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

LaneAF: Robust Multi-Lane Detection with Affinity Fields

LaneAF: Robust Multi-Lane Detection with Affinity Fields This repository contains Pytorch code for training and testing LaneAF lane detection models i

155 Dec 17, 2022
Does Oversizing Improve Prosumer Profitability in a Flexibility Market? - A Sensitivity Analysis using PV-battery System

Does Oversizing Improve Prosumer Profitability in a Flexibility Market? - A Sensitivity Analysis using PV-battery System The possibilities to involve

Babu Kumaran Nalini 0 Nov 19, 2021
Post-Training Quantization for Vision transformers.

PTQ4ViT Post-Training Quantization Framework for Vision Transformers. We use the twin uniform quantization method to reduce the quantization error on

Zhihang Yuan 61 Dec 28, 2022
PolyGlot, a fuzzing framework for language processors

PolyGlot, a fuzzing framework for language processors Build We tested PolyGlot on Ubuntu 18.04. Get the source code: git clone https://github.com/s3te

Software Systems Security Team at Penn State University 79 Dec 27, 2022
image scene graph generation benchmark

Scene Graph Benchmark in PyTorch 1.7 This project is based on maskrcnn-benchmark Highlights Upgrad to pytorch 1.7 Multi-GPU training and inference Bat

Microsoft 303 Dec 27, 2022
Implementation for Stankevičiūtė et al. "Conformal time-series forecasting", NeurIPS 2021.

Conformal time-series forecasting Implementation for Stankevičiūtė et al. "Conformal time-series forecasting", NeurIPS 2021. If you use our code in yo

Kamilė Stankevičiūtė 36 Nov 21, 2022
Official Pytorch implementation of Meta Internal Learning

Official Pytorch implementation of Meta Internal Learning

10 Aug 24, 2022
An implementation of shampoo

shampoo.pytorch An implementation of shampoo, proposed in Shampoo : Preconditioned Stochastic Tensor Optimization by Vineet Gupta, Tomer Koren and Yor

Ryuichiro Hataya 69 Sep 10, 2022
Finetune SSL models for MOS prediction

Finetune SSL models for MOS prediction This is code for our paper under review for ICASSP 2022: "Generalization Ability of MOS Prediction Networks" Er

Yamagishi and Echizen Laboratories, National Institute of Informatics 32 Nov 22, 2022
Face Synthetics dataset is a collection of diverse synthetic face images with ground truth labels.

The Face Synthetics dataset Face Synthetics dataset is a collection of diverse synthetic face images with ground truth labels. It was introduced in ou

Microsoft 608 Jan 02, 2023
A framework that constructs deep neural networks, autoencoders, logistic regressors, and linear networks

A framework that constructs deep neural networks, autoencoders, logistic regressors, and linear networks without the use of any outside machine learning libraries - all from scratch.

Kordel K. France 2 Nov 14, 2022
[Preprint] "Bag of Tricks for Training Deeper Graph Neural Networks A Comprehensive Benchmark Study" by Tianlong Chen*, Kaixiong Zhou*, Keyu Duan, Wenqing Zheng, Peihao Wang, Xia Hu, Zhangyang Wang

Bag of Tricks for Training Deeper Graph Neural Networks: A Comprehensive Benchmark Study Codes for [Preprint] Bag of Tricks for Training Deeper Graph

VITA 101 Dec 29, 2022
Anonymize BLM Protest Images

Anonymize BLM Protest Images This repository automates @BLMPrivacyBot, a Twitter bot that shows the anonymized images to help keep protesters safe. Us

Stanford Machine Learning Group 40 Oct 13, 2022
Taichi Course Homework Template

太极图形课S1-标题部分 这个作业未来或将是你的开源项目,标题的内容可以来自作业中的核心关键词,让读者一眼看出你所完成的工作/做出的好玩demo 如果暂时未想好,起名时可以参考“太极图形课S1-xxx作业” 如下是作业(项目)展开说明的方法,可以帮大家理清思路,并且也对读者非常友好,请小伙伴们多多参

TaichiCourse 30 Nov 19, 2022
gym-anm is a framework for designing reinforcement learning (RL) environments that model Active Network Management (ANM) tasks in electricity distribution networks.

gym-anm is a framework for designing reinforcement learning (RL) environments that model Active Network Management (ANM) tasks in electricity distribution networks. It is built on top of the OpenAI G

Robin Henry 99 Dec 12, 2022
Code accompanying our NeurIPS 2021 traffic4cast challenge

Traffic forecasting on traffic movie snippets This repo contains all code to reproduce our approach to the IARAI Traffic4cast 2021 challenge. In the c

Nina Wiedemann 2 Aug 09, 2022
Discord-Protect is a simple discord bot allowing you to have some security on your discord server by ordering a captcha to the user who joins your server.

Discord-Protect Discord-Protect is a simple discord bot allowing you to have some security on your discord server by ordering a captcha to the user wh

Tir Omar 2 Oct 28, 2021
InterfaceGAN++: Exploring the limits of InterfaceGAN

InterfaceGAN++: Exploring the limits of InterfaceGAN Authors: Apavou Clément & Belkada Younes From left to right - Images generated using styleGAN and

Younes Belkada 42 Dec 23, 2022
Large-scale language modeling tutorials with PyTorch

Large-scale language modeling tutorials with PyTorch 안녕하세요. 저는 TUNiB에서 머신러닝 엔지니어로 근무 중인 고현웅입니다. 이 자료는 대규모 언어모델 개발에 필요한 여러가지 기술들을 소개드리기 위해 마련하였으며 기본적으로

TUNiB 172 Dec 29, 2022
Differentiable molecular simulation of proteins with a coarse-grained potential

Differentiable molecular simulation of proteins with a coarse-grained potential This repository contains the learned potential, simulation scripts and

UCL Bioinformatics Group 44 Dec 10, 2022