The 1st Place Solution of the Facebook AI Image Similarity Challenge (ISC21) : Descriptor Track.

Overview

ISC21-Descriptor-Track-1st

The 1st Place Solution of the Facebook AI Image Similarity Challenge (ISC21) : Descriptor Track.

You can check our solution tech report from: Contrastive Learning with Large Memory Bank and Negative Embedding Subtraction for Accurate Copy Detection

setup

OS

Ubuntu 18.04

CUDA Version

11.1

environment

Run this for python env

conda env create -f environment.yml

data download

mkdir -p input/{query,reference,train}_images
aws s3 cp s3://drivendata-competition-fb-isc-data/all/query_images/ input/query_images/ --recursive --no-sign-request
aws s3 cp s3://drivendata-competition-fb-isc-data/all/reference_images/ input/reference_images/ --recursive --no-sign-request
aws s3 cp s3://drivendata-competition-fb-isc-data/all/train_images/ input/train_images/ --recursive --no-sign-request
aws s3 cp s3://drivendata-competition-fb-isc-data/all/query_images_phase2/ input/query_images_phase2/ --recursive --no-sign-request

train

Run below lines step by step.

cd exp

CUDA_VISIBLE_DEVICES=0,1,2,3 python v83.py \
  -a tf_efficientnetv2_m_in21ft1k --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --seed 9 \
  --epochs 5 --lr 0.1 --wd 1e-6 --batch-size 128 --ncrops 2 \
  --gem-p 1.0 --pos-margin 0.0 --neg-margin 1.0 \
  --input-size 256 --sample-size 1000000 --memory-size 20000 \
  ../input/training_images/
CUDA_VISIBLE_DEVICES=0,1,2,3 python v83.py \
  -a tf_efficientnetv2_m_in21ft1k --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --seed 90 \
  --epochs 10 --lr 0.1 --wd 1e-6 --batch-size 128 --ncrops 2 \
  --gem-p 1.0 --pos-margin 0.0 --neg-margin 1.0 \
  --input-size 256 --sample-size 1000000 --memory-size 20000 \
  --resume ./v83/train/checkpoint_0004.pth.tar \
  ../input/training_images/

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python v86.py \
  -a tf_efficientnetv2_m_in21ft1k --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --seed 99 \
  --epochs 7 --lr 0.1 --wd 1e-6 --batch-size 128 --ncrops 2 \
  --gem-p 1.0 --pos-margin 0.0 --neg-margin 1.0 \
  --input-size 384 --sample-size 1000000 --memory-size 20000 --weight ./v83/train/checkpoint_0005.pth.tar \
  ../input/training_images/

python v98.py \
  -a tf_efficientnetv2_m_in21ft1k --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --seed 999 \
  --epochs 3 --lr 0.1 --wd 1e-6 --batch-size 64 --ncrops 2 \
  --gem-p 1.0 --pos-margin 0.0 --neg-margin 1.0 --weight ./v86/train/checkpoint_0005.pth.tar \
  --input-size 512 --sample-size 1000000 --memory-size 20000 \
  ../input/training_images/

python v107.py \
  -a tf_efficientnetv2_m_in21ft1k --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --seed 99999 \
  --epochs 10 --lr 0.5 --wd 1e-6 --batch-size 16 --ncrops 2 \
  --gem-p 1.0 --pos-margin 0.0 --neg-margin 1.1 --weight ./v98/train/checkpoint_0001.pth.tar \
  --input-size 512 --sample-size 1000000 --memory-size 1000 \
  ../input/training_images/

The final model weight can be downloaded from here: https://drive.google.com/file/d/1ySea-NJp_J0aWvma_WmVbc3Hnwf5LHUf/view?usp=sharing You can execute inference code without run training with this model weight. To locate the model weight to suitable location, run following commands after downloaded the model weight.

mkdir -p exp/v107/train
mv checkpoint_009.pth.tar exp/v107/train/

inference

Note that faiss doesn't work with A100, so I used 4x GTX 1080 Ti for post-process.

cd exp

python v107.py -a tf_efficientnetv2_m_in21ft1k --batch-size 128 --mode extract --gem-eval-p 1.0 --weight ./v107/train/checkpoint_0009.pth.tar --input-size 512 --target-set qrt ../input/

# this script generates final prediction result files
python ../scripts/postprocess.py

Submission files are outputted here:

  • exp/v107/extract/v107_iso.h5 # descriptor track
  • exp/v107/extract/v107_iso.csv # matching track

descriptor track local evaluation score:

{
  "average_precision": 0.9479039085717805,
  "recall_p90": 0.9192546583850931
}
Comments
  • Bugs?

    Bugs?

    Congratulations! We really appreciate the work. When I run the

    python v107.py \
      -a tf_efficientnetv2_m_in21ft1k --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --seed 99999 \
      --epochs 10 --lr 0.5 --wd 1e-6 --batch-size 16 --ncrops 2 \
      --gem-p 1.0 --pos-margin 0.0 --neg-margin 1.1 --weight ./v98/train/checkpoint_0001.pth.tar \
      --input-size 512 --sample-size 1000000 --memory-size 1000 \
      ../input/training_images/
    

    I come across

    Traceback (most recent call last):                                              
      File "v107.py", line 774, in <module>
        train(args)
      File "v107.py", line 425, in train
        mp.spawn(main_worker, nprocs=ngpus_per_node, args=(ngpus_per_node, args))
      File "/home/wangwenhao/anaconda3/envs/ISC/lib/python3.7/site-packages/torch/multiprocessing/spawn.py", line 230, in spawn
        return start_processes(fn, args, nprocs, join, daemon, start_method='spawn')
      File "/home/wangwenhao/anaconda3/envs/ISC/lib/python3.7/site-packages/torch/multiprocessing/spawn.py", line 188, in start_processes
        while not context.join():
      File "/home/wangwenhao/anaconda3/envs/ISC/lib/python3.7/site-packages/torch/multiprocessing/spawn.py", line 150, in join
        raise ProcessRaisedException(msg, error_index, failed_process.pid)
    torch.multiprocessing.spawn.ProcessRaisedException: 
    
    -- Process 5 terminated with the following error:
    Traceback (most recent call last):
      File "/home/wangwenhao/anaconda3/envs/ISC/lib/python3.7/site-packages/torch/multiprocessing/spawn.py", line 59, in _wrap
        fn(i, *args)
      File "/home/wangwenhao/fbisc-descriptor-1st/exp/v107.py", line 573, in main_worker
        train_one_epoch(train_loader, model, loss_fn, optimizer, scaler, epoch, args)
      File "/home/wangwenhao/fbisc-descriptor-1st/exp/v107.py", line 595, in train_one_epoch
        labels = torch.cat([torch.tile(i, dims=(args.ncrops,)), torch.tensor(j)])
    ValueError: only one element tensors can be converted to Python scalars
    

    Do you know how to fix it? Thanks.

    opened by WangWenhao0716 14
  • data augment is wrong

    data augment is wrong

    train_dataset = ISCDataset(
        train_paths,
        NCropsTransform(
            transforms.Compose(aug_moderate),
            transforms.Compose(aug_hard),
            args.ncrops,
        ),
    )
    

    error log: apply_transform() takes from 2 to 3 positional arguments but 5 were given

    opened by AItechnology 5
  • Cannot load state dict for model

    Cannot load state dict for model

    Thanks for your amazing work. But I encounter a problem, when I use checkpoint_0009.pth.tar checkpoint,

    • When I don't remove model = nn.DataParallel(model), I encouter error:
            size mismatch for module.backbone.bn1.weight: copying a param with shape torch.Size([24]) from checkpoint, the shape in current model is 
    torch.Size([64]).
            size mismatch for module.backbone.bn1.bias: copying a param with shape torch.Size([24]) from checkpoint, the shape in current model is torch.Size([64]).
            size mismatch for module.backbone.bn1.running_mean: copying a param with shape torch.Size([24]) from checkpoint, the shape in current model is torch.Size([64]).
            size mismatch for module.backbone.bn1.running_var: copying a param with shape torch.Size([24]) from checkpoint, the shape in current model is torch.Size([64]).
            size mismatch for module.fc.weight: copying a param with shape torch.Size([256, 512]) from checkpoint, the shape in current model is torch.Size([256, 2048])
    
    • Then I remove line model = nn.DataParallel(model), the model seems to load checkpoint successfully, but I feed same input to model, the output feature vector if different for different time I run. I guess the model is not loaded successfully when load state dict, so model will use the weight initialized randomly.
    • Then I change strict=True in model.load_state_dict(state_dict=state_dict, strict=False), I encounter error RuntimeError: Error(s) in loading state_dict for ISCNet: Missing key(s) in state_dict:, I found that the key of state_dict in model and checkpoint totally diffrent even name pattern. Key of model state dict and checkpoint state dict I attached below. checkpoint.txt model.txt How can I solve the this problem?
    opened by NguyenThanhAI 2
  • Unable to reproduce Stage 1 results

    Unable to reproduce Stage 1 results

    Hi, I attempted to reproduce the Stage 1 training using your provided code, but was unable to obtain the reported muAP of 0.5831. I instead obtained this result at epoch 9 (indexed from 0):

    Average Precision: 0.49554
    Recall at P90    : 0.32701
    Threshold at P90 : -0.375733
    Recall at rank 1:  0.62448
    Recall at rank 10: 0.65961
    

    I also saw that you continued training from epoch 5, but these are the results I obtained at epoch 5:

    Average Precision: 0.47977
    Recall at P90    : 0.32501
    Threshold at P90 : -0.376619
    Recall at rank 1:  0.61409
    Recall at rank 10: 0.64903
    

    Both sets of results were obtained on the private ground truth set of Phase 1, using image size 512. Is it possible to provide some insight as to what is happening here? Thank you.

    opened by avrilwongaw 1
  • about the train output feature

    about the train output feature

    sorry to bother you again. I want train the model with a small backbone such as resnet50. Because I only have three GPU and I run with command:

    CUDA_VISIBLE_DEVICES=0,1,2 python v83.py  --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --seed 9 \
      --epochs 5 --lr 0.1 --wd 1e-6 --batch-size 96 --ncrops 2 \
      --gem-p 1.0 --pos-margin 0.0 --neg-margin 1.0 \
      --input-size 256 --sample-size 1000000 --memory-size 20000 \
    /root/zhx3/data/fb_train_data/train
    

    I find a strange problem. I test checkpoint_000{0..4}.pth.tar model. only the checkpoint_0002.pth.tar ouput different when the input is different. I mean other model will output same embedding no matter what different you input. thanks in advance. the loss log output such as:

    epoch 5:   0%|          | 0/15873 [00:00<?, ?it/s]=> loading checkpoint './v83/train/checkpoint_0004.pth.tar'
    => loaded checkpoint './v83/train/checkpoint_0004.pth.tar' (epoch 5)
    epoch 6:   0%|          | 0/15873 [00:00<?, ?it/s]epoch=5, loss=1.0154363534772417
    epoch 7:   0%|          | 0/15873 [00:00<?, ?it/s]epoch=6, loss=1.012835873522891
    
    opened by Usernamezhx 1
  • about the memory size

    about the memory size

    python v107.py \
      -a tf_efficientnetv2_m_in21ft1k --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 --seed 99999 \
      --epochs 10 --lr 0.5 --wd 1e-6 \
      --gem-p 1.0 --pos-margin 0.0 --neg-margin 1.1 --weight ./v98/train/checkpoint_0001.pth.tar \
      --input-size 512 --sample-size 1000000 --memory-size 1000 \
      ../input/training_images/
    

    why not set the --memory-size large such as 20000 ? thanks in advance

    opened by Usernamezhx 1
  • will v107 overfit for phase2?

    will v107 overfit for phase2?

    Congratulations and thanks for your sharing.

    i find v107 only use the about 5k query-ref pair (i.e. gt in phase1) as positive. How to know whether it overfits for phase2 ?

    opened by liangzimei 1
  • access denied for dataset on aws

    access denied for dataset on aws

    Thanks for you work! I have problems downloading the dataset from the given aws buckets

    $ aws s3 cp s3://drivendata-competition-fb-isc-data/all/query_images/ input/query_images/ --recursive --no-sign-request
    fatal error: An error occurred (AccessDenied) when calling the ListObjectsV2 operation: Access Denied
    

    Do I need special permissions to download the data?

    opened by sebastianlutter 0
  • Final optimizer state for the model

    Final optimizer state for the model

    Hello @lyakaap

    Thanks a lot for this work. I am trying to take this and finetune over a certain task. Is it possible you can provide the state of final optimizer after 4th stage of training. We want to try an experiment where it will be very useful.

    Thank you.

    opened by shubhamjain0594 11
Owner
lyakaap
Computer Vision, Deep Learning
lyakaap
Script that receives an Image (original) and a set of images to be used as "pixels" in reconstruction of the Original image using the set of images as "pixels"

picinpics Script that receives an Image (original) and a set of images to be used as "pixels" in reconstruction of the Original image using the set of

RodrigoCMoraes 1 Oct 24, 2021
scAR (single-cell Ambient Remover) is a package for data denoising in single-cell omics.

scAR scAR (single cell Ambient Remover) is a package for denoising multiple single cell omics data. It can be used for multiple tasks, such as, sgRNA

19 Nov 28, 2022
AirLoop: Lifelong Loop Closure Detection

AirLoop This repo contains the source code for paper: Dasong Gao, Chen Wang, Sebastian Scherer. "AirLoop: Lifelong Loop Closure Detection." arXiv prep

Chen Wang 53 Jan 03, 2023
EfficientNetv2 TensorRT int8

EfficientNetv2_TensorRT_int8 EfficientNetv2模型实现来自https://github.com/d-li14/efficientnetv2.pytorch 环境配置 ubuntu:18.04 cuda:11.0 cudnn:8.0 tensorrt:7

34 Apr 24, 2022
Decorators for maximizing memory utilization with PyTorch & CUDA

torch-max-mem This package provides decorators for memory utilization maximization with PyTorch and CUDA by starting with a maximum parameter size and

Max Berrendorf 10 May 02, 2022
AdaFocus V2: End-to-End Training of Spatial Dynamic Networks for Video Recognition

AdaFocusV2 This repo contains the official code and pre-trained models for AdaFo

79 Dec 26, 2022
The easiest way to use deep metric learning in your application. Modular, flexible, and extensible. Written in PyTorch.

News December 27: v1.1.0 New loss functions: CentroidTripletLoss and VICRegLoss Mean reciprocal rank + per-class accuracies See the release notes Than

Kevin Musgrave 5k Jan 05, 2023
ComPhy: Compositional Physical Reasoning ofObjects and Events from Videos

ComPhy This repository holds the code for the paper. ComPhy: Compositional Physical Reasoning ofObjects and Events from Videos, (Under review) PDF Pro

29 Dec 29, 2022
A pyparsing-based library for parsing SOQL statements

CONTRIBUTORS WANTED!! Installation pip install python-soql-parser or, with poetry poetry add python-soql-parser Usage from python_soql_parser import p

Kicksaw 0 Jun 07, 2022
Retinal vessel segmentation based on GT-UNet

Retinal vessel segmentation based on GT-UNet Introduction This project is a retinal blood vessel segmentation code based on UNet-like Group Transforme

Kent0n 27 Dec 18, 2022
When are Iterative GPs Numerically Accurate?

When are Iterative GPs Numerically Accurate? This is a code repository for the paper "When are Iterative GPs Numerically Accurate?" by Wesley Maddox,

Wesley Maddox 1 Jan 06, 2022
Code image classification of MNIST dataset using different architectures: simple linear NN, autoencoder, and highway network

Deep Learning for image classification pip install -r http://webia.lip6.fr/~baskiotisn/requirements-amal.txt Train an autoencoder python3 train_auto

Hector Kohler 0 Mar 30, 2022
This is the repository for Learning to Generate Piano Music With Sustain Pedals

SusPedal-Gen This is the official repository of Learning to Generate Piano Music With Sustain Pedals Demo Page Dataset The dataset used in this projec

Joann Ching 12 Sep 02, 2022
SAPIEN Manipulation Skill Benchmark

ManiSkill Benchmark SAPIEN Manipulation Skill Benchmark (abbreviated as ManiSkill, pronounced as "Many Skill") is a large-scale learning-from-demonstr

Hao Su's Lab, UCSD 107 Jan 08, 2023
Codes to pre-train T5 (Text-to-Text Transfer Transformer) models pre-trained on Japanese web texts

t5-japanese Codes to pre-train T5 (Text-to-Text Transfer Transformer) models pre-trained on Japanese web texts. The following is a list of models that

Kimio Kuramitsu 1 Dec 13, 2021
Wordle Env: A Daily Word Environment for Reinforcement Learning

Wordle Env: A Daily Word Environment for Reinforcement Learning Setup Steps: git pull [email&#

2 Mar 28, 2022
Official Code for "Constrained Mean Shift Using Distant Yet Related Neighbors for Representation Learning"

CMSF Official Code for "Constrained Mean Shift Using Distant Yet Related Neighbors for Representation Learning" Requirements Python = 3.7.6 PyTorch

4 Nov 25, 2022
Adversarial Robustness Toolbox (ART) - Python Library for Machine Learning Security - Evasion, Poisoning, Extraction, Inference - Red and Blue Teams

Adversarial Robustness Toolbox (ART) is a Python library for Machine Learning Security. ART provides tools that enable developers and researchers to defend and evaluate Machine Learning models and ap

3.4k Jan 04, 2023
Curvlearn, a Tensorflow based non-Euclidean deep learning framework.

English | 简体中文 Why Non-Euclidean Geometry Considering these simple graph structures shown below. Nodes with same color has 2-hop distance whereas 1-ho

Alibaba 123 Dec 12, 2022
A two-stage U-Net for high-fidelity denoising of historical recordings

A two-stage U-Net for high-fidelity denoising of historical recordings Official repository of the paper (not submitted yet): E. Moliner and V. Välimäk

Eloi Moliner Juanpere 57 Jan 05, 2023