search

D-Nilsson / GRFP

Code for the CVPR 2018 paper "Semantic Video Segmentation by Gated Recurrent Flow Propagation"
MIT License
56 stars 8 forks source link

The problem with the function of get_GRU_cell #6

Closed Hanhanhan11 closed 5 years ago

Hanhanhan11 commented 6 years ago

In get_GRU_cell, when to compute h = self.weights['lambda'](1-z)h_prev_reset + z*h_tilde, the wt is h_prev_reset, while in the paper, the h_prev_warped in the code represent wt, which one is right in computing the value of h in the code??

D-Nilsson commented 5 years ago

Hi, h_prev_reset in the code corresponds to r_t*w_t . See eq. 2 and 6 in the paper. In an earlier arxiv version there was no multiplication by r.

Hanhanhan11 commented 5 years ago

Hi, I see that h_prev_reset in the code corresponds to r_tw_t, but in the code, showed that h = self.weights['lambda'](1 - z)h_prev_reset + zh_tilde where self.weights['lambda']*(1 - z) should multiply h_prev_warped which corresponds to w_t in the paper.

At 2018-12-05 15:49:02, "D-Nilsson" notifications@github.com wrote:

Hi, h_prev_reset in the code corresponds to r_t*w_t . See eq. 2 and 6 in the paper. In an earlier arxiv version there was no multiplication by r.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.