Skip to content

This is the baseline of Kaggle-dog-breed-classification on Python, Keras, and TensorFlow.

Notifications You must be signed in to change notification settings

fenglf/Kaggle-dog-breed-classification

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

25 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Kaggle-dog-breed-classification

author: fenglf

Star this repository if you find it helpful, thank you!

About

This is the baseline of Kaggle-dog-breed-classification on Python, Keras, and TensorFlow.

Framework

Hardware

  • Geforce GTX 1050ti 4G
  • Intel® Core™ i5-7400 CPU
  • Memory 16G

Implemented in Keras

The repository includes 3 ways to build classifier:

  • single_model_finetune: Classify with single model finetuning.
  • gap_train: Classify with concatenating multiple models.
  • pair_train: Classify with pair of images and pair losses (Category loss plus Binary loss).

Data prepare

You can strongly improved the performance of classifier with external data Stanford dog datasets.

Train

Single_model_finetune

Fine-tune Convolutional Neural Network in Keras with ImageNet Pretrained Models

  • Choose a base model, remove the top layer and initialize it with ImageNet weights no top.
  • Add fully-connected layer to the last layer.
  • Freeze all the base model CNN layers and train the new-added fc layer.
  • Unfreeze some base model layers and finetune.

gap_train

Train with concatenating multiple ImageNet Pretrained Models' bottle-neck feature in Keras.

  • Choose several models, remove the top layer and initialize them with corresponding ImageNet weights no top.
  • Extract their bottle-neck features by data (including train data and test data) moving forward the models.
  • Concatenate the bottle-neck features of train data.
  • Add fully-connected layer to train a classifier.

pair_train

Train with pair of images and pair losses (Category loss plus Binary loss) inspired by the idea in Person Re-id.

  • Choose 2 different models or just one model, remove the top layer and initialize them with corresponding ImageNet weights no top.
  • Input two images, containing same or different (positive or negtive samples) labels, Which means whether two images belong to same class or not. In each batch, we can find some samples with the same class. So we simply swap those samples to construct positive samples.
  • Freeze all the base model(s) CNN layers, train the full connected layers for category and binary classification.
  • Unfreeze some base model(s) layers and finetune.

or

Code

Todos

  • Center loss

Reference

If you find some bug in this code, create an issue or a pull request to fix it, thanks!

Star this repository if you find it helpful, thank you!

About

This is the baseline of Kaggle-dog-breed-classification on Python, Keras, and TensorFlow.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages