Train script about deep-structured-facial-landmark-detection

[lisha-chen]

Hello, thank you for your interest. The crucial part is the loss functions used, provided in CNN_models/FAN_crf.py.
For your convenience, refer to the function "crf_forward" in CNN_models/FAN_crf.py, where you can directly use the function and "self.loss_unary" is for pre-training the CNN model, and the "self.loss_joint" is for jointly training the CNN-CRF model. Using tensorflow to take gradient w.r.t. the parameters of interest in each stage, you can follow the paper to write the training script.

[anjany]

Thanks for the hints, Lisha. Let me give this a try. I'll close this ticket for now. I hope it is okay for me to correspond with you on this issue here.