com_finetune_weights file about dannce HOT 4 CLOSED

Hello, sorry for the slow reply.

Q1: Yes. com_finetune_weights should be the path to the folder containing the pretrained weights.

Q2: demo/markerless_mouse_1/COM/weights/ contains the pretrained rat weights, which is a reasonable starting point for mice (and what we use a starting point for training our mouse COM net). However, if you want to try some mouse weights, you can use the weights from the mouse demo, demo/markerless_mouse_1/COM/train_results/.

Q3: see Q2.

Q4: com-predict will use a weights file generated by com-train (i.e., after finetuning the pre-trained COM finetune weights using your labeled data). By default, it will automatically use the weights from the best epoch (defined as the lowest validation loss) without having to change any config options (these weights will be saved in COM/train_results/). If you want to use a different set of weights for COM prediction (e.g., the weights from the final epoch, which are placed in COM/train_results/fullmodel_weights/fullmodel_end.hdf5), you can add the path to the weights file in com_predict_weights.

Q5: No, dannce is finetuned from pretrained dannce weights. Try demo/markerless_mouse_1/DANNCE/weight/weights.rat.MAX/.

Q6: You will need to set dannce_predict_model as the path to one of the weights files generated by dannce-train, analogous to the com-predict advice in Q4. dannce-predict should normally automatically use the best epoch weights generated by dannce-train, but there is a small bug that was introduced when we added checkpoint weights callbacks, such that you will currently need to point dannce_predict_model directly to your target weights file. I will push the fix as soon as I am done running tests.

As for your (5), which didn't have a question. That command is used to generate a dannce prediction file that is formatted to work with https://github.com/jessedmarshall/CAPTURE_demo.

from dannce.

Thank you so much!!! for making such helpful suggestions. Now i could set the appropriate parameters for our lab data analysis.

from dannce.

In cefd782, predict-dannce will now properly use the best epoch weights in train_results without having to set dannce_predict_model manually.

from dannce.

Thank you so much!!! for making such helpful suggestions. Now i could set the appropriate parameters for our lab data analysis.

You're welcome!

from dannce.