Add noise_scale training

lib/pairwise_transform.lua

@@ -3,7 +3,6 @@ local pairwise_transform = {}
 
 pairwise_transform = tablex.update(pairwise_transform, require('pairwise_transform_scale'))
 pairwise_transform = tablex.update(pairwise_transform, require('pairwise_transform_jpeg'))
-
-print(pairwise_transform)
+pairwise_transform = tablex.update(pairwise_transform, require('pairwise_transform_jpeg_scale'))
 
 return pairwise_transform

lib/pairwise_transform_jpeg_scale.lua

@@ -0,0 +1,175 @@
+local pairwise_utils = require 'pairwise_transform_utils'
+local iproc = require 'iproc'
+local gm = require 'graphicsmagick'
+local pairwise_transform = {}
+
+local function add_jpeg_noise_(x, quality, options)
+   for i = 1, #quality do
+      x = gm.Image(x, "RGB", "DHW")
+      x:format("jpeg"):depth(8)
+      if torch.uniform() < options.jpeg_chroma_subsampling_rate then
+	 -- YUV 420
+	 x:samplingFactors({2.0, 1.0, 1.0})
+      else
+	 -- YUV 444
+	 x:samplingFactors({1.0, 1.0, 1.0})
+      end
+      local blob, len = x:toBlob(quality[i])
+      x:fromBlob(blob, len)
+      x = x:toTensor("byte", "RGB", "DHW")
+   end
+   return x
+end
+
+local function add_jpeg_noise(src, style, level, options)
+   if style == "art" then
+      if level == 1 then
+	 return add_jpeg_noise_(src, {torch.random(65, 85)}, options)
+      elseif level == 2 or level == 3 then
+	 -- level 2/3 adjusting by -nr_rate. for level3, -nr_rate=1
+	 local r = torch.uniform()
+	 if r > 0.6 then
+	    return add_jpeg_noise_(src, {torch.random(27, 70)}, options)
+	 elseif r > 0.3 then
+	    local quality1 = torch.random(37, 70)
+	    local quality2 = quality1 - torch.random(5, 10)
+	    return add_jpeg_noise_(src, {quality1, quality2}, options)
+	 else
+	    local quality1 = torch.random(52, 70)
+	    local quality2 = quality1 - torch.random(5, 15)
+	    local quality3 = quality1 - torch.random(15, 25)
+	    return add_jpeg_noise_(src, {quality1, quality2, quality3}, options)
+	 end
+      else
+	 error("unknown noise level: " .. level)
+      end
+   elseif style == "photo" then
+      -- level adjusting by -nr_rate
+      return add_jpeg_noise_(src, {torch.random(30, 70)}, options)
+   else
+      error("unknown style: " .. style)
+   end
+end
+
+function pairwise_transform.jpeg_scale(src, scale, style, noise_level, size, offset, n, options)
+   local filters = options.downsampling_filters
+   if options.data.filters then
+      filters = options.data.filters
+   end
+   local unstable_region_offset = 8
+   local downsampling_filter = filters[torch.random(1, #filters)]
+   local blur = torch.uniform(options.resize_blur_min, options.resize_blur_max)
+   local y = pairwise_utils.preprocess(src, size, options)
+   assert(y:size(2) % 4 == 0 and y:size(3) % 4 == 0)
+   local down_scale = 1.0 / scale
+   local x
+   if options.gamma_correction then
+      local small = iproc.scale_with_gamma22(y, y:size(3) * down_scale,
+					     y:size(2) * down_scale, downsampling_filter, blur)
+      if options.x_upsampling then
+	 x = iproc.scale(small, y:size(3), y:size(2), options.upsampling_filter)
+      else
+	 x = small
+      end
+   else
+      local small = iproc.scale(y, y:size(3) * down_scale,
+				  y:size(2) * down_scale, downsampling_filter, blur)
+      if options.x_upsampling then
+	 x = iproc.scale(small, y:size(3), y:size(2), options.upsampling_filter)
+      else
+	 x = small
+      end
+   end
+   x = add_jpeg_noise(x, style, noise_level, options)
+   local scale_inner = scale
+   if options.x_upsampling then
+      scale_inner = 1
+   end
+   x = iproc.crop(x, unstable_region_offset, unstable_region_offset,
+		  x:size(3) - unstable_region_offset, x:size(2) - unstable_region_offset)
+   y = iproc.crop(y, unstable_region_offset * scale_inner, unstable_region_offset * scale_inner,
+		  y:size(3) - unstable_region_offset * scale_inner, y:size(2) - unstable_region_offset * scale_inner)
+   if options.x_upsampling then
+      assert(x:size(2) % 4 == 0 and x:size(3) % 4 == 0)
+      assert(x:size(1) == y:size(1) and x:size(2) == y:size(2) and x:size(3) == y:size(3))
+   else
+      assert(x:size(1) == y:size(1) and x:size(2) * scale == y:size(2) and x:size(3) * scale == y:size(3))
+   end
+   local batch = {}
+   local lowres_y = gm.Image(y, "RGB", "DHW"):
+      size(y:size(3) * 0.5, y:size(2) * 0.5, "Box"):
+      size(y:size(3), y:size(2), "Box"):
+      toTensor(t, "RGB", "DHW")
+   local xs = {}
+   local ys = {}
+   local lowreses = {}
+
+   for j = 1, 2 do
+      -- TTA
+      local xi, yi, ri
+      if j == 1 then
+	 xi = x
+	 yi = y
+	 ri = lowres_y
+      else
+	 xi = x:transpose(2, 3):contiguous()
+	 yi = y:transpose(2, 3):contiguous()
+	 ri = lowres_y:transpose(2, 3):contiguous()
+      end
+      local xv = image.vflip(xi)
+      local yv = image.vflip(yi)
+      local rv = image.vflip(ri)
+      table.insert(xs, xi)
+      table.insert(ys, yi)
+      table.insert(lowreses, ri)
+      table.insert(xs, xv)
+      table.insert(ys, yv)
+      table.insert(lowreses, rv)
+      table.insert(xs, image.hflip(xi))
+      table.insert(ys, image.hflip(yi))
+      table.insert(lowreses, image.hflip(ri))
+      table.insert(xs, image.hflip(xv))
+      table.insert(ys, image.hflip(yv))
+      table.insert(lowreses, image.hflip(rv))
+   end
+   for i = 1, n do
+      local t = (i % #xs) + 1
+      local xc, yc = pairwise_utils.active_cropping(xs[t], ys[t], lowreses[t],
+						    size,
+						    scale_inner,
+						    options.active_cropping_rate,
+						    options.active_cropping_tries)
+      xc = iproc.byte2float(xc)
+      yc = iproc.byte2float(yc)
+      if options.rgb then
+      else
+	 yc = image.rgb2yuv(yc)[1]:reshape(1, yc:size(2), yc:size(3))
+	 xc = image.rgb2yuv(xc)[1]:reshape(1, xc:size(2), xc:size(3))
+      end
+      table.insert(batch, {xc, iproc.crop(yc, offset, offset, size - offset, size - offset)})
+   end
+   return batch
+end
+function pairwise_transform.test_jpeg_scale(src)
+   torch.setdefaulttensortype("torch.FloatTensor")
+   local options = {random_color_noise_rate = 0.5,
+		    random_half_rate = 0.5,
+		    random_overlay_rate = 0.5,
+		    random_unsharp_mask_rate = 0.5,
+		    active_cropping_rate = 0.5,
+		    active_cropping_tries = 10,
+		    max_size = 256,
+		    x_upsampling = false,
+		    downsampling_filters = "Box",
+		    rgb = true
+   }
+   local image = require 'image'
+   local src = image.lena()
+
+   for i = 1, 10 do
+      local xy = pairwise_transform.jpeg_scale(src, 2.0, "art", 1, 128, 7, 1, options)
+      image.display({image = xy[1][1], legend = "y:" .. (i * 10), min = 0, max = 1})
+      image.display({image = xy[1][2], legend = "x:" .. (i * 10), min = 0, max = 1})
+   end
+end
+return pairwise_transform

lib/settings.lua

@@ -23,7 +23,7 @@ cmd:option("-data_dir", "./data", 'path to data directory')
 cmd:option("-backend", "cunn", '(cunn|cudnn)')
 cmd:option("-test", "images/miku_small.png", 'path to test image')
 cmd:option("-model_dir", "./models", 'model directory')
-cmd:option("-method", "scale", 'method to training (noise|scale)')
+cmd:option("-method", "scale", 'method to training (noise|scale|noise_scale)')
 cmd:option("-model", "vgg_7", 'model architecture (vgg_7|vgg_12|upconv_7|upconv_8_4x|dilated_7)')
 cmd:option("-noise_level", 1, '(1|2|3)')
 cmd:option("-style", "art", '(art|photo)')
@@ -33,13 +33,13 @@ cmd:option("-random_overlay_rate", 0.0, 'data augmentation using flipped image o
 cmd:option("-random_half_rate", 0.0, 'data augmentation using half resolution image (0.0-1.0)')
 cmd:option("-random_unsharp_mask_rate", 0.0, 'data augmentation using unsharp mask (0.0-1.0)')
 cmd:option("-scale", 2.0, 'scale factor (2)')
-cmd:option("-learning_rate", 0.0005, 'learning rate for adam')
+cmd:option("-learning_rate", 0.00025, 'learning rate for adam')
 cmd:option("-crop_size", 48, 'crop size')
 cmd:option("-max_size", 256, 'if image is larger than N, image will be crop randomly')
-cmd:option("-batch_size", 8, 'mini batch size')
-cmd:option("-patches", 16, 'number of patch samples')
-cmd:option("-inner_epoch", 4, 'number of inner epochs')
-cmd:option("-epoch", 50, 'number of epochs to run')
+cmd:option("-batch_size", 16, 'mini batch size')
+cmd:option("-patches", 64, 'number of patch samples')
+cmd:option("-inner_epoch", 1, 'number of inner epochs')
+cmd:option("-epoch", 100, 'number of epochs to run')
 cmd:option("-thread", -1, 'number of CPU threads')
 cmd:option("-jpeg_chroma_subsampling_rate", 0.0, 'the rate of YUV 4:2:0/YUV 4:4:4 in denoising training (0.0-1.0)')
 cmd:option("-validation_rate", 0.05, 'validation-set rate (number_of_training_images * validation_rate > 1)')
@@ -54,12 +54,12 @@ cmd:option("-gamma_correction", 0, 'Resizing with colorspace correction(sRGB:gam
 cmd:option("-upsampling_filter", "Box", 'upsampling filter for 2x scale training (dev)')
 cmd:option("-max_training_image_size", -1, 'if training image is larger than N, image will be crop randomly when data converting')
 cmd:option("-use_transparent_png", 0, 'use transparent png (0|1)')
-cmd:option("-resize_blur_min", 0.85, 'min blur parameter for ResizeImage')
+cmd:option("-resize_blur_min", 0.95, 'min blur parameter for ResizeImage')
 cmd:option("-resize_blur_max", 1.05, 'max blur parameter for ResizeImage')
-cmd:option("-oracle_rate", 0.0, '')
+cmd:option("-oracle_rate", 0.1, '')
 cmd:option("-oracle_drop_rate", 0.5, '')
 cmd:option("-learning_rate_decay", 3.0e-7, 'learning rate decay (learning_rate * 1/(1+num_of_data*patches*epoch))')
-cmd:option("-loss", "rgb", 'loss (rgb|y)')
+cmd:option("-loss", "y", 'loss (rgb|y)')
 
 local function to_bool(settings, name)
    if settings[name] == 1 then
@@ -92,6 +92,15 @@ if settings.save_history then
 					  settings.model_dir, settings.scale)
       settings.model_file_best = string.format("%s/scale%.1fx_model.t7",
 					       settings.model_dir, settings.scale)
+   elseif settings.method == "noise_scale" then
+      settings.model_file = string.format("%s/noise%d_scale%.1fx_model.%%d-%%d.t7",
+					  settings.model_dir,
+					  settings.noise_level,
+					  settings.scale)
+      settings.model_file_best = string.format("%s/noise%d_scale%.1fx_model.t7",
+					       settings.model_dir,
+					       settings.noise_level, 
+					       settings.scale)
    else
       error("unknown method: " .. settings.method)
    end
@@ -102,6 +111,9 @@ else
    elseif settings.method == "scale" then
       settings.model_file = string.format("%s/scale%.1fx_model.t7",
 					  settings.model_dir, settings.scale)
+   elseif settings.method == "noise_scale" then
+      settings.model_file = string.format("%s/noise%d_scale%.1fx_model.t7",
+					  settings.model_dir, settings.noise_level, settings.scale)
    else
       error("unknown method: " .. settings.method)
    end

train.lua

@@ -178,6 +178,30 @@ local function transformer(model, x, is_validation, n, offset)
 				     settings.noise_level,
 				     settings.crop_size, offset,
 				     n, conf)
+   elseif settings.method == "noise_scale" then
+      local conf = tablex.update({
+	    downsampling_filters = settings.downsampling_filters,
+	    upsampling_filter = settings.upsampling_filter,
+	    random_half_rate = settings.random_half_rate,
+	    random_color_noise_rate = random_color_noise_rate,
+	    random_overlay_rate = random_overlay_rate,
+	    random_unsharp_mask_rate = settings.random_unsharp_mask_rate,
+	    max_size = settings.max_size,
+	    jpeg_chroma_subsampling_rate = settings.jpeg_chroma_subsampling_rate,
+	    nr_rate = settings.nr_rate,
+	    active_cropping_rate = active_cropping_rate,
+	    active_cropping_tries = active_cropping_tries,
+	    rgb = (settings.color == "rgb"),
+	    gamma_correction = settings.gamma_correction,
+	    x_upsampling = not reconstruct.has_resize(model),
+	    resize_blur_min = settings.resize_blur_min,
+	    resize_blur_max = settings.resize_blur_max}, meta)
+      return pairwise_transform.jpeg_scale(x,
+					   settings.scale,
+					   settings.style,
+					   settings.noise_level,
+					   settings.crop_size, offset,
+					   n, conf)
    end
 end
 
@@ -364,6 +388,12 @@ local function train()
 					("scale%.1f_best.%d-%d.png"):format(settings.scale,
 									    epoch, i))
 		  save_test_scale(model, test_image, log)
+	       elseif settings.method == "noise_scale" then
+		  local log = path.join(settings.model_dir,
+					("noise%d_scale%.1f_best.%d-%d.png"):format(settings.noise_level, 
+										    settings.scale,
+										    epoch, i))
+		  save_test_scale(model, test_image, log)
 	       end
 	    else
 	       torch.save(settings.model_file, model:clearState(), "ascii")
@@ -375,6 +405,11 @@ local function train()
 		  local log = path.join(settings.model_dir,
 					("scale%.1f_best.png"):format(settings.scale))
 		  save_test_scale(model, test_image, log)
+	       elseif settings.method == "noise_scale" then
+		  local log = path.join(settings.model_dir,
+					("noise%d_scale%.1f_best.png"):format(settings.noise_level, 
+									      settings.scale))
+		  save_test_scale(model, test_image, log)
 	       end
 	    end
 	 end