Torch/Lua Learning Notes

---

1. nn package functions summary (Torch)

• nn.SplitTable(dim, [nInputDims])(InputTensor)
  • 输出沿着某维度dim对InputTensor分割所得的若干sub-tensor
  • nInputDims是可选参数，它可以指定对输入的tensor维度。
• nn.CAddTable()/CMulTable()({tensor1, tensor2, ...})
  • 输出tensor1，tensor2...的"和/积"，注意tensors大小都得相同
• nn.Narrow(dim, start_index, size)(InputTensor)
  • 输出InputTensor的第dim维的tensor，tensor的起始下标是start_index，长度为size。
  • e.g. 取一个二维tensor的第二行 <=> nn.Narrow(1, 2, 1)(tensor)
• nn.Linear(InputDim, OutputDim)(InputTensor)
  • nn中的全连接层,InputDim和OutputDim分别表示输入输出Tensor的维度。InputTensor只能是1-D vector 或者2-D matrix。一般情况下，InputTensor是1-D，返回的结果是1-D tensor，维度是Outputdim。
  • 但是， 当InputTensor是2-D时，默认是将tensor中的每行视为batch中的一个样本，以batch的形式输入输出，记tensor size为rows * InputDim，则输出的结果是2-D tensor，size是rows * OutputDim

---

2. Torch functions

• resize/view/reshape

  • resize never copies memory. reshape always copies memory. view never copies memory.
  • resize literally resizes the tensor. if a tensor's underlying storage is too small, it calls realloc to expand the storage to an appropriate size: https://github.com/torch/torch7/blob/master/lib/TH/generic/THTensor.c#L687 How realloc does the mallocing (i.e. whether it expands the memory in-place or allocates completely new memory) is handled at the system level. You should not rely on resize keeping the same memory if you ever change nElement of the tensor
  • reshape creates a new tensor and always copies over the memory of src tensor over.
  • view never copies, ever. worst case it errors out. view does not have an exact implementation in TH, but view's closest equivalent is this constructor: https://github.com/torch/torch7/blob/master/lib/TH/generic/THTensor.h#L42

• split(tensor, size, [dim])

  • 输出大小为size的tensors，tensors是沿着维度dim对tensor分割得到的。dim默认是1。

---

3. Some tips in Torch/Lua

• local 关键字非常重要，在一个程序段中(测试/模块)， 变量建议都设为局部变量(加上 local 关键字)，否则可能会出现很多莫名其妙的错误。 以下面这段程序为例，该程序是要读一段文本并匹配得出文本中有多少个"y"。

f= assert(io.open("test.txt"), "w")
data = f:read(5)
num = 0
print(data)
repeat
    for k in string.gmatch(data, ".") do
        if k == "y" then num = num+1 end 
    end
    data = f:read(5)
until not data
print(num)

上面的程序看似没有任何问题，但它会报错: bad argument #1 to 'gmatch' (string expected, got table. 这个错误是非常诡异的，首先，data在之前已经定义过，但由于受到全局影响，data类型发生了改变，即不是string类型（read返回的就是string）。现在在data前面预先定义local data，程序如下：

local data
f= assert(io.open("test.txt"), "w")
data = f:read(5)
local num = 0 
print(data)
repeat
    for k in string.gmatch(data, ".") do
        if k == "y" then num = num+1 end 
    end
    data = f:read(5)
until not data
print(num)

修正后，程序运行正常，问题解决。

• data是一个N维tensor, 若要访问第k维的数据块，可以这样进行：
  • data[{{},...,{start_index, end_index}, ...,{}}]
  • start_index 和 end_index 分别表示第k维数据块的起止index