Hi, I would like to use the resnet10.prototxt file in the deepstream sdk to retrain for my own object detection dataset using DIGITS. Are there any detailed instructions that I can use to go about this?
I have already created my own object detection dataset succesfully on DIGITS, I am up to the stage of running training, but am not sure how to convert resnet10.prototxt to a train.prototxt version.
This is a pruned inference model, not suitable for training.
For training, please use NVIDIA Transfer Learning Toolkit. There are pre-trained ResNet 10 models that you can start with.
Just reviewed the TLT documentation and how-to guide. If I understand correctly, I have to train my object dataset with a pre-trained ResNet 10 model, prune it and then export it, if i want to run it on the DeepStream apps. Is that correct?
My main question for now is, [b]in the DeepStream SDK, what is the type of the ResNet 10 model in the sample apps:
- Image Classification, or
- DetectNet_v2, or
- FasterRCNN, or
We will use the same type for ResNet 10 (from NGC) to train our dataset on as well, so it can run on the DeepStream apps later on without any issues.
The pre-trained resnet10 model in ngc is just a pre-trained weight for end user to start to train their own dataset.
You may use pre-trained model or may not use it. Set this in the training spec file.
Note that, the resnet10 model inside deepsteam is not trained with TLT.
TLT is designed later, which is a Python package to enable NVIDIA customers the ability to fine-tune pretrained models with customer’s own data.
The model for DeepStreamSDK is DetectNet_v2.