dyhead swin onnx model to tensorrt engine with FP16

[vedrusss]

Hi!
I'm trying to convert dyhead swin based model from onnx to tensorrt engine in fp16 mode using mmdeploy converter:
python tools/onnx2tensorrt.py configs/mmdet/detection/detection_tensorrt-fp16_static-1280x1920.py detection_tensorrt_static-1280x1920/fp16/end2end.onnx detection_tensorrt_static-1280x1920/fp16/end2end

and running in following error:

-------------- The current device memory allocations dump as below --------------
[0]:34359738368 :HybridGlobWriter in reserveRegion: at optimizer/common/globWriter.cpp: 246 idx: 28 time: 0.000158613
[0x302000000]:11324620800 :HybridGlobWriter in reserveRegion: at optimizer/common/globWriter.cpp: 224 idx: 23 time: 0.00316042
[0x7f83c0000000]:29493600 :DeviceActivationSize in reserveNetworkTensorMemory: at optimizer/common/tactic/optimizer.cpp: 4603 idx: 8 time: 0.000331153
[10/24/2022-18:37:08] [TRT] [W] Requested amount of GPU memory (34359738368 bytes) could not be allocated. There may not be enough free memory for allocation to succeed.
[10/24/2022-18:37:08] [TRT] [W] Skipping tactic 4 due to insuficient memory on requested size of 34359738368 detected for tactic 31.
Try decreasing the workspace size with IBuilderConfig::setMaxWorkspaceSize().
[10/24/2022-18:37:08] [TRT] [E] 2: [pointWiseBuilder.cpp::splitVectorizedHub::277] Error Code 2: Internal Error (Assertion !never(a.z >= 0) failed. )
Traceback (most recent call last):
File "tools/onnx2tensorrt.py", line 73, in
main()
File "tools/onnx2tensorrt.py", line 58, in main
from_onnx(
File "/root/workspace/mmdeploy/mmdeploy/backend/tensorrt/utils.py", line 225, in from_onnx
assert engine is not None, 'Failed to create TensorRT engine'
AssertionError: Failed to create TensorRT engine

The actual available GPU memory onto my device is 16Gb and it does support FP16.
I'm able to obtain the tensorrt engine for the same model in FP32 mode and cannot find out the reason why FP16 doesn't work.
Are there any nuances with using mmdeploy and mmdetection models for obtaining FP16 engines?

[ZwwWayne]

Hi @vedrusss ,
DyHead is not officially supported for ONNX export. It needs some development in MMDeploy.