I have a model with the following layers
self.conv1 = nn.Conv3d(1, 3, 3, padding=1)
self.conv2 = nn.Conv3d(3, 128, 3, padding=1, stride=2)
self.conv3 = nn.Conv3d(128, 128, 3, padding=1, stride=2)
self.conv4 = nn.Conv3d(128, 256, 3, padding=1, stride=1)
self.conv5 = nn.Conv3d(256, 512, 3, padding=1, stride=2)
self.conv6 = nn.Conv3d(512, 1024, 3, padding=1, stride=1)
self.conv7 = nn.Conv3d(1024, 1024, 3, padding=1, stride=1)
self.deconv1 = nn.ConvTranspose3d(1024, 1024, 3, 1, padding=1, output_padding=0 )
self.deconv2 = nn.ConvTranspose3d(1024, 512, 3, 2, padding=1, output_padding=0)
self.deconv3 = nn.ConvTranspose3d(512, 256, 3, 2, padding=1, output_padding=0)
self.deconv4 = nn.ConvTranspose3d(256, 128, 3, 1, padding=1, output_padding=0)
self.deconv5 = nn.ConvTranspose3d(128, 128, 3, 2, padding=1, output_padding=1)
self.deconv6 = nn.ConvTranspose3d(128, 2, 3, 1, padding=1, output_padding=0)
On passing tensor at runtime the code produces the following error with cudnn backend if output_padding is set to 1 in deconv5.
RuntimeError: cublas runtime error : library not initialized at /pytorch/torch/lib/THC/THCGeneral.c:405
.
Can someone please explain what error might be the reason for this ?