Custom input shape with nvdspreprocess plugin

[izmailoff]

Please provide complete information as applicable to your setup.

• Hardware Platform RTX3090
• DeepStream Version 6.0.1
• JetPack Version (valid for Jetson only)
• TensorRT Version 8.4
• NVIDIA GPU Driver Version 515.43.04
• Issue Type questions / new requirements
• How to reproduce the issue ? (This is for bugs. Including which sample app is using, the configuration files content, the command line used and other details for reproducing)
• Requirement details nvdspreprocess

I’m trying to create a custom input tensor for my model with shape 1;9;256;512. It should be made out of concatenation of 3 frames x 3 channels, hence dimension 9. I need to have all 3 RGB/BGR channels of the first frame to be followed by the next frame, i.e: [R1, G1, B1; R2, G2, B2; R3, G3, B3].

I’ve followed this example: DeepStream 3D Action Recognition App — DeepStream 6.1.1 Release documentation (2D and 3D action recognition). Here is my preprocess config:

[property]
gpu-id=0
enable=1
target-unique-ids=1

network-input-shape=1;9;256;512
processing-width=512
processing-height=256

# 0=NVBUF_MEM_DEFAULT 1=NVBUF_MEM_CUDA_PINNED 2=NVBUF_MEM_CUDA_DEVICE 3=NVBUF_MEM_CUDA_UNIFIED 4=NVBUF_MEM_SURFACE_ARRAY
scaling-pool-memory-type=0
# 0=NvBufSurfTransformCompute_Default 1=NvBufSurfTransformCompute_GPU 2=NvBufSurfTransformCompute_VIC
scaling-pool-compute-hw=0
scaling-buf-pool-size=8
tensor-buf-pool-size=8
scaling-filter=0


# custom-lib-path=/opt/nvidia/deepstream/deepstream-6.0/lib/gst-plugins/libcustom2d_preprocess.so
# custom-tensor-preparation-function=CustomTensorPreparation

#custom-lib-path=/opt/nvidia/deepstream/deepstream-6.0/lib/libnvds_custom_sequence_preprocess.so
custom-lib-path=../custom_sequence_preprocess/libnvds_custom_sequence_preprocess.so
custom-tensor-preparation-function=CustomSequenceTensorPreparation

#0=NCHW 1=NHWC 2=CUSTOM. We have NCHW
network-input-order=2

#0=RGB 1=BGR 2=GRAY
network-color-format=1

tensor-data-type=0
tensor-name=input_1:0

[user-configs]
# pixel-normalization-factor=0.0039215697906911373
channel-scale-factors=0.0039215697906911373;0.0039215697906911373;0.0039215697906911373
#channel-scale-factors=0.007843137;0.007843137;0.007843137
#channel-mean-offsets=127.5;127.5;127.5
# channel-mean-offsets=0;0;0
stride=3
# stride=1
subsample=0

[group-0]
src-ids=0
process-on-roi=0

I’ve tried two libraries to do preprocessing as shown in the config:

• libcustom2d_preprocess.so
• libnvds_custom_sequence_preprocess.so

The first one rejects the config with error stating that dimensions should be 3 when RGB/BGR is being used. The second library runs fine but produces input tensor with channels in a different order from desired one, i.e: [R1, R2, R3; G1, G2, G3; B1, B2, B3]. I have tested by dumping the tensor input into pgie and visualizing it.

How can I make preprocessing plugin to produce tensor in the desired order? Perhaps I need to patch the custom_sequence_preprocess or is there any other way? Maybe I can use some custom color format or any other settings? Some help would be appreciated.

Thank you

[windy60j34]

@izmailoff , deepstream-3d-action-recognitionis a example how to do sequence batching and preprocessing. It support 2 types: 2D=> [N, CxS, H, W] and 3D=>[N, C, S, H, W].
For you case, you might want to check the model input shape, it might be [N, S, H, W, C] or [N, S, H, WxC]. no matter which way, it’s not the default CustomSequenceTensorPreparation supported. you have to modify the custom-lib for your own preprocessing methods.

/opt/nvidia/deepstream/deepstream/sources/apps/sample_apps/deepstream-3d-action-recognition/custom_sequence_preprocess/sequence_image_process.cpp

and set correct input-shape in config
network-input-shape= 1;S,H,W,C

[yingliu]

There is no update from you for a period, assuming this is not an issue anymore.
Hence we are closing this topic. If need further support, please open a new one.
Thanks

[izmailoff]

Thank you for your reply and confirming my assumptions. I’ve looked at the code. It looks like I need to change the cuda *.cu file to rearrange channels.