neural-network-renderer 0.2.3

A usefull CNN/DenseNet visualization tool

neural-network-renderer

The code generating the image is writen in Python. This code generates .tex and .sty files that are directly compiled and deleted once the resulting PDF is available.

Example

Here is an example of the code to generate a simple convolutionnal-network representation:

from neural_network_renderer.architecture import Architecture
from neural_network_renderer.colors import Colors
from neural_network_renderer.layers import Conv, ConvConvRelu, Dense, DottedLines, Input, Pool, Softmax, Spacer

def main():
    Colors.Dense("lightgray")
    Colors.Softmax("lightgray")

    arch = Architecture(4 / 32)
    # first layer
    arch.add(ConvConvRelu(s_filter=64, n_filter=[32, 32], to="(5,0,0)"))
    arch.add(Pool([32, 32, 32]))
    arch.add(Spacer(width=20))

    # second layer
    arch.add(ConvConvRelu(s_filter=32, n_filter=[64, 64]))
    arch.add(Pool([16, 16, 64]))

    arch.add(Spacer())

    # third layer
    arch.add(ConvConvRelu(s_filter=16, n_filter=[64, 64]))
    arch.add(Pool([8, 8, 64], name="last_pool"))
    arch.add(Spacer())

    # GPA
    arch.add(Conv(s_filter=8, n_filter=1, name="gpa", caption="GPA"))

    # flatten
    arch.add(Dense(64, name="flatten", offset="(4,0,0)", caption="Hidden Layer"))
    arch.add(Softmax(5, 5, name="output", caption="Output", offset="(4,0,0)"))

    arch.add(DottedLines("gpa", "flatten"))
    arch.add(DottedLines("flatten", "output"))

    arch.to_pdf("output_file")


if __name__ == "__main__":
    main()