Lua + Torch 踩坑小结

1.Tensor中选择元素时索引必须用number类型，tensor与table转换

index53[torch.FloatTensor of size 2]p0.5162  0.4855  0.5177  0.5186  0.50070.4797  0.4709  0.4959  0.4905  0.4959[torch.FloatTensor of size 2x5]想要根据tensor index取tensor p中的元素,组成一个新的Tensor NEW。但是从index取出的数仍为一个tensor。解决方法：将tensor->table,再从table中提取number数据。做法如下：index=torch.totable(index)NEW=torch.tensor(2,1)for i=1,2 do    NEW[{{i},{1}}]=p[{{i},{index[i]}}]end结果：NEW0.50070.4959[torch.FloatTensor of size 2x1]

[table] torch.totable(object)

Converts a Tensor or a Storage to a lua table. Also available as methods: tensor:totable() and storage:totable(). Multidimensional Tensors are converted to a set of nested tables, matching the shape of the source Tensor.

> print(torch.totable(torch.Tensor({1, 2, 3})))
{
  1 : 1
  2 : 2
  3 : 3
}

2.分类问题，criterion的forward

error 1

error 2

/Users/zh/torch/install/bin/luajit: /Users/zh/torch/install/share/lua/5.1/nn/THNN.lua:110: Assertion `cur_target >= 0 && cur_target < n_classes' failed.  at /tmp/luarocks_nn-scm-1-2066/nn/lib/THNN/generic/ClassNLLCriterion.c:62

stack traceback:
[C]: in function 'v'
/Users/zh/torch/install/share/lua/5.1/nn/THNN.lua:110: in function 'ClassNLLCriterion_updateOutput'
.../zh/torch/install/share/lua/5.1/nn/ClassNLLCriterion.lua:41: in function 'updateOutput'

.../zh/torch/install/share/lua/5.1/nn/ParallelCriterion.lua:23: in function 'forward'

该问题是指output的种类多于target。

解决方法：

参考

以上两个error都是ClassNLLCriterion的input和target的设置(size)问题。

如下面的ClassNLLCriterion损失函数使用描述中写道：

假设有n个类别input需是一个1维的size为n的张量，target是一个类别的索引，上文1中的p和index可以分别作为input和target。

ClassNLLCriterion

criterion = nn.ClassNLLCriterion([weights])

The negative log likelihood criterion. It is useful to train a classication problem with n classes. If provided, the optional argument weights should be a 1D Tensor assigning weight to each of the classes. This is particularly useful when you have an unbalanced training set.

The input given through a forward() is expected to contain log-probabilities of each class: input has to be a 1D Tensor of size n. Obtaining log-probabilities in a neural network is easily achieved by adding a LogSoftMax layer in the last layer of your neural network. You may use CrossEntropyCriterion instead, if you prefer not to add an extra layer to your network. This criterion expects a class index (1 to the number of class) as target when calling forward(input, target) and backward(input, target).