In [1]:
from keras.models import Sequential
from keras.layers.core import Flatten, Dense, Dropout
from keras.layers.convolutional import Convolution2D, MaxPooling2D, ZeroPadding2D
from keras.optimizers import SGD
import cv2, numpy as np
import pandas as pd
In [2]:
''' basic package '''
import os
# 告訴系統要第幾張卡被看到。 Ex. 硬體總共有8張顯卡，以下設定只讓系統看到第1張顯卡
# 若沒設定，則 Tensorflow 在運行時，預設會把所有卡都佔用
# 要看裝置內顯卡數量及目前狀態的話，請在終端機內輸入 "nvidia-smi"
# 若你的裝置只有一張顯卡可以使用，可以忽略此設定
os.environ["CUDA_VISIBLE_DEVICES"] = '0'
In [3]:
train_path = "/data/examples/may_the_4_be_with_u/where_am_i/train"
test_path = "/data/examples/may_the_4_be_with_u/where_am_i/testset"
In [4]:
img_cat = pd.read_csv("/data/examples/may_the_4_be_with_u/where_am_i/mid_term_mapping.txt",
encoding="utf-8", header=None)
img_cat.columns = ["dirs","class"]
img_cat #class: 0~14 共15類
Out[4]:
In [5]:
import os
train_img_files = pd.DataFrame(columns=["dirs","files","class"])
for index, row in img_cat.iterrows():
#print(row['dirs'], row['class'])
for _root, _dirs, _files in os.walk(train_path+"/"+row['dirs']):
temp = pd.DataFrame({ 'dirs':row['dirs'], 'files': _files, 'class':row['class']})
train_img_files = pd.concat([train_img_files, temp], axis=0)
train_img_files.head(3)
Out[5]:
In [6]:
len(train_img_files)
Out[6]:
In [7]:
train_img_files = train_img_files.reset_index(drop=True)
In [8]:
import os
test_img_files = []
for _root, _dirs, _files in os.walk(test_path):
test_img_files = _files
test_img_files[0:3]
Out[8]:
In [9]:
len(test_img_files)
Out[9]:
In [10]:
from keras.preprocessing import image
IMG_WIDTH = 224
IMG_HEIGHT = 224
IMG_CHANNELS = 3
ttl_train_imgs = len(train_img_files)
ttl_test_imgs = len(test_img_files)
train_X = np.zeros((ttl_train_imgs, IMG_HEIGHT, IMG_WIDTH, IMG_CHANNELS), dtype=np.uint8)
test_X = np.zeros((ttl_test_imgs, IMG_HEIGHT, IMG_WIDTH, IMG_CHANNELS), dtype=np.uint8)
In [11]:
from keras.applications.vgg19 import preprocess_input
for index, row in train_img_files.iterrows():
img = image.load_img(train_path+"/"+row['dirs']+"/"+row['files'], target_size=(IMG_WIDTH, IMG_HEIGHT)) # for vgg input size: 224 x 224
temp = image.img_to_array(img)
temp = np.expand_dims(temp, axis=0)
train_X[index] = preprocess_input(temp)
img.close()
len(train_X)
Out[11]:
In [12]:
for index in range(len(test_img_files)):
img = image.load_img(test_path+"/"+test_img_files[index], target_size=(IMG_WIDTH, IMG_HEIGHT)) # for vgg input size: 224 x 224
temp = image.img_to_array(img)
temp = np.expand_dims(temp, axis=0)
test_X[index] = preprocess_input(temp)
img.close()
len(test_X)
Out[12]:
In [13]:
import gc
del temp
gc.collect()
Out[13]:
In [14]:
train_X = train_X.astype('float32') / 255.0
test_X = test_X.astype('float32') / 255.0
In [15]:
train_X[:,:,0] = (train_X[:,:,0]-train_X[:,:,0].mean())/train_X[:,:,0].std()
train_X[:,:,1] = (train_X[:,:,1]-train_X[:,:,1].mean())/train_X[:,:,1].std()
train_X[:,:,2] = (train_X[:,:,2]-train_X[:,:,2].mean())/train_X[:,:,2].std()
test_X[:,:,0] = (test_X[:,:,0]-test_X[:,:,0].mean())/test_X[:,:,0].std()
test_X[:,:,1] = (test_X[:,:,1]-test_X[:,:,1].mean())/test_X[:,:,1].std()
test_X[:,:,2] = (test_X[:,:,2]-test_X[:,:,2].mean())/test_X[:,:,2].std()
In [16]:
X_img_train_normalize = train_X
X_img_test_normalize = test_X
In [17]:
from keras.utils import np_utils
y_label_train_OneHot = np_utils.to_categorical(train_img_files["class"].astype('int'))
y_label_train_OneHot.shape
Out[17]:
In [19]:
from keras.applications.vgg19 import VGG19
from keras.preprocessing import image
from keras.applications.vgg19 import preprocess_input
from keras.layers import Input, Flatten, Dense, Dropout
from keras.models import Model
import numpy as np
In [20]:
model_vgg19_conv = VGG19(weights='imagenet', include_top=False, classes=15)
model_vgg19_conv.summary()
In [21]:
#Create your own input format (here 224x224x3)
input = Input(shape=(IMG_HEIGHT, IMG_WIDTH, IMG_CHANNELS),name = 'image_input')
In [22]:
#Use the generated model
output_vgg19_conv = model_vgg19_conv(input)
In [23]:
#Add the fully-connected layers
x = Flatten(name='flatten')(output_vgg19_conv)
x = Dense(512, activation='relu', name='fc1')(x)
x = Dense(256, activation='relu', name='fc2')(x)
x = Dense(15, activation='softmax', name='predictions')(x)
In [24]:
#Create your own model
my_model = Model(input=input, output=x)
In [25]:
my_model.summary()
In [26]:
from keras.models import optimizers
#, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0
opt_adam = optimizers.Adam( lr=0.001, beta_1=0.9, beta_2=0.99999, decay=0.0001 )
In [27]:
my_model.compile(loss='categorical_crossentropy', optimizer= opt_adam, metrics=['accuracy'])
In [28]:
from keras.preprocessing.image import ImageDataGenerator
print('Using real-time data augmentation.')
# This will do preprocessing and realtime data augmentation:
datagen = ImageDataGenerator(
rotation_range=30, # randomly rotate images in the range (degrees, 0 to 180)
width_shift_range=0.1, # randomly shift images horizontally (fraction of total width)
height_shift_range=0.1, # randomly shift images vertically (fraction of total height)
horizontal_flip=True, # randomly flip images
vertical_flip=True) # randomly flip images
In [29]:
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split( X_img_train_normalize, y_label_train_OneHot,
test_size = 0.3, random_state=100)
In [30]:
from keras.callbacks import EarlyStopping
# earlystop
earlystop = EarlyStopping(monitor='val_acc', patience=50, verbose=0)
# Fit the model on the batches generated by datagen.flow().
train_history = my_model.fit_generator(
datagen.flow(X_train, y_train, batch_size=32),
epochs=200, workers=-1,
validation_data=(X_test, y_test),
callbacks=[earlystop]
)
In [31]:
import matplotlib.pyplot as plt
def show_train_history(train_token,test_token):
plt.plot(train_history.history[train_token])
plt.plot(train_history.history[test_token])
plt.title('Train History')
#plt.ylabel('Accuracy')
plt.ylabel(train_token)
plt.xlabel('Epoch')
plt.legend([train_token, test_token], loc='upper left')
plt.show()
In [32]:
show_train_history('loss','val_loss')
In [33]:
show_train_history('acc','val_acc')
In [34]:
scores = my_model.evaluate(X_img_train_normalize, y_label_train_OneHot, verbose=0)
scores
Out[34]:
In [35]:
pred_train=my_model.predict(X_img_train_normalize)
In [36]:
prediction_train = np.empty(len(pred_train))
for i in range(pred_train.shape[0]):
prediction_train[i] = np.argmax(pred_train[i])
In [37]:
train_img_files["class"].astype('int').reshape(-1)
Out[37]:
In [38]:
import pandas as pd
pd.crosstab(train_img_files["class"].astype('int').reshape(-1),prediction_train,
rownames=['class'],colnames=['predict'])
Out[38]:
In [39]:
pred=my_model.predict(X_img_test_normalize)
In [40]:
pred[0]
Out[40]:
In [41]:
prediction = np.empty(len(pred))
for i in range(pred.shape[0]):
prediction[i] = np.argmax(pred[i])
In [42]:
prediction[0]
Out[42]:
In [43]:
#test_img_files
In [44]:
img_sub = pd.read_csv("/data/examples/may_the_4_be_with_u/where_am_i/img-submission.csv", encoding="utf-8")
img_sub.head(2)
Out[44]:
In [45]:
Submission = pd.DataFrame({ 'id': test_img_files, 'class': prediction }, columns=["id","class"])
Submission["class"] = Submission["class"].astype('int')
Submission.head(2)
Out[45]:
In [46]:
img_sub[img_sub["id"] == "3d766cf9e984b45cd7b6fa0550aaf2c33e1cc881c5036f907891cf76e0fea790"]
Out[46]:
In [47]:
Submission[Submission["id"] == "3d766cf9e984b45cd7b6fa0550aaf2c33e1cc881c5036f907891cf76e0fea790.jpg"]
Out[47]:
In [48]:
Submission["id"] = Submission["id"].apply(lambda x: x.replace(".jpg", ""))
In [49]:
Submission[Submission["id"] == "3d766cf9e984b45cd7b6fa0550aaf2c33e1cc881c5036f907891cf76e0fea790"]
Out[49]:
In [50]:
_test = img_sub.merge(Submission, left_on='id', right_on='id', how='outer')
In [51]:
_test.head(3)
Out[51]:
In [52]:
# Generate Submission File:
Submission[["id","class"]] = _test[["id","class_y"]]
Submission.to_csv("submission_vgg_e200_std.csv", index=False)
儲存模型架構及權重值
In [54]:
my_model.save_weights("SaveModel/VGG_Model.h5")
print("Saved model to disk")
In [55]:
model_yaml = my_model.to_yaml()
with open("SaveModel/VGG_Model.yaml", "w") as yaml_file:
yaml_file.write(model_yaml)
In [56]:
model_json = my_model.to_json()
with open("SaveModel/VGG_Model.json", "w") as json_file:
json_file.write(model_json)
In [ ]: