I tried training the model using the 02958343 data and found at least one training step had some issues:
On step 655 some values were set to zero which led to a divide by zero error in function ypr_from_campos in util/euler.py , making the loss nan and ruining the training.
I found a quick fix was just to check for this in the function and return 0,0,0 if this happened ( 0 yaw, pitch, roll).
I tried training the model using the 02958343 data and found at least one training step had some issues:
On step 655 some values were set to zero which led to a divide by zero error in function
ypr_from_camposinutil/euler.py, making the loss nan and ruining the training.I found a quick fix was just to check for this in the function and return 0,0,0 if this happened ( 0 yaw, pitch, roll).