update

Author: SerialLain3170

pix2pixHD/README.md

@@ -2,24 +2,26 @@
 
 ## Summary
 ![net](https://github.com/SerialLain3170/Line-to-Color/blob/master/pix2pixHD/network.png)
-- Coarse-to-fine GeneratorとMulti-scale discriminatorによって多段的に高解像度へ対応
-- Instance-level Feature Embeddingでインスタンス毎に違うマッピングを行う
+
+- This model is the update version of pix2pix.
+- The authors of this paper can generate high-resolution images by proposing coarse-to-fine generator and multi-scale discriminator.
 
 ## Usage
+Execute the command line below and you can pre-train glbal generator.
 ```py
 $ python pretrain.py
 ```
-でGlobal Generatorを事前に学習、そして
+
+Execute the command line below and you can train full architectures contained with local enhancer.
 ```py
 $ python train.py
 ```
-でLocal Enhancerも含めて全て学習
 
 ## Result
-私の環境で生成した例を以下に示す。
+Images generated by my development environment is below.
 ![result](https://github.com/SerialLain3170/Line-to-Color/blob/master/pix2pixHD/visualize_125.png)
 
-- バッチサイズは4
-- 最適化手法はAdam(α=0.0002, β1=0.5)
-- Multi-Scale Discriminatorによるlossの重みは10
-- まだ学習はしっかり出来ていない、というかGlobal Generatorの事前学習どうやってるんだろうか.....
+- Batch size: 4
+- Using Adam as optimizer
+- The weight of adversarial loss is 10.0
+- I think this code is not completed because no method which enables global generator to pre-train is described in this paper