yolov7-seg with 416 image size?

[aurelm95]

With the default yolov7 640px imagesize, the deeptream pipeline works perfect.

However, when I tried to export a trainerd model with an image size of 416, the masks are very bad in deepstream.

I made the exports with the following commands:

pytorch to onnx:

python .\segment\export.py --data <my custom data path> --weights <my custom pytorch model path> --batch-size 1 --device cpu --simplify --opset 14 --workspace 8 --iou-thres 0.65 --conf-thres 0.35 --include onnx --end2end --trt --cleanup --dynamic-batch --roi-align --img 416

onnx to TensorRT:

/usr/src/tensorrt/bin/trtexec --onnx=<my custom onnx model path> --saveEngine=<my custom trt model path> --workspace=8192 --fp16 --minShapes=images:1x3x416x416 --optShapes=images:1x3x416x416 --maxShapes=images:8x3x416x416 --shapes=images:1x3x416x416

is this correct? or the custom image size is not implemented?

Thanks!