HW3 Image Sentiment Classification

【问题尚未完全解决】

首先是I/O模块

读取数据并做简单的处理

# io_data.py
import os
import numpy as np
import random
from keras.utils.np_utils import to_categorical
import pandas as pd
import csv

base_dir = os.path.dirname(os.path.dirname(os.path.dirname(os.path.realpath(__file__))))
data_dir = os.path.join(base_dir,'data')

def read_dataset(mode='train',isFeat=True):
    """
    Return:
        # features: (int. list) list
        # labels: int32 2D array
        data_ids: int. list
    """
    # num_data = 0
    datas = []

    with open(os.path.join(data_dir,'{}.csv'.format(mode))) as file:
        #file = pd.read_csv(os.path.join(data_dir,'{}.csv'.format(mode)), sep=',', header=0)
        #print(file.readlines()[0:1])
        for line_id,line in enumerate(file.readlines()[1:]):
            #print('line : ', line, line_id)
            if isFeat:
                label, feat=line.split(',')
            else:
                _,feat = line.split(',')
            #print(feat)
            feat = np.fromstring(feat,dtype=int,sep=' ')
            # 归一化
            feat = feat / 255
            #print(feat)
            feat = np.reshape(feat,(48,48,1))

            if isFeat:
                datas.append((feat,int(label),line_id))
            else:
                datas.append(feat)

    random.shuffle(datas)  # shuffle outside
    if isFeat:
        feats,labels,line_ids = zip(*datas)
    else:
        feats = datas
    feats = np.asarray(feats)
    if isFeat:
        labels = to_categorical(np.asarray(labels,dtype=np.int32))
        return feats,labels,line_ids
    else:
        return feats


def dump_history(store_path,logs):
    with open(os.path.join(store_path,'train_loss'),'a') as f:
        for loss in logs.tr_losses:
            f.write('{}\n'.format(loss))
    with open(os.path.join(store_path,'train_accuracy'),'a') as f:
        for acc in logs.tr_accs:
            f.write('{}\n'.format(acc))
    with open(os.path.join(store_path,'valid_loss'),'a') as f:
        for loss in logs.val_losses:
            f.write('{}\n'.format(loss))
    with open(os.path.join(store_path,'valid_accuracy'),'a') as f:
        for acc in logs.val_accs:
            f.write('{}\n'.format(acc))

if __name__ == "__main__":
    read_dataset()

然后，设计model，原代码给出了一个easy model，training之后的正确率卡在40%左右

我在网上找了一个据说是不错的model（simple model）

# model.py
import os
from keras.models import Sequential
from keras.layers import Input,Dense,Dropout,Flatten,Activation
from keras.layers import Convolution2D,MaxPooling2D
from keras.optimizers import SGD
import tensorflow as tf
from record import *

nb_class = 7

def build_model(mode):

    """Return the Keras model for training

    Keyword arguments:
    mode: model name specified in training and predicting script

    """
    model = Sequential()
    if mode == 'easy':
        # CNN part (you can repeat this part several times)
        model.add(Convolution2D(8,3,3,border_mode='valid',input_shape=(48,48,1)))
        model.add(Activation('relu'))
        model.add(MaxPooling2D(pool_size=(2, 2)))
        model.add(Dropout(0.8))

        # Fully connected part
        model.add(Flatten())
        model.add(Dense(16))
        model.add(Activation('relu'))
        model.add(Dense(nb_class))
        model.add(Activation('softmax'))
        opt = SGD(lr=0.01,decay=0.0)

    if mode == 'simple':
        model.add(Convolution2D(32, (1, 1), strides=1, padding='same', input_shape=(48, 48, 1)))
        model.add(Activation('relu'))
        model.add(Convolution2D(32, (5, 5), padding='same'))
        model.add(Activation('relu'))
        model.add(MaxPooling2D(pool_size=(2, 2)))

        model.add(Convolution2D(32, (3, 3), padding='same'))
        model.add(Activation('relu'))
        model.add(MaxPooling2D(pool_size=(2, 2)))

        model.add(Convolution2D(64, (5, 5), padding='same'))
        model.add(Activation('relu'))
        model.add(MaxPooling2D(pool_size=(2, 2)))

        model.add(Flatten())
        model.add(Dense(2048))
        model.add(Activation('relu'))
        model.add(Dropout(0.5))
        model.add(Dense(1024))
        model.add(Activation('relu'))
        model.add(Dropout(0.5))
        model.add(Dense(nb_class))
        model.add(Activation('softmax'))
        model.summary()
        #opt = SGD(lr=0.01, decay=0.0)

    model.compile(loss='categorical_crossentropy',
                  optimizer='adam',
                  metrics=['accuracy'])
    model.summary() # show the whole model in terminal
    return model

然后开始训练数据

# train.py
#!/usr/bin/env python
# -- coding: utf-8 --

import argparse
import os
import model
from utils import *
from io_data import *
from record import *

import os
import tensorflow as tf
import keras.backend.tensorflow_backend as KTF

# 进行配置GPU
os.environ["CUDA_VISIBLE_DEVICES"]="0,1"  # 使用编号为1的GPU
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.8   # 每个GPU现存上届控制在80%以内
session = tf.Session(config=config)

# 设置session
KTF.set_session(session )

base_dir = os.path.dirname(os.path.dirname(os.path.realpath(__file__)))
exp_dir = os.path.join(base_dir,'exp')

def main():
    parser = argparse.ArgumentParser(prog='train.py',
            description='ML-Assignment training script.')
    parser.add_argument('--model', type=str,default='easy',choices=['easy','simple','strong'],
            metavar='<model>')
    parser.add_argument('--epoch',type=int,metavar='<#epoch>',default=40)
    parser.add_argument('--batch',type=int,metavar='<batch_size>',default=64)
    args = parser.parse_args()
    print(args)

    # set the path
    dir_cnt = 0
    log_path = "{}_epoch{}".format(args.model,str(args.epoch))
    log_path += '_'
    store_path = os.path.join(exp_dir,log_path+str(dir_cnt))
    while dir_cnt < 30:
        if not os.path.isdir(store_path):
            os.mkdir(store_path)
            break
        else:
            dir_cnt += 1
            store_path = os.path.join(exp_dir, log_path + str(dir_cnt))

    emotion_classifier = model.build_model(args.model)
    tr_feats,tr_labels,_ = read_dataset('train')
    dev_feats,dev_labels,_ = read_dataset('valid') # 10% of original training data
    history = History()
    emotion_classifier.fit(x = tr_feats,y = tr_labels,
            batch_size=args.batch,epochs=args.epoch,validation_split=0.1,
            callbacks=[history])
    #validation_data=(dev_feats,dev_labels)

    # 算出正确率
    train_result = emotion_classifier.evaluate(tr_feats, tr_labels)
    print('\n train Acc: ', train_result[1])

    dump_history(store_path,history)
    emotion_classifier.save(os.path.join(store_path,'model.h5'))

if __name__ == "__main__":
    main()

经过漫长的等待后，得到的在training set上的正确率大概为95%，但是在valid date中仅为55%。其中是有问题的。

训练过程中，将 acc 和 loss 记录下来

# record.py
from keras.callbacks import Callback

class History(Callback):
    def on_train_begin(self,logs={}):
        self.tr_losses=[]
        self.val_losses=[]
        self.tr_accs=[]
        self.val_accs=[]

    def on_epoch_end(self,epoch,logs={}):
        self.tr_losses.append(logs.get('loss'))
        self.val_losses.append(logs.get('val_loss'))
        self.tr_accs.append(logs.get('acc'))
        self.val_accs.append(logs.get('val_acc'))

对 testing data 进行预测

# predict.py
#!/usr/bin/env python
# -- coding: utf-8 --

import argparse
from keras.models import load_model
from termcolor import colored,cprint
from io_data import *

import os
import tensorflow as tf
import keras.backend.tensorflow_backend as KTF

# 进行配置GPU
os.environ["CUDA_VISIBLE_DEVICES"]="0,1" # 使用编号为1的GPU
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.6   # 每个GPU现存上届控制在60%以内
session = tf.Session(config=config)

# 设置session
KTF.set_session(session)


base_dir = os.path.dirname(os.path.dirname(os.path.realpath(__file__)))
exp_dir = os.path.join(base_dir,'exp')

def main():
    parser = argparse.ArgumentParser(prog='predict.py',
            description='ML-Assignment3 testing script.')
    parser.add_argument('--model',type=str,default='simple',choices=['easy','simple','strong'],
            metavar='<model>')
    parser.add_argument('--epoch',type=int,metavar='<#epoch>',default=40)
    parser.add_argument('--batch',type=int,metavar='<batch_size>',default=64)
    parser.add_argument('--idx',type=int,metavar='<suffix>', default=0)
    args = parser.parse_args()
    store_path = "{}_epoch{}_{}".format(args.model,args.epoch,args.idx)
    print(colored("\nLoading model from {}".format(store_path),'yellow',attrs=['bold']))
    model_path = os.path.join(exp_dir,store_path,'model.h5')

    emotion_classifier = load_model(model_path)
    emotion_classifier.summary()
    te_feats = read_dataset('test',False)

    ans = emotion_classifier.predict_classes(te_feats,batch_size=args.batch)

    with open('Answer','w') as f:
        f.write('id,label\n')
        for idx,a in enumerate(ans):
            f.write('{},{}\n'.format(idx,a))

if __name__ == "__main__":
    main()

结果保存在 Answer 文件中

画出model

# plot_model.py
#!/usr/bin/env python
# -- coding: utf-8 --

import os
from termcolor import colored, cprint
import argparse
from keras.models import load_model
from keras.utils.vis_utils import plot_model
import matplotlib.pyplot as plt
from PIL import Image

base_dir = os.path.dirname(os.path.dirname(os.path.realpath(__file__)))
exp_dir = os.path.join(base_dir,'exp')

def main():
    parser = argparse.ArgumentParser(prog='plot_model.py',
            description='Plot the model.')
    parser.add_argument('--model',type=str,default='simple',choices=['simple','easy','strong'],
            metavar='<model>')
    parser.add_argument('--epoch',type=int,metavar='<#epoch>',default=40)
    parser.add_argument('--batch',type=int,metavar='<batch_size>',default=64)
    parser.add_argument('--idx',type=int,metavar='<suffix>', default=0)
    args = parser.parse_args()
    store_path = "{}_epoch{}_{}".format(args.model,args.epoch,args.idx)
    print(colored("Loading model from {}".format(store_path),'yellow',attrs=['bold']))

    model_path = os.path.join(exp_dir,store_path,'model.h5')

    emotion_classifier = load_model(model_path)
    emotion_classifier.summary()
    plot_model(emotion_classifier,to_file='ml_plot.png')


if __name__ == "__main__":
    main()

最后把分好类的图片加载出来看一看

# arr_pic.py
from PIL import Image
import numpy as np
import scipy.misc
import os

base_dir = os.path.dirname(os.path.dirname(os.path.dirname(os.path.realpath(__file__))))
data_dir = os.path.join(base_dir,'data')

def output_pic(idx, num):
    with open(os.path.join(data_dir, 'test.csv')) as file:
        for line_id, line in enumerate(file.readlines()[1:200]):
            _idx, feat = line.split(',')
            if int(_idx) == idx:
                print(idx)
                feat = np.fromstring(feat, dtype=int, sep=' ')
                feat = np.reshape(feat, (48, 48))
                new_im = Image.fromarray(feat)
                #new_im.show()
                scipy.misc.imsave('C:/Users/liky/Desktop/pic_3/' + "test" + str(num) + ".jpg", new_im)
                print('save succ')
                break

if __name__ == '__main__':
    with open('F:/pycharmFile/CNN_1/Answer') as f:
        num = 0
        for line in f.readlines()[1:200] :
            if int(line.split(',')[1]) == 3: # 高兴的图片
                output_pic(int(line.split(',')[0]), num)
                num += 1
                print(line , '    ', num)