GAN? ???, ? ???? ???? ????: ? ??? ??? A to Z
?
2 likes?1,970 views
The document discusses using GANs to create an AI that can mimic a person's handwriting. It describes training a model using a two-step process: 1) training an autoencoder to reconstruct handwriting images and 2) training a GAN using the autoencoder and additional images to transfer the style of text from one sample to another. The model was trained on Google Cloud Platform using their GPUs and the trained model can successfully transfer the style of handwriting from one sample to new text in a novel way.
![/12679
[Ground Truth]
[Generated]](https://image.slidesharecdn.com/gdgdevfest2019handwritingstyler-191226150023/85/GAN-A-to-Z-79-320.jpg)
![/12685
Category vector
[0, 0, 0, 0, 1, 0, 0, ]
one-hot vector
Gaussian normal random vector
[0.253.., 0.891.., 0.412.., 0.309.., ] (128 )](https://image.slidesharecdn.com/gdgdevfest2019handwritingstyler-191226150023/85/GAN-A-to-Z-85-320.jpg)
GAN? ???, ? ???? ???? ????: ? ??? ??? A to Z
- 1. GAN? ???, ? ???? ???? ???? : ? ??? ??? ???? A to Z ???
- 2. /1262 ??? Min Jung Chung # _ _ # , # , # , # ,
- 3. /1263
- 5. /1265 # _ # _ # _ _ ,
- 6. /1266 1. - - - - , 2. - - (GAN) - - 2-step Training - 3. - ? - -
- 8. /1268 # -
- 9. /1269 # -
- 10. /12610
- 11. /12611 Content Style New Image
- 12. /12612
- 13. /12613 /
- 15. /12615 # -
- 16. /12616 # -
- 17. /12617
- 18. /12618 ? ? ? ?
- 25. /12625 # -
- 26. /12626 # -
- 27. /12627
- 28. /12628
- 33. /12633, ..
- 36. /12636 : !
- 37. /12637 # / -
- 38. /12638 # / -
- 39. /12639
- 40. /12640 ??? ???
- 43. /12643 # 1. : / , / 2. : 4. / : , 3. : ,
- 44. /12644 # - 2 step training
- 45. /12645 # - 2 step training
- 46. /12646 ,
- 48. /12648 # - GAN & Unet
- 49. /12649 # - GAN & Unet
- 50. /12650
- 51. /12651 Input : Noise Output : Target Image # Generator : Noise # Discriminator : Fake Image , Real Image (Adversarial) , Generator Real
- 52. /12652 Input : Source Image ( ) Output : Target Image ( ) # : , , Noise GAN !
- 53. /12653
- 54. /12654 Input : Source Image ( ) Output : Target Image ( - ) # " , !
- 55. /12655
- 56. /12656 Input : Source Image ( ) Output : Target Image ( ) # # Auto Encoder & GAN (pix2pix)
- 57. /12657 Input : Source Image ( ) Output : Target Image ( ) # # Auto Encoder & GAN (pix2pix)
- 58. /12658
- 61. /12661 Concatenate ( ) Category vector Encoder Decoder ,
- 62. /12662 # -
- 63. /12663 # -
- 65. /12665
- 66. /12666 Crop Resize Padding https://github.com/jeina7/Handwriting_styler/blob/master/common/utils.py
- 67. /12667 Crop Resize Padding https://github.com/jeina7/Handwriting_styler/blob/master/common/utils.py
- 68. /12668 # Pre-Training & Transfer Learning
- 69. /12669 # Pre-Training & Transfer Learning
- 70. /12670 (150 epoch)GPU : P100 (50hrs)
- 71. /12671?
- 73. /12673??
- 74. /12674 !
- 75. /12675
- 76. /12676 (350 epoch)GPU : P100 (20min)
- 83. /12683 ?
- 85. /12685 Category vector [0, 0, 0, 0, 1, 0, 0, ] one-hot vector Gaussian normal random vector [0.253.., 0.891.., 0.412.., 0.309.., ] (128 )
- 86. /12686
- 87. /12687 STEP!
- 88. /12688 STEP! ,
- 89. /12689
- 91. /126 WALKING IN THE LATENT SPACE 91 Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks, 2016
- 92. /12692 # : GCP GPU
- 93. /12693 # : GCP GPU
- 94. /12694
- 95. /12695 , ..
- 99. /12699 ? ?Wi?
- 100. /126100 GPU
- 101. /126101 (6464) (128128) 50sec 180sec 18sec - (100 iteration ) ,, 3 ! 1 10
- 104. /126104 : $300 / VM Google Cloud SDK : CUDA Jupyter Notebook ! https://jeinalog.tistory.com/8
- 105. /126105 # : ,
- 106. /126106 # : ,
- 107. /126107
- 109. /126109 , ,
- 110. /126110
- 111. /126111 # ?
- 112. /126112 # ?
- 114. /126114 https://www.facebook.com/groups/TensorFlowKR/permalink/977774049230320/ 5-6 9 , Clova OCR ? DEVIEW ,, ,, # / .
- 115. /126115 # / .https://deview.kr/2019/schedule/294# : DEVIEW 2019
- 116. /126116 ! # / .https://deview.kr/2019/schedule/294# : DEVIEW 2019
- 117. /126117 # / .https://deview.kr/2019/schedule/294# : DEVIEW 2019 Wow..
- 118. /126118 #
- 119. /126119 #
- 120. /126120 ? ,
- 121. /126121 , . . ! , , ! , !
- 122. /126122 , .. ! , ,
- 123. /126123 0 10 =0
- 124. /126124 0 10 =0 1 10 =10
- 126. /126126 Jeinas DevLog https://jeinalog.tistory.com/ GitHub https://github.com/jeina7 , Wrinie (1) - https://jeinalog.tistory.com/15 , Wrinie (2) - https://jeinalog.tistory.com/16 https://github.com/jeina7/Handwriting_styler E-mail jeina.code@gmail.com Instagram @ai.developer