Fader Networks: Manipulating Images by Sliding Attributes

Slide 1

Slide 1 text

Fader Networks: Manipulating Images by Sliding Attributes NIPS 2017 ஛Լ ᰜଠ࿠

Slide 2

Slide 2 text

No content

Slide 3

Slide 3 text

֓ཁ 1. Encoder-Decoder FrameworkͰ 2. ը૾ͷಛ௃ͷมߋΛ͢ΔϞσϧΛֶश 3. มߋ౓߹͍ͷௐ੔΋Ͱ͖Δ

Slide 4

Slide 4 text

࣮ߦྫ

Slide 5

Slide 5 text

શମߏ੒ʢޙͰ·ͨ໭ͬͯ͘Δʣ

Slide 6

Slide 6 text

ࣄલ஌ࣝ

Slide 7

Slide 7 text

Encoder-Decoder https://blog.keras.io/building-autoencoders-in-keras.html

Slide 8

Slide 8 text

GAN (Generative Adversarial Network) https://deeplearning4j.org/generative-adversarial-network

Slide 9

Slide 9 text

ຊ࿦จͷ಺༰

Slide 10

Slide 10 text

Encoder-Decoder E Ͱݩը૾ʢxʣΛEncoderͯ͠D Ͱ෮ݩͨ͠΋ͷͱ ݩը૾ʢxʣͱͰMSEʹΑΓޡࠩΛܭࢉ

Slide 11

Slide 11 text

Encoder-Decoder E Ͱݩը૾ʢxʣΛEncodeͯ͠D ͰDecodeͨ͠΋ͷͱ ݩը૾ʢxʣͱͰMSEʹΑΓޡࠩΛܭࢉ ͜ͷ··ͩͱಛ௃Ͱ͋Δy͕ແࢹ͞ΕΔͷͰ ੑผ౳ͷมߋ͕ޙ͔ΒͰ͖ͳ͍

Slide 12

Slide 12 text

Discriminator objective Discriminator͸EncoderͰಘΒΕͨը૾ͷlatent feature͔Β ͦͷαϯϓϧͷಛ௃͕ͳΜͳͷ͔ਪఆ͢Δɻ

Slide 13

Slide 13 text

Adversarial objective Encoder͸Discriminatorͷਪఆਫ਼౓ΛԼ͛ΔΑ͏ʹֶश͢Δ

Slide 14

Slide 14 text

Adversarial objective Encoder͸Discriminatorͷਪఆਫ਼౓ΛԼ͛ΔΑ͏ʹֶश͢Δ Encoder͸ը૾ͷyʹΑΒͳ͍ ීวతͳಛ௃ྔΛநग़͢Δ͜ͱ͕Ͱ͖ΔΑ͏ʹͳΔ

Slide 15

Slide 15 text

Learning algorithm

Slide 16

Slide 16 text

શମߏ੒

Slide 17

Slide 17 text

Result 1

Slide 18

Slide 18 text

Result 2

Slide 19

Slide 19 text

Result 3

Slide 20

Slide 20 text

·ͱΊ • Encoder-Decoderͷ୯७ͳϑϨʔϜϫʔΫͰը૾ͷಛ ௃ม׵ϞσϧΛֶशͨ͠ɻ • Adversarial objectiveΛಋೖ͢Δ͜ͱͰಛ௃ʹґଘ͠ͳ ͍ը૾ͷ৘ใΛ֫ಘ͢Δ͜ͱ͕Ͱ͖ͨɻ • Networkࣗମͷߏ଄͸ඇৗʹγϯϓϧ͕ͩैདྷख๏Λ্ ճΔ੒੷Λग़ͨ͠ɻ