Deep Learning + PyTorch

All 5 Phases

Week-by-Week Schedule

Topics Covered

Every subtopic below is a separate daily task in the roadmap, with hand-picked resources (YouTube videos, docs, papers) for each.

• Karpathy "Neural Networks: Zero to Hero" — Episode 1: micrograd Part 1
• Episode 1 continued — code along
• Finish micrograd, internalize backprop fully
• Karpathy Episode 2: makemore (bigram model)
• Episode 2 continued
• Karpathy Episode 3: MLP (full 3-hour video)
• Re-implement micrograd from memory — no peeking
• Activation functions: ReLU, Sigmoid, Tanh, GELU, SiLU
• Forward + backward propagation deep dive
• Loss functions: MSE, Cross-entropy, BCE
• Optimizers: SGD, Momentum, Adam, AdamW
• Regularization: Dropout, L1/L2, early stopping
• Batch norm vs Layer norm (mandatory interview topic)
• Weight init (Xavier, He) + vanishing/exploding gradients
• Tensors, devices, basic ops
• Autograd: requires_grad, .backward(), .no_grad()
• nn.Module, custom layers
• Dataset and DataLoader
• Full training loop: zero_grad → forward → loss → backward → step
• Save/load models (state_dict), checkpointing, W&B integration
• Build MLP for MNIST in PyTorch, log with W&B
• What is convolution? Filters, stride, padding
• Pooling, receptive field
• CNN architectures: LeNet, AlexNet, VGG
• ResNet + residual connections (critical)
• Transfer learning, fine-tuning pretrained models
• Build a CNN in PyTorch for CIFAR-10
• Fine-tune pretrained ResNet-18 on custom dataset + read about ViT, CLIP
• Tokenization basics: characters, words, subwords
• BPE (Byte-Pair Encoding) — how GPT tokenizes
• Karpathy tokenizer Part 2 + finish
• Word2Vec, GloVe (legacy but interview-relevant)
• Modern embeddings: sentence-transformers, BGE, OpenAI embeddings
• RNN basics + why they failed → motivation for Transformers
• LSTM + vanishing gradients (interview topic)
• "Attention Is All You Need" paper — first pass, skim
• 3Blue1Brown "But what is a GPT?" + "Attention in transformers, visually explained"
• Self-attention math: Q, K, V matrices
• Multi-head attention, positional encodings (sinusoidal, RoPE, ALiBi)
• Encoder-only (BERT), Decoder-only (GPT), Encoder-Decoder (T5) — when to use which
• Re-read "Attention Is All You Need" — careful second pass
• KV cache, Pre-LN vs Post-LN (interview topics)
• Karpathy "Let's build GPT: from scratch" — minutes 0-30
• Continue — minutes 30-60
• Continue — minutes 60-90
• Continue — minutes 90-120
• Re-watch multi-head attention section, pause and code along
• Re-implement entire nanoGPT from memory — don't copy his code
• Train mini-GPT on custom text (Shakespeare, song lyrics, your favorite book)
• HF Transformers: pipeline, AutoModel, AutoTokenizer
• Loading + using pretrained models (BERT, GPT-2)
• Tokenizers in HF, padding, attention masks
• Datasets library: load, map, filter
• Trainer API — training on custom dataset
• Fine-tune DistilBERT for sentiment classification (IMDB)
• Push model to HuggingFace Hub with model card