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

DockStream: A Docking Wrapper to Enhance De Novo Molecular Design

DockStream Description DockStream is a docking wrapper providing access to a collection of ligand embedders and docking backends. Docking execution an

AstraZeneca - Molecular AI 72 Jan 02, 2023
A framework for joint super-resolution and image synthesis, without requiring real training data

SynthSR This repository contains code to train a Convolutional Neural Network (CNN) for Super-resolution (SR), or joint SR and data synthesis. The met

83 Jan 01, 2023
LSSY量化交易系统

LSSY量化交易系统 该项目是本人3年来研究量化慢慢积累开发的一套系统,属于早期作品慢慢修改而来,仅供学习研究,回测分析,实盘交易部分未公开

55 Oct 04, 2022
[ACL 20] Probing Linguistic Features of Sentence-level Representations in Neural Relation Extraction

REval Table of Contents Introduction Overview Requirements Installation Probing Usage Citation License 🎓 Introduction REval is a simple framework for

13 Jan 06, 2023
A quantum game modeling of pandemic (QHack 2022)

Contributors: @JongheumJung, @YoonjaeChung, @GyunghunKim Abstract In the regime of a global pandemic, leaders around the world need to consider variou

Yoonjae Chung 8 Apr 03, 2022
Exploring Image Deblurring via Blur Kernel Space (CVPR'21)

Exploring Image Deblurring via Encoded Blur Kernel Space About the project We introduce a method to encode the blur operators of an arbitrary dataset

VinAI Research 118 Dec 19, 2022
Recreate CenternetV2 based on MMDET.

Introduction This project is trying to Recreate CenternetV2 based on MMDET, which is proposed in paper Probabilistic two-stage detection. This project

25 Dec 09, 2022
A state of the art of new lightweight YOLO model implemented by TensorFlow 2.

CSL-YOLO: A New Lightweight Object Detection System for Edge Computing This project provides a SOTA level lightweight YOLO called "Cross-Stage Lightwe

Miles Zhang 54 Dec 21, 2022
Async API for controlling Hue Lights

Hue API Async API for controlling Hue Lights Documentation: hue-api.nirantak.com Source: github.com/nirantak/hue-api Installation This is an async cli

Nirantak Raghav 4 Nov 16, 2022
CowHerd is a partially-observed reinforcement learning environment

CowHerd is a partially-observed reinforcement learning environment, where the player walks around an area and is rewarded for milking cows. The cows try to escape and the player can place fences to h

Danijar Hafner 6 Mar 06, 2022
Fuzzing the Kernel Using Unicornafl and AFL++

Unicorefuzz Fuzzing the Kernel using UnicornAFL and AFL++. For details, skim through the WOOT paper or watch this talk at CCCamp19. Is it any good? ye

Security in Telecommunications 283 Dec 26, 2022
The Instructed Glacier Model (IGM)

The Instructed Glacier Model (IGM) Overview The Instructed Glacier Model (IGM) simulates the ice dynamics, surface mass balance, and its coupling thro

27 Dec 16, 2022
Python suite to construct benchmark machine learning datasets from the MIMIC-III clinical database.

MIMIC-III Benchmarks Python suite to construct benchmark machine learning datasets from the MIMIC-III clinical database. Currently, the benchmark data

Chengxi Zang 6 Jan 02, 2023
This tutorial aims to learn the basics of deep learning by hands, and master the basics through combination of lectures and exercises

2021-Deep-learning This tutorial aims to learn the basics of deep learning by hands, and master the basics through combination of paper and exercises.

108 Feb 24, 2022
An original implementation of "Noisy Channel Language Model Prompting for Few-Shot Text Classification"

Channel LM Prompting (and beyond) This includes an original implementation of Sewon Min, Mike Lewis, Hannaneh Hajishirzi, Luke Zettlemoyer. "Noisy Cha

Sewon Min 92 Jan 07, 2023
ERISHA is a mulitilingual multispeaker expressive speech synthesis framework. It can transfer the expressivity to the speaker's voice for which no expressive speech corpus is available.

ERISHA: Multilingual Multispeaker Expressive Text-to-Speech Library ERISHA is a multilingual multispeaker expressive speech synthesis framework. It ca

Ajinkya Kulkarni 43 Nov 27, 2022
Optimizing Deeper Transformers on Small Datasets

DT-Fixup Optimizing Deeper Transformers on Small Datasets Paper published in ACL 2021: arXiv Detailed instructions to replicate our results in the pap

16 Nov 14, 2022
Assginment for UofT CSC420: Intro to Image Understanding

Run the code Open edge_detection.ipynb in google colab. Upload image1.jpg,image2.jpg and my_image.jpg to '/content/drive/My Drive'. chooose 'Run all'

Ziyi-Zhou 1 Feb 24, 2022
Generative vs Discriminative: Rethinking The Meta-Continual Learning (NeurIPS 2021)

Generative vs Discriminative: Rethinking The Meta-Continual Learning (NeurIPS 2021) In this repository we provide PyTorch implementations for GeMCL; a

4 Apr 15, 2022
Learning Representational Invariances for Data-Efficient Action Recognition

Learning Representational Invariances for Data-Efficient Action Recognition Official PyTorch implementation for Learning Representational Invariances

Virginia Tech Vision and Learning Lab 27 Nov 22, 2022