PyTorch Certified Associate (PTCA)

Learning plan

About 55 hours of focused study, split into 5 phases. Pick your weekly pace and the timeline adjusts.

Estimated finish: 6 weeks

What the exam covers

Source: official exam guide. Weights can change, so check before you start.

Before you start

• Comfortable Python and NumPy, and basic ML concepts (loss, gradient descent, train/validation split)
• A free Google Colab account for GPU time. A laptop with Apple Silicon also works for the MPS backend

1. Weeks 1–2 · 16 h

   Tensors, autograd and devices (largest domain)

   • Create and reshape tensors (view vs reshape, permute, broadcasting, indexing) and predict output shapes and dtypes without running code
   • Explain autograd (requires_grad, .backward(), grad accumulation, torch.no_grad, .detach) and why zero_grad is needed
   • Move models and tensors between CPU, CUDA and MPS, and fix device-mismatch errors
   • Write a correct training and evaluation loop (model.train/model.eval, optimizer.step, loss.backward) and save/load a state_dict

   Hands-on: In Colab, write a two-layer network for FashionMNIST from raw tensors and autograd only (no nn.Module or optim), then rewrite it with nn.Module, nn.CrossEntropyLoss and torch.optim.SGD, and confirm both reach similar accuracy.

   Use: PyTorch Learn the Basics tutorial series · Andrej Karpathy – Neural Networks Zero to Hero

2. Weeks 2–3 · 8 h

   Data handling

   • Write a custom Dataset (__len__/__getitem__) and use DataLoader with batch_size, shuffle, num_workers, pin_memory and collate_fn
   • Apply torchvision transforms v2 for augmentation vs evaluation, and split data with random_split or samplers
   • Spot data-leakage and class-imbalance problems (WeightedRandomSampler, class weights in the loss)

   Use: Datasets & DataLoaders tutorial · Zero to Mastery – Learn PyTorch (free book)

3. Weeks 3–4 · 12 h

   Model development with nn building blocks

   • Build models from nn.Linear, nn.Conv2d, pooling, BatchNorm, Dropout, embeddings, and nn.Sequential/ModuleList
   • Choose the right loss and output layer for regression, binary, multi-class and multi-label tasks
   • Use optimizers and LR schedulers (SGD with momentum, AdamW, StepLR, OneCycleLR) and gradient clipping
   • Fine-tune a pretrained torchvision model by freezing layers and replacing the classifier head

   Hands-on: Fine-tune a pretrained ResNet-18 on a small image dataset (for example a 5-class subset of Food-101). Freeze the backbone, train a new head with AdamW and a cosine schedule, then unfreeze the last block and compare validation accuracy.

   Use: PyTorch tutorials

4. Weeks 4–5 · 11 h

   Performance and optimisation

   • Use mixed precision (torch.autocast with bfloat16/float16 and GradScaler) and explain when each dtype is safe
   • Speed up models with torch.compile and inference_mode, and measure the gain correctly (warm-up, torch.cuda.synchronize)
   • Profile a training step with torch.profiler and find dataloader vs GPU bottlenecks
   • Explain DistributedDataParallel (process groups, DistributedSampler, torchrun) and how it differs from DataParallel and FSDP

   Hands-on: On a Colab GPU, time 100 training steps of your ResNet in fp32, then with autocast/bfloat16, then with torch.compile added. Use torch.profiler to show where the time goes in each run.

   Use: Automatic Mixed Precision docs · torch.compile tutorial · Profiler recipe · Getting started with DDP

5. Weeks 5–6 · 8 h

   Practice and review

   • Write shape-tracing drills (predict the output shape of 20 tensor operations) until you get them all right
   • Re-drill Fundamentals (38%) and Performance (26%), which together are almost two-thirds of the exam
   • Do a timed 120-minute mock exam, then review every miss in the docs

   Use: PTCA exam page (domains and policies)

You're ready to book when…

• You can write a complete train/eval loop with DataLoader, AMP and checkpointing from memory
• You can predict tensor shapes, dtypes and devices through a sequence of operations without running code
• You can explain what DDP, torch.compile and autocast each do and when they would not help

Exam tips

• The exam is 120 minutes, online-proctored multiple choice, and the $250 fee includes one retake. Eligibility lasts 12 months after purchase.
• The PyTorch docs moved from pytorch.org/docs to docs.pytorch.org. Use the current docs, because older tutorials use deprecated APIs such as torch.cuda.amp.autocast.