Feature Extraction From model

fadwa.fawzy · November 17, 2020, 7:39am

• Hardware Platform (Jetson / GPU) GTX 1650
• DeepStream Version
• TensorRT Version7.0.0.11
• NVIDIA GPU Driver Version (valid for GPU only) 450.51
• Issue Type( questions, new requirements, bugs) question

Hello;

I want to use slowfast_4x16_resnet50_kinetics400 in deepstream but as feature extractor not a classifier. This model is a 400 classes classifier, but I don’t use this output, instead I am using the output of an earlier layer. How can I achieve that in deepstream?.

The model is in MXNET, in my code I call it like that, note that feat_ext=True

from gluoncv.model_zoo import get_model
net = get_model(name='slowfast_4x16_resnet50_kinetics400', nclass=400, pretrained=True, feat_ext=True, num_segments=1, num_crop=1)

I want to do the same thing in deepstream.

AastaLLL · November 17, 2020, 8:35am

Hi,

First, please find the layer name of the feature extractor layer.
And set it as the output in the Deepstream configure.

Then, you can get the Tensor raw data with the similar API as following:

github.com

NVIDIA-AI-IOT/deepstream_python_apps/blob/2931f6b295b58aed15cb29074d13763c0f8d47be/apps/deepstream-ssd-parser/deepstream_ssd_parser.py#L273


      
              user_meta = pyds.NvDsUserMeta.cast(l_user.data)
          except StopIteration:
              break
          
          
if (
                  user_meta.base_meta.meta_type
                  != pyds.NvDsMetaType.NVDSINFER_TENSOR_OUTPUT_META
          ):
              continue
          
          
tensor_meta = pyds.NvDsInferTensorMeta.cast(user_meta.user_meta_data)
          
          
# Boxes in the tensor meta should be in network resolution which is
          # found in tensor_meta.network_info. Use this info to scale boxes to
          # the input frame resolution.
          layers_info = []
          
          
for i in range(tensor_meta.num_output_layers):
              layer = pyds.get_nvds_LayerInfo(tensor_meta, i)
              layers_info.append(layer)

Thanks.

fadwa.fawzy · November 17, 2020, 8:44am

Thank you.

Topic		Replies	Views
Do inference AND get penultimate layer for feature extraction DeepStream SDK	2	254	December 1, 2022
Deepstream app with triton server DeepStream SDK	12	505	June 22, 2022
Convert gluoncv model to engine model DeepStream SDK	8	620	October 12, 2021
How to evaluate Resnet-10 model from the DeepStream reference application DeepStream SDK	14	2718	October 12, 2021
Extract tensor data from secondary inference module DeepStream SDK deepstream	2	595	March 22, 2022
How to get `nvinfer` raw tensor output of a custom model? DeepStream SDK python , deepstream	2	64	November 9, 2024
Does the pyds api for Deepstream 5 expose the raw tensor data of output layers? DeepStream SDK nvbugs	5	1685	October 12, 2021
Retrieve the vector from output_layer.buffer DeepStream SDK	6	902	October 12, 2021
Get class_probabilities_map from segmentation task DeepStream SDK	3	608	October 12, 2021
Reading tensor data in a custom DeepStream parser. (PyTorch ---> ONNX --- > tensorrt.engine --> DeepStream) DeepStream SDK tensorrt , jetson-inference , pytorch , onnx	6	903	July 7, 2022

Feature Extraction From model

Related topics