Open dnpcs12 opened 1 month ago
Hi, thank you for your interest in our work. I reviewed our evaluation code for speed calculation and compared it with the inference code you provided. The main difference appears to be that we concatenate multiple sequences in a batch while maintaining a sliding window approach. Specifically, we use a sliding window size of 20 for per-frame output, creating multiple sliding windows as a batch, which are then fed into our network. This allows us to fully utilize the GPU memory for inference, albeit with a slight latency.
Hello, thank you for sharing your excellent project! I wanted to use the model in real-time, so I wrote the following code to perform real-time inference using the model you provided. However, the frame rate was much slower, around 0.08 ~ 0.1 s per frame, which is much slower than the 0.74ms mentioned in the paper. Did I do something wrong? I set the window size to 20 and tested it on a Nvidia RTX 3090.![image](https://github.com/user-attachments/assets/6ef66940-a762-4fee-9602-59fffe01526b)