TfMicro is a micro framework for Tensorflow training similar to Keras but easy and native to Tensorflow. Focus only on creating your TF model. All the rest TfMicro will be done for you. Multiprocessing data template, callbacks, keyboard operations, saving/loading/preloading models and much more.
- Multiprocessing data template: inherit from tfmicro.Data class to feed your data into model with multiprocessing.
- Callbacks
- ModelCheckpoint: to save models by monitor value
- Keyboard stop: to stop training by pressing 'q'
- Validation/KeyboardValidation: to make validation step when you want
- ReducingLearningRate/KeyboardLearningRate: change learning rate on the fly by pressing '+'/'-' keys.
- Testarium: connect Testarium with tfmicro together
- AccuracyCallback: evaluate accuracy by model.logits & model.labels
- FafrCallback: evaluate DET curve and EER on epoch end
- Progress bar and info while training
- Custom user indicators and progress bars
- Custom user keyboard operations: bind any key to any actions
- Saving/loading/preloading models and weights
- Tensorboard automatic support
- Production code templates: training and inference are strong separated
git clone git@github.com:makseq/tfmicro.git
cd tfmicro
python setup.py develop
See example folder.