Inference unable to place skeleton #812
-
Bug descriptionHello team, Thank you for this fantastic tool! It has been so far easy to use, however I am running into an issue where after training, I am unable to infer on new frames. When I run Predict -> Run inference after training I get this pop up. I can select any option (suggested frames, random frames, etc) and get the same pop up. Your personal set upIm running:
Here is my output from conda list: How to reproduce
|
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 4 replies
-
|
Hi @Zachary-Ip, Do you mind sharing some info about the model used? A screenshot of the Training Pipeline + Model Configuration GUI will be helpful. I'd like to check whether changing the model parameters (and/or pipeline used) will help train a more confident model/predictor. The second skeleton is the predicted skeleton - any predicted skeleton will have a confidence score associated with it. The "empty" skeleton is a result of low confidence scores for where the nodes are located... SLEAP is uncertain where any node is located in this case. I'd like to check whether changing the model parameters (and/or pipeline used) will help train a more confident model/predictor. Thanks, |
Beta Was this translation helpful? Give feedback.
-
|
Do you think that manually labeling more frames >20 would improve the confidence? |
Beta Was this translation helpful? Give feedback.
-
|
Hi @Zachary-Ip, The ideal workflow would be to just label a few frames (~5-10) for the first round as you have done. Lets try a few things:
Do you mind giving those a shot to see if it helps prediction. Otherwise, you can share your data with lmaree@salk.edu for further troubleshooting. Thanks, |
Beta Was this translation helpful? Give feedback.
-
|
These setting changes were able to help! I am now in the human-in-the-loop training. Thank you! |
Beta Was this translation helpful? Give feedback.
-
|
Awesome, glad it worked! Just a note that decreasing the input scaling will result in decreased resolution which may cause SLEAP to overlook/miss fine resolution features as seen in a similar issue. |
Beta Was this translation helpful? Give feedback.
Hi @Zachary-Ip,
Do you mind sharing some info about the model used? A screenshot of the Training Pipeline + Model Configuration GUI will be helpful. I'd like to check whether changing the model parameters (and/or pipeline used) will help train a more confident model/predictor.
The second skeleton is the predicted skeleton - any predicted skeleton will have a confidence score associated with it. The "empty" skeleton is a result of low confidence scores for where the nodes are located... SLEAP is uncertain where any node is located in this case. I'd like to check whether changing the model parameters (and/or pipeline used) will help train a more confident model/predictor.
Thanks,
Liezl