White Papers

29 CheXNet – Inference with Nvidia T4 on Dell EMC PowerEdge R7425
Script Output sample:
On completion, the script prints overall metrics and timing information over the inference
session
Average over 100 runs is 1.44041 ms (host walltime is 1.56217 ms, 99% percentile time is 1.52326).
Average over 100 runs is 1.43143 ms (host walltime is 1.54826 ms, 99% percentile time is 1.50819).
Average over 100 runs is 1.44583 ms (host walltime is 1.56766 ms, 99% percentile time is 1.54211).
Average over 100 runs is 1.43773 ms (host walltime is 1.55612 ms, 99% percentile time is 1.53363).
Average over 100 runs is 1.44332 ms (host walltime is 1.55968 ms, 99% percentile time is 1.51658).
Average over 100 runs is 1.43861 ms (host walltime is 1.56039 ms, 99% percentile time is 1.50253).
Average over 100 runs is 1.43901 ms (host walltime is 1.56038 ms, 99% percentile time is 1.55898).
Average over 100 runs is 1.43517 ms (host walltime is 1.55967 ms, 99% percentile time is 1.51555).
Average over 100 runs is 1.45124 ms (host walltime is 1.57128 ms, 99% percentile time is 1.57366).
Average over 100 runs is 1.4332 ms (host walltime is 1.55241 ms, 99% percentile time is 1.51955).
Average over 100 runs is 1.43537 ms (host walltime is 1.55512 ms, 99% percentile time is 1.50966).
• 󰇡


󰇢 󰇡


󰇛

󰇜
󰇢   󰇡


󰇢    
• Latency (msec): 1.43537
Description of files and parameters used for development:
Description
Script:
trtexec.cpp
Nvidia sample code showing the
optimized inference using
TensorRT C++ API
TensorFlow Frozen Graph:
chexnet_frozen_graph_154177
7429.pb
existing TensorFlow model
TensorFlow UFF file:
chexnet_frozen_graph_154177
7429.uff
existing TensorFlow model
converted to uff format
Input tensor name:
“input_tensor”
Input tensor name
Input tensor dimension
(NCHW):
(3,256,256)
input tensor dimensions for UFF
parser
Output tensor name:
“chexnet_sigmoid_tensor”
Output tensor name for
inference