Implementation:Tencent Ncnn Calibration Image List

Knowledge Sources	ncnn
Domains	Quantization, Model_Optimization
Last Updated	2026-02-09 00:00 GMT

Overview

Pattern for preparing calibration data files consumed by the ncnn2table calibration tool.

Description

This is a Pattern Doc documenting the calibration dataset preparation format expected by ncnn2table. The tool accepts a text file containing newline-delimited paths to calibration images, or comma-separated list files for multi-input models. Non-image calibration data can be provided as .npy files (NumPy format), parsed by the built-in npy.hpp reader.

Usage

Create an imagelist.txt file listing paths to 100-1000 representative images. For multi-input models, create separate list files and pass them comma-separated. For non-image data, generate .npy files with the correct input tensor shape and dtype.

Code Reference

Source Location

• Repository: ncnn
• File: tools/quantize/ncnn2table.cpp
• Lines: L1661-1674 (show_usage documenting expected formats)
• Also: tools/quantize/npy.hpp (NumPy file reader for non-image data)

Signature

# Generate image list file
find images/ -type f > imagelist.txt

# For multi-input models (comma-separated lists)
# input1_list.txt,input2_list.txt

Import

# No import needed — this is a file format specification
# The imagelist.txt is consumed by ncnn2table CLI tool

I/O Contract

Inputs

Name	Type	Required	Description
image files	JPEG/PNG files	Yes	Representative calibration images (100-1000)
.npy files	NumPy binary	No	Alternative for non-image model inputs

Outputs

Name	Type	Description
imagelist.txt	Text file	Newline-delimited list of image paths for ncnn2table

Usage Examples

Create Image List

# Collect calibration images (subset of validation set)
mkdir calibration_images/
cp val_images/img_{001..500}.jpg calibration_images/

# Generate image list
find calibration_images/ -name "*.jpg" -type f > imagelist.txt

# Verify
wc -l imagelist.txt
# 500 imagelist.txt

Multi-Input Model

# For models with two inputs
find rgb_images/ -type f > input1_list.txt
find depth_images/ -type f > input2_list.txt

# Pass both to ncnn2table
ncnn2table model.param model.bin input1_list.txt,input2_list.txt table.txt ...

Related Pages

Implements Principle

• Principle:Tencent_Ncnn_Calibration_Dataset_Preparation