HuaweiAscend
78 lines
3.2 KiB
Markdown
78 lines
3.2 KiB
Markdown
EN|[中文](README.md)
|
|
|
|
# img2bin
|
|
|
|
The img2bin tool generates input data required for model inference in .bin format. Currently, only preprocessing of neural network whose inputs are images is supported.
|
|
There are two types of data. The first is the image data. The second is the input data required by the model (for example, the second input of Faster R-CNN is the image shape).
|
|
Proportional resizing and zero-padding are performed on the images.
|
|
|
|
## Prerequisites
|
|
|
|
- The script can be used in the CentOS and Ubuntu environments and supports only the x86 architecture.
|
|
- The script supports Python 2 and Python 3.
|
|
- If OpenCV-Python is not installed, the script will automatically install it in the initial use.
|
|
|
|
## Obtaining the Script
|
|
|
|
1. By downloading the package
|
|
|
|
Download the script from the **tools** repository at https://gitee.com/ascend/tools to any directory on the server.
|
|
|
|
For example, the path can be **$HOME/AscendProjects/tools**.
|
|
|
|
2. By running the **git** command
|
|
|
|
Run the following command in the **$HOME/AscendProjects** directory to download code:
|
|
|
|
**git clone https://gitee.com/ascend/tools.git**
|
|
|
|
## Instructions
|
|
|
|
Go to the directory where the script is stored.
|
|
|
|
```
|
|
cd $HOME/AscendProjects/img2bin
|
|
```
|
|
|
|
### First type of data: image
|
|
|
|
- The script preprocesses all images in the directory specified by **-i** based on the parameter settings, and saves the images to the directory specified by **-o** in the name of ***file_name*****.bin**.
|
|
|
|
```
|
|
python2 img2bin.py -i ./images -w 416 -h 416 -f BGR -a NHWC -t uint8 -m [104,117,123] -c [1,1,1] -o ./out
|
|
```
|
|
|
|
```
|
|
python3 img2bin.py -i ./images -w 416 -h 416 -f BGR -a NHWC -t uint8 -m [104,117,123] -c [1,1,1] -o ./out
|
|
```
|
|
|
|
### Second type of data: text
|
|
|
|
- For the second type of data, you need to create a file. **Test.txt** is the file template, where **input_node** is the data, and shape is the image shape.
|
|
- The file name extension must be .txt.
|
|
- Only the **-i**, **-t**, and **-o** parameters are required.
|
|
- **-I** specifies the directory or file path, **-t** specifies the data type, and **-o** specifies the output directory.
|
|
|
|
```
|
|
python2 img2bin.py -i ./test.txt -t uint8 -o ./out
|
|
```
|
|
|
|
```
|
|
python3 img2bin.py -i ./test.txt -t uint8 -o ./out
|
|
```
|
|
|
|
## Parameter Description
|
|
|
|
| Parameter| Description
|
|
|----------|----------
|
|
| -i| Input directory or path. <br>**The directory cannot contain both images and .txt files. Only one type of data can be converted at a time**.
|
|
| -w| Width of the output image
|
|
| -h| Height of the output image
|
|
| -f| Color format of the output image, which can be BGR, RGB, YUV, or GRAY
|
|
| -a| Output image format, which can be NCHW or NHWC
|
|
| -t| Output data type of the image or the second data, which can be float32, uint8, int32, or uint32
|
|
| -m| Mean subtraction. Defaults to **[0,0,0]**. The sequence is consistent with the image color format. <br>When the color format is YUV, set this parameter to **[0,0]**. <br>When the color format is GRAY, set this parameter to **[0]**.
|
|
| -c| Normalization. Defaults to **[1,1,1]**. The sequence is consistent with the image color format. <br>When the color format is YUV, set this parameter to **[1,1]**. <br>When the color format is GRAY, set this parameter to **[1]**.
|
|
| -o| Output directory
|
|
|