training time

[Clarkwk]

What equipment do you use and how long is the training time？

[Jaraxxus-Me]

It depends on the model. If you just use pure motion-based PVT++, the training only takes several minutes. If you incorporate the visual cues, training is usually ~20 hours on a single A10 GPU. Compared to GPU, I think the tricky part is on CPU, the pre-processing for each batch.