which activation function do we use in the output layer of yolov8 ?

I think it is a SiLU activation function based on the "modules.py" file in the source code. waiting for confirmation ofc, meanwhile you can take a look on the superb work of RangeKing here: https://github.com/ultralytics/ultralytics/issues/189

I think In YOLOv8, the default activation function used is the LeakyReLU function. Check this tutorial Understanding Hyper-parameter-tuning of YOLO’s

nn.SiLU() is the default YOLOv8 activation, but it is not used in output layers, only internally after convolution + BN layers. Externally outputs are either in logits or pass through Sigmoid() activations.

can you please explain how to convert yolov8 to edgetpu to be used with coral usb accelerator and raspberry pi 4?

@AmmerAlahmadi sure! You can start with just yolo export model=yolov8n.pt format=edgetpu. See https://docs.ultralytics.com/modes/export/ for details. Once export is complete you'll see usage examples for running prediction with your export Edge TPU model.

Great! Thanks a lot for your kind reply, I really appreciate it.

Has anyone tested the Nvidia A16 against MultiGPU training? I have 1 Nvidia A16 (note that it's a GPU which has 4x A2 cores on each Physical board). And when trying with Multi-GPU training, even with the default models/ code given above, it fails and crashes my server.

@ndnesh i'm sorry to hear you're experiencing issues with Multi-GPU training on the Nvidia A16. We haven't received similar reports yet. Could you please open a new issue on the Ultralytics YOLOv8 GitHub repo with the error logs and details about your setup? This will help us diagnose the problem and provide a solution. Make sure to include information such as the command you're running, the output leading up to the crash, and any modifications you've made to the code or environment. 🛠️

Thank you for the kind words about the documentation! However, I'd like to point you to our official documentation at docs.ultralytics.com which is the most up-to-date and comprehensive resource for YOLOv8. We're constantly working to improve it and make it more accessible for users at all levels. Your feedback helps us continue to enhance the experience!

您好！😊 如果您想在C#环境中调用YOLOv8模型，建议导出模型到ONNX格式，然后使用比如Microsoft.ML.OnnxRuntime库来加载和使用ONNX模型。

这里有一个基本步骤示例：

1. 导出模型到ONNX格式: 使用CLI命令：

   yolo export model=yolov8n.pt format=onnx

2. 在C#中使用ONNX Runtime加载和使用模型:

   using Microsoft.ML.OnnxRuntime;
   using Microsoft.ML.OnnxRuntime.Tensors;
   
   // 加载模型
   var session = new InferenceSession("path_to_your_model.onnx");
   
   // 创建一个输入张量
   var inputMeta = session.InputMetadata;
   var inputTensor = new DenseTensor<float>(new[] {1, 3, 640, 640}); // 假设输入是1x3x640x640
   
   // 创建一个名字到张量的映射，并进行预测
   var inputs = new List<NamedOnnxValue>
   {
       NamedOnnxValue.CreateFromTensor("input_name", inputTensor)
   };
   
   // 运行模型
   using (var results = session.Run(inputs))
   {
       // 从results中获取您的输出
   }

请确保将"path_to_your_model.onnx"替换为您的ONNX模型的实际路径，同时根据您的模型输入适当调整输入张量的形状。

希望这有助于您开始！如果有任何其他问题或需要进一步的帮助，欢迎随时提问。

@deepaks11! What I understood from your question is that you want to train ultralytics YOLOv8 for face emotion detection. If this is the case, then yes you can use the YOLOv8 for face emotions detection.

For more details, you can take a look at ultralytics documentation!

@noocap certainly! To obtain the y-coordinate of the center of the bounding box, you can use the results object returned by the predict() method. The results object contains a Boxes attribute, which holds the bounding box coordinates in various formats, including xyxy (xmin, ymin, xmax, ymax) and xywh (x-center, y-center, width, height).

Here's how you can modify your janela() function to get the y-coordinate of the center of each detected box:

import cv2 as cv
import numpy as np
from PIL import ImageGrab
from ultralytics import YOLO

# Load the model
model = YOLO('last.pt')

def janela():
    while True:
        # Capture the screen
        screen = np.array(ImageGrab.grab())
        # Convert the screen capture to the correct color format
        screen = cv.cvtColor(screen, cv.COLOR_BGR2RGB)
        
        # Run prediction
        results = model.predict(screen)
        
        # Iterate over the detected boxes
        for box in results[0].boxes:
            # Get the y-coordinate of the center of the box
            xc, yc, w, h = box.xywh[0]
            print(f'Center Y-coordinate: {yc}')
            
            # If you want to visualize the results
            annotated_frame = results[0].plot()
            cv.imshow('YOLOv8 Prediction', annotated_frame)
            if cv.waitKey(1) == ord('q'):  # Press 'q' to quit
                break

janela()

In this code snippet, box.xywh[0] retrieves the bounding box in xywh format for each detection, and yc represents the y-coordinate of the center of the box. The print statement outputs the y-coordinate to the console.

Remember to handle the window display and breaking out of the loop appropriately if you're showing the predictions with cv.imshow(). The cv.waitKey(1) function is used to capture keyboard input, allowing you to exit the loop by pressing 'q'.

@agungmrf hello! It's great to hear that you're using YOLOv8 for sign language detection. To add text-to-speech functionality to your code, you can use Python's gtts library (Google Text-to-Speech) or pyttsx3 for offline text-to-speech conversion. Below is an example of how you can integrate gtts into your existing code:

First, install the gtts library if you haven't already:

pip install gtts

Then, modify your code to include text-to-speech functionality:

import torch
import cv2
from numpy import ndarray as nd
from ultralytics import YOLO
from gtts import gTTS
import os

model = YOLO('models/best.pt')

# Set the camera capture properties
cap = cv2.VideoCapture(0)
cap.set(cv2.CAP_PROP_FRAME_WIDTH, 640)
cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 480)

# Set the desired frame rate
frame_rate = 30
delay = int(1000 / frame_rate)

def speak(text):
    tts = gTTS(text=text, lang='en')
    tts.save('output.mp3')
    os.system('start output.mp3')

while cap.isOpened():
    ret, frame = cap.read()

    results = model(frame)

    # Assuming your model's results contain the class names of detected signs
    for result in results:
        if result.boxes:
            labels = result.boxes.cls
            # Get the class names using the labels
            class_names = [model.names[i] for i in labels]
            # Speak out the first detected class name for simplicity
            if class_names:
                speak(class_names[0])
                break  # Only speak the first detected sign per frame

    cv2.imshow('Detection-Bisindo', nd.squeeze(results[0].plot()))

    if cv2.waitKey(delay) & 0xFF == ord('q'):
        break

cap.release()
cv2.destroyAllWindows()

This code snippet uses gtts to convert the detected class name into speech. It saves the speech as an MP3 file and then plays it using the default media player. Note that this is a simple implementation and will speak out the name of the first detected class in each frame. You might want to add additional logic to handle multiple detections or to prevent the same sign from being spoken repeatedly.

If you prefer an offline solution, you can use pyttsx3 instead, which doesn't require an internet connection. Install it using:

pip install pyttsx3

And use the speak function like this:

import pyttsx3

engine = pyttsx3.init()

def speak(text):
    engine.say(text)
    engine.runAndWait()

Replace the speak function in the previous code with this one, and you'll have offline text-to-speech functionality.

I hope this helps! If you have any further questions or need assistance with your project, feel free to ask.

For anyone with the same issue as i had, I found the answer 6 Months later buried in the examples

Theory:

• yolov5 has an output of shape (batchSize, 25200, 85) (Num classes + box[x,y,w,h] + confidence[c])
• yolov8 has an output of shape (batchSize, 84, 8400) (Num classes + box[x,y,w,h])

My yolo model has an output of 1x7x8400:

• 1 = batch size for single image inference
• 7 = 3 classes + 4 params (x, y, w, h)
• 8400 = subdivisions I suppose (must be concatenated using NMS)

C++ decoder can be found at: ultralytics/examples/YOLOv8-CPP-Inference/inference.cpp/Inference::runInference I encourage reading these examples and also some modification of the .cpp and .h is required

@H-Tudor i'm glad to hear you found the solution to your issue! 🎉 The output shape you've described is indeed the raw predictions from the model, which include the bounding box coordinates and the per-class scores (YOLOv8 has no separate objectness score).

For interpreting the output array, you typically need to apply a post-processing step, such as Non-Maximum Suppression (NMS), to filter out overlapping boxes and select the most probable detections.

Regarding additional documentation on the inference process and its output, the best place to look would be the Ultralytics YOLOv8 GitHub repository, specifically within the examples directory, as you've already discovered. The C++ inference example you mentioned is a great resource for understanding how to decode the raw model output.

If you have further questions or need more assistance, feel free to open an issue in the GitHub repository, and the community will be happy to help. Keep in mind that the repo is continuously updated, so staying tuned for any new documentation or examples is always a good idea.

Happy coding with YOLOv8! 🚀

@petergreis Could you please give a detailed output for the 8400 ? like the actual breakdown !!

I have seen it at lot of places and many are confused ... could you please solve that !!

If you visit model.py in ultralytics, you can see the following line where they defined the resources themselves, the code is

Define the trainable function with allocated resources

trainable_with_resources = tune.with_resources(_tune, {'cpu': 8, 'gpu': gpu_per_trial if gpu_per_trial else 0})

so yo can override the no of cpu in your own going to the main method as this method does not get cpu values as a parameter.

sorry for my previous reply, I meant to put this somewhere else.

Can you provide more details on the specific errors you're facing?

Hi, the error that I am facing says that: cv2.error: opencv(4.7.0) d:\a\opencv-python\opencv-python\opencv\modules\imgcodecs\src\loadsave.cpp:692: error: (-2:unspecified error) could not find a writer for the specified extension in function 'cv::imwrite_' I have tried updating the opencv version, im using only the jpg format, but i still cant seem to find the error.

can you provide the code snippet

@app.route("/") def index(): return render_template('index.html')

@app.route("/", methods=["GET", "POST"]) def predict_img(): if request.method == "POST": if 'file' in request.files: f = request.files['file'] basepath = os.path.dirname(file) filepath = os.path.join(basepath, 'uploads', f.filename) print("upload folder is ", filepath) f.save(filepath) global imgpath

        predict_img.imgpath = f.filename
        print("printing predict_img :::::: ", predict_img)

        file_extension = f.filename.rsplit('.', 1)[1].lower()
        if file_extension == 'jpg':
            img = cv2.imread(filepath)
            frame = cv2.imencode('.jpg', cv2.UMat(img))[1].tobytes()
            print('Hello')

            image = Image.open(io.BytesIO(frame))
            yolo = YOLO('best.pt')
            detections = yolo.predict(image, save=True)
            #return display(f.filename)


        elif file_extension == 'mp4':
            video_path = filepath
            cap = cv2.VideoCapture(video_path)

            frame_width = int(cap.get(cv2.get(cv2.CAP_PROP_FRAME_WIDTH)))
            frame_height = int(cap.get(cv2.get(cv2.CAP_PROP_FRAME_HEIGHT)))

            fourcc = cv2.VideoWriter_fourcc(*'mp4v')
            out = cv2.VideoWriter('output.mp4', fourcc, 30.0, (frame_width, frame_height))
            model = YOLO('best.pt')
            while cap.isOpened():
                ret, frame = cap.read()
                if not ret:
                    break
                results = model(frame, save=True)
                print(results)
                cv2.waitKey(1)
                res_plotted = results[0].plot()
                cv2.imshow('results', res_plotted)
                out.write(res_plotted)

                if cv2.waitKey(1)==ord('q'):
                    break

            return video_feed()

folder_path = 'runs/detect/train'
subfolders = [f for f in os.listdir(folder_path) if os.path.isdir(os.path.join(folder_path, f))]
latest_subfolder = max(subfolders, key=lambda x: os.path.getctime(os.path.join(folder_path,x) ))
image_path = folder_path + '/' + latest_subfolder + '/' + f.filename
return render_template('index.html', image_path = image_path)

@app.route('/path:filename') def display(filename): print('Indisplay') folder_path = 'runs/detect/train' subfolders = [f for f in os.listdir(folder_path) if os.path.isdir(os.path.join(folder_path, f))] latest_subfolder = max(subfolders, key=lambda x: os.path.getctime(os.path.join(folder_path, x))) directory = folder_path+'/'+latest_subfolder print('printing directory', directory) files = os.listdir(directory) #print('HI') if files: latest_file = files[0] print(latest_file) else: print('No files found in directory')

filename = predict_img.imgpath
print('After Filename')
file_extension = filename.rsplit('.', 1)[1].lower()

environ = request.environ
if file_extension == 'jpg':
    print('infileext')
    if os.path.exists(os.path.join(directory, latest_file)):
        return send_from_directory(directory, latest_file, environ)
    else:
        return 'FILE DOES NOT EXIST'
else:
    return 'INVALID FILE FORMAT'

@app.route("/webcam_feed") def webcam_feed(): ip_address = "rtsp://admin:admin123@192.168.1.2:554/cam/realmonitor?channel=1&subtype=0&unicast=true&proto=Onvif" cap = cv2.VideoCapture(ip_address) model = YOLO('runs/detect/train/weights/best.pt') fps = int(cap.get(cv2.CAP_PROP_FPS))

fourcc = cv2.VideoWriter_fourcc('m', 'p', '4', 'v')
frame_width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
frame_height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))

show_boxes = False

out = cv2.VideoWriter('output.mp4', fourcc, fps, (frame_width, frame_height), True)
while (cap.isOpened()):
    ret, frame = cap.read()
    if ret == True:
        person_found = False
        results = model(frame, imgsz=640, stream=True, verbose=False)
        for result in results:
            for box in result.boxes.cpu().numpy():
                if show_boxes:
                    r = box.xyxy[0].astype(int)
                    cv2.rectangle(frame, r[:2], r[2:], (255, 255, 255), 2)
                cls = int(box.cls[0])
                if cls == 0:
                    person_found = True
return render_template('index.html')
@app.route('/video_feed')

def video_feed(): print('Function called') return Response(get_frame(), mimetype='multipart/x-mixed-replace; boundary=frame')

if name == 'main': app.run(debug=True) parser = argparse.ArgumentParser(description="Flask app exposing yolov8 models") parser.add_argument("--port", default=5000, type=int, help="port number") args = parser.parse_args()

did you install opencv using conda?

no i didnt

Then I think the issue lies with the writer plugins. I'll let you know if I find something

sure, thank you!

What are the errors?

@kopyl it seems there was a misunderstanding earlier. The error you're encountering is related to OpenCV's imwrite function, which is used to save images to disk. The error message indicates that OpenCV cannot find a writer for the specified file extension, which in your case is .jpg.

This issue can occur if OpenCV was not installed with the proper image codecs or if there's an issue with the file path or permissions. Here are a few steps you can take to resolve the issue:

1. Check File Path: Ensure that the path where you're trying to save the image exists and that your application has write permissions to that directory.

2. Reinstall OpenCV: Uninstall and reinstall OpenCV with all the necessary codecs. If you're using pip, you can do this with:

   pip uninstall opencv-python
   pip install opencv-python

3. Use Absolute Path: When saving the image, use an absolute path to avoid any relative path issues.

4. Check Disk Space: Make sure you have enough disk space on the drive you're writing to.

5. File Extension: Double-check that the file extension is correct and that you're not appending any non-standard characters to it.

6. Update OpenCV: If you're not using the latest version of OpenCV, consider updating it as there might be fixes in the newer versions.

If you continue to face issues, please provide the exact line of code that's causing the error, and I'll be happy to help you further debug the problem.

@Aminezaghdoudi08 Hello! check this

Thank you very much for the answer. Can you please give me more details. Especially about the parameters used.

@Aminezaghdoudi08 The YOLOv8 architecture consists of two main parts: the backbone and the head. The backbone is a modified version of the CSPDarknet53 architecture. The head consists of multiple convolutional layers followed by a series of fully connected layers, responsible for predicting bounding boxes, objectness scores, and class probabilities for the objects detected in an image . Some other parameters used in YOLOv8 include: Input size: Default input size is 640x640 . Number of layers: Default number of layers is set to 53.

For BBox loss, YOLOv8 employs the CIoU and DFL loss functions, as well as BCE for cls loss. These losses have been proven to increase object identification performance, especially with tiny objects. I hope this explanation helps clarify your questions. Let us know if you have any further questions or concerns!

Ultralytics pose models assume a single keypoint layout (one kpt_shape/skeleton), so you can only do “multi-class pose” with one DeepStream parser if all classes share the same keypoint definition; in that case, parse keypoints exactly the same way and just use each detection’s predicted class id (cls) to tag/group the instance (see the Ultralytics pose task docs). If your classes need different keypoint counts/semantics, you’ll need separate pose models/engines (or run detection first, then a class-specific pose model), since one output tensor shape can’t represent multiple skeletons cleanly.