retrain the ssd_mobilenet_v2_coco model with only four classes

I’m a newbie working with jetson tx2 and in the area of machine learning.

following this tutorials https://github.com/NVIDIA-AI-IOT/tf_trt_models#od_train I can run the demo successfully, but I just need four classes of totally 90 classes, for example, ‘person’, ‘car’, ‘dog’, ‘bicycle’, how can I retrain the model ssd_mobilenet_v2 with dataset coco (hopefully by tensorrt optimized). Does anyone has this kind of experiences or any tutorials?

Hi,

Jetson is designed for fast inference so you will need a desktop environment to apply the training work.
For ssd_mobilenet_v2, you can retrain it with the TensorFlow tutorial here:
https://github.com/tensorflow/models/tree/master/research/object_detection

If the desktop environment is not an option for you, it’s recommended to silent the unwanted class as an workaround.
Thanks.

Hi AastaLLL,

Thanks for your reply. I would like to silent the unwanted class temporarily. I downloaded the model ssd_mobilenet_v1_coco from here https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md. File strukture is like this:

saved_model
-----variables
-----saved_model.pb

checkpoint
frozen_inference_graph.pb
model.ckpt.data-00000-of-000001
model.ckpt.index
model.ckpt.meta
pipeline.config

So do you have any idea which file should I modify to comment or silent the unwanted class?

Hi,

Sorry for the late reply.

You can remove the unwanted class in the label.
Location should be ${HOME}/Model/ResNet_18/labels.txt.

Thanks.

Hi, AastaLL

I wanna do the same thing.
Which folder do you mean?
Is this from the Jetson-inference source?
Thank you

Hi Ayuchan,

Please open a new topic for your issue. Thanks