zhangsong1234

模型是否支持FP16

1

将pth转为onnx，然后将onnx直接转为tensorrt后（FP32），人脸图片修复正常，onnx转tensorrt为 FP16时，推理出的结果变成了一张纯色图片

PyCUDA ERROR: The context stack was not empty upon module cleanup

1

graph.decode 输入的input_path为直播流时，当直播流突然断开，bmf 会coredump: [2024-03-23 04:12:10.668] [info] node:c_ffmpeg_encoder 2 scheduler 1 [2024-03-23 04:14:31.322] [info] node id:0 decode flushing [2024-03-23 04:14:31.322] [info] node id:0 Process node end [2024-03-23 04:14:31.364] [info] node id:0...

## Python Stack ignored Stack trace (most recent call last): #5 Object "/usr/bin/python3.8", at 0x5d6065, in _PyObject_MakeTpCall #4 Object "/usr/bin/python3.8", at 0x5d5498, in PyCFunction_Call #3 Object "/home/dan/zs/cuda118/bmf/output/bmf/lib/_hmp.cpython-38-x86_64-linux-gnu.so", at 0x7fe59311d0f4, in...

[rtsp @ 0x7f7d7a7f8980] max delay reached. need to consume packet [rtsp @ 0x7f7d7a7f8980] RTP: missed 6 packets

3

[rtsp @ 0x7f7d7a7f8980] max delay reached. need to consume packet [rtsp @ 0x7f7d7a7f8980] RTP: missed 2 packets [rtsp @ 0x7f7d7a7f8980] max delay reached. need to consume packet [rtsp @ 0x7f7d7a7f8980]...

trtexec --onnx=truck.onnx --saveEngine=truck-detect.engine --exportProfile=model_gn.json --fp16 --useDLACore=0 --allowGPUFallback --useSpinWait --separateProfileRun > model_gn.log [W] [TRT] Layer '/model.24/Constant_1_output_0' (CONSTANT): Unsupported on DLA. Switching this layer's device type to GPU. [W] [TRT] Layer '(Unnamed...

输入为rtsp直播流时，有时会出现av_read_frame(input_fmt_ctx_, &pkt)返回AVERROR_EOF

3

同时起多个bmf脚本，每个脚本拉取一路直播流，过程中有时会出现av_read_frame(input_fmt_ctx_, &pkt)返回AVERROR_EOF，导致程序终止

使用torch.from_dlpack(vf.reformat(rgb).frame().plane(0))方法，会增加很多显存

2

人脸检测模型中使用torch.from_dlpack(vf.reformat(rgb).frame().plane(0))方法后显存明显增加了很多，输入是4K图片，使用与不使用这个方法显存差了500M左右