Skip to content

param and model file structure

Jump to bottom
wiki-sync-bot edited this page Sep 10, 2026 · 1 revision

net.param

example

7767517
3 3
Input         input    0 1 data 0=4 1=4 2=1
InnerProduct  ip       1 1 data fc 0=10 1=1 2=80
Softmax       softmax  1 1 fc prob 0=0

overview

[magic]
  • magic number : 7767517
[layer count] [blob count]
  • layer count : count of the layer line follows, should be exactly the count of all layer names
  • blob count : count of all blobs, usually greater than or equals to the layer count

layer line

each layer must occupy exactly one physical line, including all blob names and parameters; do not split a layer across lines or put multiple layers on the same line

Custom DataReader::scan() implementations and C API ncnn_datareader_t::scan callbacks must follow scanf conversion and input consumption rules, including field widths and scansets. Parameter parsing uses %1023[^\r\n] to read up to 1023 characters without skipping leading whitespace or consuming CR/LF, repeating the scan for longer lines. A successful scanset conversion must append a null terminator and return 1. Returning 0 for an unsupported format can be interpreted as an empty parameter list and silently select default parameter values.

[layer type] [layer name] [input count] [output count] [input blobs] [output blobs] [layer specific params]
  • layer type : type name, such as Convolution Softmax etc
  • layer name : name of this layer, must be unique among all layer names
  • input count : count of the blobs this layer needs as input
  • output count : count of the blobs this layer produces as output
  • input blobs : name list of all the input blob names, separated by space, must be unique among input blob names of all layers
  • output blobs : name list of all the output blob names, separated by space, must be unique among output blob names of all layers
  • layer specific params : key=value pair list, separated by space

layer param

0=1 1=2.5 -23303=2,2.0,3.0

key index should be unique in each layer line, pair can be omitted if the default value used

the meaning of existing param key index can be looked up at operation-param-weight-table

  • integer or float key : index 0 ~ 31
  • integer value : int
  • float value : float
  • integer array or float array key : -23300 minus index 0 ~ 31
  • integer array value : [array size],int,int,...,int
  • float array value : [array size],float,float,...,float

Use a decimal point or exponent when generating floating-point scalar values, including integral values, for example 1=6.0, 1=6e0, or 1=0.0. When loading text parameters, the float getter also converts integer spellings such as 1=6 and 1=0 to 6.0f and 0.0f. The int getter does not convert floating-point parameters to integers.

Keep floating-point spellings when converting models with ncnn2mem: binary scalar parameters do not retain integer/float type tags, and the converter writes integer spellings as integer bit patterns without this numeric conversion.

Use a decimal point or exponent for every element of a floating-point array, including integral values, for example -23303=2,1.0,2.0. Mixed integer and float element spellings within an array are not defined by the format.

A zero-length array such as -23300=0 explicitly supplies an empty array. Array getters return that empty array even when a nonempty default is supplied; omitting the parameter returns the default.

In modern ncnn param file

  • array could be represented as 3=2.0,3.0 that is much more human friendly
  • string typed value: 4=hello and the string is no longer than 255

net.bin

  +---------+---------+---------+---------+---------+---------+
  | weight1 | weight2 | weight3 | weight4 | ....... | weightN |
  +---------+---------+---------+---------+---------+---------+
  ^         ^         ^         ^
  0x0      0x80      0x140     0x1C0

the model binary is the concatenation of all weight data, each weight buffer is aligned by 32bit

weight buffer

[flag] (optional)
[raw data]
[padding] (optional)
  • flag : unsigned int, little-endian, indicating the weight storage type, 0 => float32, 0x01306B47 => float16, 0x01348B83 => bfloat16, otherwise => quantized int8, may be omitted if the layer implementation forced the storage type explicitly
  • raw data : raw weight data, little-endian, float32 data or float16 data or bfloat16 data or quantized table and indexes depending on the storage type flag
  • padding : padding space for 32bit alignment, may be omitted if already aligned

Clone this wiki locally