Project Profile

LadaGAN–EfficientNetV2

從人臉生成、可追蹤 synthetic dataset 到判別器評估的完整深度學習實驗。

Completed
Status
Completed
Portfolio
Complete portfolio
Role
Completed deep-learning course project
GitHub Read overview
GANDeep LearningComputer VisionEvaluation

LadaGAN–EfficientNetV2 是一個已完成的深度學習實驗。它把生成模型、synthetic dataset 建立與判別器評估串成一條可檢查的 pipeline，並保留失敗模型與資料切分限制，而不是只展示最漂亮的 sample。

專案範圍

  • 在 CelebA 上訓練 LadaGAN 並保留多個 checkpoint
  • 由 checkpoint 產生 synthetic faces，再建立 real/fake 判別資料
  • 使用 EfficientNetV2-B1 檢查生成分布差異與評估邊界

這個 Profile 是完整作品集中的 compact case file；模型演進、圖表與評估 caveats 放在相關 Overview Article。

What changed or was learned

可驗證成果

A traceable synthetic-image pipeline

The experiment preserves 30 checkpoints, 150,000 synthetic faces, and the real/fake dataset construction used to train the discriminator.

A result with an explicit evidence boundary

The documented 30,000-image evaluation reports 29,972 correct predictions, while keeping split and data-pipeline limitations visible.

Traceable sources

Evidence Trail

  1. Overview Article

    The complete experiment record, including failed baselines, checkpoint grids, and evaluation caveats.

    Inspect source

  2. Source repository

    Training, generation, dataset, and EfficientNetV2 evaluation scripts.

    Inspect source

  3. Original technical report

    The HackMD report behind the preserved experiment figures and metrics.

    Inspect source

Writing