Mininglamp Technology releases free open source TensorBoard.cpp to promote pre-training of large models

Recently, Mininglamp Technology Group has implemented the C interface of TensorBoard, a machine learning visualization tool, which furtherenriches the C-based large model project toolset and enables large model pre-training process monitoring More convenient and efficient, it accelerates the pre-training process of large models in the marketing field. The tool is open source on Github.

TensorBoard is a machine learning visualization tool developed by Google and is often used to monitor various indicators of the machine learning process. Zhao Liang, senior technical director of Mininglamp Technology, said: “In the process of large model training, data monitoring is an important dimension, and TensorBoard visualizes various parameters and results in the model, such as recording the training process of the large model

Loss changes, PPL changes in the verification set, learning rate changes, Token consumption, single-step parameter update delay and other indicators can help analyze the training status, discover problems that arise during the training process and take timely intervention measures. , to improve the training process and effect of large models."

Minglue Technology’s C interface TensorBoard tool page is open source

Previously, TensorBoard only supported the Python language interface. This time, Mininglamp Technology implements TensorBoard through C,

will further enrich the tool set of large model projects based on C,

will greatly improve the efficiency of model training and monitoring, accelerate the model training process, and after rewriting the interface The tool will display training indicators through multi-dimensional data patterns, including scalar, histogram, image, image collection, audio, text and other data patterns. The toolkit is shared through the github project Tensorboard.cpp, helping more researchers and developers to participate in and accelerate the research and development process of large models, and promote the application exploration of artificial intelligence in multiple fields.

Minglue Technology has two open source toolkits on G

ithub: ASR-BlockFormer and tensorboard.cpp

## Minglue Technology Group CTO Hao Jie said: "We must make it under the requirements of more efficiency and lower cost

Large models in the marketing field use adaptive technology to improve the capabilities of large models. A good industry large model needs to have the logic and language smoothness of a general large model. At the same time, it also needs to achieve what general large models do not have. In a certain industry Authenticity and professionalism within or in specific fields. Based on the massive industry data accumulated by Mininglamp Technology over the past 17 years, we start from the actual needs of customers and use the huge data and knowledge base to conduct enhanced training to meet the diversification of customers. Tasks and scenario requirements. With the support of training monitoring visualization tools, we will increase the training speed, discover problems in time, and create a more reliable and better industry model for customers."

The above is the detailed content of Mininglamp Technology releases free open source TensorBoard.cpp to promote pre-training of large models. For more information, please follow other related articles on the PHP Chinese website!