What it does

This repo is essentially a cookbook for running standard TensorFlow image classification and object-detection models on NVIDIA Jetson boards (TX1, TX2). It provides download helpers, graph-freezing utilities, and the exact tensorflow.contrib.tensorrt incantations needed to convert FP32 graphs into FP16 TensorRT-optimized ones. The models themselves come straight from the TensorFlow models repository; this project is the glue and the benchmark numbers.

The interesting bit

The value is in the tedious parts someone already suffered through: matching TensorFlow’s checkpoint formats to TensorRT’s graph expectations, and documenting which configs actually work. The README includes hard TX2 benchmark data—Inception v1 drops from 22.9 ms to 7.36 ms, SSD MobileNet from 72.9 ms to 50.5 ms—gathered in MAX-N mode with jetson_clocks.sh, so you know the ceiling, not the fantasy.

Key highlights

• Covers both classification (Inception, ResNet, MobileNet families) and object detection (SSD MobileNet, SSD Inception v2)
• Provides download_* and build_* Python helpers to automate the checkpoint → frozen graph → TRT graph pipeline
• Benchmarks compare vanilla TensorFlow FP32 against TensorRT FP16 on identical TX2 hardware
• Includes Jupyter notebooks for end-to-end classification and detection examples
• Supports custom training workflows by pointing back at TensorFlow models repo checkpoints

Caveats

• Requires JetPack 3.2 and TensorFlow 1.7+ with TensorRT support; the install path involves manual pip wheel downloads from NVIDIA dev forums
• Object detection support is narrow: only SSD MobileNet v1 and SSD Inception v2 are benchmarked, and the README warns that untested configs “may vary” in performance
• Some object-detection config files ship with near-zero NMS thresholds (1e-8), which the README notes can hammer CPU post-processing unless you manually edit them to ~0.3

Verdict

Grab this if you’re already committed to TensorFlow + Jetson and want a verified, copy-paste path to TensorRT optimization. Skip it if you’re on newer hardware (JetPack 4.x+), need PyTorch, or expect turnkey training—this is an inference optimization reference, not a framework.