yolov5/tutorials/tips_for_best_training_results/

[giscus[bot]]

yolov5/tutorials/tips_for_best_training_results/

Our comprehensive guide provides insights on how to train your YOLOv5 system to get the best mAP. Master dataset preparation, model selection, training settings, and more.

https://docs.ultralytics.com/yolov5/tutorials/tips_for_best_training_results/

[Halafz7]

It would be nice if you add tutorial how to train custom model (such as YOLOv3-quantized or something)

[glenn-jocher]

@Halafz7 I would recommend a visit to https://docs.ultralytics.com/modes/train for training code examples with YOLOv8 :)

[thomasjubin]

hello team,
I am doing a project object dictection on road damages as part of my thesis and pretty new to this field, I have 4 classes and total of close to 6000 images including augmented images, these are the results obtained after training using yoloV5s for 64 epochs using google colab,
YOLOv5s summary: 157 layers, 7020913 parameters, 0 gradients, 15.8 GFLOPs
Class Images Instances P R mAP50 mAP50-95: 100% 12/12 [00:17<00:00, 1.44s/it]
all 731 2466 0.549 0.495 0.475 0.227
Potholes 731 558 0.593 0.579 0.553 0.258
AligatorC 731 800 0.755 0.696 0.739 0.435
LongitudanalC 731 791 0.452 0.425 0.353 0.14
TransverseC 731 317 0.396 0.279 0.254 0.0775

there is a class imbalances for the dataset and instances per class is as follow
pothole: 6000
AligatorC : 11000
longitudanalC: 8000
transverseC:3000

1. How can I further improve the model performance
   2 .The images are resized to 640x640 , but I have images in 3 different resolution. should I train again in the
   respective native resolution to obtain better results or should I use higher model version?
2. How can this be done
   ",Always train from a local dataset. Mounted or network drives like Google Drive will be very slow".

[glenn-jocher]

Hello,

Here are some suggestions to further improve the performance of your model based on your scenario:

1. Addressing Class Imbalance: You might want to try techniques like adjusting the class weights in the loss function or using more advanced sampling methods like SMOTE to balance your dataset during training.

2. Image Resolution: Training at native resolution can help if the dataset includes a lot of fine details that are relevant to your classes. However, this would increase computational costs. Given that you're using Google Colab, resources might be limited, so using a stronger version of YOLO, such as YOLOv5m or YOLOv5l, might provide a better improvement while keeping the resolution at 640x640.

3. Local versus Mounted Dataset: Training from a local dataset helps speed up the training process as data transfer rates are much higher. If you are bound to use online platforms like Colab, ensure your data is packaged neatly in a way that minimizes I/O time, perhaps using zip files and then unzipping them directly into the environment's local storage.

For additional tips, consider reviewing our detailed guide on achieving the best training results with YOLOv5: Tips for Best Training Results.

Good luck with your thesis! Let us know if you have more questions. 🚀

[kishore-work-hard]

My question is how to quantize a yolo v8 model? Pruning and quantization. Is there any tutorials ?

[glenn-jocher]

Currently, YOLOv8 does not natively support model quantization or pruning directly through the provided tools or documentation. For quantization and pruning, you might need to explore third-party libraries or frameworks that support these operations on PyTorch models, such as torch.quantization for quantization and torch.nn.utils.prune for pruning.

For a basic example of applying quantization in PyTorch, you can start with something like this:

import torch
from ultralytics import YOLO

# Load your model
model = YOLO('yolov8n.pt').to('cpu')

# Specify quantization configuration
model.fuse()  # Fuse Conv2d + BatchNorm2d layers if applicable
model.qconfig = torch.quantization.get_default_qconfig('fbgemm')

# Prepare and convert the model to quantized version
torch.quantization.prepare(model, inplace=True)
torch.quantization.convert(model, inplace=True)

This is a basic approach and might need adjustments based on your specific requirements. For more detailed guidance on quantization and pruning, you might need to refer to the PyTorch documentation or other specialized resources. If you have further questions or need more specific advice, feel free to ask!

[beatsea20]

Hey, how can I apply data augmentation to increase the size of dataset before training the model. I was confused about the annotations of those augmented images. Do I have to annotate them again?

[glenn-jocher]

@beatsea20 to apply data augmentation to your dataset before training, you can use libraries like Albumentations or TensorFlow's ImageDataGenerator. These tools can generate augmented images along with their corresponding annotations, so you don't need to manually re-annotate them. YOLOv8 also supports data augmentation directly through its configuration file, where you can specify various augmentation techniques. For detailed guidance, please refer to our data preprocessing guide.

[beatsea20]

@glenn-jocher Thanks, kindly help me with two things.

1. if the model is loaded by 'model = YOLO("yolov8n.yaml")' how do I edit the configuration file?
2. I'm trying to train yolov8n on custom dataset. I tried using both the pretrained model and training from scratch. Both ways the model fails to predict anything on new image.
   Since the dataset is small (150 images), I thought using the pretrained model and higher epochs (as suggested by you in an issues on github) but even then there's no predictions by the model.

for training:
from ultralytics import YOLO
model = YOLO("yolov8n.pt")
results = model.train(data="config.yaml", epochs=100, batch=32, verbose=True)

for prediction:
from ultralytics import YOLO
model = YOLO("\yolo\runs\detect\train4\weights\best.pt")

img_path = "path_to_img"
results = model.predict(source=img_path, save=True)
for result in results:
result.save_txt('\yolo\runs\detect\train4\output.txt', save_conf=True)
boxes = result.boxes
print(boxes) #returns empty lists

[glenn-jocher]

To edit the configuration file when loading the model with YOLO("yolov8n.yaml"), you can modify the YAML file directly before loading it. Ensure you save the changes before using it in your script.

Regarding your training issue, small datasets like yours (150 images) can be challenging. Using pretrained weights and increasing epochs is a good approach, but also consider data augmentation to artificially increase dataset size and variety. Ensure your annotations are accurate and consistent. If the model still fails to predict, try validating your dataset and training process for any inconsistencies. For more detailed guidance, please refer to our data preprocessing guide.

[beatsea20]

@glenn-jocher I'm using the 'ultralytics' python package, so there's no 'yolov8n.yaml' file in the directory.

[glenn-jocher]

If you're using the 'ultralytics' Python package and don't have a 'yolov8n.yaml' file, you can still modify the model configuration directly in your script. For training issues, ensure your dataset is well-annotated and consider using data augmentation to increase its size. If the model fails to predict, verify the annotations and training process. For more detailed guidance, please refer to our data preprocessing guide.

[danegarrido]

If I want to include background images (images without the object of interest) in my custom dataset to fine-tune a YOLO model, should the .txt files containing the labels be empty or not exist at all?

[glenn-jocher]

@danegarrido for background images, you simply omit the .txt files—no label file is required when there are no objects to annotate.