CS Course Learning

【李宏毅】2024大语言模型课程

课程学习

课程链接：https://speech.ee.ntu.edu.tw/~hylee/genai/2024-spring.php

Bilibili相关视频链接：https://www.bilibili.com/video/BV1XS411w7qr

GPT: Autoregressive model

In-context Learning

1. Chain of Thoughts (CoT)
2. Tree of Thoughts (ToT)
3. Algorithm of Thoughts (AoT)
4. ....

使用工具：

1. 搜寻引擎 Retrieval Augmented Generation (RAG)
2. 写程序 Program of Thought (PoT)
3. 文字生图 DALL-E

Explainable ML:

1. Local Explanation
   • Saliency Map
   • SmoothGrad (improved Saliency Map)
   • Integrated Gradient(IG)
2. Global Explanation

Three steps of LLM training:

1. Pre-train -> Foundation model
2. Instruction Fine-tuning (Supervised Learning)
3. Reinforcement Learning from Human Feedback (RLHF)

Seq2seq:

• Syntactic Parsing (文法分析)

• Multi-label Classification (区别于 Multi-class Classification)

  An object can belong to multiple classes

• Object Detection

Tranformer:

• Self-attention
• Cross-attention

Copy Mechanism => Summarization

• Pointer Network

Attention Decoder

• Greedy Decoding (每次都选择输出概率最大的token)
• Bean Search
• Sampling (more creative, randomness is needed for decoder when generating)

作业总结

【李宏毅】2023机器学习系列课程

课程链接：https://speech.ee.ntu.edu.tw/~hylee/ml/2023-spring.php

课程学习

能够使用工具的AI：

• WebGPT
• Toolformer

作业总结

trainer

# trainer
n_epochs = config['n_epochs']
criterion = nn.MSELoss(reduction='mean')  # define loss function
optimizer = torch.optim.SGD(model.parameters(), lr=config['learning_rate'], momentum=0.7)  # define optimizerfor epoch in range(n_epochs):# trainmodel.train()loss_record = []for X, y in train_loader:optimizer.zero_grad()X, y = X.to(device), y.to(device)pred = model(X)loss = criterion(pred, y)loss.backward()optimizer.step()loss_record.append(loss.detach().item())  # loss value of a batch : loss.detach().item()mean_train_loss = sum(loss_record) / len(loss_record)# evaluatemodel.eval()loss_record = []with torch.no_grad():for X, y in valid_loader:X, y = X.to(device), y.to(device)pred = model(X)loss = criterion(pred, y)loss_record.append(loss.detach().item())  # loss value of a batch : loss.detach().item()mean_eval_loss = sum(loss_record) / len(loss_record)

tensorboard

.from torch.utils.tensorboard import SummaryWriterwriter = SummaryWriter() # Writer of tensoboard.
writer.add_scalar('Loss/train', mean_train_loss, step)
"""
def add_scalar(tag: Any,  # 图表的名称scalar_value: Any,  # 纵坐标取值global_step: Any | None = None,  # 横坐标取值walltime: Any | None = None,new_style: bool = False,double_precision: bool = False
)
"""

【ETH】2020 Digital Design and Computer Architecture

课程链接：https://safari.ethz.ch/digitaltechnik/spring2020/doku.php?id=start

课程视频链接：https://www.youtube.com/playlist?list=PL5Q2soXY2Zi_FRrloMa2fUYWPGiZUBQo2

课程学习

作业总结

【UCB】2020 Structure and Interpretation of Computer Programs

课程链接：https://web.archive.org/web/20210104105406/https://cs61a.org/

课程视频链接：https://www.bilibili.com/video/BV1s3411G7yM/

课程学习

作业总结