How do I download the model in Ultralyrics HUB? Encountered downloading error.

[KDLPro]

Search before asking

• I have searched the HUB issues and discussions and found no similar questions.

Question

Right now, I'm training a model with the help of Ultralytics HUB. Here's the progress so far...

However, I encounter issues when continuing from previous checkpoint as earlier, the notebook was disconnected from the HUB so I tried restarting it:

The checkpoint saved in local storage is in epoch 68 though. How do I solve this?

Additional

No response

[UltralyticsAssistant]

👋 Hello @KDLPro, thank you for raising an issue about the Ultralytics HUB 🚀! Please check out our HUB Docs for more detailed information:

• Quickstart. Begin training and deploying YOLO models with HUB in seconds.
• Models: Training and Exporting. Get guidance on downloading, training, and exporting your models.
• Projects: Creating and Managing. Discover how to manage your models efficiently through projects.

It sounds like you might be encountering some issues with the model checkpoint. If this is a 🐛 Bug Report, please provide screenshots and steps to reproduce the problem to help us assist you better. Ensuring a Minimum Reproducible Example (MRE) will be very helpful.

If this is a ❓ Question, share as much relevant information as possible about your dataset, model, and environment.

This is an automated response, and an Ultralytics engineer will also assist you soon. We appreciate your patience and understanding! 😊

[sergiuwaxmann]

@KDLPro I just looked at your model and it looks like you can resume training from epoch 95. Can you try resuming again?

[KDLPro]

Oh I just resumed the model and it was done! Still it was weird that Roboflow is waiting for connection.

Maybe due to this?

[sergiuwaxmann]

@KDLPro Something is indeed strange with this model.
Do you mind training it again (create new model)?

[KDLPro]

I'll try when I have time. But I do have results for all 100 epochs. Honestly, the app has problems when the system is disconnected and reconnected again.

[KDLPro]

Here are the results btw @sergiuwaxmann:

[KDLPro]

It might be a bug or smthn @sergiuwaxmann related to Internet connection

[sergiuwaxmann]

@KDLPro Indeed, we noticed this as well and will try to improve this feature.
The resume works perfectly with our Cloud Training or if the environment doesn't change (e.g., local training - resumed from the same environment).

[KDLPro]

I see, also apparently the model is missing from my library as well.

[KDLPro]

Also, how do I calculate model accuracy?

[sergiuwaxmann]

@KDLPro What do you mean the model is missing from your library? Model mAP is shown in the model list.

[KDLPro]

It's missing, however I can access through the link directly

[KDLPro]

Ahh so mAP is equivalent to model accuracy?

[pderrenger]

Hello @KDLPro! Yes, mAP (mean Average Precision) is often used as a measure of model accuracy in object detection tasks. It evaluates how well the model predicts bounding boxes and classifies objects.

If you're experiencing issues with your model not appearing in the library, try refreshing the page or clearing your cache. If the problem persists, it might be a temporary glitch. You can still access it directly through the link, which is good.

Feel free to reach out if you have more questions or need further assistance. 😊

[KDLPro]

Gotcha, but which is more commonly used if you talk about model accuracy only? Since there's mAP50 and mAP50-95. Do I have to calculate model accuracy by taking the average of the two values @pderrenger?

[pderrenger]

Hello! When discussing model accuracy in object detection, mAP @KDLPro and mAP@50-95 are both important metrics:

• mAP@50: This measures precision and recall at an IoU threshold of 0.5. It's often used for a quick assessment of model performance.
• mAP@50-95: This is a more comprehensive metric, averaging mAP across multiple IoU thresholds (from 0.5 to 0.95 in increments of 0.05). It provides a more detailed view of the model's accuracy.

Typically, mAP@50-95 is considered a more robust measure of accuracy as it evaluates the model's performance across various levels of overlap. You don't need to average the two; instead, use them to understand different aspects of your model's performance.

If you have further questions, feel free to ask! 😊

[KDLPro]

Gotcha, thank you very much!

[KDLPro]

Also, I have one final question. How would I be able to improve the performance of the model given that there is underfitting in the object loss and minor overfitting in the box loss?

[pderrenger]

Hello! To improve your model's performance, especially with underfitting in object loss and minor overfitting in box loss, consider the following strategies:

1. Data Augmentation: Enhance your dataset with techniques like flipping, rotation, and scaling to increase diversity and help the model generalize better.

2. Learning Rate Adjustment: Experiment with different learning rates. A learning rate that's too high can cause underfitting, while a lower rate might help the model converge better.

3. Regularization: Implement techniques like dropout or L2 regularization to reduce overfitting.

4. Model Architecture: Try using a more complex model if your current one is too simple, or simplify it if it's too complex for your dataset.

5. More Data: If possible, increase the size of your training dataset to provide more examples for the model to learn from.

6. Hyperparameter Tuning: Adjust other hyperparameters such as batch size, optimizer, and epochs to find the optimal configuration.

Feel free to experiment with these suggestions and see which combination works best for your specific case. If you have more questions, just let me know! 😊