It's been a long time , Never blogged ( Low voice bb, I'm busy with my private work recently ). School is about to begin , harm , I'm going back to work ！！！

In this tutorial, write a pytorch Version of yolov3 testing , Using the flame detection dataset , The effect is as follows ：

So you can make a fire prediction ,yolov3 It's really fragrant , This time it's about github One of the pytorch Implementation version , The effect is good .

that , Next , Just come with me and practice ！！！

One 、 Environmental requirements

Old rules , If a worker wants to do a good job, he must sharpen his tools first , Set up the environment ！！

• Python： 3.7.4
• Tensorflow-GPU 1.14.0
• Keras: 2.2.4
• numpy:1.17.4

It is suggested to use anaconda To quickly build a virtual environment , Very fast ！！！

Two 、 Data set preparation

Collect fire pictures from the Internet , And use labelimg Annotate , Get the labeled image and the location information .

as follows ：

everything , Start coding！！！！

3、 ... and 、Pytorch Version of YoloV3

Pytorch_Yolov3

1. Install the module

stay requirements.txt It contains what you need this time python modular .

• numpy
• torch==1.2.0
• torchvision==0.4.0
• matplotlib
• tensorflow==1.13.2
• tensorboardX==2.0
• terminaltables
• pillow
• tqdm

You can use pip install -i https://pypi.tuna.tsinghua.edu.cn/simple -r requirements.txt To install the required modules .

2. Download the required weight file

Linux Under the platform ,cd weights/, And then run bash download_weights.sh file , You can download the required weight information .

Windows Under the platform , It can be edited directly download_weigths.sh file , Copy the model links in it , Open download in Explorer .

After downloading , stay weights The contents of the document are as follows ：

3. Modify the configuration file

Linux Under the platform , function cd config/ Catalog , And then run bash create_custom_model.sh <num-classes> among Is the class parameter , Modify according to your needs , Here I change it to 1.

Windows Under the platform , To configure git Of bin After the variables in the directory , function sh.exe file . after cd config After the catalog , function sh create_custom_model.sh <num-classes> that will do

After execution , modify custom.data, Modify its configuration information .

4. Configure this time yolov3 Data format

The key is coming. , The key is coming. , The key is coming. ！！！

In the github Next , For custom data , It is not stated that , It's just a stroke . But it's time to yolov3 The required data format for the version of is the same as voc Format and coco The format is not the same . One for each picture txt Label information . The first column is category information , The next four columns are standardized annotation information . among label It's a category in data/custom/classes.names The index of , <> Represents the scaling factor after scaling

• <1>*w = (xmax-xmin)/2 + xmin
• <2>*h = (ymax-ymin)/2 + ymin
• <3> = (xmax-xmin)/w
• <4> = (ymax-ymin)/h

here github No data conversion provided , Two new ones here Annotations and JPEGImages Folder , Will be ready for pictures and xml Tag information in it .

And then run voc2yolov3 file , Generate train.txt and valid.txt file information , Divide the dataset into , Save the image path in two txt In file .

And then run voc_annotation.py Yes xml Tag information for processing , Deal with it as follows txt File form

And remember to modify classes.names The class name of , And copy pictures to images In file . namely

Okay , The data format is finished ！！！

Now you can start training .

5. function train.py

# Training orders 
python train.py --model_def config/yolov3-custom.cfg --data_config config/custom.data --pretrained_weights weights/darknet53.conv.74
# For additional parameters, see train.py file 
# Start training where you left off 
python train.py --model_def config/yolov3-custom.cfg --data_config config/custom.data --pretrained_weights checkpoints/yolov3_ckpt_99.pth --epoch

If there is a warning solution UserWarning: indexing with dtype torch.uint8 is now deprecated, please use a dtype torch.bool instead.

stay model.py Calculate the location of the loss In about 192 Add the following two sentences to the left and right of the line ：

obj_mask=obj_mask.bool() # convert int8 to bool

noobj_mask=noobj_mask.bool() #convert int8 to bool

The running process is shown in the figure ：

It can be done by tensorboard To see what's going on .tensorboard --logdir='logs\'

6. test result

Ding Dong , Ding Dong , It's done right away ！！！

python detect.py --image_folder data/imgs/ --weights_path checkpoints/yolov3_ckpt_99.pth --model_def config/yolov3-custom.cfg --class_path data/custom/classes.names

Run the above , It will be right data/imgs The images under the file are predicted , And save the prediction results to output/imgs Under the document

If it's in GPU Training on your computer , stay CPU It's predicted on the computer , It needs to be modified model.load_state_dict(torch.load(opt.weights_path, map_location='cpu'))

Okay , It's done ！！！ Stand up flag！！！ One more tomorrow ！！！