Deploying a Custom SSD MobileNet Model on the NVIDIA Jetson Nano

Deploying a Custom SSD MobileNet Model on the NVIDIA Jetson Nano using TensorFlow Object Detection API

Embedded devices, such as NVIDIA Jetson Nano, enables powerful AI-based applications in real-time at very low power and cost. However, deploying deep learning models to these devices requires model optimization, which is not always straightforward. In this post, we explain how we deployed a retrained SSD MobileNet TensorFlow model on an NVIDIA Jetson Nano development kit using the latest version of the TensorFlow Object Detection API. We also investigate the errors that we encountered during the procedure and how we solved each one. The resulting code is available here on the Neuralet’s GitHub repository.

The workflow

NVIDIA TensorRT is a programmable inference accelerator that facilitates high-performance inference on NVIDIA GPUs. TensorRT takes a trained neural network as input and generates a TensorRT engine, a highly optimized runtime engine that performs inference efficiently. TensorRT-based applications perform up to 40x faster than CPU-only platforms during inference. TensorRT SDK provides FP16 and INT8 optimizations to perform low-latency inference with little to no degradation in model accuracy.

TensorRT takes your network definition, performs several optimizations, and generates a TensorRT engine to optimize your model for inference. The optimizations include platform-specific optimizations, layer optimizations, operation optimizations, etc. Visit TensorRT Developer Guide by NVIDIA to learn more about TensorRT and how it performs optimizations.

TensorRT optimization on GoogLeNet architecture.
Figure 1. The Inception module from the GoogLeNet architecture, before and after TensorRT layer optimizations (source: NVIDIA Developer Blog).

Models trained with TensorFlow can be deployed on Jetson Nano in two ways: you can either use TensorFlow with TensorRT (TF-TRT) or convert the TensorFlow model to UFF (Universal Framework Format) and generate a TensorRT execution engine from that. The full documentation on how to use TensorFlow with TensorRT (TF-TRT) is available here. We will focus on the second approach for this post as it outperforms TF-TRT.

To construct a TensorRT engine, first, download and install the TensorFlow Object Detection API. The easiest way to get this API installed is to use a Docker container. You can run Neuralet’s Docker container by following the instructions here to install TensorFlow Object Detection API with minimal effort.

After getting the API installed, you need to download a trained checkpoint from the TensorFlow object detection model zoo, where you can explore various pre-trained models. For this tutorial, we used the SSD MobileNet V2 COCO model. You can train the network starting from the downloaded checkpoint using your training data. Before starting the training procedure, you might need to make some changes to the corresponding config file (find it here). We changed the number of classes to one object class, i.e., pedestrian to use this model in our Smart Social Distancing application.

When training is completed, export the trained model to a frozen inference graph using the tool. Now we need to: 1- convert the frozen graph to a UFF file and 2- generate a TensorRT execution engine from the UFF file.

How to generate a TensorRT engine and what are the steps required.
Figure 2. The steps required for generating a TensorRT engine.

Constructing a TensorRT engine by following the steps illustrated in Figure 1 may seem straightforward enough, but it is not. The problem is that the latest version of the TensorFlow Object Detection API is not compatible with the tools and configuration files required for UFF parsing and generating a TensorRT engine from a UFF file. Therefore, you have to decide between using a relatively old version of the Object Detection API (from early 2018) without all the new features and optimizations, or use the latest version of the Object Detection API and try to fix all the compatibility issues that will arise during the procedure. Fortunately, it is not that difficult to fix the compatibility issues, so let’s do that.

In the next section, we explain how to perform the TensorRT conversion process and describe the problems we encountered during the conversion process as well as the fix to each one. We were able to successfully convert a custom SSD MobileNet model to UFF and build a TensorRT engine by applying the below fixes to each problem.

Software used:

The following fixes are also tested using the TensorRT version using an NGC Docker container on an AMD64 node with all the steps required.

Generating TensorRT engine

You can use NVIDIA’s convert-to-uff conversion tool to convert a TensorFlow .pb file to UFF. You should call this tool with a -p flag and a config file that includes all custom plugins and graph modification operations that will be embedded in the generated UFF file. You can then load the generated UFF file into the TensorFlow UFF Parser to build the execution engine.

In this post, we used the single program setup instead of using the convert-to-uff tool and referred to the tensorrt_demos as a base for our experiments.

Installing the prerequisites

To perform the conversion process, you need to have Jetpack 4.3, TensorFlow 1.15, and PyCuda installed on the Jetson Nano. You can either install each requirement separately or run Neuralet’s Docker container to get all the required dependencies installed on your device in just a few minutes.

This Docker container installs all the prerequisites, including PyCuda, and runs the script from the tensorrt_demos repository to create a symbolic link to and patch the Graph Surgeon. If you want to install the dependencies separately without running the Docker container, you can use to check if PyCuda is installed on your device correctly. The Docker container will then run the script to generate the TensorRT engine with the specified config file.


To use the Neuralet’s TensorRT Generation Tool, you should first add your custom model specs to the config file that matches your model. We have provided sample config files for SSD MobileNet v1 and v2 here.

You may need to customize the number of classes by changing the num_classes parameter. In our example for the Smart Social Distancing application, this parameter is set to two since the model contains pedestrian and background classes. The model input order should be equal to the “NMS” node input order in the .pbtxt file. Here is an example of how your config file may look like:

Name = ssd_mobilenet_v2_coco
Input = /repo/ssd_mobilenet_v2_coco.pb
; Path to the Input Frozen Inference Graph
TmpUFF = /repo/tmp_ssd_mobilenet_v2_coco.uff
; Path for writing the output UFF model
OutputBin = /repo/TRT_ssd_mobilenet_v2_coco.bin
; Path for writing the output TensorRT Engine
NumberOfClasses = 91
; number of classes plus one for the background class
MinSize = 0.2
MaxSize = 0.95
InputOrder = 1,0,2
; Order of `loc_data`, `conf_data`, and `priorbox_data` of the model, which is set equal to the `NMS` node input order in the `.pbtxt` file
InputDims = 3,300,300
; Input Dimension of the model
DownloadPath =
; Model Download Path

Path = /repo/

Conflicts and errors

Since we wanted to use the latest version of the TensorFlow Object Detection API, there were a few compatibility issues that we needed to take care of. We made a few changes to the script to work around all these issues. If you are curious about how we solved each error, read the rest of this section.

FusedBatchNormV3 error

[TensorRT] ERROR: UffParser: Validator error: FeatureExtractor/MobilenetV2/layer_19_2_Conv2d_4_3x3_s2_256/BatchNorm/FusedBatchNormV3: Unsupported operation _FusedBatchNormV3

The FusedBatchNormV3 operation is introduced in TensorFlow version 1.15.0 and is not supported by the TensorRT 6 UFF parser. We replaced this operation with FusedBatchNorm using Graph Surgeon, as explained here to solve this error.

AddV2 error

[TensorRT] ERROR: UffParser: Validator error: FeatureExtractor/MobilenetV2/expanded_conv_15/add: Unsupported operation _AddV2

Similar to the previous conflict, this error was raised because AddV2 is not supported by the TensorRT 6. To fix this issue, we replaced AddV2 with Add, as shown here.

Cast error

[TensorRT] ERROR: UffParser: Validator error: Cast: Unsupported operation _Cast

This error was raised because the model included an Input operation that is not supported by the TensorRT 6, which is the Cast operation. TensorFlow version 1.15 has replaced this operation with the toFloat operation. Therefore, we should add Cast to the namespace_plugin_map as a custom layer as done below (you can learn more about custom layers here):

namespace_plugin_map = {
"MultipleGridAnchorGenerator": PriorBox,
"Postprocessor": NMS,
"Preprocessor": Input,
"Cast": Input,
"ToFloat": Input,
"image_tensor": Input,
"MultipleGridAnchorGenerator/Concatenate": concat_priorbox, # for 'ssd_mobilenet_v1_coco'
"Concatenate": concat_priorbox,
"concat": concat_box_loc,
"concat_1": concat_box_conf

GridAnchor error

[libprotobuf FATAL /externals/protobuf/aarch64/10.0/include/google/protobuf/repeated_field.h:1408] CHECK failed: (index) < (current_size_):

The UFF parser raised this error because the UFF file did not provide an input element for the GridAnchor node. To work around this problem, we used the Graph Surgeon to define a constant input tensor and set that as the input for the GridAnchor node:

def parse_gridAnchor(graph):

data = np.array([1, 1], dtype=np.float32)
anchor_input = gs.create_node("AnchorInput", "Const", value=data)
graph.find_nodes_by_op("GridAnchor_TRT")[0].input.insert(0, "AnchorInput")

return graph

All of the errors and conflicts above are fixed here on Neuralet’s GitHub repo. You can use the provided Docker container and customize the config file to generate a TensorRT engine on Jetson Nano from your SSD MobileNet model.


Deploying a custom deep learning model on embedded devices is a challenging task. In this tutorial, we went through deploying a custom SSD MobileNet model on Jetson Nano and explained some issues we faced when trying to convert a frozen graph retrained by the latest version of the TensorFlow Object Detection API to a UFF file using TensorRT, as well as the fixes we applied to those problems. Check out Neuralet’s GitHub repo for more.

Further readings

Share This

Leave a Reply

Your email address will not be published. Required fields are marked *

Get In Touch

Have a question? Send us a message and we will respond as soon as possible.