Bug when using annotation assistant with by custom uploaded model

Feedback
1

As the title suggests, when I use the uploaded model to perform object detection on images, I can get normal results.

However, when I use this model for annotating assist, I cannot annotate normally (even without a “0 annotations applied” bubble, just nothing happened). There is no response on the page, and the developer option shows that there is a bug on the page, as shown in the figure below (the image to annotate is exactly the same the the image above):

image
image1847×944 195 KB

However, when I switched to the MS COCO model for annotating assist, this issue did not occur again. The “0 annotations applied” bubble popped up on the page, and everything was normal.

So, I would like to ask what causes this problem?

2

This is the exact same image, and the deployed model mentioned above can detect the target normally.

image
image1833×948 115 KB

3

This is the full log on the browser’s console

graph_executor.js:131 Uncaught (in promise) Error: This execution contains the node 'StatefulPartitionedCall/assert_equal_1/Assert/AssertGuard/branch_executed/_219', which has the dynamic op 'Merge'. Please use model.executeAsync() instead. Alternatively, to avoid the dynamic ops, specify the inputs [Identity]
    at Le.compile (graph_executor.js:131:19)
    at Le.execute (graph_executor.js:172:33)
    at Pe.execute (graph_model.js:311:38)
    at e.<anonymous> (index.ts:175:41)
    at roboflow.js:2:1249565
    at Object.next (roboflow.js:2:1249670)
    at roboflow.js:2:1248582
    at new Promise (<anonymous>)
    at r (roboflow.js:2:1248327)
    at e.infer (roboflow.js:2:1251050)
    at e.<anonymous> (index.ts:495:17)
    at roboflow.js:2:1249565
    at Object.next (roboflow.js:2:1249670)
    at roboflow.js:2:1248582
    at new Promise (<anonymous>)
    at r (roboflow.js:2:1248327)
    at e.detect (roboflow.js:2:1253680)
    at Object.detect (main.js:371:34)
    at PixiApplication.ts:766:23
    at lodash.js:2792:43
    at i (helpers.js:90:17)

I suspect that this might be due to compatibility issues between tfjs running in the browser and TensorFlow (or other frameworks) running on the backend, such as the frontend version of tfjs being outdated compared to the backend version. Therefore, this issue may only occur on relatively new models.

I just filed a bug report abort this problem: Bug when using annotation assistant with by custom uploaded model · Issue #12 · roboflow/external-bugtracker · GitHub

4

@narugo1992 thank you so much for your detailed report! I’m getting this transferred to our internal bug tracker