Describe the problem
Hi! I’ve converted a custom tensorflow model to .uff for inference on Jetson device using TensorRT python. This model was trained to detect number of fingers held up.
However, the .uff model always predicts the first class (ZERO) when inferencing with 0.1667 confidence.
I suspect there is something wrong with the way I am pre-processing my input image for the model. Could you please advise?
I have used rescale(1. /255) while training. Applying the same preprocessing during inference does not help.
Code used to run inference on .uff model - https://drive.google.com/open?id=1TeDQ4jtnCZ8_KEMowGfyYK8VCEuWpwu8
Original Tensorflow model (.h5) - https://drive.google.com/open?id=1ooLM1910AlJPsQQXUDgixSzgd7HB0TTY
Frozen TF graph (using Nvidia docs example) - https://drive.google.com/open?id=18CBrLnOLz8EQhL0HDxoPD6p2RjIKJvx8
Final converted .uff model - https://drive.google.com/open?id=1_E1aoY-njy5k8C_m8WVacOAcMyItw-R7
Test image for inference - https://drive.google.com/open?id=17xo09oAmKVPoCJ3FErQFwJGWSQND4zcs
Provide details on the platforms you are using:
Linux distro and version: Ubuntu 18.04 LTS
GPU type: GTX 1060, Jetson Nano, TX2.
nvidia driver version: 425
CUDA version: 10.0
CUDNN version: 7.5
Python version: v3.6.8
Tensorflow version: 1.13.1
TensorRT version: 5GA
If Jetson, OS, hw versions: Jetson Nano Ubuntu with jetpack 4.2