1_1 = tf.get_variable('conv1_1', shape=[3, 3, 3, 64], initializer=tf.contrib.keras.initializers.he_normal())
b_conv1_1 = bias_variable([64])
output = tf.nn.relu(norm(conv2d(x, W_conv1_1) + b_conv1_1,norm_type='group',is_train = True,i='1_1'))
print(output)
W_conv1_2 = tf.get_variable('conv1_2', shape=[3, 3, 64, 64], initializer=tf.contrib.keras.initializers.he_normal())
b_conv1_2 = bias_variable([64])
output = tf.nn.relu(norm(conv2d(output, W_conv1_2) + b_conv1_2,norm_type='group',is_train = True,i='1_2'))
output = max_pool(output, 2, 2, "pool1")
print(output)
W_conv2_1 = tf.get_variable('conv2_1', shape=[3, 3, 64, 128], initializer=tf.contrib.keras.initializers.he_normal())
b_conv2_1 = bias_variable([128])
output = tf.nn.relu(norm(conv2d(output, W_conv2_1) + b_conv2_1,norm_type='group',is_train = True,i='2_1'))
print(output)
W_conv2_2 = tf.get_variable('conv2_2', shape=[3, 3, 128, 128], initializer=tf.contrib.keras.initializers.he_normal())
b_conv2_2 = bias_variable([128])
output = tf.nn.relu(norm(conv2d(output, W_conv2_2) + b_conv2_2,norm_type='group',is_train = True,i='2_2'))
output = max_pool(output, 2, 2, "pool2")
print(output)
W_conv3_1 = tf.get_variable('conv3_1', shape=[3, 3, 128, 256], initializer=tf.contrib.keras.initializers.he_normal())
b_conv3_1 = bias_variable([256])
output = tf.nn.relu( norm(conv2d(output,W_conv3_1) + b_conv3_1,norm_type='group',is_train = True,i='3_1'))
print(output)
W_conv3_2 = tf.get_variable('conv3_2', shape=[3, 3, 256, 256], initializer=tf.contrib.keras.initializers.he_normal())
b_conv3_2 = bias_variable([256])
output = tf.nn.relu(norm(conv2d(output, W_conv3_2) + b_conv3_2,norm_type='group',is_train = True,i='3_2'))
print(output)
W_conv3_3 = tf.get_variable('conv3_3', shape=[3, 3, 256, 256], initializer=tf.contrib.keras.initializers.he_normal())
b_conv3_3 = bias_variable([256])
output = tf.nn.relu( norm(conv2d(output, W_conv3_3) + b_conv3_3,norm_type='group',is_train = True,i='3_3'))
print(output)
W_conv3_4 = tf.get_variable('conv3_4', shape=[3, 3, 256, 256], initializer=tf.contrib.keras.initializers.he_normal())
b_conv3_4 = bias_variable([256])
output = tf.nn.relu(norm(conv2d(output, W_conv3_4) + b_conv3_4,norm_type='group',is_train = True,i='3_4'))
output = max_pool(output, 2, 2, "pool3")
print(output)
W_conv4_1 = tf.get_variable('conv4_1', shape=[3, 3, 256, 512], initializer=tf.contrib.keras.initializers.he_normal())
b_conv4_1 = bias_variable([512])
output = tf.nn.relu(norm(conv2d(output, W_conv4_1) + b_conv4_1,norm_type='group',is_train = True,i='4_1'))
print(output)
W_conv4_2 = tf.get_variable('conv4_2', shape=[3, 3, 512, 512], initializer=tf.contrib.keras.initializers.he_normal())
b_conv4_2 = bias_variable([512])
output = tf.nn.relu(norm(conv2d(output, W_conv4_2) + b_conv4_2,norm_type='group',is_train = True,i='4_2'))
print(output)
W_conv4_3 = tf.get_variable('conv4_3', shape=[3, 3, 512, 512], initializer=tf.contrib.keras.initializers.he_normal())
b_conv4_3 = bias_variable([512])
output = tf.nn.relu(norm(conv2d(output, W_conv4_3) + b_conv4_3,norm_type='group',is_train = True,i='4_3'))
print(output)
W_conv4_4 = tf.get_variable('conv4_4', shape=[3, 3, 512, 512], initializer=tf.contrib.keras.initializers.he_normal())
b_conv4_4 = bias_variable([512])
output = tf.nn.relu(norm(conv2d(output, W_conv4_4) + b_conv4_4,norm_type='group',is_train = True,i='4_4'))
output = max_pool(output, 2, 2)
print(output)
W_conv5_1 = tf.get_variable('conv5_1', shape=[3, 3, 512, 512], initializer=tf.contrib.keras.initializers.he_normal())
b_conv5_1 = bias_variable([512])
output = tf.nn.relu(norm(conv2d(output, W_conv5_1) + b_conv5_1,norm_type='group',is_train = True,i='5_1'))
print(output)
W_conv5_2 = tf.get_variable('conv5_2', shape=[3, 3, 512, 512], initializer=tf.contrib.keras.initializers.he_normal())
b_conv5_2 = bias_variable([512])
output = tf.nn.relu(norm(conv2d(output, W_conv5_2) + b_conv5_2,norm_type='group',is_train = True,i='5_2'))
print(output)
W_conv5_3 = tf.get_variable('conv5_3', shape=[3, 3, 512, 512], initializer=tf.contrib.keras.initializers.he_normal())
b_conv5_3 = bias_variable([512])
output = tf.nn.relu(norm(conv2d(output, W_conv5_3) + b_conv5_3,norm_type='group',is_train = True,i='5_3'))
print(output)
W_conv5_4 = tf.get_variable('conv5_4', shape=[3, 3, 512, 512], initializer=tf.contrib.keras.initializers.he_normal())
b_conv5_4 = bias_variable([512])
output = tf.nn.relu(norm(conv2d(output, W_conv5_4) + b_conv5_4,norm_type='group',is_train = True,i='5_4'))
print(output)
# output = tf.contrib.layers.flatten(output)
output = tf.reshape(output, [-1, 2*2*512])
print(output)
W_fc1 = tf.get_variable('fc1', shape=[2048, 4096], initializer=tf.contrib.keras.initializers.he_normal())
b_fc1 = bias_variable([4096])
output = tf.nn.relu(norm(tf.matmul(output, W_fc1) + b_fc1,norm_type='batch',is_train = True,i='fc1') )
output = tf.nn.dropout(output, keep_prob)
print(output)
W_fc2 = tf.get_variable('fc7', shape=[4096, 4096], initializer=tf.contrib.keras.initializers.he_normal())
b_fc2 = bias_variable([4096])
output = tf.nn.relu(norm(tf.matmul(output, W_fc2) + b_fc2,norm_type='batch',is_train = True,i='fc2'))
output = tf.nn.dropout(output, keep_prob)
print(output)
W_fc3 = tf.get_variable('fc3', shape=[4096, 12], initializer=tf.contrib.keras.initializers.he_normal())
b_fc3 = bias_variable([12])
output = tf.nn.relu(norm(tf.matmul(output, W_fc3) + b_fc3,norm_type='batch',is_train = True,i='fc3'))
print(output)
# output = tf.reshape(output,[-1,10])
# loss function: cross_entropy
# train_step: training operation
cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=output))
l2 = tf.add_n([tf.nn.l2_loss(var) for var in tf.trainable_variables()])
train_step = tf.train.MomentumOptimizer(learning_rate, momentum_rate, use_nesterov=True).\
minimize(cross_entropy + l2 * weight_decay)
correct_prediction = tf.equal(tf.argmax(output, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
4.调参
主要涉及的参数：
#全局one-hot编码空间 label_binarizer = "" #初始的定义的数据 class_num = 12 image_size = 32 img_channels = 3 iterations = 72 batch_size = 24 total_epoch = 100 weight_decay = 0.0003 dropout_rate = 0.5 momentum_rate = 0.9 log_save_path = './vgg_19_logs' model_save_path = './model/' batch_size_test = 4
目前比赛还在进行中，调参有待提高
5.目前成绩：
当日排名175 综合排名602（截至10.4日一共2950支队伍，本科生第一次参加，第一次提交，轻喷）
关于我
Github:https://github.com/Qinxianshen
CSDN: https://blog.csdn.net/Qin_xian_shen
个人博客: http://saijiadexiaoqin.cn/
Gitchat:https://gitbook.cn/gitchat/author/59ef0b02a276fd1a69094634
哔哩哔哩：https://space.bilibili.com/126021651/#/
微信公众号：松爱家的小秦
更多LIVE：