Can model convert to torchscript? about mask2former HOT 8 OPEN

same question, also want to export to torchscript, did you made it ?

from mask2former.

same question，anyone can share successful export case

from mask2former.

I tried the following method.

1. use ms_deform_attn_core_pytorch instead of MSDeformAttnFunction.apply.
   

   Mask2Former/mask2former/modeling/pixel_decoder/ops/modules/ms_deform_attn.py

   Lines 116 to 121 in 9b0651c

   	try:
   	output = MSDeformAttnFunction.apply(
   	value, input_spatial_shapes, input_level_start_index, sampling_locations, attention_weights, self.im2col_step)
   	except:
   	# CPU
   	output = ms_deform_attn_core_pytorch(value, input_spatial_shapes, sampling_locations, attention_weights)

2. use detectron2 deploy tool.
   --format=torchscript
   --export-method=tracing
   using sample-image
   https://github.com/facebookresearch/detectron2/blob/7c2c8fb168a2093ce06a531c1208fba48d2984ec/tools/deploy/export_model.py

succeed in export but torchscript model got different results from pytorch model.
Can someone please share a successful method?

from mask2former.

Same issue here. Do you have any updates? Have you tried re-training with ms_deform_attn_core_pytorch(Must be slow) and exporting after?

from mask2former.

Any update? Would like to export the model to any format for deployment

from mask2former.

I have been troubled by this problem for a long time, has anyone solved it? thanks

from mask2former.

I tried the following method.

1. use ms_deform_attn_core_pytorch instead of MSDeformAttnFunction.apply.
   

   Mask2Former/mask2former/modeling/pixel_decoder/ops/modules/ms_deform_attn.py

   Lines 116 to 121 in 9b0651c

   	try:
   	output = MSDeformAttnFunction.apply(
   	value, input_spatial_shapes, input_level_start_index, sampling_locations, attention_weights, self.im2col_step)
   	except:
   	# CPU
   	output = ms_deform_attn_core_pytorch(value, input_spatial_shapes, sampling_locations, attention_weights)

2. use detectron2 deploy tool.
   --format=torchscript
   --export-method=tracing
   using sample-image
   https://github.com/facebookresearch/detectron2/blob/7c2c8fb168a2093ce06a531c1208fba48d2984ec/tools/deploy/export_model.py

succeed in export but torchscript model got different results from pytorch model. Can someone please share a successful method?

have you solved this problem? i tried you method,and export the model successfully,but the exported model cann't inference on image. @Osd-ysck

from mask2former.

After 3 years, is there a way to export this model to ONNX or TorchScript?
M2F is supported in Huggingface now, can we simply load M2F trained with Detectron2 with Huggingface? With HF, we would be able to export to ONNX for fast inference.

from mask2former.