Visual analysis of ncnn param file and bin model

param file

Content

first line 【7767517】 Version information

The second line 【79 87】layer Number and blob Count

layer Number finger ：input、Convolution、BatchNorm、ReLU... Equal number

In fact, it is to remove the number of remaining lines in the first two lines when opening the file namely 81-2=79

         blob Number finger   The number of structural branches generated in the middle .

bin File Visualization

use netron open

param File with the bin Model network structure correspondence

The specific correspondence is as follows Convolution  layer ：

Convolution   ： Layer type
Conv_0        ： Layer name           
1： Enter the number of data structures
1： Number of output data structures
input ： Network input layer name
123 ： Network output layer name output（ Some structure layer names are replaced by numbers ）
0=64： Number of output channels -num_output
1=7 、11=7： Convolution kernel size -kernel_h/kernel_w
2=1、 12=1 ： expand -dilation_h/dilation_w
3=2、 13=2 ： step -stride_h/stride_w
4=3、 14=3 ：pad-left /pad-right
15=3、 16=3 :pad-top/pad-bottom
5=0 :bias-term
6=9408（7*7*64*3) Total number of parameters k*k*in*out

With key=value Formal description , among 1=7,11=7 Both represent convolution kernels , the second key value +10 To distinguish , step 、pad similar

It is essentially a convolution layer function parameter

 Such as nn.Conv2d function 
 Initialize parameters 
def __init__(
    self,
    in_channels: int,
    out_channels: int,
    kernel_size: _size_2_t,
    stride: _size_2_t = 1,
    padding: Union[str, _size_2_t] = 0,
    dilation: _size_2_t = 1,
    groups: int = 1,
    bias: bool = True,
    padding_mode: str = 'zeros',  # TODO: refine this type
    device=None,
    dtype=None
)