Tefla: Deep Learning library, a Higher level API for TensorFlow

Tefla is built on top of Tensorflow. It provides higher level access to tensorflow's features. Inerface, Easy to build complex models.

Tefla features:

    . Support for data-sets, data-augmentation

    . easy to define complex deep models

    . single and multi GPU training

    . various prediction fnctions including ensembling of models

    . different metrics for performance measurement\

    . custom losses

    . learning rate schedules, polynomial, step, validation_loss based

TensorFlow Installation

Tefla requires Tensorflow(version >=r1.0)

Additional Requirements

pip install scipy
pip install git+https://github.com/lucasb-eyer/pydensecrf.git

Tefla Installation

For the latest stable version:

pip install tefla

for current version installation:

pip install git+https://github.com/n3011/tefla.git

For Developer / TO Work with source and modifying source code:

git clone https://github.com/n3011/tefla.git
cd tefla
export PYTHONPATH=.

Documentation

Tefla Docs

Examples

Many examples available with recent deep learning research integration

1. Gumbel Softmax
2. Unrolled_GAN
3. Spatial Transoformer Network
4. LSTM
5. DATASETS
6. ImageNet classificatrion

Pretrained Weights

1. VGG19 Trained on ImageNet Note: (Model is converted from original Caffe version
2. Inception_Resnet_v2 trained on ImageNet Note: model is converted from original tf version, trained by Google brain

Tefla Models

Recent deep convolutional models are easy to implement using TEFLA

1. Recent Models

Getting Started

1. Its as easy as

>>>from tefla.core.layers import conv2d
>>>convolved = conv2d(input, 48, False, None)

2 a. Data Directory structure for using normal images

|-- Data_Dir
|   |-- training_image_size (eg. training_256, for 256 image size)
|   |-- validation_image_size (eg. validation_256, for 256 image size)
|   |-- training_labels.csv
|   |-- validation_labels.csv

2 b. TFRecords support available using tefla/dataset class

1. [Train v2](https://github.com/n3011/tefla/blob/master/tefla/trainv2.py)

Run training:

python tefla/train.py --model models/alexnet.py --training_cnf models/multiclass_cnf.py --data_dir /path/to/data/dir (as per instructions 2.a)

3. Mnist example gives a overview about Tefla usages

image_size =(32, 32)
crop_size = (28, 28)
def model(is_training, reuse):
    common_args = common_layer_args(is_training, reuse)
    conv_args = make_args(batch_norm=True, activation=prelu, **common_args)
    fc_args = make_args(activation=prelu, **common_args)
    logit_args = make_args(activation=None, **common_args)

    x = input((None, height, width, 1), **common_args)
    x = conv2d(x, 32, name='conv1_1', **conv_args)
    x = conv2d(x, 32, name='conv1_2', **conv_args)
    x = max_pool(x, name='pool1', **common_args)
    x = dropout(x, drop_p=0.25, name='dropout1', **common_args)
    x = fully_connected(x, n_output=128, name='fc1', **fc_args)
    x = dropout(x, drop_p=0.5, name='dropout2', **common_args)
    logits = fully_connected(x, n_output=10, name="logits", **logit_args)
    predictions = softmax(logits, name='predictions', **common_args)

    return end_points(is_training)

training_cnf = {
    'classification': True,
    'validation_scores': [('validation accuracy', util.accuracy_wrapper), ('validation kappa', util.kappa_wrapper)],
    'num_epochs': 50,
    'lr_policy': StepDecayPolicy(
        schedule={
            0: 0.01,
            30: 0.001,
        }
    )
}
util.init_logging('train.log', file_log_level=logging.INFO, console_log_level=logging.INFO)

trainer = SupervisedTrainer(model, training_cnf, classification=training_cnf['classification'])
trainer.fit(data_set, weights_from=None, start_epoch=1, verbose=1, summary_every=10)

Contributions

Welcome to the first release of Tefla, if you find any bug, please report it in the GitHub issues section.

Improvements and requests for new features are more than welcome! Do not hesitate to twist and tweak Tefla, and send pull-requests.

License

MIT License

Note: This project BASE is jointly developed with Artelus team: www.github.com/litan/tefla. Both projects are evolving independently, with a cross-pollination of ideas.