zxg

Results 4 comments of zxg

you are using a layer that caffe does not support,just replace it and then test?

@AlphaNext @SoulCoderCN label读取并resize之后应该是0或者1的二值图,但是在经过self.im_gradient函数后取值范围应该是(0, ]也就是大于零,1.5这个取值应该是凭经验取吧。还有就是loss会出现nan的问题,总的损失包括了iou损失和交叉熵损失,我把他们单独拿出来测试,发现问题出现在iou损失,当iou损失降低到一定值时就会出现nan,根据iou损失的定义,把问题定位到了im_gradient函数,返回值使用了tf.sqrt,这个函数会有结果为nan的现象出现,我把代码修改为tf.sqrt(tf.add(tf.square(gx)+1e-8,tf.square(gy)+1e-8)+1e-8)解决了这个问题。

@eghouti @liuzhuang13 @jychoi118 In original metatune.data, gpus = 1,2,3,4, you should change it according to your device(mine: gpus = 0,1,2,3 please set the correct gpu id) . It might help...

"x = torch.zeros((b, d_in, n), device=deltaA.device)" is out of the loop, I think it is init the hidden state with 0(x = torch.zeros((b, d_in, n), device=deltaA.device)) @Ykiiii @ZK-Zhou