calculate class var and std

CNN-Pred-Feat/visualizeFeatures.lua

@@ -2,7 +2,7 @@
 
 -- contact:
 -- Min-Hung (Steve) Chen at <cmhungsteve@gatech.edu>
--- Last updated: 02/01/2018
+-- Last updated: 02/03/2018
 
 require 'xlua'
 require 'torch'
@@ -23,8 +23,12 @@ op:option{'-idS', '--idSplit', action='store', dest='idSplit',
           help='which split used for testing', default=1}     
 op:option{'-idV', '--idVideo', action='store', dest='idVideo',
           help='video id of the extracted video', default=0}  
-op:option{'-s', '--save', action='store', dest='save',
+op:option{'-sP', '--savePath', action='store', dest='savePath',
           help='full path to the saved images', default='feature_images/'}
+op:option{'-s', '--save', action='store', dest='save',
+          help='save the images or not', default='No'}
+op:option{'-n', '--normalize', action='store', dest='normalize',
+          help='normalize the features or not', default='Yes'}
 
 opt,args = op:parse()
 
@@ -39,20 +43,58 @@ print('featurePath: '..opt.featurePath)
 print('Database: '..opt.nameDatabase)
 print('Modality: '..opt.dataModality)
 print('split #: '..idSplit)
+if opt.save == 'Yes' then
+	print('will save images')
+end
+if opt.normalize == 'Yes' then
+	print('will normalize the features')
+end
+
+-- load the label
+ucf101Label = require './ucf-101'
+table.sort(ucf101Label)
 
 ----------------------------------------------
 -- 				  Load Data 			    --
 ----------------------------------------------
 feat = torch.load(opt.featurePath..opt.nameDatabase..'/data_feat_test_'..opt.dataModality..'_centerCrop_25f_sp'..idSplit..'.t7')
+-- normalize the features 
+if opt.normalize == 'Yes' then
+	featMats = (feat.featMats-torch.mean(feat.featMats))/torch.std(feat.featMats)
+else
+	featMats = feat.featMats
+end
 
 if idVideo > 0 then
 	print(feat.name[idVideo])
-	test = feat.featMats[idVideo]
-	image.save(opt.save..'test_'..feat.name[idVideo]..'.png', test)
-else 
-	for i=1,feat.featMats:size(1) do
-		print(feat.name[i])
-		test = feat.featMats[i]
-		image.save(opt.save..'test_'..feat.name[i]..'.png', test)
+	test = featMats[idVideo]
+
+	var_test = torch.mean(torch.var(test,2))
+	std_test = torch.mean(torch.std(test,2))
+	print(var_test, std_test)
+	if opt.save == 'Yes' then
+		image.save(opt.savePath..'test_'..feat.name[idVideo]..'.png', test)
+	end
+else
+	var_std_allClass = torch.zeros(#ucf101Label,2)
+	for j=1, #ucf101Label do
+		print(ucf101Label[j])
+		var_allVideo = torch.Tensor()
+		std_allVideo = torch.Tensor()
+		for i=1,featMats:size(1) do
+			if string.match(feat.name[i], ucf101Label[j]) then
+				-- print(feat.name[i])
+				test = featMats[i]
+				var_test = torch.mean(torch.var(test,2))
+				std_test = torch.mean(torch.std(test,2))
+				var_allVideo = torch.cat(var_allVideo, torch.Tensor({var_test}))
+				std_allVideo = torch.cat(std_allVideo, torch.Tensor({std_test}))
+			end
+		end
+		var_std_allClass[j][1] = torch.mean(var_allVideo)
+		var_std_allClass[j][2] = torch.mean(std_allVideo)
 	end
+	print('variance & standard deviation of each class for '..opt.dataModality)
+	print(var_std_allClass)
+	torch.save('var_std_'..opt.dataModality..'.txt', var_std_allClass, 'ascii')
 end