summary2

发表于 更新于 分类于
本文字数： 1.7k 阅读时长 ≈ 2 分钟

公开课

Machine Learning 2021 Spring：

  • Input as vector
    • One-Hot encoding
    • Word Embedding
  • Sequence Labeling
    • Fully-Connected Network
    • Self-attention
    • Multi-head Self-attention
    • Positional Encoding + Self-attention

PyTorch Fundamentals by MicroSoft

  • Natural Language Tasks （Text Classification、Intent Classification、Sentiment Analysis、Named Entity Recognition、Keyword extraction、Text Summarization ）

  • Text Classfication with PyTorch (torchtext)

    • Representing text as Tensors : (Character-level, Word-level )

    • Dataset : AG_NEWS (4 categories)

    • Bag of words text representation

      • Training Bow classifier (simple linear layer) val_acc: 90%

      • Term Frequency Inverse Document Frequency (TF-IDF)

    • Embeddings text represent

      • word index in vocab as Input, padding minibatch into same length

      • Training Embedding classifier (Embedding layer, linear layer)

      • python train_loss: 0.131295 train_acc: 0.959406 val_loss: 0.258924 val_acc: 0.916708

      • Variable-Length Sequence Representation，offset vector without padding batch

    • Using Pre-Trained Semantic Embeddings: Word2Vec

      • val_acc: 92.67%

    • using RNN

      • use padded data loader
      • model ( Embedding layer + RNN + linear layer)
      • Val_acc : 90%

    • Long Short Term Memory (LSTM)

      • Packed sequences
      • Val_acc : 91%

    • Using BERT for text classification

      • training can not be finished by cpu

  • Generate text with LSTM

    • Building character vocabulary
    • training generative LSTM
    • Soft text generation and temperature

CVPR 2021 论文分享会

  • Session1 图像生成

    • Information Bottleneck Disentanglement for Identity Swapping（换脸）

    • Leveraging Line-point Consistent to Preserve Structure for Wide Parallax Image Stitching

    • Facelnpainter: High Fidelity Face Adaptation to Heterogeneous Domains

  • Session2 图像处理

    • Deep Homography for Efficient Stereo Image Compression（CVPR-2021 oral）
    • Learning Scalable \(\ell_\infty\)-constrained Near-lossless Image Compression via Joint Lossy Image and Residual Compression

  • Session3 底层视觉

    • Neighbor2 Neighbor Self-supervised Denoising from Single Noisy Images
    • Deep Animation Video Interpolation in the Wild