https://github.com/cgarciae/irs2

https://github.com/cgarciae/irs2

Last synced: 10 days ago
JSON representation

• Host: GitHub
• URL: https://github.com/cgarciae/irs2
• Owner: cgarciae
• Created: 2017-09-03T01:29:59.000Z (over 7 years ago)
• Default Branch: master
• Last Pushed: 2017-10-15T17:10:46.000Z (about 7 years ago)
• Last Synced: 2024-11-07T09:39:35.602Z (2 months ago)
• Language: Jupyter Notebook
• Size: 170 MB
• Stars: 0
• Watchers: 6
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# Keras Inception-V4

Keras implementation of Google's inception v4 model with ported weights!

As described in:

[Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning (Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, Alex Alemi)](https://arxiv.org/abs/1602.07261)

Note this Keras implementation tries to follow the [tf.slim definition](https://github.com/tensorflow/models/blob/master/slim/nets/inception_v4.py) as closely as possible.

Pre-Trained weights for this Keras model can be found here (ported from the tf.slim ckpt): https://github.com/kentsommer/keras-inceptionV4/releases

You can evaluate a sample image by performing the following (weights are downloaded automatically):

* ```$ python evaluate_image.py```

```

Loaded Model Weights!

Class is: African elephant, Loxodonta africana

Certainty is: 0.868498

```

# News

5/23/2017:

* Enabled support for both Theano and Tensorflow (again... :neckbeard:)

* Added useful training parameters 

  * l2 regularization added to conv layers

  * Variance Scaling initialization added to conv layers

  * Momentum value updated for batch_norm layers

* Updated pre-processing to match paper (subtracts 0.5 instead of 1.0 :fire:)

* Minor code changes and cleanup is also included in the recent changes

# Performance Metrics (@Top5, @Top1)

Error rate on non-blacklisted subset of ILSVRC2012 Validation Dataset (Single Crop):

* Top@1 Error: 19.54%

* Top@5 Error: 4.88%

These error rates are actually slightly lower than the listed error rates in the paper:

* Top@1 Error: 20.0%

* Top@5 Error: 5.0%