some questions

[longzeyilang]

Thank you for good job. some question as follows:
(1) I train my own data. the raw image size: 128*128. wo revised model to get big feature map.
(2) self._unfold2d(x, ws=8), the 8 is fixed-parameters, but the keypoint_position_loss and coordinate_classification_loss still have 8, is equal param, must modify at the same time?
(3) the generateRandomTPS has grid (8,6), what is the meaning? and it relates to (2) ws=8
(4) acc_f is always nan in my training, is it normal?
Loss: 7.6677 acc_c0 0.314 acc_c1 0.156 acc_f: nan loss_c: 5.476 loss_f: 8.002 loss_kp: 0.065 #matches_c: 64 loss_kp_pos: 16.468 acc_kp_pos: 0.030:

[longzeyilang]

I train my data, the result so bad. and the train result is not same as test. the code not right?

[longzeyilang]

would you please answer me?

[guipotje]

Hi @longzeyilang,

(1) Couldn't understand whats the question, could you please reformulate?
(2) You should modify the keypoint head and the alike_distill_loss.
(3) These are control points for the TPS warper , we will use them to apply geometric transformations to augment the image.
(4) Yes, in the beginning this is normal, because there is a confidence threshold in the loss that prevents the optimization in too low confident matches. After a few iterations the loss should be stable.