一、获取GPU信息
require 'cutorch'print(
cutorch.getDeviceProperties(cutorch.getDevice()) )
二、GPU和CPU之间的数据互通
1.构建GPU数据
t1 = torch.CudaTensor(100):fill(0.5)
t2 = torch.CudaTensor(100):fill(1)
t2 = torch.CudaTensor(100):fill(1)
t1:add(t2)
2.GPU=>CPU
t1_cpu = t1:float()
3.CPU=>GPU
t1:zero()
t1[{}] = t1_cpu -- copies the data back to the GPU, with no new alloc
t1[{}] = t1_cpu -- copies the data back to the GPU, with no new alloc
t1_new = t1_cpu:cuda() -- allocates a new tensor
三、nn使用GPU
Module参数是Torch的默认类型,如果需要Cuda的module,需要把默认类型修改为CUda。
1.构建基于Cuda的nn
require 'cunn'
-- we define an MLP
mlp = nn.Sequential()
mlp:add(nn.Linear(ninput, 1000))
mlp:add(nn.Tanh())
mlp:add(nn.Linear(1000, 1000))
mlp:add(nn.Tanh())
mlp:add(nn.Linear(1000, 1000))
mlp:add(nn.Tanh())
mlp:add(nn.Linear(1000, noutput))
-- and move it to the GPU:
mlp = nn.Sequential()
mlp:add(nn.Linear(ninput, 1000))
mlp:add(nn.Tanh())
mlp:add(nn.Linear(1000, 1000))
mlp:add(nn.Tanh())
mlp:add(nn.Linear(1000, 1000))
mlp:add(nn.Tanh())
mlp:add(nn.Linear(1000, noutput))
-- and move it to the GPU:
mlp:cuda()
输入数据需要首先转化为cuda
-- input
input = torch.randn(ninput)
-- retype and feed to network:
result = mlp:forward( input:cuda() )
-- the result is a CudaTensor, if your loss is CPU-based, then you will-- need to bring it back:
input = torch.randn(ninput)
-- retype and feed to network:
result = mlp:forward( input:cuda() )
-- the result is a CudaTensor, if your loss is CPU-based, then you will-- need to bring it back:
result_cpu = result:float()
可以增加Copy层自动进行数据类型转化
-- we put the mlp in a new container:
mlp_auto = nn.Sequential()
mlp_auto:add(nn.Copy('torch.FloatTensor', 'torch.CudaTensor'))
mlp_auto:add(mlp)
mlp_auto:add(nn.Copy('torch.CudaTensor', 'torch.FloatTensor'))
mlp_auto = nn.Sequential()
mlp_auto:add(nn.Copy('torch.FloatTensor', 'torch.CudaTensor'))
mlp_auto:add(mlp)
mlp_auto:add(nn.Copy('torch.CudaTensor', 'torch.FloatTensor'))