search

dontLoveBugs / CSPN_monodepth

Unofficial Faster PyTorch implementation of Convolutional Spatial Propagation Network
67 stars 16 forks source link

CSPN_ours的问题 #5

Open Homay0118 opened 5 years ago

Homay0118 commented 5 years ago

我在调用x = conv2d(x, kernel, kernel_size=K, stride=1, padding=K // 2, dilation=1, native_impl=True)时，native_impl 取True时能backward，取False时不能backward。两者运算有本质的什么不同吗？