Question about numbers of Yolo Network parameters

Course Q&A Deep Learning Specialization Convolutional Neural Networks
1

Hi,

On the first programming assignment of the CNN class, we display the model summary of Yolo.

image
image1101×374 49.8 KB

I ave some difficulties to understand the number of parameters.
First, just by looking at the model summary can we compute the number of parameters ? It seems that without kernel size of conv layer for example, we cannot compute the # of parameters.

Second, if I’m not mistaken, the formula for getting the number of parameters is the following:

param_number = output_channel_number * (input_channel_number * kernel_height * kernel_width + 1)

If I apply that here:
param_number = 32 * ( 3 * 3 * 3 +1) = 865

Did I miss something ?

Thank you

2

First, just by looking at the model summary can we compute the number of parameters ? It seems that without kernel size of conv layer for example, we cannot compute the # of parameters.

Under ‘./yad2k/models’, you can find two files, “keras_darknet19.py” and “keras_yolo.py”. You can find parameters for Darknet in “keras_darknet19.py”.

If I apply that here:
param_number = 32 * ( 3 * 3 * 3 +1) = 865

Did I miss something ?

Your formula is right. But, as you see in darknet file, Conv2D is created with “use_bias=False”.
So, number of parameters for the first conv2D layer is

32 x (3 * 3 * 3) = 864

Hope this clarifies.

1 Like
3

Ok thank for the clarification ! So without taking a look at the model, we cannot compute the number of parameters just by looking at the summary right ?