ssd_inception_v2_coco(custom trained model) about tensorrt_demos

[jkjung-avt]

There are version dependencies among these things, so I wouldn't recommend you to use the latest 'tensorflow/models' code to train your own model.

• Object detection API (the code gets updated very frequently, and the latest code on GitHub probably only works with the latest version of tensorflow
• tensorflow
• UFF & TensorRT

Reference: jkjung-avt/hand-detection-tutorial#24 (comment)

[SahilChachere]

There are version dependencies among these things, so I wouldn't recommend you to use the latest 'tensorflow/models' code to train your own model.

• Object detection API (the code gets updated very frequently, and the latest code on GitHub probably only works with the latest version of tensorflow
• tensorflow
• UFF & TensorRT

Reference: jkjung-avt/hand-detection-tutorial#24 (comment)

@jkjung-avt - I have trained my model in tensorflow==1.14.0 and in another machine, so on jetson nano also i have installed tf==1.14.0, if my tf version in training machine and nano is different will it work?

[jkjung-avt]

Reference: jkjung-avt/hand-detection-tutorial#24 (comment)

One very key thing is to use an older version ('6518c1c') of object detection API. For example,

git clone https://github.com/tensorflow/models.git
cd models
git checkout 6518c1c

I have trained my model in tensorflow==1.14.0 and in another machine, so on jetson nano also i have installed tf==1.14.0, if my tf version in training machine and nano is different will it work?

I've tested the following and it worked. I cannot guarantee other combination would also work though.

• ssd_mobilenet_v1_egohands and ssd_mobilenet_v2_egohands trained with tensorflow-1.12.0
• optimize the 2 models with TensorRT 6 (JetPack-4.3) and tensorflow-1.15.0 on Jetson Nano
• run inference with TensorRT 6 on Jetson Nano

[PythonImageDeveloper]

Hi @jkjung-avt
1- Has the TensorRT 6 more speed than TensorRT 5?
2- You tested the models with jetpack 4.3 and TensorFlow-1.15.0 on jetson. For training you use Tensorflow-1.12.x?

[jkjung-avt]

1. TensorRT 6 engines run slightly faster than TensorRT 5 (on the Jetson Nano I tested).
2. Yes.

[jkjung-avt]

@SahilChachere, alternatively you could check out this post on NVIDIA developer forum: https://devtalk.nvidia.com/default/topic/1069027/tensorrt/parsing-gridanchor-op-_gridanchor_trt-protobuf-repeated_field-h-1408-check-failed-index-lt-current_size_-/post/5415537/#5415537

The author says the fix works for the latest tensorflow object detection API.

[SahilChachere]

@SahilChachere, alternatively you could check out this post on NVIDIA developer forum: https://devtalk.nvidia.com/default/topic/1069027/tensorrt/parsing-gridanchor-op-_gridanchor_trt-protobuf-repeated_field-h-1408-check-failed-index-lt-current_size_-/post/5415537/#5415537

The author says the fix works for the latest TensorFlow object detection API.

@jkjung-avt, will I have to train model again with the latest TensorFlow object detection API? or are there are any tweaks to change/Parameters adjustments(without training possible?) for the already trained model in the latest TensorFlow object detection API?

[jkjung-avt]

I don't have more information about the fix mentioned in that NVIDIA forum post.

On the other hand, if you plan to retrain your SSD model, I suggest you to use my hand-detection-tutorial code. Here's another reference which might help: jkjung-avt/hand-detection-tutorial#24

[SahilChachere]

I don't have more information about the fix mentioned in that NVIDIA forum post.

On the other hand, if you plan to retrain your SSD model, I suggest you to use my hand-detection-tutorial code. Here's another reference which might help: jkjung-avt/hand-detection-tutorial#24

@jkjung-avt yes, the only problem I am facing now is in last step of conversion, .uff and .pbtxt files are written but script stops in the last step of conversion i.e. .bin file.

[SahilChachere]

I don't have more information about the fix mentioned in that NVIDIA forum post.

On the other hand, if you plan to retrain your SSD model, I suggest you to use my hand-detection-tutorial code. Here's another reference which might help: jkjung-avt/hand-detection-tutorial#24

@jkjung-avt - you have mentioned that you trained using tensorflow==1.12.0 but the problem persists with tf==1.12.0 that it is not compatible with Cuda 10.0 and I have installed Cuda 10.0. I will have to downgrade to Cuda version 9.0

[jkjung-avt]

I build tensorflow-1.12.0 on my x86_PC by myself. CUDA 10.0 works.

[SahilChachere]

Thank you for all the help, I have resolved my issue by providing correct input_order = [1, 2, 0] but before that, I trained models using models@6518c1c version only, it doesn't support trained models using - models repo which I was using, need resolve dependency issues.