necatiincekara/gradient-check

Gradient Diagnostics

Quantum layer'lardaki gradient sorunlarini teshis et. V6'nin 0% accuracy sorununun ana sebebi gradient vanishing - bu skill bunu tespit eder.

Adimlar

1. Model dosyalarini oku: src/model.py, src/trainable_quantum_model.py, improved_quantum_circuit.py
2. Diagnostik scripti olustur ve calistir:

import torch
import sys
sys.path.insert(0, '.')

def analyze_gradients(model, sample_input):
    model.train()
    output = model(sample_input)
    loss = output.sum()
    loss.backward()

    print("=== GRADIENT FLOW ANALYSIS ===")
    quantum_grads, classical_grads = [], []

    for name, param in model.named_parameters():
        if param.grad is not None:
            norm = param.grad.norm().item()
            mean = param.grad.abs().mean().item()
            status = "VANISHING" if norm < 1e-7 else ("EXPLODING" if norm > 10 else "OK")
            print(f"{status:10s} | {name:40s} | norm={norm:.2e} mean={mean:.2e}")

            bucket = quantum_grads if any(k in name.lower() for k in ['quantum', 'qnode', 'quanv']) else classical_grads
            bucket.append(norm)
        else:
            print(f"{'NO GRAD':10s} | {name:40s}")

    if quantum_grads:
        print(f"\nQuantum grad avg: {sum(quantum_grads)/len(quantum_grads):.2e}")
    if classical_grads:
        print(f"Classical grad avg: {sum(classical_grads)/len(classical_grads):.2e}")

3. Analiz et:

   • Vanishing: norm < 1e-7
   • Exploding: norm > 10
   • Barren plateau: Quantum parametrelerde uniform sifira yakin
   • Quantum vs Classical gradient orani

4. Cozum oner:

   • Gradient clipping, residual connections, ayri learning rate
   • Circuit depth azaltma, parameter initialization
   • /architecture ile mimari degisiklik

Bilinen Sorunlar

| Versiyon | Feature Map | Gradient Durumu | |----------|------------|-----------------| | V4 (8x8) | 16 calls | Stabil, yavas ogrenme | | V6 (6x6) | 9 calls | TAMAMEN COLLAPSE - 0% accuracy |