Capture a video from the camera on a platform (etc. Jetson Xavier) and the compressed and encoded frames are transmitted to another paltform (etc. Jetson TX), where the frames are decoded and reconstructed in real time.
- python 3.6.9
- pycuda 2022.1
- tensorrt 7.1.3.4
- opencv 4.1.1
- grpc 1.48.2
- compressai 1.2.4
- torch 1.9.0
- torchvision 1.10.0
- numpy 1.19.5
- make sure the ip address works.
- make sure
.trt
files match the platform. - Folders
int8/flt_onnx
generate.onnx
models with executingImplicitQ/flt_*.py
directly. - Folders
Int8LIC_*2/127
build Int8 mode TensorRT engines (.trt
files) from.onnx
models, where2
means/2
and folder127
means-127
;build_*_trt.py
builds Float32 mode TensorRT engines from.onnx
models. - The input size is fixed, which is
[1, 3, 256, 256]
. - TX2 platform can not support Int8 mode of TensorRT.
flt_decoder_cmp.py
&flt_encoder_cmp.py
: floating-point pretrained models withflt_*.onnx
&flt_*_TX/XA.trt
.decoder_cmp.py
&encoder_cmp.py
: int8 quantized models withInt8LIC_*.onnx
&Int8LIC_*_TX/XA.trt
.flt_decoder_cmp.py
&decoder_cmp.py
useD
&pD
, whileflt_encoder_cmp.py
&encoder_cmp.py
useE
&pD
.
python3 flt_decoder_cmp.py --ip 192.168.1.188:50051
python3 flt_encoder_cmp.py --ip 192.168.1.188:50051
- The ip addresses on two machines should keep the same.