changed CMakeLists

CMakeLists.txt

@@ -10,7 +10,7 @@ foreach(target ctc normalizer)
 	set_target_properties(${target}
 		PROPERTIES PREFIX ""
 		SUFFIX ".so"
-		LIBRARY_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/../${target}"
+		LIBRARY_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/../"
 	)
 	target_link_libraries(${target} TH luaT luajit)
 endforeach(target)

ctc.lua

@@ -150,7 +150,6 @@ function ctc.getCTCCost(outputTable, target)
 
 	fvs = ctc.getForwardVariable(outputTable, alignedTable, target)
 
-
 	-- calculate backwardVariable
 
 	bvs = ctc.getBackwardVariable(outputTable, alignedTable, target)

loader.lua

@@ -1,4 +1,6 @@
+require 'image'
 require 'codec'
+require 'normalizer'
 utf8 = require 'utf8'
 
 Loader = {
@@ -27,11 +29,34 @@ function Loader:new(o)
 	return o
 end
 
+function Loader.__getNormalizedImage(src)
+	local im = image.load(src, 1)
+
+	if im:dim() == 3 then
+		im = im[1]
+	end
+
+	output = torch.Tensor()
+
+	w = im:size()[2]
+	h = im:size()[1]
+
+	ones = torch.ones(h, w)
+
+	im = ones - im
+
+	normalizer.normalize(im, output)
+
+	return output
+end
+
 function Loader:load(file)
 	self.samples = {}
 	local f = assert(io.open(file, "r"))
 	for line in f:lines() do
 		local src = line
+		local im = Loader.__getNormalizedImage(src)
+
 		local gt = src:gsub(".png", ".gt.txt")
 		local cf = assert(io.open(gt, "r"))
 		local gt = cf:read("*line")
@@ -45,7 +70,7 @@ function Loader:load(file)
 
 		end
 
-		table.insert(self.samples, {src = src, gt = gt})
+		table.insert(self.samples, {src = src, gt = gt, img = im})
 	end
 	f:close()
 
@@ -56,7 +81,7 @@ function Loader:load(file)
 	self.codec_obj = nil
 	self.weights = nil
 
-	return self.samples
+	-- return self.samples
 end
 
 function Loader:pick()
@@ -74,8 +99,8 @@ function Loader:pickWithWeight()
 
 		self.p = torch.zeros(#self.samples)
 		local i = 0
-		self.p:apply(function() 
-			i = i + 1 
+		self.p:apply(function()
+			i = i + 1
 			return torch.normal(1.0 / self.weights[i], 1.0 / self.weights[i] / 3.0) 
 		end)
 	end

main.lua

@@ -3,14 +3,15 @@ require 'rnn'
 require 'image'
 require 'optim'
 
+require 'loader'
 require 'ctc_log'
 require 'utils/decoder'
 
 mnist = require 'mnist'
 
 DROPOUT_RATE = 0.4
 
-local input_size = 28
+local input_size = 64
 local hidden_size = 100
 local class_num = 10
 
@@ -36,12 +37,27 @@ state = {
 	momentum = 0.5
 }
 
+loader = Loader()
+loader:load("1.txt")
+codec = loader:codec()
+
+local sample = loader:pick()
+local im = sample.img
+local target = codec:encode(sample.gt)
+
+raw = image.load(sample.src, 1)
+
+print(raw[1])
+
+print(im)
+
+--[[
 for i = 1, 100000 do
-	local no = torch.random() % 100 + 1
-	local sample = mnist.traindataset()[no]
-	local im = sample.x:double():t()
-	local target = torch.Tensor{sample.y + 1}
-
+	local sample = loader:pick()
+	local im = sample.img
+	local target = codec:encode(sample.gt)
+
+	print(im)
 	
 	local feval = function(params)
 		net:forget()
@@ -51,8 +67,8 @@ for i = 1, 100000 do
 		loss, grad = ctc.getCTCCostAndGrad(outputTable, target)
 	
 		if i % 20 == 0 then
-			print(target[1] - 1)
-			print(decoder.decodeTable(outputTable))
+			print(sample.gt)
+			print(decoder.best_path_decode(outputTable))
 			print(loss)
 		end
 	
@@ -67,3 +83,6 @@ for i = 1, 100000 do
 	
 	optim.sgd(feval, params, state)
 end
+
+
+]]

normalizer.cc

@@ -114,14 +114,14 @@ static double bilinear(double * in, int w, int h, double x, double y) {
 	xt = xt > w - 1 ? w - 1 : xt;
 	yt = yt > h - 1 ? h - 1 : yt;
 
-	// printf("(%d, %d)\n", xi, yi);
+	printf("(%d, %d)\n", xi, yi);
 
 	double p00 = in[yi * w + xi];
 	double p01 = in[yt * w + xi];
 	double p10 = in[yi * w + xt];
 	double p11 = in[yt * w + xt];
 
-	return p00 * (1.0 - xf) * (1.0 - yf) + p10 * xf * (1.0 - yf) + p01 * (1.0 - xf) * yf + p11 * xf * yf;
+	p00 * (1.0 - xf) * (1.0 - yf) + p10 * xf * (1.0 - yf) + p01 * (1.0 - xf) * yf + p11 * xf * yf;
 
 }
 
@@ -174,7 +174,7 @@ static void normalize
 
 	THDoubleTensor_resize2d(out, target_height, target_width);
 
-	printf("scale = %.4f\n", scale);
+	// printf("scale = %.4f\n", scale);
 
 	double * outData = THDoubleTensor_data(out);
 

normalizer.lua → test_normalizer.lua

@@ -1,5 +1,5 @@
 require 'image'
-require 'normalizer/normalizer'
+require 'normalizer'
 
 im = image.load("bq01_006.png", 1)
 

utils/decoder.lua

@@ -12,7 +12,7 @@ function decoder.best_path_decode(outputTable, codec)
 		local max_val, max = torch.max(outputTable[i], 1)
 		max = max[1]
 
-		if max = class_num then
+		if max == class_num then
 			if last_max ~= -1 and last_max_class ~= nil then
 				table.insert(result, last_max_class)
 				last_max = -1