Fader Networks: Manipulating Images by Sliding Attributes

Fader Networks: Manipulating Images by Sliding Attributes for RINKO

Transcript

1. Fader Networks: Manipulating Images by Sliding Attributes NIPS 2017 ஛Լ ᰜଠ࿠
3. ֓ཁ 1. Encoder-Decoder FrameworkͰ 2. ը૾ͷಛ௃ͷมߋΛ͢ΔϞσϧΛֶश 3. มߋ౓߹͍ͷௐ੔΋Ͱ͖Δ
4. ࣮ߦྫ
5. શମߏ੒ʢޙͰ·ͨ໭ͬͯ͘Δʣ
6. ࣄલ஌ࣝ
7. Encoder-Decoder https://blog.keras.io/building-autoencoders-in-keras.html
8. GAN (Generative Adversarial Network) https://deeplearning4j.org/generative-adversarial-network
9. ຊ࿦จͷ಺༰
10. Encoder-Decoder E Ͱݩը૾ʢxʣΛEncoderͯ͠D Ͱ෮ݩͨ͠΋ͷͱ ݩը૾ʢxʣͱͰMSEʹΑΓޡࠩΛܭࢉ
11. Encoder-Decoder E Ͱݩը૾ʢxʣΛEncodeͯ͠D ͰDecodeͨ͠΋ͷͱ ݩը૾ʢxʣͱͰMSEʹΑΓޡࠩΛܭࢉ ͜ͷ··ͩͱಛ௃Ͱ͋Δy͕ແࢹ͞ΕΔͷͰ ੑผ౳ͷมߋ͕ޙ͔ΒͰ͖ͳ͍
12. Discriminator objective Discriminator͸EncoderͰಘΒΕͨը૾ͷlatent feature͔Β ͦͷαϯϓϧͷಛ௃͕ͳΜͳͷ͔ਪఆ͢Δɻ
13. Adversarial objective Encoder͸Discriminatorͷਪఆਫ਼౓ΛԼ͛ΔΑ͏ʹֶश͢Δ
14. Adversarial objective Encoder͸Discriminatorͷਪఆਫ਼౓ΛԼ͛ΔΑ͏ʹֶश͢Δ Encoder͸ը૾ͷyʹΑΒͳ͍ ීวతͳಛ௃ྔΛநग़͢Δ͜ͱ͕Ͱ͖ΔΑ͏ʹͳΔ
15. Learning algorithm
16. શମߏ੒
17. Result 1
18. Result 2
19. Result 3
20. ·ͱΊ • Encoder-Decoderͷ୯७ͳϑϨʔϜϫʔΫͰը૾ͷಛ ௃ม׵ϞσϧΛֶशͨ͠ɻ • Adversarial objectiveΛಋೖ͢Δ͜ͱͰಛ௃ʹґଘ͠ͳ ͍ը૾ͷ৘ใΛ֫ಘ͢Δ͜ͱ͕Ͱ͖ͨɻ • Networkࣗମͷߏ଄͸ඇৗʹγϯϓϧ͕ͩैདྷख๏Λ্ ճΔ੒੷Λग़ͨ͠ɻ