Get very small mrcnn_mask values in model.detect.

[Xuxin1995]

I trained a mrcnn model using my own dataset. Then I tested this model, no mask found. I printed some variables, and found that the predicted mask values is much smaller than 0.5 as follows:

mask in unmold_mask: (581, 1078) [[1.1965359e-06 1.2571201e-06 1.3177042e-06 ... 2.4339261e-05
  2.3220206e-05 2.2101149e-05]
 [1.3065624e-06 1.3727173e-06 1.4388726e-06 ... 2.6577356e-05
  2.5355397e-05 2.4133440e-05]
 [1.4165887e-06 1.4883146e-06 1.5600408e-06 ... 2.8815450e-05
  2.7490591e-05 2.6165731e-05]
 ...
 [4.2061267e-05 4.4190954e-05 4.6320634e-05 ... 3.7213843e-07
  3.5502848e-07 3.3791852e-07]
 [3.8794366e-05 4.0758638e-05 4.2722910e-05 ... 3.4323443e-07
  3.2745339e-07 3.1167238e-07]
 [3.5527464e-05 3.7326325e-05 3.9125182e-05 ... 3.1433041e-07
  2.9987831e-07 2.8542621e-07]]

A full_mask matrix with all "False" value based on this mask. (threshold values：0.5) So what's wrong with my training process? Maybe somebody can help me with this ?

[raulperezalejo]

take a look at TensorBoard to see how was your training, those graphs may give you some clues