whiteboxlayer 0.1.13.post2

TensorFlow based custom layers

---

White-Box-Layer is a Python module for deep learning built on top of TensorFlow and is distributed under the MIT license.

The project was started in May 2021 by YeongHyeon Park.
This project does not limit for participation.
Contribute now!

Installation

Dependencies

whiteboxlayer requires:

• Numpy: 1.18.5
• Scipy: 1.4.1
• TensorFlow: 2.3.0

User installation

You can install the white-box-layer via simple command as below.

$ pip install whiteboxlayer

Development

We welcome new contributors of all experience levels. The white-box-layer community goals are to be helpful, welcoming, and effective. The Development Guide has detailed information about contributing code, documentation, tests, and more. We've included some basic information in this README.

Example

Example for Convolutional Neural Network

An example of constructing a convolutional neural network is covered. The relevant source code is additionally provided following links.

• Python script
• Jupyter notebook

Define TensorFlow based module

class Neuralnet(tf.Module):

    def __init__(self, **kwargs):
        super(Neuralnet, self).__init__()

        self.who_am_i = kwargs['who_am_i']
        self.dim_h = kwargs['dim_h']
        self.dim_w = kwargs['dim_w']
        self.dim_c = kwargs['dim_c']
        self.num_class = kwargs['num_class']
        self.filters = kwargs['filters']

        self.layer = wbl.Layers()

        self.forward = tf.function(self.__call__)

    @tf.function
    def __call__(self, x, verbose=False):

        logit = self.__nn(x=x, name=self.who_am_i, verbose=verbose)
        y_hat = tf.nn.softmax(logit, name="y_hat")

        return logit, y_hat

    def __nn(self, x, name='neuralnet', verbose=True):

        for idx, _ in enumerate(self.filters[:-1]):
            if(idx == 0): continue
            x = self.layer.conv2d(x=x, stride=1, \
                filter_size=[3, 3, self.filters[idx-1], self.filters[idx]], \
                activation='relu', name='%s-%dconv' %(name, idx), verbose=verbose)
            x = self.layer.maxpool(x=x, ksize=2, strides=2, \
                name='%s-%dmp' %(name, idx), verbose=verbose)

        x = tf.reshape(x, shape=[x.shape[0], -1], name="flat")
        x = self.layer.fully_connected(x=x, c_out=self.filters[-1], \
                activation='relu', name="%s-clf0" %(name), verbose=verbose)
        x = self.layer.fully_connected(x=x, c_out=self.num_class, \
                activation=None, name="%s-clf1" %(name), verbose=verbose)

        return x

Initializing module

model = Neuralnet(\
    who_am_i="CNN", \
    dim_h=28, dim_w=28, dim_c=1, \
    num_class=10, \
    filters=[1, 32, 64, 128])

dummy = tf.zeros((1, model.dim_h, model.dim_w, model.dim_c), dtype=tf.float32)
model.forward(x=dummy, verbose=True)

Results

Conv (CNN-1conv) (1, 28, 28, 1) -> (1, 28, 28, 32)
MaxPool (CNN-1mp) (1, 28, 28, 32) > (1, 14, 14, 32)
Conv (CNN-2conv) (1, 14, 14, 32) -> (1, 14, 14, 64)
MaxPool (CNN-2mp) (1, 14, 14, 64) > (1, 7, 7, 64)
FC (CNN-clf0) (1, 3136) -> (1, 128)
FC (CNN-clf1) (1, 128) -> (1, 10)
Conv (CNN-1conv) (1, 28, 28, 1) -> (1, 28, 28, 32)
MaxPool (CNN-1mp) (1, 28, 28, 32) > (1, 14, 14, 32)
Conv (CNN-2conv) (1, 14, 14, 32) -> (1, 14, 14, 64)
MaxPool (CNN-2mp) (1, 14, 14, 64) > (1, 7, 7, 64)
FC (CNN-clf0) (1, 3136) -> (1, 128)
FC (CNN-clf1) (1, 128) -> (1, 10)