DeepLearningSchool

Part I

1. Transfer learning

• task: img classification (Simpsons dataset)
• architectures:
  • alexnet
  • vgg16
  • inception_v3
  • resnext101_32x8d
  • efficientnet-b4
  • mobilenet_v3_large
  • googlenet

2. Semantic segmentation

• task: binary semantic segmetation of medical imgs (PH2Dataset)
• SegNet
• IoU metric
• U-Net
• loss:
  • BCE
  • Dice
  • Focal
  • Lovasz

3. Autoencoders

• tasks:
  • reconstruction
  • sampling
  • img augs (smiles, glasses)
  • denoising
  • search for similar imgs
• architectures:
  • vanilla autoencoder
  • variational autoencoder
    • loss: log_likelihood + KL_divergence
  • conditional VAE (generating numbers)

Denoising (AE)

Mnist generating (VAE)

4. Generative adversarial networks

5. Project 3D ML

Project tasks:

• find a person in a video
• get it's body pose
• take a garment and change it's pose according to bosy pose
• render the garment in a video

3dml.mp4

Part II

1. Simple embeddings

• ranking task (Stackoverflow questions, cosine similarity)
• Hits@K metric
• Word2Vec
• tokenization
• normalization
  • stemming
  • lemmatization

2. Embeddings

• task: semantic classification of Tweets
• Average embedding
• Embeddings for unknown words
  • context
  • TF-IDF

3. RNN

• classification task: IMDB reviews (positive/negative)
• GRU
• LSTM
• CNN

4. Language modeling

• task: Part-Of-Speech Tagging
• HiddenMarkovModel
• NLTK
• Rnnmorph
• BiLSTMTagger

5. Neural Machine Translation

• task: russian to english
• Seq2Seq
  • Encoder RNN
  • Decoder RNN
    • Attention
    • Teacher forcing
  • Bleu metric

Source: в доме имеется кухня .
Target original: the unit is fitted with a kitchen .
Target generated: the unit is equipped with a kitchen .

6. Transformers

• task: classification (sentiment analysis of tweets)

GPT

• HuggingFace
• fine-tuning
• attention maps
• lr scheduler with warmup

Attention map:

BERT

• fine-tuning

7. Summatization
8. Audio
9. Project