Please provide complete information as applicable to your setup.
• Hardware Platform (Jetson / GPU) x86 RTX-3060
• DeepStream Version 6.1
• JetPack Version (valid for Jetson only) N/A
• TensorRT Version 8.2.5.1
• NVIDIA GPU Driver Version (valid for GPU only) 525.125.06
• Issue Type( questions, new requirements, bugs) questions
• 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( This is for new requirement. Including the module name-for which plugin or for which sample application, the function description)
I am working on custom pre-processing transform implemented in PyTorch inference module, which is producing correct results for the given input tensor, however when converted the PyTorch model to native trt engine it’s not producing the equivalent results, there appears to be issue with the way deepstreem handles the pre-processing, below is PyTorch pre-processing function,
def data_transform(model):
# transforms needed for shufflenet
if model == 'shufflenet':
np_transforms = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
])
return np_transforms
##########################################################################
# read/process image and apply transformation
def read_img(frame, np_transforms):
small_frame = cv2.resize(frame, (224, 224), cv2.INTER_AREA)
small_frame = cv2.cvtColor(small_frame, cv2.COLOR_BGR2RGB)
small_frame = Image.fromarray(small_frame)
small_frame = np_transforms(small_frame).float()
small_frame = small_frame.unsqueeze(0)
small_frame = small_frame.to(device)
return small_frame
Below is config file settings for the sgie classifier , but it’s not producing the correct results as that of the PyTorch inference.
[property]
gpu-id=0
net-scale-factor=0.007843137255
model-engine-file=../weights/tensorrt/shufflenet_fp16_rtx-3050.engine
#batch-size=1
# 0=NCHW, 1=NHWC, 2=CUSTOM
network-input-order=0
# 0=FP32 and 1=INT8 mode
network-mode=2
num-detected-classes=1
process-mode=2
model-color-format=0
gie-unique-id=5
operate-on-gie-id=1
operate-on-class-ids=80
is-classifier=1
output-blob-names=output
classifier-async-mode=0
classifier-threshold=0.10
#input-dims=3;80;160
process-mode=2
parse-classifier-func-name=NvDsInferParseCustomNVFire
custom-lib-path=../plugins/libs/libnvdsanalytics_fire.so
#scaling-filter=0
#scaling-compute-hw=0
[class-attrs-all]
# Define pre-processing parameters for 'shufflenet'
pre-processing-parameters=0,0,0,0;0,0,0,0
NOTE: The post-processing function implemented for deepstreem is working correctly as the output of the notwork is single output value with sigmoid activation, hence it doesn’t looks to be an issue with the post-processing module. The Model is tested with FP16 and FP32 precision’s but the results are almost same in both the cases.