Pooling throws an exception in Tegra TX1 #1906
Description
Hi all, great work with pytorch. I'm trying to use pytorch in a Jetson TX1, but running into some issues. One of them is pretty simple: pooling doesn't seem to work.
First, some details:
- Jetson TX1, a 64-bit ARM architecture with a tegra TX1
- Ubuntu 16.04 aarch64
- CUDA 8.0, CUDNN 5105
- python 2.7
- Built from source (conda not available yet) from at least two git commits - 7c24a, Jun 23 and one from roughly a week earlier.
Currently compiling the 0.1.12 tag - will update when done.Tag 1.12 (ccdf54) actually works correctly, with and without cuda.
Here's a simple example:
import torch
import torch.nn as nn
from torch.autograd import Variable
net = nn.MaxPool2d(2)#.cuda()
#net = nn.AvgPool2d(2)#.cuda()
xd = torch.FloatTensor(1, 3, 448, 448)
x = Variable(xd, volatile=True)#.cuda() `
y = net.forward(x)
print(y.data.mean())when using MaxPool2d, I get
terminate called after throwing an instance of 'THArgException'
what(): can't expand an empty tensor at /home/ubuntu/apps/pytorch/torch/lib/TH/generic/THTensor.c:295
Aborted
when using AvgPool2d, oddly I get a different error: AvgPool2d works as expected.
Since I'm not too experienced with the internals, I'm not sure what changes matter, but I changed a few things (using cuda or cpu, volatile or not, the tensor size, the pooling parameters) and always got the same result. As a very very approximate workaround I've been using extra striding in my convolutions, but obviously that's not ideal ;).