aim-dection-issue / yolov4_tensorflow

implement yolov4 with pure tensorflow and you can training net on your own dataset. YOLOv4的纯tensorflow实现,可训练自己的模型.

YOLOv4_tensorflow | 中文说明

• Implement yolov4 with pure tensorflow
• implemented the part of data augment strategies
• continuous update the code

• [email protected]

introductions

• run the following command.

python val.py

• if have no error, it's ok

convert yolov4.weights to fit our code

• refer to this weights convert file, I'm already converted yolov4.weights to this project.
• you can download yolov4.weights from baidu cloud Extraction code: wm1j
• put yolov4.weights into the "yolo_weights" folder, and run the command.

python convert_weight.py
python test_yolo_weights.py

• the ckpt weights file wound exits in the 'yolo_weights' folder(exists in the baidu cloud)
• and you'll see some images like this, it seems perfect

* * *

• the weights_name.txt contains all model layer's name of the network

train on VOC2007 and VOC2012

• open config.py and modify the voc_root_dir to the root of your VOC dataset, modify the voc_dir_ls to the name of VOC dataset witch you want to train
  
• just like this

path_to_voc_root_dir
        voc_dir_ls[0] (VOC_2007)
        |       |Annotations
        |       |JPEGImages
        |       |...
        voc_dir_ls[1] (VOC_2012)
        |        |Annotations
        |        |JPEGImages
        |        |...
        others_folder_name

• run the command

python train_voc.py

• put test images into voc_test_pic folder, and run the following command after the model training finished.
  

python val_voc.py

• it's the result of our code training(input_size:416*416, batch_size:2, lr:2e-4, optimizer:momentum) for a day(364999 steps), not bad

* *

• all configuration parameters are in the config.py, you can modify them according to your actual situation
• it's the image of loss value, and seems that the lr is too lower(2e-4), we should set it larger.

python show_loss.py 20 300

*

train with own dataset

• The jpg image and the corresponding json file which marked with labelme are stored in the folder ./data/JPEGImages, just like what I do in the ./data/JPEGImages folder
• and then, go to the folder ./data, execute the following python command, it automatically generates label files and train.txt

python generate_labels.py

• excute the python command, to get anchor box

python k_means.py

• open config.py, write the anchor box to line 6, just like this

anchors = 12,19, 19,27, 18,37, 21,38, 23,38, 26,39, 31,38, 39,44, 67,96

• and, now, modify the content in data/train.names to the category name that you need to train, and change the class_num in config.py to your own category number.
• all configuration parameters are in the config.py, you can modify them according to your actual situation
• ok, that's all, execute the command

python train.py

• puts images into test_pic folder, and run the following command after the model training finished.

python val.py

• this image is the result of training 5000 steps (25 minutes) with 123 pictures, it looks not bad.

convert ckpt model to pb model

• open ckpt2pb.py , and modify 'ckpt_file_dir', 'class_num', 'anchors'. run.

python ckpt2pb.py

• you will see a pb model named 'model.pb' in 'ckpt_file_dir'
• if you want to use pb model, you will learn more from 'val_pb.py'
• run python file

python val_pb.py

• you will see the detection results with pb model
• you can download pb model from baidu cloud Extraction code: wm1j

some tips with config.py and train the model

1. the parameters of width and height in config.py should be 608, but i have not a powerful GPU, that is why i set them as 416
2. actually, you can replace mish activation with leaky_relu to save GPU memory, and then you can set batch_size from 2 to 4
3. learning rate do not set too large
4. when the loss value is Nan, please lower your learning rate.

my device

GPU : 1660ti (ASUS) 6G
CPU : i5 9400f
mem : 16GB
os : ubuntu 18.04
cuda: 10.2
cudnn : 7
python : 3.6.9
tensorflow-gpu:1.14.0
numpy : 1.18.1
opencv-python : 4.1.2.30