ConvNet Drawer

Python script for illustrating Convolutional Neural Networks (CNN). Inspired by the draw_convnet project [1].

Models can be visualized via Keras-like (Sequential) model definitions. The result can be saved as SVG file or pptx file!

Requirements

python-pptx (if you want to save models as pptx)

pip install python-pptx

Keras (if you want to convert Keras sequential model)

pip install keras

matplotlib (if you want to save models via matplotlib)

pip install matplotlib

Usage

Write a script to define and save a model. An example of visualizing AlexNet [2] is as follows.

Write and save convnet_drawer.Model

from convnet_drawer import Model, Conv2D, MaxPooling2D, Flatten, Dense
from pptx_util import save_model_to_pptx
from matplotlib_util import save_model_to_file

model = Model(input_shape=(227, 227, 3))
model.add(Conv2D(96, (11, 11), (4, 4)))
model.add(MaxPooling2D((3, 3), strides=(2, 2)))
model.add(Conv2D(256, (5, 5), padding="same"))
model.add(MaxPooling2D((3, 3), strides=(2, 2)))
model.add(Conv2D(384, (3, 3), padding="same"))
model.add(Conv2D(384, (3, 3), padding="same"))
model.add(Conv2D(256, (3, 3), padding="same"))
model.add(MaxPooling2D((3, 3), strides=(2, 2)))
model.add(Flatten())
model.add(Dense(4096))
model.add(Dense(4096))
model.add(Dense(1000))
save as svg file
model.save_fig("example.svg")
save as pptx file
save_model_to_pptx(model, "example.pptx")
save via matplotlib
save_model_to_file(model, "example.pdf")

Result:

The other examples can be found here.

Convert Keras sequential model

Keras sequential model can be converted to

convnet_drawer.Model

from keras_util import convert_drawer_model
from keras_models import AlexNet
from pptx_util import save_model_to_pptx
from matplotlib_util import save_model_to_file

get Keras sequential model
keras_sequential_model = AlexNet.get_model()
model = convert_drawer_model(keras_sequential_model)
save as svg file
model.save_fig("example.svg")
save as pptx file
save_model_to_pptx(model, "example.pptx")
save via matplotlib
save_model_to_file(model, "example.pdf")

Supported Layers

• Conv2D

  • Conv2D(filters=None, kernel_size=None, strides=(1, 1), padding="valid")

  • e.g.

    Conv2D(96, (11, 11), (4, 4)))

• Deconv2D

  • Deconv2D(filters=None, kernel_size=None, strides=(1, 1), padding="valid")

  • e.g.

    Deconv2D(256, (3, 3), (2, 2)))

• MaxPooling2D, AveragePooling2D

  • MaxPooling2D(pool_size=(2, 2), strides=None, padding="valid")

  • e.g.

    MaxPooling2D((3, 3), strides=(2, 2))

  • If

    strides = None

    , stride is set to be

    pool_size

    .
• GlobalAveragePooling2D

  • GlobalAveragePooling2D()

• Flatten

  • Flatten()

• Dense

  • Dense(units)

  • e.g.

    Dense(4096)

Visualization Parameters

Visualization Parameters can be found in config.py. Please adjust these parameters before model definition (see LeNet.py). The most important parameter is

channel_scale = 3 / 5

c

c_

c_ = math.pow(c, channel_scale)

If the maximum channel size is small (e.g. 512), please increase

channel_scale

Check how the other parameters works:

Default Values

theta = - math.pi / 6
ratio = 0.7
bounding_box_margin = 10
inter_layer_margin = 50
text_margin = 10
channel_scale = 3 / 5
text_size = 14
one_dim_width = 4
line_color_feature_map = (0, 0, 0)
line_color_layer = (0, 0, 255)
text_color_feature_map = (0, 0, 0)
text_color_layer = (0, 0, 0)

TODOs

• [x] Implement padding option for Conv2D and Pooling layers.
• [x] Add some effects to Dense layer (and Flatten / GlobalAveragePooling2D).
• [ ] Automatically calibrate the scale of feature maps for better visibility.
• [x] Move hard-coded parameters to a config file or options.
• [x] Refactor Layer classes.
• [x] ~~Draw with matplotlib? for other formats.~~ The model is now directly saved as a pptx file.

Results

LeNet

AlexNet

ZFNet

VGG16

AutoEncoder

AlexNet saved by matplotlib with plt.xkcd()

References

[1] https://github.com/gwding/draw_convnet

[2] A. Krizhevsky, I. Sutskever, and G. E. Hinton, "ImageNet Classification with Deep Convolutional Neural Networks," in Proc. of NIPS, 2012.