我使用 keras - vgg16，經過多次調整，最後上傳 Kaggle 準確率為 0.8744，DATA可以從 https://www.kaggle.com/c/cifar-10/data 下載， 結果不算好，如果有更好的方法可以提供給我，在這先說聲謝謝
MY GPU is GTX-1070
首先，下圖是其中一張的 image
馬的圖片，大小為 32x32x3，我們將訓練 vgg 16 model ，學會辨識
import numpy as np
np.random.seed(100)
import os, cv2, random
import pandas as pd
%matplotlib inline
import matplotlib.pyplot as plt
import keras.utils.np_utils as kutils
import re
from keras.models import Sequential
from keras.layers import Input, Dropout, Flatten, Convolution2D, MaxPooling2D, ZeroPadding2D, Dense, Activation, Conv2DTranspose
from keras.layers import Conv2D
import keras
import datetime
import math
import gcos.chdir('/home/linsam/kaggle/cifar_10') # 設定資料夾位置
train_labels = pd.read_csv('trainLabels.csv')# 讀進 labels
# 抓取 train 資料夾中所有檔案名稱
train_image_path = ['train/'+i for i in os.listdir('train/')]
# 抓取 test 資料夾中所有檔案名稱
test_image_path = ['test/'+i for i in os.listdir('test/')]
# 印出 image 數量
print('train_image_amount : ' + str( len(train_image_path) ) )
print('test_image_amount : ' + str( len(test_image_path) ) )# input train data
train_image,train_labels_int,train_labels_class = (
input_train_data(train_image_path,train_labels) )
# input test data
test_image,index = input_test_data(test_image_path)print( train_image.shape )
n = int( len(train_image)*1 )
sub_train_x = train_image[:n]
sub_train_y = train_labels_int[:n]
sub_test_x = train_image[n:]
sub_test_y = train_labels_int[n:]# function 雖然有寫 cnn two/three layer, 但是結果沒有 vgg 好
#====================== step 6 : build vgg16 model =============================
from keras.applications.vgg16 import VGG16
from keras.applications.vgg19 import VGG19
from keras.models import Model
from keras.optimizers import SGD# VGG16 可以換成 VGG19, 但是不保證較好
model_vgg16_conv = VGG16(weights='imagenet',
include_top=False)
#model_vgg16_conv = VGG16(weights = None,
# include_top=False) # set input shape, 32*32*3
# 使用 vgg16 or 其他人的 model, 需要給定自己 image 的 shape
input = Input( shape=(32,32,3),name = 'image_input' )
#Use the generated model
output_vgg16_conv = model_vgg16_conv(input)#Add the fully-connected layers
x = Flatten(name='flatten')(output_vgg16_conv)
x = Dense(10, activation='softmax', name='predictions')(x)
#Create your own model
model = Model(input=input, output=x)sgd = SGD(lr=1e-5, momentum=0.2, nesterov=True)
model.compile(optimizer = sgd, #sgd
loss='categorical_crossentropy',
metrics = ['accuracy'])train_history = model.fit(sub_train_x,# train x ( feature )
sub_train_y,# train y ( label or target )
validation_split = 0.2,# use 20% data to validation
epochs = 10,# run 10 times
batch_size = 128,# 128 data/times
verbose = 1, # print process
shuffle = False)
show_train_history(train_history)train_table,tem11 = compare_corr_per(sub_train_x,sub_train_y)
# 如果沒有進行 cross_validation, 以下將不用執行
#test_table,tem21 = compare_corr_per(sub_test_x,sub_test_y)
model.save_weights('vgg16_temp_alldata.h5')sgd = SGD(lr=1e-4, momentum=0.5, nesterov=True)
model.compile(optimizer = sgd,
loss='categorical_crossentropy',
metrics = ['accuracy'])
train_history = model.fit(sub_train_x,# train x ( feature )
sub_train_y,# train y ( label or target )
validation_split = 0.2,# catch 20% data to validation
epochs = 10,# run 10 times
batch_size = 128,# 128 data/times
verbose = 1, # print process
shuffle = False)
train_table,tem = compare_corr_per(sub_train_x,sub_train_y)# 計算正確率
test_table,tem2 = compare_corr_per(sub_test_x,sub_test_y)# 計算正確率sgd = SGD(lr=1e-3, momentum=0.8, nesterov=True)
model.compile(optimizer = sgd, #sgd
loss='categorical_crossentropy',
metrics = ['accuracy'])
train_history = model.fit(sub_train_x,# train x ( feature )
sub_train_y,# train y ( label or target )
#validation_split = 0.1,# catch 20% data to validation
epochs = 10,# run 50 times
batch_size = 128,# 128 data/times
verbose = 1, # print process
shuffle = False)
train_table,tem = compare_corr_per(sub_train_x,sub_train_y)# 計算正確率
test_table,tem2 = compare_corr_per(sub_test_x,sub_test_y)# 計算正確率model_vgg16_conv = VGG16(weights='imagenet', # 一樣使用 keras 提供的 weight
include_top=False)
model_vgg16_conv.summary()
#_______________________________________________________________
#Layer (type) Output Shape Param #
#=================================================================
#input_20 (InputLayer) (None, None, None, 3) 0
#_______________________________________________________________
#block1_conv1 (Conv2D) (None, None, None, 64) 1792
#_______________________________________________________________
#block1_conv2 (Conv2D) (None, None, None, 64) 36928
#_______________________________________________________________
#block1_pool (MaxPooling2D) (None, None, None, 64) 0
#_______________________________________________________________
#block2_conv1 (Conv2D) (None, None, None, 128) 73856
#_______________________________________________________________
#block2_conv2 (Conv2D) (None, None, None, 128) 147584
#_______________________________________________________________
#block2_pool (MaxPooling2D) (None, None, None, 128) 0
#_______________________________________________________________
#block3_conv1 (Conv2D) (None, None, None, 256) 295168
#_______________________________________________________________
#block3_conv2 (Conv2D) (None, None, None, 256) 590080
#_______________________________________________________________
#block3_conv3 (Conv2D) (None, None, None, 256) 590080
#_______________________________________________________________
#block3_pool (MaxPooling2D) (None, None, None, 256) 0
#_______________________________________________________________
#block4_conv1 (Conv2D) (None, None, None, 512) 1180160
#_______________________________________________________________
#block4_conv2 (Conv2D) (None, None, None, 512) 2359808
#_______________________________________________________________
#block4_conv3 (Conv2D) (None, None, None, 512) 2359808
#_______________________________________________________________
#block4_pool (MaxPooling2D) (None, None, None, 512) 0
#_______________________________________________________________
#block5_conv1 (Conv2D) (None, None, None, 512) 2359808
#_______________________________________________________________
#block5_conv2 (Conv2D) (None, None, None, 512) 2359808
#_______________________________________________________________
#block5_conv3 (Conv2D) (None, None, None, 512) 2359808
#_______________________________________________________________
#block5_pool (MaxPooling2D) (None, None, None, 512) 0
#=================================================================
#Total params: 14,714,688
#Trainable params: 14,714,688
#Non-trainable params: 0input = Input(shape=(32,
32,3),
name = 'image_input')
#Use the generated model
output_vgg16_conv = model_vgg16_conv(input)
#Add the fully-connected layers
x = Flatten()(output_vgg16_conv)
x = Dense(256,activation = 'relu')(x)
x = Dropout(0.5)(x)
x = Dense(10, activation='sigmoid')(x)
#Create your own model
model = Model(input=input, output=x)
model.summary()# look my CNN architecture#_______________________________________________________________
#
#_______________________________________________________________
#Layer (type) Output Shape Param #
#=================================================================
#image_input (InputLayer) (None, 32, 32, 3) 0
#_______________________________________________________________
# *** 這裡是原始 vgg16 結構, 可以由 Param 數量 14714688 得知
#vgg16 (Model) multiple 14714688
#_______________________________________________________________
#flatten_20 (Flatten) (None, 512) 0
#_______________________________________________________________
#dense_37 (Dense) (None, 256) 131328
#_______________________________________________________________
#dropout_18 (Dropout) (None, 256) 0
#_______________________________________________________________
#dense_38 (Dense) (None, 10) 2570
#=================================================================
#Total params: 14,848,586
#Trainable params: 14,848,586
#Non-trainable params: 0
#_______________________________________________________________一樣先進行小 lr training，比較特別的是，這裡 loss 使用 binary_crossentropy，以下提供我 train 好的 weight vgg16_temp_alldata.h5
sgd = SGD(lr=1e-5, momentum=0.9)
model.compile(optimizer = sgd, #sgd
loss='binary_crossentropy',
metrics = ['accuracy'])
# use train_history, that can get model history
# 如果沒有GPU, 可以使用我 train 好的 weight
# model.load_weights('vgg16_temp_alldata.h5')
train_history = model.fit(sub_train_x,# train x ( feature )
sub_train_y,# train y ( label or target )
#validation_split = 0.2,# catch 20% data to validation
# 不進行 validation_split, 由於 data 多, 可以 train 的更好
# 使用 validation_split 可以看出收斂性
epochs = 10,# run 20 times
batch_size = 128,# 128 data/times
verbose = 1, # print process
shuffle = False)
#show_train_history(train_history)# plot training hisrtory
train_table,tem11 = compare_corr_per(sub_train_x,sub_train_y)
# 如果沒有進行 cross_validation, 以下將不用執行
#test_table,tem21 = compare_corr_per(sub_test_x,sub_test_y)
model.save_weights('vgg16_temp_alldata.h5')再次訓練，150次，以下提供我 train 好的 weight vgg16_temp3_alldata.h5，可以藉由 model.load_weights 讀取
model.load_weights('vgg16_temp_alldata.h5')
# 這是並沒有 1e-5 -> 1e-4 -> 1e-3, 直接 1e-5 -> 1e-3, loss 也改為 categorical_crossentropy
sgd = SGD(lr=1e-3, momentum=0.9)
model.compile(optimizer = sgd, #sgd
loss='categorical_crossentropy',
metrics = ['accuracy'])
# 如果沒有GPU, 可以使用我 train 好的 weight
# model.load_weights('vgg16_temp3_alldata.h5')
train_history = model.fit(sub_train_x,# train x ( feature )
sub_train_y,# train y ( label or target )
#validation_split = 0.2,# catch 20% data to validation
# 不進行 validation_split, 由於 data 多, 可以 train 的更好
# 使用 validation_split 可以看出收斂性
epochs = 150,# run times 150
batch_size = 128,# 128 data/times
verbose = 1, # print process
shuffle = False)
#show_train_history(train_history)# plot training hisrtory
train_table,tem13 = compare_corr_per(sub_train_x,sub_train_y)
#test_table,tem13 = compare_corr_per(sub_test_x,sub_test_y)
model.save_weights('vgg16_temp3_alldata.h5')#28final_pred = model.predict(test_image,verbose=1)cate_pred = change_prob_to_cate(final_pred)
output = {'id':index,
'label':cate_pred}
output = pd.DataFrame(output)
output = output.sort_values('id')
output.to_csv('output.csv',index=False)model = Sequential()
#------------------------------------------------
# build convolution 1
model.add(Conv2D(
filters = 32,# random build 32 filter(濾波器) weight
kernel_size = (3,3),# 濾波器 is 3*3 size
input_shape = (32,32,3),# image.shape is 32*32*3
activation = 'relu',# 激活 function is relu
padding='same'# image size is same
))
model.add(Dropout(rate = 0.25))# drop 25% net, avoid overfitting
# build pool, (2,2) is dimension reduction,
# (2,2) --dimension reductioin ->1
# ex: shape 32*32 -> 16*16, but filter doesn't change, it still is 32
model.add(MaxPooling2D(pool_size = (2,2)))
#------------------------------------------------
# build convolution 2
model.add(Conv2D(
# filters 要*2, 32 -> 64, 如果沒*2, 那 1->1 沒有效果
filters = 64,# random build 64 filter(濾波器) weight
kernel_size = (3,3),# 濾波器 is 3*3
# this doesn't need input_shape, it is auto catch conv 1
activation = 'relu',# 激活 function is relu
padding='same'# image size is same
))
model.add(Dropout(rate = 0.25))# drop 25% net, avoid overfitting
# build pool 2
model.add(MaxPooling2D(pool_size = (2,2)))
#------------------------------------------------
# build flatten(平坦層), that is one dimension
# 64 (filters) * 8 * 8 = 4096, 8 is 32(shape) -> pool -> 16 -> pool ->8
model.add(Flatten())
model.add(Dropout(rate = 0.25))# drop 25% net, avoid overfitting
# build Dense( 隱藏層 )
model.add(Dense(1024,activation = 'relu'))
model.add(Dropout(rate = 0.25))# drop 25% net, avoid overfitting
# build output, 10 is ten categories, softmax is output probability
model.add(Dense(10,activation = 'softmax'))