cluster-rnn

GloVeEmbedding.lua (6534B)

---

 
 function split(inputstr, sep)
     if sep == nil then
         sep = "%s"
     end
     local t={} ; i=1
     for str in string.gmatch(inputstr, "([^"..sep.."]+)") do
         t[i] = str
         i = i + 1
     end
     return t
 end
 
 function numkeys(T)
     local count = 0
     for _ in pairs(T) do count = count + 1 end
     return count
 end
 
 local GloVeEmbedding, parent = torch.class('GloVeEmbedding', 'nn.LookupTable')
 
 function GloVeEmbedding:__init(word2idx, embedding_size, data_dir)
     -- you need glove embeddings in the directory ./util/glove/
     -- download them from http://nlp.stanford.edu/projects/glove/
     local embedding_file = 'util/glove/vectors.6B.200d.txt' 
     local file_embedding_size = 200
     self.vocab_size = numkeys(word2idx)
     self.word2idx = word2idx
     parent.__init(self, self.vocab_size, embedding_size)
     print("loading glove vectors")
     self.embedding_size = embedding_size
     --load glove vectors as a tensor from a .t7 file if it exists, otherwise generate that .t7 file
     local vocab_embedding_file = path.join(data_dir, "glove_" .. self.vocab_size .. "x" .. embedding_size .. ".t7")
     --vocab_embedding_file = "lstm-glove-w.t7"
     --(if you want to load a different word vector here, just swap the vocab_embedding_file file name with something else)
     if path.exists(vocab_embedding_file) then
         self.weight = torch.load(vocab_embedding_file):clone()
     else
         w = self:parseEmbeddingFile(embedding_file, file_embedding_size, word2idx)
         if file_embedding_size ~= embedding_size then
             w = torch.mm(w, torch.rand(file_embedding_size, embedding_size))
         end
         self.weight = w:contiguous()
         torch.save(vocab_embedding_file, self.weight)
     end
     print("loaded glove vectors")
 end
 
 function GloVeEmbedding:parseEmbeddingFile(embedding_file, file_embedding_size, word2idx)
     local word_lower2idx = {}
     local loaded = {}
     local weight = torch.Tensor(self.vocab_size, file_embedding_size)
     for word, idx in pairs(word2idx) do
         word_lower2idx[word:lower()] = idx
     end
 
     for line in io.lines(embedding_file) do
         local parts = split(line, " ")
         local word = parts[1]
         if word_lower2idx[word] then
             local idx = word_lower2idx[word]
             for i=2, #parts do
                 weight[idx][i-1] = tonumber(parts[i])
             end
             loaded[word] = true
         end
     end
     for word, idx in pairs(word2idx) do
         if not loaded[word:lower()] then
             print("Not loaded: " .. word:lower())
             for i=1, file_embedding_size do
                 weight[idx][i] = torch.normal(0, 0.9) --better way to do this?
             end
         end
     end
     return weight
 end
 
 function GloVeEmbedding:updateOutput(input)
     return parent.updateOutput(self, input:contiguous())
 end
 
 function GloVeEmbedding:accGradParameters(input, gradOutput, scale)
     return parent.accGradParameters(self, input:contiguous(), gradOutput:contiguous(), scale)
 end
 
 
 local GloVeEmbeddingFixed, parent = torch.class('GloVeEmbeddingFixed', 'GloVeEmbedding')
 
 function GloVeEmbeddingFixed:accGradParameters(input, gradOutput, scale)
     return nil
 end
 function GloVeEmbeddingFixed:parameters()
     return {}, {}
 end
 
 --[[
 function GloVeEmbeddingProject:__init(word2idx, embedding_size, data_dir)
     local embedding_file = 'util/glove/vectors.6B.200d.txt'
     local file_embedding_size = 200
     self.vocab_size = numkeys(word2idx)
     parent.__init(self, self.vocab_size, embedding_size)
     print("loading glove vectors")
     self.embedding_size = embedding_size
     local vocab_embedding_file = path.join(data_dir, "glove_" .. self.vocab_size .. "x" .. embedding_size .. ".t7")
     --print("loading pretrained word vectors")
     --vocab_embedding_file = "glove_embeddings_pretrained1862x200-4sitelow-descs.t7"
     --vocab_embedding_file = "glove_embeddings_pretrained1704x200-5site-title.t7"
     if path.exists(vocab_embedding_file) then
         self.weight = torch.load(vocab_embedding_file):clone()
     else
         w = self:parseEmbeddingFile(embedding_file, file_embedding_size, word2idx)
         if file_embedding_size ~= embedding_size then
             w = torch.mm(w, torch.rand(file_embedding_size, embedding_size))
         end
         self.weight = w:contiguous()
         torch.save(vocab_embedding_file, self.weight)
     end
     print("self.weight size")
     print(self.weight:size())
     print("loaded glove vectors")
 end
 
 
 
 local GloVeEmbeddingFixed, parent = torch.class('GloVeEmbeddingFixed', 'nn.Module')
 function GloVeEmbeddingFixed:__init(word2idx, embedding_file, embedding_size, data_dir)
     self.vocab_size = numkeys(word2idx)
     parent.__init(self, self.vocab_size, embedding_size)
     print("loading glove vectors")
     self.embedding_size = embedding_size
     local vocab_embedding_file = path.join(data_dir, "glove_" .. self.vocab_size .. "x" .. embedding_size .. ".t7")
     print("loading pretrained word vectors")
     vocab_embedding_file = "glove_embeddings_pretrained1704x200.t7" --pretrained vectors
     local loaded = {}
     if path.exists(vocab_embedding_file) then
         self.weight = torch.load(vocab_embedding_file)
     else
         self.weight = torch.Tensor(self.vocab_size, embedding_size)
         local word_lower2idx = {}
         for word, idx in pairs(word2idx) do
             word_lower2idx[word:lower()] = idx
         end
 
         for line in io.lines(embedding_file) do
             local parts = split(line, " ")
             local word = parts[1]
             if word_lower2idx[word] then
                 local idx = word_lower2idx[word]
                 for i=2, #parts do
                     self.weight[idx][i-1] = tonumber(parts[i])
                 end
                 loaded[word] = true
             end
         end
         for word, idx in pairs(word2idx) do
             if not loaded[word:lower()] then
                 print("Not loaded: " .. word:lower())
                 for i=1, self.embedding_size do
                     self.weight[idx][i] = torch.normal(0, 0.9) --better way to do this?
                 end
             end
         end
         torch.save(vocab_embedding_file, self.weight)
     end
     print("loaded glove vectors")
 end
 
 function GloVeEmbeddingFixed:updateOutput(input)
   print("GloVeEmbeddingFixed input size: " .. input:size())
   self.output:resize(input:size(1), self.embedding_size):zero()
   local longInput = input:long()
   self.output:copy(self.weight:index(1, longInput))
   return self.output
 end
 ]]--