reception_field_visualization

A visual tool for CNN feature map pixel reception field: Constructing the net by operators just like in deep-learning framewok. Then use Visualizer to generate the reception-field image of layer you wanted. A simple example can be find in main function.

Prerequisites

• Python
• OpenCV (pip install opencv-python)

Operator Supported

• conv(input_, kernel=3, stride=1) : Standred Convolution Layer.
• deconv(input_, kernel=3, stride=1) : Standred Deconvolution Layer.
• dilated_conv(input_, rate=1) : Dilated Convolution Layer.

Visualizer

• rect_size: The Pixel num of Rect in Image.
• line_width: The Width of Separator Lines.
• visual(input_): Generate Visible Images and Show.
• save(path): Save Visible Images in 'path'.
• show(): Show Visible Images.

Visualization Result

Convolution Layer (kernel = 3, stride = 1)

Dilated Convolution Layer (rate = [1, 2, 5, 1, 2, 7])

TODO

• User-defined dilated_conv kernel size.
• Wider Painting Color Range.
• Real Reception_field Size Compute API.
• Real Pixel Visual Image.(Bug fix)