maggiez0138
commented
1 year ago Hello, I have not reproduced the income of the batch you mentioned. My test machine environment is as follows：
The model I tested is as follows:
My evaluation script is as follows：
trtexec --loadEngine=./weights/swin_tiny_patch4_window7_224_batch16.engineMy test results are as follows:The specific test screenshots are as follows:
Because, I'm more curious, why are my results so different from yours? same graphics hardware.
Looking forward to your reply and good luck.
Actually, the throughput should be batchsize*(throughput of trtexec output). So the performance is as expected. But my perforce is a little better than you, maybe it is because the GPU workload(GPU is busy with other tasks), or GPU frequency, it is really complexed. Overrall, with bigger batchsize(from 1 to 32), the throughput should increase.
Hello, I have not reproduced the income of the batch you mentioned. My test machine environment is as follows：
The model I tested is as follows:
My evaluation script is as follows：
The specific test screenshots are as follows:
trtexec --loadEngine=./weights/swin_tiny_patch4_window7_224_batch16.engineMy test results are as follows:Because, I'm more curious, why are my results so different from yours? same graphics hardware.
Looking forward to your reply and good luck.