The Code for ICME2019 Grand Challenge: Short Video Understanding (Single Model Ranks 6th)

README
Issues 5

Introduction

Implement of the paper "Multi-modal Representation Learning for Short Video Understanding and Recommendation"

Short Video Understanding Challenge

Introduction of ICME2019 Grand Challenge refers to the website

1. Requirement

scikit-learn
tqdm
pandas
numpy
scipy
tensorFlow=1.12.0 (≥1.4 and ≠1.5 or 1.6)
128G memory and 1 GPU

2. Download Dataset

cd data
bash download.sh
cd ..

3. Preprocess data

python preprocess.py

4. Extract features

python extract_features.py

5.Convert format of data

python convert_format.py

6. Train the model

python run.py

1 Open More issues

Closed

extract_features

郭大，我最近在学习你的工程代码，目前遇到几点疑问？烦请解答在计算title_mean时，出现
mean=0 # {"item_id": 1488899, "title_features": {"314": 1, "3499": 1, "5285": 1, "174": 1}} for t1,t2 in zip(item[0].split(),item[1]): mean+=int(t1)*t2 mean/=sum(item[1]) tract2_title.txt文件中的title_features的key表示的是什么？value应该是词频吧，为啥你能将key和value相乘，然后去求mean，这里不理解。

在var_mean部分，你使用了傅里叶变换，用它的作用或者意义是啥呢？ data=train_df[['uid','author_id']].append(test_df[['uid','author_id']]) group=data[['uid','author_id']].groupby('uid') group=group.apply(lambda x:np.var(np.fft.fft(list(Counter(list(x['author_id'])).values())))) train_df['uid_num_of_author_fft_var']=train_df['uid'].apply(lambda x:group[x]) test_df['uid_num_of_author_fft_var']=test_df['uid'].apply(lambda x:group[x]) print(train_df[train_df['finish']==1]['uid_num_of_author_fft_var'].mean()) print(train_df[train_df['finish']==0]['uid_num_of_author_fft_var'].mean())

opened Jun 4, 2020 by chen849157649 8

Closed

day=[i*7//len(train_df) for i in range(len(train_df))]==>全0 list

郭大，split_data()里面这行代码是不是有误？请问您是按照什么划分训练验证集的呢？

opened Nov 21, 2019 by TenYearMan 2

Closed

xdeepfm

你好，我是新手，看了下里面的模型，似乎不是xdeepFM，与xdeepFM相比speed与AUC的trade-off哪个更好，有做过比较吗？谢谢

opened May 26, 2020 by chen849157649 1

Closed

extract_features

郭大，你好，我最近在研读你的代码，有问题不懂 def w2v(train_df,test_df,f,flag,L): # word2vec 即通过一个嵌入空间使得语义上相似的单词在该空间内距离很近 # 在这里是，利用word2vec来表示用户观看的相似作品在该空间内距离很近 print("w2v:",f) sentence=[] dic={} day=0 # 分别对f='item_id', f='author_id' for item in train_df[['day','uid',f]].values: if day!=item[0]: for key in dic: # 将第0天用户看过的全部放入sentence sentence.append(dic[key]) dic={} day=item[0] print(day) # dic统计用户uid 看过哪些作品 try: # 第1天开始这句似乎不会执行了 dic[item[1]].append(str(item[2])) except: dic[item[1]]=[str(item[2])] for key in dic: sentence.append(dic[key])

为啥将初始第0天要区别处理呢？从第1天开始try 内部的似乎不会运行

opened Jun 3, 2020 by chen849157649 0

Open