Tsinghua Dogs classification with
Deep Metric Learning

1. Introduction

Tsinghua Dogs dataset

Tsinghua Dogs is a fine-grained classification dataset for dogs, over 65% of whose images are collected from people's real life. Each dog breed in the dataset contains at least 200 images and a maximum of 7,449 images. For more info, see dataset's homepage.

Following is the brief information about the dataset:

Number of categories: 130
Number of training images: 65228
Number of validating images: 5200

Variation in Tsinghua Dogs dataset. (a) Great Danes exhibit large variations in appearance, while (b) Norwich terriers and (c) Australian terriers are quite similar to each other. (Source)

Deep metric learning

Deep metric learning (DML) aims to measure the similarity among samples by training a deep neural network and a distance metric such as Euclidean distance or Cosine distance. For fine-grained data, in which the intra-class variances are larger than inter-class variances, DML proves to be useful in classification tasks.

Goal

In this projects, I use deep metric learning to classify dog images in Tsinghua Dogs dataset. Those loss functions are implemented:

I also evaluate models' performance on some common metrics:

2. Benchmarks

Architecture: Resnet-50 for feature extractions.
Embedding size: 128.
Batch size: 48.
Number of epochs: 100.
Online hard negatives mining.
Augmentations:
- Random horizontal flip.
- Random brightness, contrast and saturation.
- Random affine with rotation, scale and translation.

	MAP	[email protected]	[email protected]	[email protected]	Top-5	NMI
Triplet loss	73.85%	74.66%	73.90	73.00%	93.76%	0.82
Proxy-NCA loss	89.10%	90.26%	89.28%	87.76%	99.39%	0.98
Proxy-anchor loss
Soft-triple loss

3. Visualization

Proxy-NCA loss

Confusion matrix on validation set

T-SNE on validation set

Similarity matrix of some images in validation set

Each cell represent the L2 distance between 2 images.
The closer distance to 0 (blue), the more similar.
The larger distance (green), the more dissimilar.

Triplet loss

Confusion matrix on validation set

T-SNE on validation set

Similarity matrix of some images in validation set

Each cell represent the L2 distance between 2 images.
The closer distance to 0 (blue), the more similar.
The larger distance (green), the more dissimilar.

4. Train

4.1 Install dependencies

# Create conda environment
conda create --name dml python=3.7 pip
conda activate dml

# Install pytorch and torchvision
conda install -n dml pytorch torchvision cudatoolkit=10.2 -c pytorch

# Install faiss for indexing and calulcating accuracy
# https://github.com/facebookresearch/faiss
conda install -n dml faiss-gpu cudatoolkit=10.2 -c pytorch

# Install other dependencies
pip install opencv-python tensorboard torch-summary torch_optimizer scikit-learn matplotlib seaborn requests ipdb flake8 pyyaml

4.2 Prepare Tsinghua Dogs dataset

PYTHONPATH=./ python src/scripts/prepare_TsinghuaDogs.py --output_dir data/

Directory data should be like this:

data/
└── TsinghuaDogs
    ├── High-Annotations
    ├── high-resolution
    ├── TrainAndValList
    ├── train
    │   ├── 561-n000127-miniature_pinscher
    │   │   ├── n107028.jpg
    │   │   ├── n107031.jpg
    │   │   ├── ...
    │   │   └── n107218.jp
    │   ├── ...
    │   ├── 806-n000129-papillon
    │   │   ├── n107440.jpg
    │   │   ├── n107451.jpg
    │   │   ├── ...
    │   │   └── n108042.jpg
    └── val
        ├── 561-n000127-miniature_pinscher
        │   ├── n161176.jpg
        │   ├── n161177.jpg
        │   ├── ...
        │   └── n161702.jpe
        ├── ...
        └── 806-n000129-papillon
            ├── n169982.jpg
            ├── n170022.jpg
            ├── ...
            └── n170736.jpeg

4.3 Train model

Train with proxy-nca loss

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=./ python src/main.py --train_dir data/TsinghuaDogs/train --test_dir data/TsinghuaDogs/val --loss proxy_nca --config src/configs/proxy_nca_loss.yaml --checkpoint_root_dir src/checkpoints/proxynca-resnet50

Train with triplet loss

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=./ python src/main.py --train_dir data/TsinghuaDogs/train --test_dir data/TsinghuaDogs/val --loss tripletloss --config src/configs/triplet_loss.yaml --checkpoint_root_dir src/checkpoints/tripletloss-resnet50

Run PYTHONPATH=./ python src/main.py --help for more detail about arguments.

If you want to train on 2 gpus, replace CUDA_VISIBLE_DEVICES=0 with CUDA_VISIBLE_DEVICES=0,1 and so on.

If you encounter out of memory issues, try reducing classes_per_batch and samples_per_class in src/configs/triplet_loss.yaml or batch_size in src/configs/your-loss.yaml

5. Evaluate

To evaluate, directory data should be structured like this:

data/
└── TsinghuaDogs
    ├── train
    │   ├── 561-n000127-miniature_pinscher
    │   │   ├── n107028.jpg
    │   │   ├── n107031.jpg
    │   │   ├── ...
    │   │   └── n107218.jp
    │   ├── ...
    │   ├── 806-n000129-papillon
    │   │   ├── n107440.jpg
    │   │   ├── n107451.jpg
    │   │   ├── ...
    │   │   └── n108042.jpg
    └── val
        ├── 561-n000127-miniature_pinscher
        │   ├── n161176.jpg
        │   ├── n161177.jpg
        │   ├── ...
        │   └── n161702.jpe
        ├── ...
        └── 806-n000129-papillon
            ├── n169982.jpg
            ├── n170022.jpg
            ├── ...
            └── n170736.jpeg

Plot confusion matrix

PYTHONPATH=./ python src/scripts/visualize_confusion_matrix.py --test_images_dir data/TshinghuaDogs/val/ --reference_images_dir data/TshinghuaDogs/train -c src/checkpoints/proxynca-resnet50.pth

Plot T-SNE

PYTHONPATH=./ python src/scripts/visualize_tsne.py --images_dir data/TshinghuaDogs/val/ -c src/checkpoints/proxynca-resnet50.pth

Plot similarity matrix

PYTHONPATH=./ python src/scripts/visualize_similarity.py  --images_dir data/TshinghuaDogs/val/ -c src/checkpoints/proxynca-resnet50.pth

6. Developement

.
├── __init__.py
├── README.md
├── src
│   ├── main.py  # Entry point for training.
│   ├── checkpoints  # Directory to save model's weights while training
│   ├── configs  # Configurations for each loss function
│   │   ├── proxy_nca_loss.yaml
│   │   └── triplet_loss.yaml
│   ├── dataset.py
│   ├── evaluate.py  # Calculate mean average precision, accuracy and NMI score
│   ├── __init__.py
│   ├── logs
│   ├── losses
│   │   ├── __init__.py
│   │   ├── proxy_nca_loss.py
│   │   └── triplet_margin_loss.py
│   ├── models  # Feature extraction models
│   │   ├── __init__.py
│   │   └── resnet.py
│   ├── samplers
│   │   ├── __init__.py
│   │   └── pk_sampler.py  # Sample triplets in each batch for triplet loss
│   ├── scripts
│   │   ├── __init__.py
│   │   ├── prepare_TsinghuaDogs.py  # download and prepare dataset for training and validating
│   │   ├── visualize_confusion_matrix.py
│   │   ├── visualize_similarity.py
│   │   └── visualize_tsne.py
│   ├── trainer.py  # Helper functions for training
│   └── utils.py  # Some utility functions
└── static
    ├── proxynca-resnet50
    │   ├── confusion_matrix.jpg
    │   ├── similarity.jpg
    │   ├── tsne_images.jpg
    │   └── tsne_points.jpg
    └── tripletloss-resnet50
        ├── confusion_matrix.jpg
        ├── similarity.jpg
        ├── tsne_images.jpg
        └── tsne_points.jpg

7. Acknowledgement

@article{Zou2020ThuDogs,
    title={A new dataset of dog breed images and a benchmark for fine-grained classification},
    author={Zou, Ding-Nan and Zhang, Song-Hai and Mu, Tai-Jiang and Zhang, Min},
    journal={Computational Visual Media},
    year={2020},
    url={https://doi.org/10.1007/s41095-020-0184-6}
}

Dogs classification with Deep Metric Learning using some popular losses

Related tags

Overview

Tsinghua Dogs classification with Deep Metric Learning

1. Introduction

Tsinghua Dogs dataset

Deep metric learning

Goal

2. Benchmarks

3. Visualization

Proxy-NCA loss

Confusion matrix on validation set

T-SNE on validation set

Similarity matrix of some images in validation set

Triplet loss

Confusion matrix on validation set

T-SNE on validation set

Similarity matrix of some images in validation set

4. Train

4.1 Install dependencies

4.2 Prepare Tsinghua Dogs dataset

4.3 Train model

5. Evaluate

Plot confusion matrix

Plot T-SNE

Plot similarity matrix

6. Developement

7. Acknowledgement

Owner

QuocThangNguyen

Send text to girlfriend in the morning

The Official PyTorch Implementation of DiscoBox.

Introduction to Statistics and Basics of Mathematics for Data Science - The Hacker's Way

Motion and Shape Capture from Sparse Markers

Simple object detection app with streamlit

📚 Papermill is a tool for parameterizing, executing, and analyzing Jupyter Notebooks.

Cereal box identification in store shelves using computer vision and a single train image per model.

Deep Hedging Demo - An Example of Using Machine Learning for Derivative Pricing.

Multi-Scale Vision Longformer: A New Vision Transformer for High-Resolution Image Encoding

Generic image compressor for machine learning. Pytorch code for our paper "Lossy compression for lossless prediction".

Official Implementation (PyTorch) of "Point Cloud Augmentation with Weighted Local Transformations", ICCV 2021

Neural Caption Generator with Attention

[ECCV 2020] Reimplementation of 3DDFAv2, including face mesh, head pose, landmarks, and more.

A CV toolkit for my papers.

Neural Message Passing for Computer Vision

HODEmu, is both an executable and a python library that is based on Ragagnin 2021 in prep.

LIMEcraft: Handcrafted superpixel selectionand inspection for Visual eXplanations

KoCLIP: Korean port of OpenAI CLIP, in Flax

'Solving the sampling problem of the Sycamore quantum supremacy circuits

PyTorch implementation of "PatchGame: Learning to Signal Mid-level Patches in Referential Games" to appear in NeurIPS 2021

Tsinghua Dogs classification with
Deep Metric Learning