Curriculum & Milestones

The 20 modules

Four parts, each one asking the same question of a new piece of the stack: can I build this from scratch, and can I build it well?

Part I: Foundations (01–08)

#	Module	What you build
01	Tensor	Your core data structure, with autodiff
02	Activations	ReLU, Sigmoid, Softmax
03	Layers	Linear layers, the Module system
04	Losses	MSE, CrossEntropy
05	DataLoader	Efficient data batching
06	Autograd	Automatic differentiation
07	Optimizers	SGD, Adam
08	Training	A complete training loop

Part II: Vision (09)

#	Module	What you build
09	Convolutions	Conv2d, MaxPool2d, CNNs

Part III: Language (10–13)

#	Module	What you build
10	Tokenization	Text processing
11	Embeddings	Token & positional embeddings
12	Attention	Multi-head attention
13	Transformers	Complete transformer blocks

Part IV: Optimization (14–20)

#	Module	What you build
14	Profiling	Performance analysis
15	Quantization	Precision reduction
16	Compression	Pruning & distillation
17	Acceleration	Hardware-aware optimization (kernel fusion, etc.)
18	Memoization	KV-cache
19	Benchmarking	Performance measurement
20	Capstone	Ties the whole framework together

Each module lives at data/src/NN_name/NN_name.py, that's the file you actually edit. tren module start NN turns it into a notebook stub; tren module complete NN tests it and exports it into the real trentorch package.

The 6 historical milestones

As you complete modules, you unlock milestones: recreations of landmark ML moments, run entirely on the framework you wrote:

Year	Milestone	What it proves	Unlocked by
1958	Perceptron	Binary classification with gradient descent	Modules 1–3
1969	XOR Crisis	A single-layer perceptron cannot solve XOR, and why that mattered	Modules 1–3
1986	MLP Revival	Hidden layers + backprop solve the "impossible" XOR problem, then scale to real digit data	Modules 1–8
1998	CNN Revolution	CNNs beat MLPs on image classification, LeNet-style	Modules 1–9
2017	Transformer Era	Attention learns sequence reversal, copying, and conditional tasks	Modules 1–8, 11–13
2018+	MLPerf Olympics	Production-style optimization: compress and accelerate your own network	Full stack

Run one with:

tren milestone run 04        # or: tren milestone run cnn

Each milestone script checks its own required modules are actually done before running, and for milestones with multiple parts (like CNN Revolution's TinyDigits vs. CIFAR-10 tracks), --part N picks which one.

tren milestone list          # what's unlocked, what's next
tren milestone info 04       # details on one milestone
tren milestone timeline      # the whole arc, as a progress view

Not toy demos

These milestones aren't scaled-down illustrations, they're the real historical experiments (perceptron convergence, XOR's non-linearity, LeNet-scale CNN training, attention learning real sequence tasks), run against real data, using real optimizers and real backprop, all implemented by you in the modules above. The one place CI takes a shortcut is trimming demo epoch counts under automated verification, since a real student running the milestone interactively still gets the full experience; see Architecture for that distinction.