From 4cd74d95e07f14e35d6337d1a63239ffa04dcfc5 Mon Sep 17 00:00:00 2001
From: macournoyer <macournoyer@gmail.com>
Date: Wed, 16 Dec 2015 23:07:08 -0500
Subject: [PATCH] Make model output all sorted word IDs, not just top one
 (argmax).

Also add --debug flag to eval script to output probabilities.
---
 eval.lua    | 34 +++++++++++++++++++++++++++-------
 seq2seq.lua | 30 ++++++++++++++++--------------
 2 files changed, 43 insertions(+), 21 deletions(-)

diff --git a/eval.lua b/eval.lua
index 42066e1..08e0016 100644
--- a/eval.lua
+++ b/eval.lua
@@ -1,11 +1,13 @@
 require 'neuralconvo'
 local tokenizer = require "tokenizer"
 local list = require "pl.List"
+local options = {}
 
 if dataset == nil then
   cmd = torch.CmdLine()
   cmd:text('Options:')
   cmd:option('--cuda', false, 'use CUDA. Training must be done on CUDA')
+  cmd:option('--debug', false, 'show debug info')
   cmd:text()
   options = cmd:parse(arg)
 
@@ -24,15 +26,26 @@ if model == nil then
   model = torch.load("data/model.t7")
 end
 
--- Word ID tensor to words
-function t2w(t)
+-- Word IDs to sentence
+function pred2sent(wordIds, i)
   local words = {}
+  i = i or 1
 
-  for i = 1, t:size(1) do
-    table.insert(words, dataset.id2word[t[i]])
+  for _, wordId in ipairs(wordIds) do
+    local word = dataset.id2word[wordId[i]]
+    table.insert(words, word)
   end
 
-  return words
+  return tokenizer.join(words)
+end
+
+function printProbabilities(wordIds, probabilities, i)
+  local words = {}
+
+  for p, wordId in ipairs(wordIds) do
+    local word = dataset.id2word[wordId[i]]
+    print(string.format("%-23s(%4d%%)", word, probabilities[p][i] * 100))
+  end
 end
 
 function say(text)
@@ -44,7 +57,14 @@ function say(text)
   end
 
   local input = torch.Tensor(list.reverse(wordIds))
-  local output = model:eval(input)
+  local wordIds, probabilities = model:eval(input)
+
+  print(">> " .. pred2sent(wordIds))
 
-  print(">> " .. tokenizer.join(t2w(torch.Tensor(output))))
+  if options.debug then
+    for i = 1, 4 do
+      print(string.rep("-", 30))
+      printProbabilities(wordIds, probabilities, i)
+    end
+  end
 end
diff --git a/seq2seq.lua b/seq2seq.lua
index eda7e26..1b1f422 100644
--- a/seq2seq.lua
+++ b/seq2seq.lua
@@ -91,15 +91,6 @@ function Seq2Seq:train(input, target)
   return Edecoder
 end
 
-local function argmax(t)
-  local max = t:max()
-  for i = 1, t:size(1) do
-    if t[i] == max then
-      return i
-    end
-  end
-end
-
 local MAX_OUTPUT_SIZE = 20
 
 function Seq2Seq:eval(input)
@@ -109,20 +100,31 @@ function Seq2Seq:eval(input)
   self.encoder:forward(input)
   self:forwardConnect(input:size(1))
 
+  local predictions = {}
+  local probabilities = {}
+
   -- Forward <go> and all of it's output recursively back to the decoder
   local output = self.goToken
-  local outputs = {}
   for i = 1, MAX_OUTPUT_SIZE do
-    local predictions = self.decoder:forward(torch.Tensor{output})
-    output = argmax(predictions[1])
+    local prediction = self.decoder:forward(torch.Tensor{output})[1]
+    -- prediction contains the probabilities for each word IDs.
+    -- The index of the probability is the word ID.
+    local prob, wordIds = prediction:sort(1, true)
+
+    -- First one is the most likely.
+    output = wordIds[1]
+
+    -- Terminate on EOS token
     if output == self.eosToken then
       break
     end
-    table.insert(outputs, output)
+
+    table.insert(predictions, wordIds)
+    table.insert(probabilities, prob)
   end 
 
   self.decoder:forget()
   self.encoder:forget()
 
-  return outputs
+  return predictions, probabilities
 end