/pycr

Pytorch for Classification and Retrieval.

Primary LanguagePython

PyCR

Pytorch for Classification and Retrieval
Pipeline

Installation

Please refer to get_started.md for installation.

How to use

TODO

Classification

Data

Dataset Load

  • openCV
  • PIL

Sampler

  • Imbalance

Augmentation


BackBone

(fork from TIMM)

  • TIMM

Layer

  • dropblock
  • Non local
  • DCN
  • SKNet
  • CBAM
  • RGA Module

Aggregation

  • GAP
  • GMP
  • GAP + GMP
  • GeM(w/o Fix)
  • SoftPool

Head

  • BNNeck Head
  • Identity Head
  • Reduction Head

Loss

  • CE
  • FocalLoss
  • Norm Softmax
  • Triplet
  • CosFace
  • ArcFace
  • SnapMix
  • EqLoss
  • Circle

Other

  • TensorBoard
  • EMA
  • Multi-Scale Training
  • w/o Multi-GPUs
  • TTA



Retrieval

Dimension Process

  • Normalize
  • PCA

Feature Enhance

  • DBA
  • QE(w/o weight)

Distance Metric(By Faiss)

  • Euclidean
  • Cosine

Re-rank

  • K-reciprocal
  • K-reciprocal++



Step by Step




ToDo

  • SA Modules
  • Noisy Label
  • More Data Augmentaions
  • More Convenient



References

[1] FastReID
[2] PyRetri
[3] mmdetection