Some problems about training the parameters in CRF

MiguelMonteiro / CRFasRNNLayer

Conditional Random Fields as Recurrent Neural Networks (Tensorflow)

MIT License

82 stars 19 forks source link

Some problems about training the parameters in CRF #6

Closed jingyzhang closed 6 years ago

jingyzhang commented 6 years ago

Hi~ When I couples the CRF on the top of some popular segmentation mdels (such as U-Net), can the training process be done just via back-propagation in Tensorflow? I'm worry about the trainable parameters in the CRF cannot be updated and trained, although this problem would not happen in theory. Also, i have met the same problems with you when testing on some medical images. Do you find some reasons?

MiguelMonteiro commented 6 years ago

What parameters are you talking about ? The weights of the CRF are trained through backprop. Do you have some code to reproduce your problem?

Best,

Miguel