Initial commit

corloc.lua

@@ -0,0 +1,29 @@
+dofile('opts.lua')
+dofile('util.lua')
+dofile('dataset.lua')
+
+opts.SCORES_FILES = #arg >= 1 and arg or {opts.PATHS.SCORES_PATTERN:format('trainval')}
+
+loaded = hdf5_load(opts.SCORES_FILES[1], {'subset', 'rois', 'labels', 'output'})
+outputs = {}
+
+for i = 1, #opts.SCORES_FILES do
+	outputs_i = hdf5_load(opts.SCORES_FILES[i], 'outputs')
+	for output_field, scores in pairs(outputs_i) do
+		outputs[output_field] = {}
+		for exampleIdx = 1, #scores do
+			outputs[output_field][exampleIdx] = (outputs[output_field][exampleIdx] or scores[exampleIdx]:clone():zero()):add(scores[exampleIdx]:div(#opts.SCORES_FILES))
+		end
+	end
+end
+
+res = {training_MAP = dataset_tools.meanAP(loaded.output, loaded.labels)}
+for output_field, scores in pairs(outputs) do
+	res[output_field] = {by_class = {}, _mean = corloc(dataset[loaded.subset], {scores, loaded.rois})}
+	for classLabelInd, classLabel in ipairs(classLabels) do
+		res[output_field].by_class[classLabels[classLabelInd]] = corloc(dataset[loaded.subset], {scores, loaded.rois}, classLabelInd)
+	end
+end
+
+json_save(opts.PATHS.CORLOC, res)
+print('result in ' .. opts.PATHS.CORLOC)

dataset.lua

@@ -0,0 +1,10 @@
+if opts.DATASET == 'VOC2007' or opts.DATASET == 'VOC2012' then
+	dataset_tools = dofile('pascal_voc.lua')
+	classLabels = dataset_tools.classLabels
+	numClasses = dataset_tools.numClasses
+end
+
+dataset = torch.load(opts.PATHS.DATASET_CACHED)
+
+dofile('parallel_batch_loader.lua')
+dofile('example_loader.lua')

detection_mAP.lua

@@ -0,0 +1,47 @@
+dofile('opts.lua')
+dofile('util.lua')
+dofile('dataset.lua')
+threads = require 'threads'
+
+overlap_threshold = 0.4
+score_threshold = 1e-3 / (2 * 5)
+output_field = 'output_prod'
+local subset = 'test'
+
+opts.SCORES_FILES = #arg >= 1 and arg or {opts.PATHS.SCORES_PATTERN:format(subset)}
+rois = hdf5_load(opts.SCORES_FILES[1], 'rois')
+
+scores = {}
+for i = 1, #opts.SCORES_FILES do
+	scores_i = hdf5_load(opts.SCORES_FILES[i], 'outputs/' .. output_field)
+	for exampleIdx = 1, #scores_i do
+		scores[exampleIdx] = (scores[exampleIdx] or scores_i[exampleIdx]:clone():zero()):add(scores_i[exampleIdx]:div(#opts.SCORES_FILES))
+	end
+end
+
+local detrespath = dataset_tools.package_submission(opts.PATHS.DATA, dataset, opts.DATASET, subset, 'comp4_det', rois, scores, nms_mask(rois, scores, overlap_threshold, score_threshold))
+local opts = opts
+
+if dataset['test'].objectBoxes == nil then
+	print('detection mAP cannot be computed for ' .. opts.DATASET .. '. Quitting.')
+	print('VOC submission saved in ' .. detrespath)
+	os.exit(0)
+end
+
+res = {[output_field] = {_mean = nil, by_class = {}}}
+APs = torch.FloatTensor(numClasses):zero()
+
+jobQueue = threads.Threads(numClasses)
+for classLabelInd, classLabel in ipairs(classLabels) do
+	jobQueue:addjob(function()
+		os.execute(('matlab -nodisplay -nojvm -nosplash -r "classLabel = \'%s\'; cd(\'%s\'); addpath(\'VOCcode\'); VOCinit; VOCopts.testset = \'%s\'; VOCopts.detrespath = \'%s\'; [rec, prec, ap] = VOCevaldet_fast(VOCopts, \'comp4\', classLabel, false); dlmwrite(sprintf(VOCopts.detrespath, \'resu4\', classLabel), ap); quit;"'):format(classLabel, paths.dirname(opts.PATHS.VOC_DEVKIT_VOCYEAR), subset, detrespath))
+		return tonumber(io.open(detrespath:format('resu4', classLabel)):read('*all'))
+	end, function(ap) res[output_field].by_class[classLabel] = ap; APs[classLabelInd] = ap; end)
+end
+jobQueue:synchronize()
+os.execute('[ -t 1 ] && reset')
+
+res[output_field]._mean = APs:mean()
+
+json_save(opts.PATHS.DETECTION_MAP, res)
+print('result in ' .. opts.PATHS.DETECTION_MAP)

example_loader.lua

@@ -0,0 +1,124 @@
+local ExampleLoader, parent = torch.class('ExampleLoader')
+
+function ExampleLoader:__init(dataset, example_loader_opts, channel_order, rgb_pixel_means)
+	self.scales = {{608, 800}, {496, 656}, {400, 544}, {720, 960}, {864, 1152}}
+	--self.scales = {{608, 800}, {368, 480}, {432, 576}, {528, 688}, {656, 864}, {912, 1200}}
+
+	self.channel_order = {3, 2, 1} -- BGR because we use Caffe-derived models
+	self.rgb_pixel_means = {122.7717, 115.9465, 102.9801}
+	self.example_loader_opts = example_loader_opts
+	self.dataset = dataset
+end
+
+local function table2d(I, J, elem_generator)
+	local res = {}
+	for i = 1, I do
+		res[i] = {}
+		for j = 1, J do
+			res[i][j] = elem_generator(i, j)
+		end
+	end
+	return res
+end
+
+local function subtract_mean(dst, src, channel_order, rgb_pixel_means)
+	for c = 1, 3 do
+		dst[c]:copy(src[channel_order[c]]):add(-rgb_pixel_means[channel_order[c]])
+	end
+end
+
+local function rescale(img, max_height, max_width)
+	--local height_width = math.max(dhw_rgb:size(3), dhw_rgb:size(2))
+	--local im_scale = target_height_width / height_width
+	local scale_factor = max_height / img:size(2)
+	if torch.round(img:size(3) * scale_factor) > max_width then
+		scale_factor = math.min(scale_factor, max_width / img:size(3))
+	end
+
+	return image.scale(img, math.min(max_width, img:size(3) * scale_factor), math.min(max_height, img:size(2) * scale_factor))
+end
+
+local function flip(images_j, rois_j)
+	image.hflip(images_j, images_j)
+	rois_j:select(2, 1):mul(-1):add(images_j:size(3))
+	rois_j:select(2, 3):mul(-1):add(images_j:size(3))
+
+	local tmp = rois_j:select(2, 1):clone()
+	rois_j:select(2, 1):copy(rois_j:select(2, 3))
+	rois_j:select(2, 3):copy(tmp)
+end
+
+local function insert_dummy_dim1(...)
+	for _, tensor in ipairs({...}) do
+		tensor:resize(1, unpack(tensor:size():totable()))
+	end
+end
+
+function ExampleLoader:makeBatchTable(batchSize, isTrainingPhase)
+	local o = self:getPhaseOpts(isTrainingPhase)
+	local num_jittered_copies = isTrainingPhase and 2 or (1 + (o.hflips and 2 or 1) * o.numScales)
+
+	return table2d(batchSize, num_jittered_copies, function() return {torch.FloatTensor(), torch.FloatTensor(), torch.FloatTensor()} end)
+end
+
+function ExampleLoader:loadExample(exampleIdx, isTrainingPhase)
+	local o = self:getPhaseOpts(isTrainingPhase)
+
+	local labels_loaded = self.dataset[o.subset]:getLabels(exampleIdx)
+	local rois_loaded = self.dataset[o.subset]:getProposals(exampleIdx)
+	local jpeg_loaded = self.dataset[o.subset]:getJpegBytes(exampleIdx)
+	local scales = o.scales or self.scales
+	local channel_order, rgb_pixel_means = self.channel_order, self.rgb_pixel_means
+
+	local scale_inds = isTrainingPhase and {0, torch.random(1, o.numScales)} or torch.range(0, o.numScales):totable()
+	local hflips = isTrainingPhase and (o.hflips and torch.random(0, 1) or 0) or (o.hflips and 2 or 0) -- 0 is no_flip, 1 is do_flip, 2 is both
+	local rois_perm = isTrainingPhase and torch.randperm(rois_loaded:size(1)) or torch.range(1, rois_loaded:size(1))
+
+	return function(indexInBatch, batchTable)
+		image = image or require 'image'
+		local img_original = image.decompressJPG(jpeg_loaded, 3, 'byte')
+		local height_original, width_original = img_original:size(2), img_original:size(3)
+
+		local rois_scale0 = rois_loaded:index(1, rois_perm:sub(1, math.min(rois_loaded:size(1), o.numRoisPerImage)):long())
+		rois_scale0[1]:copy(torch.FloatTensor{0, 0, width_original - 1, height_original - 1, 0.0}:sub(1, rois_scale0:size(2)))
+
+		for j, scale_ind in ipairs(scale_inds) do
+			local images, rois, labels = unpack(batchTable[indexInBatch][j])
+
+			local img_scaled = scale_ind == 0 and img_original:clone() or rescale(img_original, scales[scale_ind][1], scales[scale_ind][2])
+			local width_scaled, height_scaled = img_scaled:size(3), img_scaled:size(2)
+
+			subtract_mean(images:resize(img_scaled:size()), img_scaled, channel_order, rgb_pixel_means)
+			rois:cmul(rois_scale0, torch.FloatTensor{{width_scaled / width_original, height_scaled / height_original, width_scaled / width_original, height_scaled / height_original, 1.0}}:narrow(2, 1, rois_scale0:size(2)):contiguous():expandAs(rois_scale0))
+			labels:resize(labels_loaded:size()):copy(labels_loaded)
+
+			if hflips == 1 then
+				flip(images, rois)
+			elseif scale_ind ~= 0 and hflips == 2 then
+				local jj = #batchTable[indexInBatch] - j + 2
+				local images_flipped, rois_flipped, labels_flipped = unpack(batchTable[indexInBatch][jj])
+				images_flipped:resizeAs(images):copy(images)
+				rois_flipped:resizeAs(rois):copy(rois)
+				labels_flipped:resizeAs(labels):copy(labels)
+				flip(images_flipped, rois_flipped)
+				insert_dummy_dim1(images_flipped, rois_flipped, labels_flipped)
+			end
+
+			insert_dummy_dim1(images, rois, labels)
+		end
+
+		collectgarbage()
+	end
+end
+
+function ExampleLoader:getNumExamples(isTrainingPhase)
+	return self.dataset[self:getSubset(isTrainingPhase)]:getNumExamples()
+end
+
+function ExampleLoader:getPhaseOpts(isTrainingPhase)
+	return isTrainingPhase and self.example_loader_opts['training'] or self.example_loader_opts['evaluate']
+end
+
+function ExampleLoader:getSubset(isTrainingPhase)
+	return self:getPhaseOpts(isTrainingPhase).subset
+end