QQ Browser 2021 AI Algorithm Competition Track 1 1st Place Program

Overview

2021_QQ_AIAC_Tack1_1st

QQ浏览器2021AI算法大赛赛道一 第1名 方案

paper :

环境

python==3.7.10
torch==1.7.1
transformers==4.5.1
pretrain 需要显存>=24GB 内存>=100GB

数据下载

(1) 视频数据集
视频数据集在官网下载 https://algo.browser.qq.com/
预期主办方会开源数据集,开源后会将地址补上
下载后放到 ./input/data 文件夹
tag_list 为标签的 top1w,官方 baseline 中提供,放到同一文件夹

(2) 预训练模型
预训练模型使用了 https://huggingface.co/hfl/chinese-roberta-wwm-ext-large
请使用 python3 -u download_pretrain_model.py 下载

步骤代码

(1) 预训练 + finetune
脚本命令:sh train.sh
时间算力:单模在 1 张 a100 上大约需要 pretrain(2 day),finetune(2 hour)
输出文件:每个单模的 checkpoint 保存在 jobN/model_finetune_1.pth
备注:各个单模间没有前后依赖关系,每个任务需要一张单卡,有多卡可以并行训练各个单模

(2) 代码结构说明
download_pretrain_model.py : 下载预训练模型的脚本
ensemble.py : 融合的脚本
job1-job6 : 六个模型训练任务,其文件结构完全一致,各 job 之间主要差别在预训练设置上
注:job1在赛后额外补充了一些代码注释
jobN/pretrain.py 预训练脚本
jobN/finetune.py finetune脚本
jobN/data 数据预处理部分,包含 dataset、mask token 等
jobN/config 包含 pretrain 与 finetune 的一些超参配置
jobN/qqmodel/qq_uni_model.py 模型定义

简介

简要介绍的 ppt 请参考 Introduction.pdf

模型简介

多模态模型结构与参数量和 Bert-large 一致,
layer=24, hidden_size=1024, num_attention_heads=16。
其输入为[CLS] Video_frame [SEP] Video_title [SEP]。
frame_feature 通过 fc 降维为 1024 维,与 text 的 emb 拼接。
Input_emb -> TransformerEncoder * 24 -> Pooling -> Fc -> Video_emb

预训练

预训练采用了 Tag classify, Mask language model, Mask frame model 三个任务

(1) Video tag classify 任务
tag 为人工标注的视频标签,pointwise 和 pairwise 数据集合中提供。
和官方提供的 baseline 一致,我们采用了出现频率前1w 的tag 做多标签分类任务。
Bert 最后一层的 [CLS] -> fc 得到 tag 的预测标签,与真实标签计算 BCE loss

(2) Mask language model 任务
与常见的自然语言处理 mlm 预训练方法相同,对 text 随机 15% 进行 mask,预测 mask 词。
多模态场景下,结合视频的信息预测 mask 词,可以有效融合多模态信息。

(3) Mask frame model 任务
对 frame 的随机 15% 进行 mask,mask 采用了全 0 的向量填充。
考虑到 frame 为连续的向量,难以类似于 mlm 做分类任务。
借鉴了对比学习思路,希望 mask 的预测帧在整个 batch 内的所有帧范围内与被 mask 的帧尽可能相似。
采用了 Nce loss,最大化 mask 帧和预测帧的互信息

(4) 多任务联合训练
预训练任务的 loss 采用了上述三个任务 loss 的加权和,
L = L(tag) * 1250 / 3 + L(mlm) / 3.75 + L(mfm) / 9
tag 梯度量级比较小,因此乘以了较大的权重。
注:各任务合适的权重对下游 finetune 的效果影响比较大。

(5) 预训练 Setting
初始化:bert 初始化权重来自于在中文语料预训练过的开源模型 https://huggingface.co/hfl/chinese-roberta-wwm-ext-large
数据集:预训练使用了 pointwise 和 pairwise 集合,部分融合模型中加上了 test 集合(只有 mlm 和 mfm 任务)
超参:batch_size=128, epoch=40, learning_rate=5e-5, scheduler=warmup_with_cos_decay, warum_ratio=0.06
注:预训练更多的 epoch 对效果提升比较大,从10 epoch 提升至 20 epoch 对下游任务 finetune 效果提升显著。

Finetune

(1) 下游任务
视频 pair 分别通过 model 得到 256维 embedding,两个 embedding 的 cos 相似度与人工标注标签计算 mse

(2) Finetune header
实验中发现相似度任务中,使用 mean_pooling 或者 attention_pooling 聚合最后一层 emb 接 fc 层降维效果较好。

(3) Label normalize
评估指标为 spearman,考查预测值和实际值 rank 之间的相关性,因此对人工标注 label 做了 rank 归一化。
即 target = scipy.stats.rankdata(target, 'average')

(4) Finetune Setting
数据集:训练集使用了 pairwise 中 (id1%5!=0) | (id2%5 !=0) 的部分约 6.5w,验证集使用了(id1%5==0) & (id2%5==0) 的部分约 2.5k
超参:batch_size=32, epoch=10, learning_rate=1e-5, scheduler=warmup_with_cos_decay, warum_ratio=0.06

Ensemble

(1) 融合的方法
采用了 weighted concat -> svd 降维 方法进行融合。实验中发现这种方法降维效果折损较小。
concat_vec = [np.sqrt(w1) * emb1, np.sqrt(w2) * emb2, np.sqrt(w3) * emb3 ...]
svd_vec = SVD(concat_vec, 256)

(2) 融合的模型
最终的提交融合了六个模型。 模型都使用了 bert-large 这种结构,均为迭代过程中产出的模型,各模型之间只有微小的 diff,各个模型加权权重均为 1/6。
下面表格中列出了各模型的diff部分,验证集mse,验证集spearman

jobid ensemble-weight detail val-spearman val-mse
job1 1/6 base 0.886031 0.028813
job2 1/6 预训练tag分类任务为mean_pooling+fc 0.884257 0.029493
job3 1/6 预训练任务无 mfm 0.883843 0.029248
job4 1/6 预训练数据为 (point + pair)shuf-40epoch => pair-5epoch 0.885397 0.029059
job5 1/6 预训练数据为 (point-shuf => pair-shuf => test-shuf)-32epoch 0.885795 0.028866
job6 1/6 预训练 mlm/mfm mask概率调整为25% 0.886289 0.029039

(3) 单模型的效果与融合的效果
单模的测试集成绩约在 0.836
融合两个模型在 0.845
融合三个模型在 0.849
融合五个模型在 0.852

CO-PILOT: COllaborative Planning and reInforcement Learning On sub-Task curriculum

CO-PILOT CO-PILOT: COllaborative Planning and reInforcement Learning On sub-Task curriculum, NeurIPS 2021, Shuang Ao, Tianyi Zhou, Guodong Long, Qingh

Shuang Ao 1 Feb 18, 2022
This repository contains several image-to-image translation models, whcih were tested for RGB to NIR image generation. The models are Pix2Pix, Pix2PixHD, CycleGAN and PointWise.

RGB2NIR_Experimental This repository contains several image-to-image translation models, whcih were tested for RGB to NIR image generation. The models

5 Jan 04, 2023
Ἀνατομή is a PyTorch library to analyze representation of neural networks

Ἀνατομή is a PyTorch library to analyze representation of neural networks

Ryuichiro Hataya 50 Dec 05, 2022
CS583: Deep Learning

CS583: Deep Learning

Shusen Wang 2.6k Dec 30, 2022
Very Deep Convolutional Networks for Large-Scale Image Recognition

pytorch-vgg Some scripts to convert the VGG-16 and VGG-19 models [1] from Caffe to PyTorch. The converted models can be used with the PyTorch model zo

Justin Johnson 217 Dec 05, 2022
ColossalAI-Examples - Examples of training models with hybrid parallelism using ColossalAI

ColossalAI-Examples This repository contains examples of training models with Co

HPC-AI Tech 185 Jan 09, 2023
Code and data for ACL2021 paper Cross-Lingual Abstractive Summarization with Limited Parallel Resources.

Multi-Task Framework for Cross-Lingual Abstractive Summarization (MCLAS) The code for ACL2021 paper Cross-Lingual Abstractive Summarization with Limit

Yu Bai 43 Nov 07, 2022
Job-Recommend-Competition - Vectorwise Interpretable Attentions for Multimodal Tabular Data

SiD - Simple Deep Model Vectorwise Interpretable Attentions for Multimodal Tabul

Jungwoo Park 40 Dec 22, 2022
Open source repository for the code accompanying the paper 'Non-Rigid Neural Radiance Fields Reconstruction and Novel View Synthesis of a Deforming Scene from Monocular Video'.

Non-Rigid Neural Radiance Fields This is the official repository for the project "Non-Rigid Neural Radiance Fields: Reconstruction and Novel View Synt

Facebook Research 296 Dec 29, 2022
Quantization library for PyTorch. Support low-precision and mixed-precision quantization, with hardware implementation through TVM.

HAWQ: Hessian AWare Quantization HAWQ is an advanced quantization library written for PyTorch. HAWQ enables low-precision and mixed-precision uniform

Zhen Dong 293 Dec 30, 2022
Official repository for "On Improving Adversarial Transferability of Vision Transformers" (2021)

Improving-Adversarial-Transferability-of-Vision-Transformers Muzammal Naseer, Kanchana Ranasinghe, Salman Khan, Fahad Khan, Fatih Porikli arxiv link A

Muzammal Naseer 47 Dec 02, 2022
A proof of concept ai-powered Recaptcha v2 solver

Recaptcha Fullauto I've decided to open source my old Recaptcha v2 solver. My latest version will be opened sourced this summer. I am hoping this proj

Nate 60 Dec 20, 2022
GazeScroller - Using Facial Movements to perform Hands-free Gesture on the system

GazeScroller Using Facial Movements to perform Hands-free Gesture on the system

2 Jan 05, 2022
[ICCV2021] 3DVG-Transformer: Relation Modeling for Visual Grounding on Point Clouds

3DVG-Transformer This repository is for the ICCV 2021 paper "3DVG-Transformer: Relation Modeling for Visual Grounding on Point Clouds" Our method "3DV

22 Dec 11, 2022
A PyTorch Toolbox for Face Recognition

FaceX-Zoo FaceX-Zoo is a PyTorch toolbox for face recognition. It provides a training module with various supervisory heads and backbones towards stat

JDAI-CV 1.6k Jan 06, 2023
Ros2-voiceroid2 - ROS2 wrapper package of VOICEROID2

ros2_voiceroid2 ROS2 wrapper package of VOICEROID2 Windows Only Installation Ins

Nkyoku 1 Jan 23, 2022
Official repository for Hierarchical Opacity Propagation for Image Matting

HOP-Matting Official repository for Hierarchical Opacity Propagation for Image Matting 🚧 🚧 🚧 Under Construction 🚧 🚧 🚧 🚧 🚧 🚧   Coming Soon   

Li Yaoyi 54 Dec 30, 2021
code for the ICLR'22 paper: On Robust Prefix-Tuning for Text Classification

On Robust Prefix-Tuning for Text Classification Prefix-tuning has drawed much attention as it is a parameter-efficient and modular alternative to adap

Zonghan Yang 12 Nov 30, 2022
Synthesize photos from PhotoDNA using machine learning 🌱

Ribosome Synthesize photos from PhotoDNA. See the blog post for more information. Installation Dependencies You can install Python dependencies using

Anish Athalye 112 Nov 23, 2022