玩命加载中...
<** | 修改校对：SofaSofa TeamM |
<hr>
### 1. 加载数据
```python
# --- 参数设置 ---
# 手动查看测试集中的图像，并且增加了五个异形，放入训练集
num_anno=5
epochs=100
SEED=0
# 考虑到异性并不多，所以设置如下权重，来解决非平衡下的分类
class_weight={0:1.,1:1.,2:10.}
```
```python
# 加载训练集
import numpy as np
import matplotlib.pyplot as plt
np.random.seed(0)
train_data_raw = np.genfromtxt('train.csv', delimiter=',')
train_data=train_data_raw[1:,1:]
[n_train,d]=train_data.shape
y_train=train_data[:,d-1]
x_train=train_data[:,:-1]
[n_train,d]=x_train.shape
sz=40
x_train=x_train.reshape([n_train,sz,sz])
# 加载测试集
test_data_raw = np.genfromtxt('test.csv', delimiter=',')
test_data=test_data_raw[1:,:]
x_test=test_data[:,1:]
n_test=x_test.shape[0]
x_test=x_test.reshape([n_test,sz,sz])
```
## 2. 数据处理
数据处理中我们先对图像进行预处理，步骤包括
- 利用中值滤波(median filter)进行降噪
- 利用阈值分割法(threshold segmentation)生成掩膜(binary mask)
- 利用形态闭合(morphology closing)来填充图中的小洞
```python
from skimage.restoration import (denoise_tv_chambolle, denoise_bilateral,
denoise_wavelet, estimate_sigma)
from scipy import ndimage, misc
import statistics
import scipy
def my_preprocessing(I,show_fig=False):
I_median=ndimage.median_filter(I, size=5)
mask=(I_median<statistics.mode(I_median.flatten()));
I_out=scipy.ndimage.morphology.binary_closing(mask,iterations=2)
if(np.mean(I_out[15:25,15:25].flatten())<0.5):
I_out=1-I_out
if show_fig:
fig= plt.figure(figsize=(8, 4))
plt.gray()
plt.subplot(2,4,1)
plt.imshow(I)
plt.axis('off')
plt.title('Image')
plt.subplot(2,4,2)
plt.imshow(I_median)
plt.axis('off')
plt.title('Median filter')
plt.subplot(2,4,3)
plt.imshow(mask)
plt.axis('off')
plt.title('Mask')
plt.subplot(2,4,4)
plt.imshow(I_out)
plt.axis('off')
plt.title('Closed mask')
fig.tight_layout()
plt.show()
return I_out
I_out=my_preprocessing(x_test[5],True);
x_train_prc=np.zeros_like(x_train)
x_test_prc=np.zeros_like(x_test)
for i in range(n_train):
x_train_prc[i]=my_preprocessing(x_train[i])
for i in range(n_test):
x_test_prc[i]=my_preprocessing(x_test[i])
```
此外，我们还人工从测试集中挑选了训练集中并未出现过的标签的五个样本，加入训练集中。
```python
from sklearn.preprocessing import MultiLabelBinarizer
from sklearn.utils import shuffle
# 五个异形在原数据中对应的下标
anno_idx=np.array([4949, 4956, 4973, 4974, 4988])
anno_idx=anno_idx[::-1]
anno_idx_add=anno_idx[:num_anno]
x_train_prc_anno=x_train_prc
y_train_anno=y_train.reshape([-1,1])
x_add=x_test_prc[anno_idx_add]
y_add=np.ones([num_anno,1])*2
# 对异形进行过采样
for i in range(4000/num_anno):
x_train_prc_anno=np.append(x_train_prc_anno,x_add,axis=0)
y_train_anno=np.append(y_train_anno,y_add,axis=0)
x_train_prc_anno,y_train_anno = shuffle(x_train_prc_anno, y_train_anno, random_state=0)
mlb1=MultiLabelBinarizer()
y_train_mlb=mlb1.fit_transform(y_train_anno)
```
### 3. 训练卷积神经网络
卷积神经网络(CNN)一般包括卷积层、池化层、全连接层。有时为了防止过拟合，我们也会加入dropout层。
在数据样本量较小时，我们也会进行数据增广。
```python
from keras.callbacks import TensorBoard
from keras.layers import Dense, Dropout, MaxPooling2D, Flatten, Convolution2D
from keras.models import Sequential
from keras.optimizers import Adam
from keras import backend as K
from keras.preprocessing.image import ImageDataGenerator
import tensorflow as tf
tf.random.set_random_seed(SEED)
session_conf = tf.ConfigProto(intra_op_parallelism_threads=1, inter_op_parallelism_threads=1)
sess = tf.Session(graph=tf.get_default_graph(), config=session_conf)
K.set_session(sess)
def built_model():
n_filter=32;
model = Sequential()
model.add(Convolution2D(filters=n_filter,
kernel_size=(5, 5),
input_shape=(40, 40, 1),
activation='relu'))
model.add(Convolution2D(filters=n_filter,
kernel_size=(5,5),
activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Convolution2D(filters=n_filter,
kernel_size=(5,5),
activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Convolution2D(filters=n_filter,
kernel_size=(3, 3),
activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.5))
model.add(Flatten())
model.add(Dense(units=128,
activation='relu'))
model.add(Dense(3, activation='softmax')) # Final Layer using Softmax
model.compile(loss='categorical_crossentropy',
optimizer=Adam(lr=0.0003),
metrics=['accuracy'])
model.summary()
return model
def train_model(x_train,y_train, x_test, batch_size=64, epochs=20, model=None,
class_weight={0:1.,1:1.,2:10.}):
if np.ndim(x_train)<4:
x_train=np.expand_dims(x_train,3)
x_test=np.expand_dims(x_test,3)
if model is None:
model = built_model()
datagen = ImageDataGenerator(
rotation_range=180,
width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True
)
# 训练模型的同时进行数据增广
history=model.fit_generator(datagen.flow(x_train, y_train, batch_size=batch_size),
steps_per_epoch=len(x_train) / batch_size, epochs=epochs,
class_weight=class_weight,
validation_data=datagen.flow(x_train, y_train, batch_size=batch_size),
validation_steps=1)
print("Loss on training and testing.")
plt.plot(history.history['loss'], label='train')
plt.plot(history.history['val_loss'], label='valid')
plt.legend()
plt.show()
pred_prob_train = model.predict(x_train, batch_size=batch_size, verbose=1)
pred_train = np.array(pred_prob_train > 0.5).astype(int)
pred_prob_test = model.predict(x_test, batch_size=batch_size, verbose=1)
pred_test = np.array(pred_prob_test > 0.5).astype(int)
y_test_hat=pred_test[:,1]+pred_test[:,2]*2;
y_train_hat=pred_train[:,1]+pred_train[:,2]*2;
return y_train_hat,y_test_hat,history
```
Using TensorFlow backend.
```python
y_train_hat,y_test_hat,history=train_model(x_train_prc_anno, y_train_mlb,x_test_prc,
epochs=epochs,batch_size=64,class_weight=class_weight)
```
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_1 (Conv2D) (None, 36, 36, 32) 832
_________________________________________________________________
conv2d_2 (Conv2D) (None, 32, 32, 32) 25632
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 16, 16, 32) 0
_________________________________________________________________
conv2d_3 (Conv2D) (None, 12, 12, 32) 25632
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 6, 6, 32) 0
_________________________________________________________________
conv2d_4 (Conv2D) (None, 4, 4, 32) 9248
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 (None, 2, 2, 32) 0
_________________________________________________________________
dropout_1 (Dropout) (None, 2, 2, 32) 0
_________________________________________________________________
flatten_1 (Flatten) (None, 128) 0
_________________________________________________________________
dense_1 (Dense) (None, 128) 16512
_________________________________________________________________
dense_2 (Dense) (None, 3) 387
=================================================================
Total params: 78,243
Trainable params: 78,243
Non-trainable params: 0
_________________________________________________________________
Epoch 1/100
156/156 [==============================] - 4s 26ms/step - loss: 2.2816 - acc: 0.3979 - val_loss: 1.1561 - val_acc: 0.4688
Epoch 2/100
156/156 [==============================] - 2s 15ms/step - loss: 1.4215 - acc: 0.5093 - val_loss: 0.7161 - val_acc: 0.6562
Epoch 3/100
156/156 [==============================] - 2s 14ms/step - loss: 0.7764 - acc: 0.7068 - val_loss: 0.3253 - val_acc: 0.8594
Epoch 4/100
156/156 [==============================] - 2s 13ms/step - loss: 0.5159 - acc: 0.8224 - val_loss: 0.4067 - val_acc: 0.8281
Epoch 5/100
156/156 [==============================] - 2s 13ms/step - loss: 0.3667 - acc: 0.8801 - val_loss: 0.0952 - val_acc: 0.9688
... ...
Epoch 96/100
156/156 [==============================] - 2s 13ms/step - loss: 0.0201 - acc: 0.9949 - val_loss: 9.9051e-05 - val_acc: 1.0000
Epoch 97/100
156/156 [==============================] - 2s 13ms/step - loss: 0.0207 - acc: 0.9941 - val_loss: 0.0071 - val_acc: 1.0000
Epoch 98/100
156/156 [==============================] - 2s 13ms/step - loss: 0.0263 - acc: 0.9945 - val_loss: 6.1150e-04 - val_acc: 1.0000
Epoch 99/100
156/156 [==============================] - 2s 15ms/step - loss: 0.0200 - acc: 0.9945 - val_loss: 0.0016 - val_acc: 1.0000
Epoch 100/100
156/156 [==============================] - 3s 17ms/step - loss: 0.0300 - acc: 0.9934 - val_loss: 0.0015 - val_acc: 1.0000
Loss on training and testing.
10000/10000 [==============================] - 0s 46us/step
5191/5191 [==============================] - 0s 47us/step
### 5. 提交结果
最后的结果可以提交到<a href="http://sofasofa.io/competition.php?id=9#c4" target="_blank">这里</a>，来看看精度如何哦！
```python
import pandas as pd
submit = pd.read_csv('sample_submit.csv')
submit['y'] = y_test_hat
submit.to_csv('my_CNN_prediction.csv', index=False)
```
<ul class="pager">
<li class="next"><a href="../../tutorials.php"><b><i class="fa fa-graduation-cap" aria-hidden="true"></i> 学完咯！</b></a></li>
</ul>
]]>
