Image caption project

規範(主要寫給hsuanchia看的)

1. 能寫中文就寫中文， 除了專有名詞跟大家都有共識的詞彙，否則一律使用中文
   • 包含commit, 註解...etc
2. Code 盡量寫的簡潔，否則一定要寫註解，不寫沒有人看得懂你在衝三小
3. 每次在 Colab 上改完一個段落，就更新上傳到 GitHub
4. 訓練好的Model請放到Models的資料夾下面，傳上來的model請務必要用commit來記錄訓練資訊
   • Model命名規則: Base on 哪篇paper的架構_000n ex.Visualattention_0001
   • 訓練資訊包含: 使用多少資料, epochs

Dataset

• MSCOCO 2017

Feature Map & Captions (pickle)

• link
• 忘了補上的新 dataset ...

Load

import pickle
with open('output_500.pkl', 'rb') as fp:
  data = pickle.load(fp)

data

[
    {
        'filename': '000000415840.jpg',
        'feature': NUMPY_ARRAY,
        'captions': [
            'A silver and red train traveling down a busy city street.',
            'A grey train passes on a city street.',
            'Electric rail cars move along a track between automobile traffic lanes. ',
            'A silver and orange bus is on a city street.',
            'The trolley for the San Francisco Zoo is on its tracks.'
        ]
    }
]

Architecture

Seq2seq

Soft Attention

Show-and-tell with visual attention (By 老師)