Python / PyTorch Fundamentals Interview Questions
What optimizers does PyTorch provide and how do you configure them?
Optimizers update model parameters based on computed gradients. PyTorch provides all major optimizers in torch.optim. Choosing and configuring the right optimizer significantly affects training speed and final performance.
| Optimizer | Key feature | Typical use |
|---|---|---|
| SGD | Simple, supports momentum and weight decay | Computer vision with lr scheduling |
| Adam | Adaptive lr per param; momentum + RMSProp | Default for NLP, general purpose |
| AdamW | Adam with decoupled weight decay | Transformers, fine-tuning (recommended over Adam) |
| RMSprop | Adaptive lr without momentum | RNNs |
| Adagrad | Accumulates squared gradients; rare today | Sparse features |
| LBFGS | Second-order quasi-Newton; very slow | Small networks, physics-informed NNs |
import torch, torch.nn as nn, torch.optim as optim model = nn.Linear(128, 10) # Γ’ΒΒΓ’ΒΒ SGD with momentum and weight decay opt_sgd = optim.SGD( model.parameters(), lr=0.1, momentum=0.9, # Nesterov-style acceleration weight_decay=1e-4, # L2 regularisation nesterov=True, ) # Γ’ΒΒΓ’ΒΒ Adam opt_adam = optim.Adam( model.parameters(), lr=1e-3, betas=(0.9, 0.999), # (ΓΒ²1, ΓΒ²2) Γ’ΒΒ momentum terms eps=1e-8, weight_decay=0, # Adam + L2 is suboptimal Γ’ΒΒ use AdamW! ) # Γ’ΒΒΓ’ΒΒ AdamW (preferred for transformers) opt_adamw = optim.AdamW( model.parameters(), lr=1e-3, weight_decay=0.01, # decoupled from gradient update ) # Γ’ΒΒΓ’ΒΒ Per-layer learning rates opt_layerwise = optim.Adam([ {"params": model.weight, "lr": 1e-4}, # slower for weight {"params": model.bias, "lr": 1e-3}, # faster for bias ]) # Γ’ΒΒΓ’ΒΒ Standard training step opt = optim.AdamW(model.parameters(), lr=1e-3) for x, y in [(torch.randn(32,128), torch.randint(0,10,(32,)))]: opt.zero_grad() # 1. clear old gradients loss = nn.CrossEntropyLoss()(model(x), y) # 2. forward loss.backward() # 3. backward opt.step() # 4. update parameters
More Related questions...