Dataset used for training sample models

Accelerated Computing Intelligent Video Analytics TAO Toolkit
1

Hi Guys,

I am trying to train a custom 6 class classifier using VOC / COCO dataset. However, I am not able to achieve similar performance as the sample model shared along with the deepstream sample application.

What is the dataset which was used to train the model. Is the framework detectnet_v2 ?

Please help me out.

Thanks.

2

Hi neophyte1,
The model in official DS sample is essentially a DetectNet_v2 model with Resnet10 as backbone.
The training dataset is from NV internal only.

Furthermore, DS model is pruned(6M).So, it is necessary to prune your own model to similar ratio and retrain as TLT mentioned.

3

Hi Morganh,

Thanks for your response. I am trying to train DetectNet_v2 model with Resnet10 backbone. With your help on other threads, I was able to achieve an improvement in average precision values for 3 class detector. However, I do not have a reference to compare the value against. Would it be possible for you to share average precision values achieved by NVIDIA for the sample model shipped with Deepstream SDK ? That would be great help as it will set a target for me to compare the performance of my models against.

Thanks.

4

Hi neophyte1,
Actually I have also no reference result since I have never run the experiment — to reproduce the DS model with a tlt resnet10 detectnet_v2 pretrained model. The Deepstream caffe model is trained using detectnet_v2 with Resnet10 network and internal data. But the training was not performed using TLT because TLT is designed later.
Many users see the performance of DS model and want to train their dataset. So, TLT is designed for end-users to train their own dataset.

5

Hi Morganh,

Would it be possible to know the order of number of images / instances of each class in the dataset?

Also, how was the training of the Deepstream caffe model done? Can you please throw light on that? Can we use the same method too?

Thanks

6

Actually I do not know more details about the generation process about deepstream reanet10 caffe model.
I just know the model in official DS sample is essentially a DetectNet_v2 model with Resnet10 as backbone.
The training dataset is from NV internal only.
I can confirm the training process was not performed via tlt. Because tlt is designed later than DS, as I mentioned in previous comments.