zcdliuwei / efficientnet-tf2

A reusable implementation of EfficientNet in TensorFlow 2.0 and Keras

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool


A TensorFlow 2.0 implementation of EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks, aka EfficientNet.


EfficientNet is still one of the most efficient architectures for image classification. Considering that TensorFlow 2.0 has already hit version beta1, I think that a flexible and reusable implementation of EfficientNet in TF 2.0 might be useful for practitioners.


I implemented a running mean and standard deviation calculation with Welford algorithm, which eliminates the problem of loading the whole dataset into the memory. Normalizer class, calculating the mean and standard deviation, is also used as a preprocessing_function argument to tf.keras.preprocessing.image.ImageDataGenerator.


  1. conda create -n effnet python=3.6.8
  2. conda activate effnet
  3. git clone https://github.com/monatis/effnet-tf2.git
  4. cd efficientnet-tf2
  5. python -m pip install -r requirements.gpu.txt # Change to requirements.cpu.txt if you're not using GPU.


train_dir and validation_dir directories should contain a subdirectory for each class in the dataset. Then run:

  • python train.py --train_dir /path/to/training/images --validation_dir /path/to/validation/images
  • See model/ directory for training output.

run python train.py --help to see all the options.


  • Share model architecture and a training script.
  • Implement export to saved model.
  • Implement command line arguments to configure data augmentation.
  • Share an inference script.
  • Implement mean and STD normalization.
  • Implement confusion matrix.
  • Implement export to TFLite for model inference.
  • Share an example Android app using the exported TFLite model.



ezoic increase your site revenue


A reusable implementation of EfficientNet in TensorFlow 2.0 and Keras


Language:Python 100.0%