ONNX opset version 12 is not supported

Deploying the model to another embedded platform, but the embedded platform only supports opset <=12, so setting the opset_version in export.py to 12 gets the error.

Hi @valenbase , thank you for your interest in LightGlue-ONNX!

There were several reasons as to why opset_version=16 is used as the default export:

SuperPoint uses the GridSample operation, for which support was introduced in version 16
ONNX support for aten::scaled_dot_product_attention in LightGlue requires version 14
LightGlue has now been updated to use unflatten operations in the original repo, requiring version 13; this fork still keeps the einops.rearrange, but I'm not sure when I'll be syncing this fork with those changes.

However, if you don't mind using DISK instead of SuperPoint, I managed to export the DISK extractor and LightGlue to opset version 12 (with --dynamic):

I hope you find these models helpful!

thank you!

@fabio-sim can you give me the model for --end2end

Here you go:

disk_lightglue_end2end_opset12.onnx

In case anyone would like to reproduce LightGlue ONNX with opset 12 compatibility, these are the changes required (using release v0.1.3 as a reference point):

Modify LightGlue's Attention module forward to always use the einsum branch instead of the scaled_dot_product_attention branch (e.g., by removing the if branch):

LightGlue-ONNX/lightglue_onnx/lightglue.py

Lines 66 to 76 in e82a1a4

    
           def forward(self, q, k, v) -> torch.Tensor: 
        
               if ( 
        
                   hasattr(F, "scaled_dot_product_attention") 
        
                   and not torch.is_autocast_enabled() 
        
               ): 
        
                   q, k, v = [x.contiguous() for x in [q, k, v]] 
        
                   return F.scaled_dot_product_attention(q, k, v) 
        
               else: 
        
                   s = self.s 
        
                   attn = F.softmax(torch.einsum("bnid,bnjd->bnij", q, k) * s, -1) 
        
                   return torch.einsum("bnij,bnjd->bnid", attn, v)

Modify the opset_version in export.py from 16 to 12. Note that there are multiple occurrences of torch.onnx.export().
Run the export:

python export.py --extractor_type disk --dynamic --end2end

	def forward(self, q, k, v) -> torch.Tensor:
	if (
	hasattr(F, "scaled_dot_product_attention")
	and not torch.is_autocast_enabled()
	):
	q, k, v = [x.contiguous() for x in [q, k, v]]
	return F.scaled_dot_product_attention(q, k, v)
	else:
	s = self.s
	attn = F.softmax(torch.einsum("bnid,bnjd->bnij", q, k) * s, -1)
	return torch.einsum("bnij,bnjd->bnid", attn, v)