Closed JawnWB closed 7 years ago
Hi, SpatialBatchNormalization returns NaN.
A simple test: net=nn.Sequential() net.add(nn.SpatialConvolution(1, 1, 5, 5)) net:add(nn.SpatialBatchNormalization(1,1e-3)) t=torch.Tensor(1, 16, 16) net:forward(t) -- returns finite values in CPU mode net:cl() net:forward(t:cl()) -- returns NaN net:double() net:forward(t) -- also returns NaN
however, net:get(2).weight & .bias are >0 and 0 respectively....
Hi, SpatialBatchNormalization returns NaN.
A simple test: net=nn.Sequential() net.add(nn.SpatialConvolution(1, 1, 5, 5)) net:add(nn.SpatialBatchNormalization(1,1e-3)) t=torch.Tensor(1, 16, 16) net:forward(t) -- returns finite values in CPU mode net:cl() net:forward(t:cl()) -- returns NaN net:double() net:forward(t) -- also returns NaN
however, net:get(2).weight & .bias are >0 and 0 respectively....