Skip to content

A Tensorflow retrieval (space embedding) baseline. Metric learning baseline on CUB and Stanford Online Products.

License

Notifications You must be signed in to change notification settings

ahmdtaha/tf_retrieval_baseline

Repository files navigation

Tensorflow Retrieval Baseline

This repository provides a retrieval/space embedding baseline using multiple retrieval datasets and ranking losses. This code is based on triplet-reid repos.

Evaluation Metrics

  1. Normalized Mutual Information (NMI)
  2. Recall@K

Deep Fashion In-shop Retrieval Evaluation

All the following experiments assume a training mini-batch of size 60. The architecture employed is the one used in In Defense of the Triplet Loss for Person Re-Identification but ResNet is replaced by a DenseNet169. Optimizer: Adam, Number of iterations = 25K

Method Normalized Margin NMI R@1 R@4 # of classes #samples per class
Semi-Hard Yes 0.2 0.902 87.43 95.42 10 6
Hard-Negative No 1.0 0.904 88.38 95.74 10 6
Lifted Structured No 1.0 0.903 87.32 95.59 10 6
N-Pair Loss No N/A 0.903 89.12 96.13 30 2
Angular Loss Yes N/A 0.8931 84.70 92.32 30 2
Custom Contrastive Loss Yes 1.0 0.826 44.09 67.17 15 4

CUB200-2011 Retrieval Evaluation

Mini-batch size=120. Architecture: Inception_Net V1. Optimizer: Momentum. Number of iterations = 10K

Method Normalized Margin NMI R@1 R@4 # of classes #samples per class
Semi-Hard Yes 0.2 0.587 49.03 73.43 20 6
Hard Negatives No 1.0 0.561 46.55 71.03 20 6
Lifted Structured No 1.0 0.502 35.26 59.82 20 6
N-Pair Loss No N/A 0.573 46.52 59.26 60 2
Angular Loss Yes N/A 0.546 45.50 68.43 60 2
Custom Contrastive Loss Yes 1.0 0.476 37.27 62.39 30 4

Stanford Online Products Retrieval Evaluation

Mini-batch size=120. Architecture: Inception_Net V1. Optimizer: Adam. Number of iterations = 30K

Method Normalized Margin NMI R@1 R@4 # of classes #samples per class
Semi-Hard Yes 0.2 0.893 71.22 81.77 20 6
Hard Negatives No 1.0 0.895 72.03 82.55 20 6
Lifted Structured No 1.0 0.889 68.26 79.72 20 6
N-Pair Loss No N/A 0.893 72.60 82.59 60 2
Angular Loss Yes N/A 0.878 60.30 72.78 60 2
Custom Contrastive Loss Yes 1.0 0.825 19.05 32.28 30 4

Requirements

  • Python 3+ [Tested on 3.4.7 / 3.7]
  • Tensorflow 1 and TF 2.0 [Tested on 1.8 / 1.14 / 2.0]

Code Setup

  1. Update the directories' paths in constants.py.
  2. Use train.py and train_tf2.py for TF 1.X and TF 2.X, respectively.
  3. Use embed.py and embed_tf2.py for TF 1.X and TF 2.X, respectively.
  4. eval.py.

Supported Ranking losses

  • Triplet Loss with hard mining - 'hard_triplet'
  • Triplet Loss with semi-hard mining - 'semi_hard_triplet'
  • Lifted Structure Loss - 'lifted_loss'
  • N-pairs loss - 'npairs_loss'
  • Angular loss - 'angular_loss'
  • Contrastive loss - 'contrastive_loss'

Keep an eye on ranking/__init__.py for new ranking loss

Recommeneded Setting for each loss

Method Setting
Semi-Hard L2-Norm Yes, Margin =0.2
Hard Negatives L2-Norm No , Margin =1.0
Lifted Structured L2-Norm No , Margin =1.0
N-Pair Loss L2-Norm No , Margin =N/A
Angular Loss L2-Norm Yes, Margin =N/A
Custom Contrastive Loss L2-Norm Yes, Margin =1.0

Wiki

  • [Done] Explain the fast contrastive loss sampling procedure
  • [Done] The contrastive loss in the repos is customized to avoid nan during training. When the anchor and positive belong to the same class and the distance between their embeddings is near zero, the derivative turns into nan. Lei Mao provides a nice detailed mathematical explanation for this issue.

TODO

  • [TODO] bash script for train, embed and then eval
  • [TODO] Evaluate space embedding during training.
  • [TODO] After supporting TF 2.0 (eager execution), It become easier to support more losses -- Maybe add Margin loss.

Misc Notes

  • I noticed that some methods depend heavily on training parameters like the optimizer and number of iterations. For example, the semi-hard negative performance drops significantly on CUB-dataset if Adam optimizer is used instead of Momentum! The number of iterations seems also matter for this small dataset.
  • The Tensorflow 2.0 implementation uses more memory even when disabling the eager execution. I tested the code with a smaller batch size -- ten classes and five samples per class. After training for 10K iterations, the performance achieved is NMI=0.54, R@1=42.64, R@4=66.52.

Release History

  • 0.0.1
    • CHANGE: Jan 8, 2020. Update code to support Tensorflow 2.0
    • CHANGE: Dec 31, 2019. Update code to support Tensorflow 1.14
    • First Commit: May 24, 2019. Code tested on Tensorflow 1.8