Deepstream 6.1
I am training a ResNet18 model for classification using PyTorch 2.5.1 and converting it to ONNX with ONNX 1.17.0 and onnxruntime-gpu 1.21.0. However, the result returns only None. What is the most recommended training and conversion method to ONNX?
resnet18_colors_21_03_last_base.zip (39.7 MB)
config_sgie_classification_color.txt (866 Bytes)
Is it trained by TAO Toolkit?
I am not training with the TAO Toolkit.
I am trying to run the Deepstream 7.0 application to obtain inferences from two models in a cascade. The first model is a YoloV8, which performs plate detection, and in cascade, I plan to use a classification model that identifies the color of these plates.
I am attempting to run the Deepstream application to get inferences from two cascading models. The first model is YoloV8, which detects the plates, and it is working fine. However, the classification model for identifying the color of the detected plates is not functioning as expected.
For training the classification model, I am using the ResNet18 architecture with the PyTorch 2.5.1 library. After training, I am converting the model to the ONNX format using the ONNX 1.17.0 and onnxruntime-gpu 1.21.0 libraries.
However, when I run the Deepstream application (ONNX model resnet18_colors_21_03_last_base.zip (39.7 MB)
, model configuration config_sgie_classification_color.txt (866 Bytes)
, labels file
cores_17_03.txt (65 Bytes)
), it detects the plate but does not classify the detected plate, returning the value ‘None,’ as shown in the image
.
The results I am getting from the classification are as follows:
[…]
while l_obj is not None:
try:
# Casting l_obj.data to pyds.NvDsObjectMeta
obj_meta = pyds.NvDsObjectMeta.cast(l_obj.data)
except StopIteration:
break
tracker_id = obj_meta.object_id
class_label = obj_meta.obj_label
# Itera sobre a lista de metadados do classificador (modelo secundário)
print('tracker_id:',tracker_id)
print('class_label:',class_label)
l_classifier = obj_meta.classifier_meta_list
print('l_classifier:',l_classifier)
while l_classifier is not None:
try:
classifier_meta = pyds.NvDsClassifierMeta.cast(l_classifier.data)
except StopIteration:
break
[…]
So, it is not related to TAO. It is a topic to deploy model in deepstream.
Maybe you can check if net-scale-factor and offsets in config_sgie_classification_color.txt match the normalization used during training.
Check that the ONNX model’s output layer name matches the output-blob-names in the config file.
Confirm cores_17_03.txt contains the exact class labels in the correct order with one label per line.
cc @yingliu to reassign.