• Hardware Platform (Jetson / GPU) GPU Tesla T4.
• DeepStream Version 5.1
• JetPack Version (valid for Jetson only) –
• TensorRT Version 7.2.2
• NVIDIA GPU Driver Version (valid for GPU only) 460.91.03
• Issue Type( questions, new requirements, bugs) bug
Approbation of a custom secondary classifier model without a custom parsing library causes Caught SIGSEGV
The error was discovered by launching the following pipeline with the custom secondary classifier model A:
gst-launch-1.0 filesrc location="/root/barrier_special_raw.mp4" ! qtdemux ! h264parse ! "video/x-h264,stream-format=byte-stream" ! nvv4l2decoder ! m2.sink_0 nvstreammux name=m2 width=1920 height=1080 batch-size=1 batched-push-timeout=4000000 ! nvinfer config-file-path="/opt/nvidia/deepstream/deepstream-5.1/samples/configs/deepstream-app/config_infer_primary.txt" ! nvtracker tracker-width=480 tracker-height=272 gpu_id=0 ll-lib-file="/opt/nvidia/deepstream/deepstream-5.1/lib/libnvds_mot_iou.so" ! nvinfer config-file-path="/root/models/A/A_ds_config.txt" ! fakesink
After some investigation it was determined that this error is caused by the second nvinfer and emerges in its gstnvinfer_meta_utils.cpp module in void merge_classification_output (GstNvInferObjectHistory & history, GstNvInferObjectInfo &new_result) function because of the unconditional strdup at line 207:
attr.attributeLabel = strdup(attr.attributeLabel);
Somehow this error appears only if the nvtracker is presented in the pipeline and vanishes when it is removed.
Recompiling the nvinfer plugin with the following patch fixes the issue:
--- gstnvinfer_meta_utils.cpp 2022-04-06 22:15:34.878814358 +0300
+++ gstnvinfer_meta_utils_new.cpp 2022-04-06 22:07:04.661095894 +0300
@@ -277,7 +277,9 @@
history.cached_info.attributes.assign (new_result.attributes.begin (),
new_result.attributes.end ());
for (auto &attr : history.cached_info.attributes) {
- attr.attributeLabel = strdup(attr.attributeLabel);
+ if (attr.attributeLabel) {
+ attr.attributeLabel = strdup(attr.attributeLabel);
+ }
}
history.cached_info.label.assign (new_result.label);
}
The error is also absent when the custom parsing library is applied for the A model. But still though it is inconvenient to test a custom model applicability/launch-ability when a custom parsing library is not written yet.
The mentioned A model here is the custom secondary classifier that cannot be attached or reported because of the NDA, but here are its input/output parameters:
Input layer: [3;224;224] // RGB image 224x224
Output layer 0: [5] // 5 classes
Output layer 1: [6] // 6 classes
Suggesting to apply the mentioned above patch permanently in the future DeepStream releases.