Customizing Deep Learning FAQ

1.When using Custom Deep Learning Classification, If prompted to "Please go to (service pack address) to download the complete service package," please click the link to download and install it before launching the feature.

• Please click the link to download and install the complete service pack before launching the feature.

2.When using Custom Deep Learning Classification, if prompted "Please update the graphics card driver (version required > 465.89), you can download the latest driver from (graphics card driver address)".

• Please click on the graphics card driver address to download and install the latest version of the graphics card driver. This will resolve the issue.

3.If the software is upgraded online or reinstalled after uninstallation, does the deep learning extension package need to be reinstalled?

• The deep learning extension package needs to be reinstalled.

4.When using Custom Deep Learning Classification, if you get "The code cannot continue because nvml.dll was not found. Reinstalling the program may solve this problem".

• Please first check if the graphics card driver is installed. If you have confirmed that you have installed the latest version of the graphics card driver and the issue still persists, you can add the path C:\Program Files\NVIDIA Corporation\NVSMI (the path may vary based on your graphics card driver installation) to the system environment variable Path.

5.When running the training module, I receive the following message: "GPU driver version is too old. Please update to version 465.89 or above."

• Please upgrade your graphics card driver to at least version 465.89.

6.When running the training module, I get the message "NVIDIA GPU not found or driver not installed."

• Please make sure you have an NVIDIA GPU installed and that the graphics card driver is properly installed

7.When using deep learning related functions, if the following prompt is printed in the log window, but the computer has an NVIDIA graphics card

• To check if there is any problem related to NVIDIA graphics card driver, you can use win+r key combination to hit the run screen of Windows and type cmd to open the command prompt window. In the Command Prompt window, type nvidia-smi and press Enter. If the interface prompts "nvidia-smi is not an internal or external command, nor is it a program that can be run", it proves that the current computer has an NVIDIA graphics card, but the graphics card driver is abnormal, so it is recommended to download NVIDIA drivers from NVIDIA Driver Download page to download the corresponding driver.

8.When labeling training data, should the number of the same category be uniform across all the data?

• The number of the same category must be the same, otherwise a good training model cannot be obtained. If there are data with different categories, you can use the classification function in LiDAR360MLS software to number all the data with the same categories and then train them.

9.If I already have a trained model and now need to continue training it with new data, what should I do?

• It is possible to use both old and new data as training sets and then retrain or use the existing model as a pre-trained model to continue training.

10.When using the Custom Deep Learning Classification module, I choose to train 100 rounds but why does it ends the training early without reaching 100 rounds?

• The module has an early stop function to save training time by stopping training when the mIoU of the validation data is not improved after 10 consecutive rounds of training.

11.When using the Custom Deep Learning Classification module and I terminate the training early, how do I get a report on the training sets?

• There is a button to print the report in the training screen, click Print Report to print after terminating the training.

12.When performing the inference process with a trained model, where can I find the data after inference?

• After inference is finished, the source data will be overwritten. If the source data has labels information, it will be replaced. So it is recommended to backup the inferred data before inference. If it is image data, it can be found in the latest inference folder in the same directory as the image.

13.What are the functions of software related to deep image learning? What is the difference?

• Road signs: Built-in extraction models for China and the United States, as well as support for custom labeling or synthetic samples to train models for other regions.
• Road facilities: Built-in extraction models for China, such as street signs, street lights, facilities, etc., and also supports custom labeling or synthetic samples to train models for other facilities you want to extract.
• Road damage: built-in damage detection such as cracks, manhole covers, potholes and other types of damage detection model, while supporting custom labeling training own damage data model
• Image de-privacy: built-in automatic fuzzy face, license plate model, this task, the current function does not support the use of custom-trained models for the time being
• Automatic MASK: Built-in color assignment used in the mask file to automatically generate mask model, the current function does not support the use of custom-trained models for the time being
• Generic target detection: If the customer wants to identify any feature through the image, for example, facilities or pipeline scenes want to extract tees, bolts, etc., any target detection task, you can customize the labeling, training, model inference, and through the inverse projection point cloud to get to identify the location of the target (point layer storage).
• General semantic segmentation: If the customer wants to identify and segment any features through images, such as various broken segmentation needs, any image segmentation tasks, you can customize the labeling, training, model inference and through the inverse projection point cloud to identify the shape of the target (face layer storage).