-
Notifications
You must be signed in to change notification settings - Fork 148
Description
Hello @nicknochnack ,
I tried to make adjustments to the code and run on the entire GRID corpus for a potential research using the DLib to automatically capture. However, I realized that on the training process, there are some damages I encounter on the files (frames) with some damages like this below:
[mpeg1video @ 0x7f3a580934c0] Warning MVs not available 1545/5950 [======>.......................] - ETA: 14:57:24 - loss: 4.0864[mpeg1video @ 0x7f39b81f6000] ac-tex damaged at 22 17 [mpeg1video @ 0x7f39b81f6000] Warning MVs not available 2764/5950 [============>.................] - ETA: 10:48:35 - loss: 4.1598[mpeg1video @ 0x7f36a0083cc0] ac-tex damaged at 22 17 [mpeg1video @ 0x7f36a0083cc0] Warning MVs not available 3598/5950 [=================>............] - ETA: 7:58:38 - loss: 4.3521[mpeg1video @ 0x7f35ac00f240] ac-tex damaged at 22 17 [mpeg1video @ 0x7f35ac00f240] Warning MVs not available 5304/5950 [=========================>....] - ETA: 2:11:28 - loss: 4.5888[mpeg1video @ 0x7f33f41f9700] ac-tex damaged at 22 17 [mpeg1video @ 0x7f33f41f9700] Warning MVs not available 5708/5950 [===========================>..] - ETA: 49:15 - loss: 4.5972^C
Could there be any reason for this. Currently on the 29th iteration for the training process with some interesting probable loss values. Moreover, what better evaluation metrics can be made aside maybe Word Error Rate, especially if looking at the direction where it can be used as visemes for an articulation for a Realistic Humanoid robots. Thank you and looking forward to your input.
