Skip to content

verified-classification.md

Latest commit

 

History

History
87 lines (57 loc) · 4.58 KB

verified-classification.md

File metadata and controls

87 lines (57 loc) · 4.58 KB
Raw

Verified matrix classification and probabilities

Compiler baseline: Scarp v0.0.5-alpha.1. These three complete programs and their stdout are mirrored from the tagged compiler's examples/ directory. The documentation checker verifies their order and exact visible content in both English and Hindi; CI executes each with the pinned Rust interpreter and compares its stdout byte-for-byte.

Numerical tile and vector are specialized, immutable, finite-float types. These examples illustrate numerical operations, not a general AI model training framework or GPU backend. Start with verified vectors and tile inference if the types are new.

1. Class scores and tie-breaking

Create matrix_classification.srp with this full source:

bind weights = tile([1.0, 0.0, 0.0, 2.0, -1.0, 1.0], 3, 2);
bind scores = matvec(weights, [2.0, 1.0]);
// Equal best scores choose the first class.
print(scores, argmax(scores), maxval(scores));

Expected stdout:

[2.0, 2.0, -1.0] 0 2.0

A three-row, two-column tile maps two input features into three scores. The top two scores are tied at 2.0, and the fixture confirms argmax selects index 0 rather than index 1; maxval returns the best score.

You can validate the original file with scarp check matrix_classification.srp and run it with scarp run matrix_classification.srp.

Pinned compiler source: matrix_classification.srp and expected stdout.

2. Probabilities and cross-entropy

Create matrix_probabilities.srp with this full source:

bind weights = tile([1.0, 0.0, 0.0, 1.0], 2, 2);
bind scores = matvec(weights, [0.0, 0.0]);
print(softmax(scores));
bind loss = crossentropy(scores, 1);
print(loss > 0.6931 and loss < 0.6932);

Expected stdout:

[0.5, 0.5]
true

A two-by-two identity tile applied to a zero vector produces equal scores. softmax therefore prints [0.5, 0.5]. crossentropy(scores, 1) produces a positive loss close to 0.6931; this particular fixture checks the range rather than printing a platform-dependent decimal.

You can validate the original file with scarp check matrix_probabilities.srp and run it with scarp run matrix_probabilities.srp.

Pinned compiler source: matrix_probabilities.srp and expected stdout.

3. Turn a selected class into a one-hot target

Create matrix_targets.srp with this full source:

bind scores = matvec(tile([1.0, 0.0, 0.0, 2.0, -1.0, 1.0], 3, 2), [2.0, 1.0]);
print(onehot(len(scores), argmax(scores)));

Expected stdout:

[1.0, 0.0, 0.0]

len(scores) specifies the number of classes, while argmax(scores) selects the predicted class. onehot constructs a vector containing 1.0 at that position and 0.0 at the others. This example uses the same three-score geometry as the first example.

You can validate the original file with scarp check matrix_targets.srp and run it with scarp run matrix_targets.srp.

Pinned compiler source: matrix_targets.srp and expected stdout.

What to watch for

  • Shapes: matvec(weights, features) requires a tile width equal to the vector length. Incorrect shapes produce a diagnostic, not automatic broadcasting.
  • Domains: the softmax and crossentropy kernels have input rules, checked numerical behavior and runtime limits. See built-ins and numerical diagnostics.
  • Class indices: argmax selects the first maximal element for a tie in this published example. Do not confuse the position of the class with the value of its score.
  • Output: the results above are compiler-owned fixtures, not guesses. Altered inputs can produce different scores or errors.

See runtime budgets, modules tutorial, and feature status for release boundaries.

::: warning Experimental prerelease Scarp v0.0.5-alpha.1 does not promise stable language compatibility. Avoid copying these examples into older versions without checking available features. :::