Skip to content

Latest commit

 

History

History
87 lines (57 loc) · 6.41 KB

verified-classification.md

File metadata and controls

87 lines (57 loc) · 6.41 KB

सत्यापित मैट्रिक्स वर्गीकरण और संभावनाएँ

कंपाइलर संस्करण: Scarp v0.0.5-alpha.1। इस पृष्ठ के तीनों पूरे प्रोग्राम और stdout, प्रकाशित compiler के examples/ फ़ोल्डर से लिए गए हैं। English तथा Hindi दोनों पृष्ठों में source और output की जाँच होती है; CI वास्तव में pinned Rust interpreter चलाकर output का byte-for-byte मिलान करती है।

Scarp के tile और vector विशेष immutable numerical types हैं। ये उदाहरण गणना समझाते हैं; इन्हें पूरा AI training framework या GPU backend न समझें। शुरू में वैक्टर और टाइल inference पढ़ सकते हैं।

1. क्लास स्कोर और बराबरी की स्थिति

matrix_classification.srp फ़ाइल बनाकर पूरा कोड लिखें:

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));

अपेक्षित stdout:

[2.0, 2.0, -1.0] 0 2.0

तीन rows और दो columns वाली tile, दो input features से तीन class scores बनाती है। पहले दो scores 2.0 पर बराबर हैं। प्रकाशित fixture के अनुसार argmax पहला index 0 चुनता है; maxval सबसे बड़ा score दिखाता है।

जाँचने के लिए scarp check matrix_classification.srp और चलाने के लिए scarp run matrix_classification.srp कमांड दें।

प्रकाशित सोर्स: matrix_classification.srp और stdout fixture।

2. संभावनाएँ और cross-entropy

matrix_probabilities.srp फ़ाइल बनाकर पूरा कोड लिखें:

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);

अपेक्षित stdout:

[0.5, 0.5]
true

दो-बाई-दो identity tile और zero vector से बराबर scores मिलते हैं। इसलिए softmax का output [0.5, 0.5] है। crossentropy(scores, 1) से लगभग 0.6931 loss मिलता है; fixture floating result की पूरी decimal value छापने के बजाय एक दायरा जाँचती है।

जाँचने के लिए scarp check matrix_probabilities.srp और चलाने के लिए scarp run matrix_probabilities.srp कमांड दें।

प्रकाशित सोर्स: matrix_probabilities.srp और stdout fixture।

3. चुनी हुई class को one-hot target बनाना

matrix_targets.srp फ़ाइल बनाकर पूरा कोड लिखें:

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)));

अपेक्षित stdout:

[1.0, 0.0, 0.0]

len(scores) classes की कुल संख्या देता है और argmax(scores) चुना हुआ index देता है। onehot उस स्थान पर 1.0 तथा बाकी स्थानों पर 0.0 वाला vector बनाता है। इसमें पहले उदाहरण की तरह तीन scores हैं।

जाँचने के लिए scarp check matrix_targets.srp और चलाने के लिए scarp run matrix_targets.srp कमांड दें।

प्रकाशित सोर्स: matrix_targets.srp और stdout fixture।

किन बातों का ध्यान रखें?

  • आकार: matvec(weights, features) में tile की width और vector की length बराबर हों। गलत आकार पर error आएगा; अपने-आप broadcasting नहीं होती।
  • Numerical domain: softmax और crossentropy में input तथा numerical limits जाँचे जाते हैं। Built-ins और डायग्नोस्टिक्स देखें।
  • Class index: इस सत्यापित fixture में बराबर सर्वाधिक score होने पर argmax पहला index चुनता है। Index और score अलग चीजें हैं।
  • Output: ऊपर दिए outputs प्रकाशित compiler के हैं। स्वयं inputs बदलने पर परिणाम या diagnostic बदल सकता है।

आगे रनटाइम बजट, मॉड्यूल और फ़ीचर स्थिति पढ़ें।

::: warning प्रयोगात्मक प्रीरिलीज़ Scarp v0.0.5-alpha.1 स्थिर compatibility की गारंटी नहीं देता। दूसरे संस्करण में उदाहरण आज़माने से पहले फीचर सत्यापित करें। :::