AI Training System
Csmart-Digit Training
Last updated
Csmart-Digit Training
Last updated
The training system consists of command-line software, without a graphical interface, designed for the creation and refinement of neural networks used in the Csmart-Digit Desktop software. It can be installed on a dedicated computer or, initially, on the same computer used for image capture if the client chooses to conduct training in-house.
However, it is recommended that the installation of the Training System, the execution of training processes, and the export of models to .onnx files be performed in a dedicated environment. This process is not straightforward and requires prior technical knowledge. When conducted by the Csmart team, the procedure follows the diagram presented below:
If the client chooses to perform this process internally, it is recommended to maintain a dedicated environment for the training server, with remote access available. This will allow the Csmart team to provide the necessary support. The model training process typically takes between 8 and 24 uninterrupted hours, depending on the dataset size.
Csmart Training comes with its installer and is designed for installation in a local environment, without requiring internet access except for system updates or downloading base neural networks necessary for developing customized networks.
Minimum Hardware Requirements:
Processor: Intel Core i7-13650HX (13th generation)
RAM: 32GB DDR5 (2x16GB) 5600MT/s
Storage: 2TB PCIe NVMe M.2 SSD
Graphics Card: Nvidia GeForce RTX 4070 12GB GDDR6X
Operating System: Windows 11 or Linux
Conditions for Installation and Use:
The computer must be exclusively dedicated to use as a training server. The requirements for this purpose, as well as the installation location, must be appropriate to ensure the performance and reliability demanded by such intensive computational tasks.
The storage and lifecycle management of models, as well as the datasets used for training, will be the responsibility of the client. Proper dimensioning of storage systems and data backups is recommended. It is important to note that a typical training model uses, on average, 100,000 images, but this number can go up to 1 million images per model.
