Exec_CharLM_CTC.lua

require 'nn'
require 'cudnn'
require 'cunn'
require 'cutorch'
local threads = require 'threads'
threads.Threads.serialization('threads.sharedserialize')
require 'CTC_CLM_NN_lang_multithread'
require 'SequenceError'
local final_threading = torch.class('Final_lua_threading')
local cmd = torch.CmdLine()

cmd:option('-threads',16, 'Number of threads for execution')
cmd:option('-log_probs_arr','./Test_With_Char_lm/predictions_log_vec.t7','Path where log probabilties for test data are stored')
cmd:option('-target_label','./Test_With_Char_lm/targets_vec.t7','Path where ground truth labels are stored')
cmd:option('-predictions_without_lm','./Test_With_Char_lm/predicts_vec.t7','Path where predictions generated by model without language model are stored')
cmd:option('-alpha',1.25,'alpha value based on Character level beam search algorithm')
cmd:option('-beta',0,'beta value based on Character level beam search algorithm')
cmd:option('-beam_val',40, 'Value of beam' )
cmd:option('-dictionary','dictionary', 'path to dictionary')
cmd:option('-path','/home/sbp3624/CTCSpeechRecognition/Test_With_Char_lm/', 'path to store results')

local opt = cmd:parse(arg)

function final_threading:init(nthreads)
  self.pool = threads.Threads(nthreads,function() require 'CTC_CLM_NN_lang_multithread' 
      end)
end

-- get log probabilities 
local predictions_arr = torch.load(opt.log_probs_arr)
local targets = torch.load(opt.target_label)
local predicts_without_lm = torch.load(opt.predictions_without_lm)

local numberOfSamples = #predictions_arr
--numberOfSamples=16
local alpha = opt.alpha
local beta = opt.beta
local beam_val = opt.beam_val
local sent
--local ctc_clm_obj = CTC_CLM_NN_lang_multithread('dictionary')
local final_sent_tbl={}
local nthreads = opt.threads
final_threading:init(nthreads)
local tbl_objs={}

--create objecs equal to number of threads 
for i = 1,nthreads do 
  tbl_objs[i] = CTC_CLM_NN_lang_multithread(opt.dictionary)
end


-- execute parallely all the sentences and get WER,CER using Character Language model 

local start_time = os.time()
for i = 1,numberOfSamples do
  final_threading.pool:addjob(function()
      print (__threadid,i)
      sent = tbl_objs[__threadid]:decode_beam_search(predictions_arr[i], alpha, beta, beam_val)
      return i, sent
    end,
    function(i, sent)
      final_sent_tbl[i] = sent
    end
    )
end

final_threading.pool:synchronize()
final_threading.pool:terminate()

local end_time = os.time()
local targetTranscript, predictTranscript_lm, predictTranscript
local CER , WER
local cumCER = 0
local cumWER = 0
local evaluationPredictions = {}
local prev_CER,prev_WER
local prev_cumCER = 0
local prev_cumWER = 0

-- Find Final Scores
for i = 1,numberOfSamples do
  targetTranscript = targets[i]
  predictTranscript_lm = final_sent_tbl[i]
  predictTranscript = predicts_without_lm[i]
  CER = SequenceError:calculateCER(targetTranscript, predictTranscript_lm)
  WER = SequenceError:calculateWER(targetTranscript, predictTranscript_lm)
  cumCER = cumCER + CER
  cumWER = cumWER + WER
          
  table.insert(evaluationPredictions, { wer = WER * 100, cer = CER * 100, target = targetTranscript, prediction = predictTranscript_lm , prediction_without = predictTranscript})
    
  prev_CER = SequenceError:calculateCER(targetTranscript, predictTranscript)
  prev_WER = SequenceError:calculateWER(targetTranscript, predictTranscript)
  prev_cumCER = prev_cumCER + prev_CER
  prev_cumWER = prev_cumWER + prev_WER

end

local function comp(a, b) return a.wer < b.wer end
table.sort(evaluationPredictions, comp)
-- path to save results
local path = opt.path
local suffix = '_'..os.date('%Y-%m-%d_%H-%M-%S')
 
for index, eval in ipairs(evaluationPredictions) do
  local f = assert(io.open(path..'Evaluation'..suffix..'.log', 'a'))
  f:write(string.format("WER = %.2f | CER = %.2f | Text = \"%s\" | Predict = \"%s\" | Predict_without_lm = \"%s\"\n",eval.wer, eval.cer, eval.target, eval.prediction,eval.prediction_without))
  f:close()
end
  

local averageWER = cumWER / numberOfSamples
local averageCER = cumCER / numberOfSamples
local prev_averageWER = prev_cumWER / numberOfSamples
local prev_averageCER = prev_cumCER / numberOfSamples
local f = assert(io.open(path..'Evaluation'..suffix..'.log', 'a'))
f:write(string.format("Average WER = %.2f | CER = %.2f", averageWER * 100, averageCER * 100))
    f:close()

print (string.format('Average WER = %.2f Average CER = %.2f',averageWER * 100, averageCER * 100))

print (string.format('Average prev_WER = %.2f Average prev_CER = %.2f',prev_averageWER * 100, prev_averageCER * 100))

print (string.format('Total Time %.2f',(end_time-start_time)/60))