Closed indygit closed 1 month ago
Opening SAM demo, click left mouse button on the image ... Is reloading the decoder model necessary?
@ibelem - Is this behavior intentional? 🤔 https://microsoft.github.io/webnn-developer-preview/demos/segment-anything/
@indygit @fdwr This is known and intentional behavior.
This demo supports multiple point coordinates segmentation, we need to update num_points and re-create ort session for WebNN provider as WebNN doesn't support the dynamic shape model. CC @Honry
Please refer to https://github.com/microsoft/webnn-developer-preview/blob/main/demos/segment-anything/index.js#L223C12-L224C60
Thanks Belem. That is good learning for me.
Opening SAM demo, click left mouse button on the image, the following message shows in the log window:
[09:24:55] [Load] ONNX Runtime Execution Provider: webnn [09:24:55] [Load] ONNX Runtime EP device type: gpu [09:24:55] [Load] Loading SAM ViT-B Decoder (FP16) · 15.7MB [09:24:55] [Load] SAM ViT-B Decoder (FP16) load time: 31.90ms [09:24:55] [Session Create] Creating SAM ViT-B Decoder (FP16) [09:24:55] [Session Create] SAM ViT-B Decoder (FP16) create time: 517.30ms
Is reloading the decoder model necessary?
Also once the decoder is reloaded, segmentation doesn't work as well as when the SAM page is initially loaded. (I am using a Windows 11 PC with NVidia card).
Should we disable the decoder reloading?
======================================================================
I know we can hover the mouse to trigger the segmentation inference and clicking the left mouse button saves the segmented image (very useful feature)