Table of contents
1. pytorch environment construction
3. Download according to version
1. pytorch environment construction
On the premise of installing anaconda
In the terminal of the compiler pycharm
1. Create a new environment
cuda create -n pytorch1.6_cuda10.2 python=3.7
//Create the compilation environment of python3.7pytorch1.6
2. Activate the environment
conda activate pytorch1.6_cuda10.2
3. Download according to version
conda install pytorch==1.6.0 torchvision==0.7.0 cpuonly -c pytorch
//Only install the cpu version of pytorch1.6
(I’m not sure if you can install a GPU in this place. Try using the CPU first to see the training effect)
The file in txt contains the environment needed now
Set a requirements.txt
pip install -r requirements.txt -i ttps://pypi.doubanio.com/simple/
There is a problem when calling some of the remaining functions or function names.
Directly use pycharm to set some of the available packages directly pip
Or pip install scipy -i Simple Index
Change the thing in red to the function name such as yaml
2. Installation of labelimg
1.win+R and enter cmd to open the terminal
conda activate pytorch1.6_cuda10.2 (activate environment)
2. Direct pip install labelimg
3. Data processing part
1. rename data file
Make sure your file is in jpg or png format. If there is any inconsistency, it cannot be called in the python file.
Data processing part
Correct file path
Before the file is jpg, modify the '.jpg' after if
For JPG files, j=2 is modified and continued after the previous file.
① When labeling, you can use an ordered sequence of images without labeling the pictures, but you must use labelimg to pass all pictures, otherwise the images and annotation will appear at voc_label and cannot correspond.
②Data enhancement can be simply processed through opencv or rotated
Place the folders according to images and annation and follow the yolo steps again.
2. Data enhancement
● Embed the data enhancement module into the model
● Perform data enhancement in the Dataset dataset
Or you can customize functions to enhance the data
Such as: adjusting image saturation
#visualize(image, saturated)
image = tf.expand_dims(images[3]*255, 0)
saturated = tf.image.adjust_saturation(image, 3)
plt.figure(figsize=(8, 8))
for i in range(9):
augmented_image = aug_img(saturated)
ax = plt.subplot(3, 3, i + 1)
plt.imshow(augmented_image[0].numpy().astype("uint8"))
plt.axis("off")
Crop images, etc.
image = tf.expand_dims(images[3]*255, 0)
cropped = tf.image.central_crop(image, central_fraction=0.5)
plt.figure(figsize=(8, 8))
for i in range(9):
augmented_image = aug_img(cropped)
ax = plt.subplot(3, 3, i + 1)
plt.imshow(augmented_image[0].numpy().astype("uint8"))
plt.axis("off")4. Yolov4 training process
1.data in the data
Change the name and data of the class in the file
For example:
6 types of fruits
Apple
Mango
Banana
nongfushanquan
toothbrush
wanglaoji
2. Run the kmeans file in the file
The above problem indicates that the size of the labels is uniform and the k clustering algorithm cannot be used to adjust the k value to 2 or 3 (normally 6)
After running, open the kmeans.txt file
copy result
3. Open yolo-tiny in the cfg file
Shortcut key ctrl+f to search yolo
anchors = 84,112, 273,273,84,112, 273,273,84,112, 273,273
classes=6
Filter of the previous layer (number of categories + 5 ) *3 a priori box
filters=33
There are two places in yolo that need to be changed
4. Run makeTxt.py
Separate training and validation sets
5.voc_label.py
Among them, modify classes as data content
And correct the file path in list_file.write
Run the voc_label.py file
6. You can start the training program
Correct the default content for the current data set and training requirements
Note that you can use /// to change the path or +r in front of the path //
train.py
parser.add_argument('--cfg', type=str, default='cfg/csdarknet53s-BIFPN-spp-CA-GTSDB.cfg', help='model.yaml path')
parser.add_argument('--data', type=str, default='data/GTSDB.yaml', help='data.yaml path')
5. Rent GPU
If you don’t have the training conditions, you can rent an online GPU. The tutorial is as follows:
- Baidu search gpushare
- Create pytorch environment
3. Click File Site Manager to log in
4. Transfer the compressed file to the hy-tmp folder
Put the requirements file in first and save it later.
5.
Open jupyterlab terminal
create environment
create create -n py python=3.7
conda activate py
conda install pytorch==1.7.0 torchvision==0.8.0 torchaudio==0.7.0 cudatoolkit=10.2 -c pytorch
cd /hy-tmp #1s ##Open the folder
pip install -r requirements.txt installs the required files
Remember to add pyyaml and scipy inside
Unzip: unzip 1.zip
Open the folder: cd tea #1s
Run the code python train.py