Do

试过`noobj_mask[b,anch_ious>ignore_threshold,gi,gj]=0`，conf_loss在训练过程中上下跳动的很大。我觉得有可能是由于confidence是shared weight，所以将cell[bs, best_n, gi, gj]的confidence weight调高的同时，其他cell的同一个confidence weight又被调低了。不知道否将shared confidence weight变成一个个独立的全连接weight可以解决这个问题。

@xuzheyuan624 确实是反了。。。我再试一下，太感谢你了

maybe you could replace transposed convolutions with nearest-neighbor upsampling to address checkerboard artifacts.

Hi @mark22 , tested it on my iOS10 - 5.5.0.GA - iPad Air 2 simulator with the code below, didn't met any problems: ``` javascripts var POP = require("guy.mcdooooo.tipop"); function...

@iMaciah It means you should create the Ti.UI.ScrollView in advance in order to use this module. you could create a metadata.json file in your project root directory, and put the...

I have the same issue. any idea on how to solve it?