first commit

nagadomi · May 16, 2015 · 1273b36 · 1273b36
commit 1273b36
Show file tree

Hide file tree

Showing 19 changed files with 7,313 additions and 0 deletions.
diff --git a/.gitignore b/.gitignore
@@ -0,0 +1,2 @@
+*~
+cache/*.png
diff --git a/assets/index.html b/assets/index.html
@@ -0,0 +1,78 @@
+<html>
+  <head>
+    <style type="text/css" rel="http://cdnjs.cloudflare.com/ajax/libs/normalize/3.0.3/normalize.min.css"></style>
+    <style type="text/css">
+    body {
+      margin: 1em 2em 1em 2em;
+      background: LightGray;
+      width: 640px;
+    }
+    fieldset {
+      margin-top: 1em;
+      margin-bottom: 1em;
+    }
+    .about {
+      font-size: 0.8em;
+      margin: 1em 0 1em 0;
+    }
+    .help {
+      font-size: 0.8em;
+      margin: 1em 0 1em 0;
+    }
+    </style>
+    <script src="http://ajax.googleapis.com/ajax/libs/jquery/2.1.3/jquery.min.js"></script>
+    <script type="text/javascript">
+    function clear_file() {
+      var new_file = $("#file").clone();
+      new_file.change(clear_url);
+      $("#file").replaceWith(new_file);
+    }
+    function clear_url() {
+      $("#url").val("")
+    }
+    $(function (){
+      $("#url").change(clear_file);
+      $("#file").change(clear_url);
+    })
+    </script>
+  </head>
+  <body>
+    <h1>waifu2x</h1>
+    <div class="header">
+      <a href="https://github.com/you">
+	<img style="position: absolute; top: 0; left: 540; border: 0;" src="https://camo.githubusercontent.com/a6677b08c955af8400f44c6298f40e7d19cc5b2d/68747470733a2f2f73332e616d617a6f6e6177732e636f6d2f6769746875622f726962626f6e732f666f726b6d655f72696768745f677261795f3664366436642e706e67" alt="Fork me on GitHub" data-canonical-src="https://s3.amazonaws.com/github/ribbons/forkme_right_gray_6d6d6d.png">
+      </a>
+      <a href="index.ja.html">ja</a>/<a href="index.html">en</a>
+    </div>
+    <div class="about">
+      Single-Image Super-Resolution for anime/fan-arts using Deep Convolutional Neural Networks.
+    </div>
+    <form action="/api" method="POST" enctype="multipart/form-data" target="_blank">
+      <fieldset>
+	<legend>Image</legend>
+	<div>
+	  URL: <input id="url" type="text" name="url" size="64"/> or
+	</div>
+	<div>
+	  FILE: <input id="file" type="file" name="file"/>
+	</div>
+	<div class="help">
+	  Limits: FileSize: 2MB, Noise Reduction: 2560x2560px, Upscaling: 1280x1280px
+	</div>
+      </fieldset>
+      <fieldset>
+	<legend>Noise Reduction (expect JPEG Artifact)</legend>
+	<label><input type="radio" name="noise" value="0"> None</label>
+	<label><input type="radio" name="noise" value="1" checked="checked"> Low</label>
+	<label><input type="radio" name="noise" value="2"> High</label>
+      </fieldset>
+      <fieldset>
+	<legend>Upscaling</legend>
+	<label><input type="radio" name="scale" value="0" checked="checked"> None</label>
+	<label><input type="radio" name="scale" value="1"> 1.6x</label>
+	<label><input type="radio" name="scale" value="2"> 2x</label>
+      </fieldset>
+      <input type="submit"/>
+    </form>
+  </body>
+</html>
diff --git a/assets/index.ja.html b/assets/index.ja.html
@@ -0,0 +1,85 @@
+<html lang="ja">
+  <head>
+    <meta charset="UTF-8">
+    <style type="text/css" rel="http://cdnjs.cloudflare.com/ajax/libs/normalize/3.0.3/normalize.min.css"></style>
+    <style type="text/css">
+    body {
+      margin: 1em 2em 1em 2em;
+      background: LightGray;
+      width: 640px;
+    }
+    fieldset {
+      margin-top: 1em;
+        margin-bottom: 1em;
+    }
+    .about {
+      font-size: 0.8em;
+      margin: 1em 0 1em 0;
+    }
+    .help {
+      font-size: 0.8em;
+      margin: 1em 0 1em 0;
+    }
+    </style>
+    <script src="http://ajax.googleapis.com/ajax/libs/jquery/2.1.3/jquery.min.js"></script>
+    <script type="text/javascript">
+    function clear_file() {
+      var new_file = $("#file").clone();
+      new_file.change(clear_url);
+      $("#file").replaceWith(new_file);
+    }
+    function clear_url() {
+      $("#url").val("")
+    }
+    $(function (){
+      $("#url").change(clear_file);
+      $("#file").change(clear_url);
+    })
+    </script>
+  </head>
+  <body>
+    <h1>waifu2x</h1>
+    <div class="header">
+      <a href="https://github.com/you">
+	<img style="position: absolute; top: 0; left: 540; border: 0;" src="https://camo.githubusercontent.com/a6677b08c955af8400f44c6298f40e7d19cc5b2d/68747470733a2f2f73332e616d617a6f6e6177732e636f6d2f6769746875622f726962626f6e732f666f726b6d655f72696768745f677261795f3664366436642e706e67" alt="Fork me on GitHub" data-canonical-src="https://s3.amazonaws.com/github/ribbons/forkme_right_gray_6d6d6d.png">
+      </a>
+    </div>
+    <div class="about">
+      深層畳み込みニューラルネットワークによる二次元画像のための超解像システム.
+    </div>
+    <form action="/api" method="POST" enctype="multipart/form-data" target="_blank">
+      <fieldset>
+	<legend>Image</legend>
+	<div>
+	  URL: <input id="url" type="text" name="url" size="64"/> or
+	</div>
+	<div>
+	  FILE: <input id="file" type="file" name="file"/>
+	</div>
+	<div class="help">
+	  制限: サイズ: 2MB, ノイズ除去: 2560x2560px, 拡大: 1280x1280px
+	</div>
+      </fieldset>
+      <fieldset>
+	<legend>ノイズ除去 (JPEGノイズを想定)</legend>
+	<label><input type="radio" name="noise" value="0"> なし</label>
+	<label><input type="radio" name="noise" value="1" checked="checked"> 弱</label>
+	<label><input type="radio" name="noise" value="2"> 強</label>
+      </fieldset>
+      <fieldset>
+	<legend>拡大</legend>
+	<label><input type="radio" name="scale" value="0" checked="checked"> なし</label>
+	<label><input type="radio" name="scale" value="1"> 1.6x</label>
+	<label><input type="radio" name="scale" value="2"> 2x</label>
+      </fieldset>
+      <input type="submit" value="実行"/>
+    </form>
+    <div class="help">
+      <ul>
+	<li>なし/なしで入力画像を変換せずに出力する。ブラウザのタブで変換結果を比較したい人用。</li>
+	<li>JPEG画像であれば劣化がないように見えてもノイズ除去弱を推奨。</li>
+	<li>マンガスキャンの拡大はスクリーントーンが謎の模様に再構成されるため非対応。</li>
+      </ul>
+    </div>
+  </body>
+</html>
diff --git a/cache/.gitkeep b/cache/.gitkeep
diff --git a/cleanup_model.lua b/cleanup_model.lua
@@ -0,0 +1,69 @@
+require 'cunn'
+require 'cudnn'
+require './lib/LeakyReLU'
+
+torch.setdefaulttensortype("torch.FloatTensor")
+
+-- ref: https://github.com/torch/nn/issues/112#issuecomment-64427049
+local function zeroDataSize(data)
+   if type(data) == 'table' then
+      for i = 1, #data do
+	 data[i] = zeroDataSize(data[i])
+      end
+   elseif type(data) == 'userdata' then
+      data = torch.Tensor():typeAs(data)
+   end
+   return data
+end
+
+-- Resize the output, gradInput, etc temporary tensors to zero (so that the
+-- on disk size is smaller)
+local function cleanupModel(node)
+   if node.output ~= nil then
+      node.output = zeroDataSize(node.output)
+   end
+   if node.gradInput ~= nil then
+      node.gradInput = zeroDataSize(node.gradInput)
+   end
+   if node.finput ~= nil then
+      node.finput = zeroDataSize(node.finput)
+   end
+   if tostring(node) == "nn.LeakyReLU" then
+      if node.negative ~= nil then
+	 node.negative = zeroDataSize(node.negative)
+      end
+   end
+   if tostring(node) == "nn.Dropout" then
+      if node.noise ~= nil then
+	 node.noise = zeroDataSize(node.noise)
+      end
+   end
+   -- Recurse on nodes with 'modules'
+   if (node.modules ~= nil) then
+     if (type(node.modules) == 'table') then
+	for i = 1, #node.modules do
+	   local child = node.modules[i]
+	   cleanupModel(child)
+	end
+     end
+   end
+
+   collectgarbage()
+end
+
+local cmd = torch.CmdLine()
+cmd:text()
+cmd:text("cleanup model")
+cmd:text("Options:")
+cmd:option("-model", "./model.t7", 'path of model file')
+cmd:option("-iformat", "binary", 'input format')
+cmd:option("-oformat", "binary", 'output format')
+
+local opt = cmd:parse(arg)
+local model = torch.load(opt.model, opt.iformat)
+if model then
+   cleanupModel(model)
+   torch.save(opt.model, model, opt.oformat)
+else
+   error("model not found")
+end
diff --git a/lib/LeakyReLU.lua b/lib/LeakyReLU.lua
@@ -0,0 +1,30 @@
+if nn.LeakyReLU then
+   return
+end
+local LeakyReLU, parent = torch.class('nn.LeakyReLU','nn.Module')
+
+function LeakyReLU:__init(negative_scale)
+   parent.__init(self)
+   self.negative_scale = negative_scale or 0.333
+   self.negative = torch.Tensor()
+end
+
+function LeakyReLU:updateOutput(input)
+   self.output:resizeAs(input):copy(input):abs():add(input):div(2)
+   self.negative:resizeAs(input):copy(input):abs():add(-1.0, input):mul(-0.5*self.negative_scale)
+   self.output:add(self.negative)
+
+   return self.output
+end
+
+function LeakyReLU:updateGradInput(input, gradOutput)
+   self.gradInput:resizeAs(gradOutput)
+   -- filter positive
+   self.negative:sign():add(1)
+   torch.cmul(self.gradInput, gradOutput, self.negative)
+   -- filter negative
+   self.negative:add(-1):mul(-1 * self.negative_scale):cmul(gradOutput)
+   self.gradInput:add(self.negative)
+
+   return self.gradInput
+end
diff --git a/lib/image_loader.lua b/lib/image_loader.lua
@@ -0,0 +1,73 @@
+local gm = require 'graphicsmagick'
+require 'pl'
+
+local image_loader = {}
+
+function image_loader.decode_float(blob)
+   local im = image_loader.decode_byte(blob)
+   if im then
+      im = im:float():div(255)
+   end
+   return im
+end
+function image_loader.encode_png(tensor)
+   local im = gm.Image(tensor, "RGB", "DHW")
+   im:format("png")
+   return im:toBlob()
+end
+function image_loader.decode_byte(blob)
+   local load_image = function()
+      local im = gm.Image()
+      im:fromBlob(blob, #blob)
+      -- FIXME: How to detect that a image has an alpha channel?
+      if blob:sub(1, 4) == "\x89PNG" or blob:sub(1, 3) == "GIF" then
+	 -- merge alpha channel
+	 im = im:toTensor('float', 'RGBA', 'DHW')
+	 local w2 = im[4]
+	 local w1 = im[4] * -1 + 1
+	 local new_im = torch.FloatTensor(3, im:size(2), im:size(3))
+	 -- apply the white background
+	 new_im[1]:copy(im[1]):cmul(w2):add(w1)
+	 new_im[2]:copy(im[2]):cmul(w2):add(w1)
+	 new_im[3]:copy(im[3]):cmul(w2):add(w1)
+	 im = new_im:mul(255):byte()
+      else
+	 im = im:toTensor('byte', 'RGB', 'DHW')
+      end
+      return im
+   end
+   local state, ret = pcall(load_image)
+   if state then
+      return ret
+   else
+      return nil
+   end
+end
+function image_loader.load_float(file)
+   local fp = io.open(file, "rb")
+   local buff = fp:read("*a")
+   fp:close()
+   return image_loader.decode_float(buff)
+end
+function image_loader.load_byte(file)
+   local fp = io.open(file, "rb")
+   local buff = fp:read("*a")
+   fp:close()
+   return image_loader.decode_byte(buff)
+end
+local function test()
+   require 'image'
+   local img
+   img = image_loader.load_float("./a.jpg")
+   if img then
+      print(img:min())
+      print(img:max())
+      image.display(img)
+   end
+   img = image_loader.load_float("./b.png")
+   if img then
+      image.display(img)
+   end
+end
+--test()
+return image_loader
diff --git a/lib/iproc.lua b/lib/iproc.lua
@@ -0,0 +1,35 @@
+local gm = require 'graphicsmagick'
+local image = require 'image'
+local iproc = {}
+
+function iproc.sample(src, width, height)
+   local t = "float"
+   if src:type() == "torch.ByteTensor" then
+      t = "byte"
+   end
+   local im = gm.Image(src, "RGB", "DHW")
+   im:sample(math.ceil(width), math.ceil(height))
+   return im:toTensor(t, "RGB", "DHW")
+end
+function iproc.scale(src, width, height, filter)
+   local t = "float"
+   if src:type() == "torch.ByteTensor" then
+      t = "byte"
+   end
+   filter = filter or "Box"
+   local im = gm.Image(src, "RGB", "DHW")
+   im:size(math.ceil(width), math.ceil(height), filter)
+   return im:toTensor(t, "RGB", "DHW")
+end
+function iproc.padding(img, w1, w2, h1, h2)
+   local dst_height = img:size(2) + h1 + h2
+   local dst_width = img:size(3) + w1 + w2
+   local flow = torch.Tensor(2, dst_height, dst_width)
+   flow[1] = torch.ger(torch.linspace(0, dst_height -1, dst_height), torch.ones(dst_width))
+   flow[2] = torch.ger(torch.ones(dst_height), torch.linspace(0, dst_width - 1, dst_width))
+   flow[1]:add(-h1)
+   flow[2]:add(-w1)
+   return image.warp(img, flow, "simple", false, "clamp")
+end
+
+return iproc