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

Fine-grained Control of Image Caption Generation with Abstract Scene Graphs

Faster R-CNN pretrained on VisualGenome This repository modifies maskrcnn-benchmark for object detection and attribute prediction on VisualGenome data

Shizhe Chen 7 Apr 20, 2021
PyTorch implementation for our paper Learning Character-Agnostic Motion for Motion Retargeting in 2D, SIGGRAPH 2019

Learning Character-Agnostic Motion for Motion Retargeting in 2D We provide PyTorch implementation for our paper Learning Character-Agnostic Motion for

Rundi Wu 367 Dec 22, 2022
[arXiv22] Disentangled Representation Learning for Text-Video Retrieval

Disentangled Representation Learning for Text-Video Retrieval This is a PyTorch implementation of the paper Disentangled Representation Learning for T

Qiang Wang 49 Dec 18, 2022
Complete the code of prefix-tuning in low data setting

Prefix Tuning Note: 作者在论文中提到使用真实的word去初始化prefix的操作(Initializing the prefix with activations of real words,significantly improves generation)。我在使用作者提供的

Andrew Zeng 4 Jul 11, 2022
Keras implementation of Normalizer-Free Networks and SGD - Adaptive Gradient Clipping

Keras implementation of Normalizer-Free Networks and SGD - Adaptive Gradient Clipping

Yam Peleg 63 Sep 21, 2022
Poisson Surface Reconstruction for LiDAR Odometry and Mapping

Poisson Surface Reconstruction for LiDAR Odometry and Mapping Surfels TSDF Our Approach Table: Qualitative comparison between the different mapping te

Photogrammetry & Robotics Bonn 305 Dec 21, 2022
Prototypical Cross-Attention Networks for Multiple Object Tracking and Segmentation, NeurIPS 2021 Spotlight

PCAN for Multiple Object Tracking and Segmentation This is the offical implementation of paper PCAN for MOTS. We also present a trailer that consists

ETH VIS Group 328 Dec 29, 2022
[NAACL & ACL 2021] SapBERT: Self-alignment pretraining for BERT.

SapBERT: Self-alignment pretraining for BERT This repo holds code for the SapBERT model presented in our NAACL 2021 paper: Self-Alignment Pretraining

Cambridge Language Technology Lab 104 Dec 07, 2022
The final project of "Applying AI to 2D Medical Imaging Data" of "AI for Healthcare" nanodegree - Udacity.

Pneumonia Detection from X-Rays Project Overview In this project, you will apply the skills that you have acquired in this 2D medical imaging course t

Omar Laham 1 Jan 14, 2022
Citation Intent Classification in scientific papers using the Scicite dataset an Pytorch

Citation Intent Classification Table of Contents About the Project Built With Installation Usage Acknowledgments About The Project Citation Intent Cla

Federico Nocentini 4 Mar 04, 2022
Unsupervised Discovery of Object Radiance Fields

Unsupervised Discovery of Object Radiance Fields by Hong-Xing Yu, Leonidas J. Guibas and Jiajun Wu from Stanford University. arXiv link: https://arxiv

Hong-Xing Yu 148 Nov 30, 2022
A series of convenience functions to make basic image processing operations such as translation, rotation, resizing, skeletonization, and displaying Matplotlib images easier with OpenCV and Python.

imutils A series of convenience functions to make basic image processing functions such as translation, rotation, resizing, skeletonization, and displ

Adrian Rosebrock 4.3k Jan 08, 2023
Discovering Interpretable GAN Controls [NeurIPS 2020]

GANSpace: Discovering Interpretable GAN Controls Figure 1: Sequences of image edits performed using control discovered with our method, applied to thr

Erik Härkönen 1.7k Jan 03, 2023
SurvITE: Learning Heterogeneous Treatment Effects from Time-to-Event Data

SurvITE: Learning Heterogeneous Treatment Effects from Time-to-Event Data SurvITE: Learning Heterogeneous Treatment Effects from Time-to-Event Data Au

14 Nov 28, 2022
Measuring and Improving Consistency in Pretrained Language Models

ParaRel 🤘 This repository contains the code and data for the paper: Measuring and Improving Consistency in Pretrained Language Models as well as the

Yanai Elazar 26 Dec 02, 2022
AI grand challenge 2020 Repo (Speech Recognition Track)

KorBERT를 활용한 한국어 텍스트 기반 위협 상황인지(2020 인공지능 그랜드 챌린지) 본 프로젝트는 ETRI에서 제공된 한국어 korBERT 모델을 활용하여 폭력 기반 한국어 텍스트를 분류하는 다양한 분류 모델들을 제공합니다. 본 개발자들이 참여한 2020 인공지

Young-Seok Choi 23 Jan 25, 2022
The Pytorch code of "Joint Distribution Matters: Deep Brownian Distance Covariance for Few-Shot Classification", CVPR 2022 (Oral).

DeepBDC for few-shot learning        Introduction In this repo, we provide the implementation of the following paper: "Joint Distribution Matters: Dee

FeiLong 116 Dec 19, 2022
Digan - Official PyTorch implementation of Generating Videos with Dynamics-aware Implicit Generative Adversarial Networks

DIGAN (ICLR 2022) Official PyTorch implementation of "Generating Videos with Dyn

Sihyun Yu 147 Dec 31, 2022
People log into different sites every day to get information and browse through these sites one by one

HyperLink People log into different sites every day to get information and browse through these sites one by one. And they are exposed to advertisemen

0 Feb 17, 2022
ConE: Cone Embeddings for Multi-Hop Reasoning over Knowledge Graphs

ConE: Cone Embeddings for Multi-Hop Reasoning over Knowledge Graphs This is the code of paper ConE: Cone Embeddings for Multi-Hop Reasoning over Knowl

MIRA Lab 33 Dec 07, 2022