GAN(Generative Adversarial Networks)ã‚’å¦ç¿’ã•ã›ã‚‹éš›ã®14ã®ãƒ†ã‚¯ãƒ‹ãƒƒã‚¯ - Qiita※ã“ã®è¨˜äº‹ã¯"How to Train a GAN?" at NIPS2016ã‚’ã€ã“ã“ã‚’å‚考ã«ã—ã¦ã€ç§ãªã‚Šã«è§£é‡ˆã—ã¦ã€é‡è¦ãã†ãªéƒ¨åˆ†ã‚’引用ã—ã¦ç¿»è¨³ã—ãŸã‚‚ã®ã§ã™ã€‚å½¹ã«ç«‹ã¤ã“ã¨ãŒå¤šã‹ã£ãŸã®ã§å…±æœ‰è‡´ã—ã¾ã™ã€‚ ※GANã®èª¬æ˜Žã¯çœç•¥ã—ã¾ã™ã€‚G=Generatorã€D=Discriminatorã¨ã—ã¦èª¬æ˜Žã—ã¾ã™ã€‚ ※ãã‚Œãžã‚Œã®ãƒ†ã‚¯ãƒ‹ãƒƒã‚¯ã«æ ¹æ‹ ã¯ã‚ã¾ã‚Šæ›¸ã„ã¦ãªã„ã§ã™ãŒã€ã»ã¨ã‚“ã©ãŒè«–æ–‡ã«æ›¸ã„ã¦ã‚ã£ãŸã€Œã“ã†ã™ã‚Œã°ã†ã¾ãã„ã£ãŸã€ã¨ã„ã†ã‚‚ã®ã§ã™ã€‚GANã®å¦ç¿’ãŒã†ã¾ãã„ã‹ãªã„ã¨ãã«è©¦ã—ã¦ã¿ã‚‹ã¨è‰¯ã„ã¨æ€ã„ã¾ã™ã€‚ 1. 入力をæ£è¦åŒ– (Normalize the inputs) ・Dã®Input(=Gã®å‡ºåŠ›)ã¨ãªã‚‹ç”»åƒã‚’[-1,1]ã®ç¯„囲ã«æ£è¦åŒ–ã™ã‚‹ã€‚ ・Gã®å‡ºåŠ›ãŒ[-1,1]ã¨ãªã‚‹ã‚ˆã†ã«ã€Outputã®ã¨ã“ã‚ã‚’Tanhã«ã™ã‚‹ã€‚ 2. Gã®Loss関数を修æ£ã™ã‚‹ (A modified loss function)
https://www.clouderaworldtokyo.com/session-download/C2-Cloudera_Tokyo_2016_DMM_submit%20.pdf
教師ãªã—ç”»åƒã®ãƒ™ã‚¯ãƒˆãƒ«åŒ–ã¨ã€ãƒ™ã‚¯ãƒˆãƒ«ã‹ã‚‰ã‚¿ã‚°ã‚’予想ã—ãŸã‚Šé¡žä¼¼åº¦ã‚’計算ã—ãŸã‚Šã™ã‚‹ - ã«ã»ã‚“ã”ã®ã‚Œã‚“ã—ã‚…ã†
ã€Day-17】DeepLearning系ライブラリã€ã€ŽKerasã€ã®ä½¿ã„æ–¹ã¾ã¨ã‚(2.x対応版) - プãƒã‚¯ãƒ©ã‚·ã‚¹ãƒˆ
DeepLearningç ”ç©¶ 2016å¹´ã®ã¾ã¨ã‚ - Qiita
ゼãƒã‹ã‚‰Deepã¾ã§å¦ã¶å¼·åŒ–å¦ç¿’ - Qiita
WaveNet: A Generative Model for Raw Audio | DeepMind
TensorFlow Playgroundã¯ãƒ‹ãƒ¥ãƒ¼ãƒ©ãƒ«ãƒãƒƒãƒˆã‚’ç†è§£ã™ã‚‹ã®ã«ãŠã‚¹ã‚¹ãƒ¡ - 西尾泰和ã®ã¯ã¦ãªãƒ€ã‚¤ã‚¢ãƒªãƒ¼
DeepLearningを使ã£ãŸå®Ÿè£…ã‚’çºã‚ã¦ã¿ãŸ - ã®ã‚“ã³ã‚Šã—ã¦ã„るエンジニアã®æ—¥è¨˜
Deep Learningã®éŽåŽ»ã¨æœªæ¥ ～黒é”è¡“ã‹ã‚‰ã®è„±å´ã¸å‘ã‘ã¦ï½ž
ツイートã‹ã‚‰å¥½ã¿ã®ã‚¿ã‚¤ãƒ—ã®å¥³ã®åを見ã¤ã‘るセンãƒãƒ¡ãƒ³ãƒˆåˆ†æžã¨Deep Learning | æ ªå¼ä¼šç¤¾ã‚µã‚¤ãƒãƒ¼ã‚¨ãƒ¼ã‚¸ã‚§ãƒ³ãƒˆ
{{#tags}}- {{label}}
{{/tags}}