Delete skeptic_test.py with huggingface_hub

skeptic_test.py

@@ -1,215 +0,0 @@
-import warnings
-warnings.filterwarnings('ignore')
-from safetensors.torch import load_file
-import torch
-
-model = load_file('neural_computer.safetensors')
-
-def heaviside(x):
-    return (x >= 0).float()
-
-def int_to_bits(val):
-    return torch.tensor([(val >> (7-i)) & 1 for i in range(8)], dtype=torch.float32)
-
-def eval_xor_bool(a, b):
-    inp = torch.tensor([float(a), float(b)], dtype=torch.float32)
-    w1_or = model['boolean.xor.layer1.neuron1.weight']
-    b1_or = model['boolean.xor.layer1.neuron1.bias']
-    w1_nand = model['boolean.xor.layer1.neuron2.weight']
-    b1_nand = model['boolean.xor.layer1.neuron2.bias']
-    w2 = model['boolean.xor.layer2.weight']
-    b2 = model['boolean.xor.layer2.bias']
-    h_or = heaviside(inp @ w1_or + b1_or)
-    h_nand = heaviside(inp @ w1_nand + b1_nand)
-    hidden = torch.tensor([h_or.item(), h_nand.item()])
-    return int(heaviside(hidden @ w2 + b2).item())
-
-def eval_xor_arith(inp, prefix):
-    w1_or = model[f'{prefix}.layer1.or.weight']
-    b1_or = model[f'{prefix}.layer1.or.bias']
-    w1_nand = model[f'{prefix}.layer1.nand.weight']
-    b1_nand = model[f'{prefix}.layer1.nand.bias']
-    w2 = model[f'{prefix}.layer2.weight']
-    b2 = model[f'{prefix}.layer2.bias']
-    h_or = heaviside(inp @ w1_or + b1_or)
-    h_nand = heaviside(inp @ w1_nand + b1_nand)
-    hidden = torch.tensor([h_or.item(), h_nand.item()])
-    return heaviside(hidden @ w2 + b2).item()
-
-def eval_full_adder(a, b, cin, prefix):
-    inp_ab = torch.tensor([a, b], dtype=torch.float32)
-    ha1_sum = eval_xor_arith(inp_ab, f'{prefix}.ha1.sum')
-    w_c1 = model[f'{prefix}.ha1.carry.weight']
-    b_c1 = model[f'{prefix}.ha1.carry.bias']
-    ha1_carry = heaviside(inp_ab @ w_c1 + b_c1).item()
-    inp_ha2 = torch.tensor([ha1_sum, cin], dtype=torch.float32)
-    ha2_sum = eval_xor_arith(inp_ha2, f'{prefix}.ha2.sum')
-    w_c2 = model[f'{prefix}.ha2.carry.weight']
-    b_c2 = model[f'{prefix}.ha2.carry.bias']
-    ha2_carry = heaviside(inp_ha2 @ w_c2 + b_c2).item()
-    inp_cout = torch.tensor([ha1_carry, ha2_carry], dtype=torch.float32)
-    w_or = model[f'{prefix}.carry_or.weight']
-    b_or = model[f'{prefix}.carry_or.bias']
-    cout = heaviside(inp_cout @ w_or + b_or).item()
-    return int(ha2_sum), int(cout)
-
-def add_8bit(a, b):
-    carry = 0.0
-    result = 0
-    for i in range(8):
-        s, carry = eval_full_adder(float((a >> i) & 1), float((b >> i) & 1), carry, f'arithmetic.ripplecarry8bit.fa{i}')
-        result |= (s << i)
-    return result, int(carry)
-
-def xor_8bit(a, b):
-    result = 0
-    for i in range(8):
-        bit = eval_xor_bool((a >> i) & 1, (b >> i) & 1)
-        result |= (bit << i)
-    return result
-
-def and_8bit(a, b):
-    result = 0
-    w = model['boolean.and.weight']
-    bias = model['boolean.and.bias']
-    for i in range(8):
-        inp = torch.tensor([float((a >> i) & 1), float((b >> i) & 1)], dtype=torch.float32)
-        out = int(heaviside(inp @ w + bias).item())
-        result |= (out << i)
-    return result
-
-def or_8bit(a, b):
-    result = 0
-    w = model['boolean.or.weight']
-    bias = model['boolean.or.bias']
-    for i in range(8):
-        inp = torch.tensor([float((a >> i) & 1), float((b >> i) & 1)], dtype=torch.float32)
-        out = int(heaviside(inp @ w + bias).item())
-        result |= (out << i)
-    return result
-
-def not_8bit(a):
-    result = 0
-    w = model['boolean.not.weight']
-    bias = model['boolean.not.bias']
-    for i in range(8):
-        inp = torch.tensor([float((a >> i) & 1)], dtype=torch.float32)
-        out = int(heaviside(inp @ w + bias).item())
-        result |= (out << i)
-    return result
-
-def gt(a, b):
-    a_bits, b_bits = int_to_bits(a), int_to_bits(b)
-    w = model['arithmetic.greaterthan8bit.comparator']
-    return 1 if ((a_bits - b_bits) @ w).item() > 0 else 0
-
-def lt(a, b):
-    a_bits, b_bits = int_to_bits(a), int_to_bits(b)
-    w = model['arithmetic.lessthan8bit.comparator']
-    return 1 if ((b_bits - a_bits) @ w).item() > 0 else 0
-
-def eq(a, b):
-    return 1 if (gt(a,b) == 0 and lt(a,b) == 0) else 0
-
-print('=' * 70)
-print('SKEPTICAL NERD TESTS')
-print('=' * 70)
-
-failures = []
-
-print('\n[1] IDENTITY LAWS')
-for a in [0, 1, 127, 128, 255, 170, 85]:
-    r, _ = add_8bit(a, 0)
-    if r != a: failures.append(f'A+0: {a}')
-    if xor_8bit(a, 0) != a: failures.append(f'A^0: {a}')
-    if and_8bit(a, 255) != a: failures.append(f'A&255: {a}')
-    if or_8bit(a, 0) != a: failures.append(f'A|0: {a}')
-print('  28 tests')
-
-print('\n[2] ANNIHILATION LAWS')
-for a in [0, 1, 127, 128, 255]:
-    if and_8bit(a, 0) != 0: failures.append(f'A&0: {a}')
-    if or_8bit(a, 255) != 255: failures.append(f'A|255: {a}')
-    if xor_8bit(a, a) != 0: failures.append(f'A^A: {a}')
-print('  15 tests')
-
-print('\n[3] INVOLUTION (~~A = A)')
-for a in [0, 1, 127, 128, 255, 170]:
-    if not_8bit(not_8bit(a)) != a: failures.append(f'~~A: {a}')
-print('  6 tests')
-
-print('\n[4] TWOS COMPLEMENT: A + ~A + 1 = 0')
-for a in [0, 1, 42, 127, 128, 255]:
-    not_a = not_8bit(a)
-    r1, _ = add_8bit(a, not_a)
-    r2, _ = add_8bit(r1, 1)
-    if r2 != 0: failures.append(f'twos comp: {a}')
-print('  6 tests')
-
-print('\n[5] CARRY PROPAGATION (worst case)')
-cases = [(255, 1, 0), (127, 129, 0), (1, 255, 0), (128, 128, 0), (255, 255, 254)]
-for a, b, exp in cases:
-    r, _ = add_8bit(a, b)
-    if r != exp: failures.append(f'carry: {a}+{b}={r}, expected {exp}')
-print('  5 tests')
-
-print('\n[6] COMMUTATIVITY')
-pairs = [(17, 42), (0, 255), (128, 127), (1, 254), (170, 85)]
-for a, b in pairs:
-    r1, _ = add_8bit(a, b)
-    r2, _ = add_8bit(b, a)
-    if r1 != r2: failures.append(f'add commute: {a},{b}')
-    if xor_8bit(a, b) != xor_8bit(b, a): failures.append(f'xor commute: {a},{b}')
-    if and_8bit(a, b) != and_8bit(b, a): failures.append(f'and commute: {a},{b}')
-    if or_8bit(a, b) != or_8bit(b, a): failures.append(f'or commute: {a},{b}')
-print('  20 tests')
-
-print('\n[7] DE MORGAN')
-for a, b in [(0, 0), (0, 255), (255, 0), (255, 255), (170, 85)]:
-    lhs = not_8bit(and_8bit(a, b))
-    rhs = or_8bit(not_8bit(a), not_8bit(b))
-    if lhs != rhs: failures.append(f'DM1: {a},{b}')
-    lhs = not_8bit(or_8bit(a, b))
-    rhs = and_8bit(not_8bit(a), not_8bit(b))
-    if lhs != rhs: failures.append(f'DM2: {a},{b}')
-print('  10 tests')
-
-print('\n[8] COMPARATOR EDGE CASES')
-cmp_tests = [
-    (0, 0, 0, 0, 1), (0, 1, 0, 1, 0), (1, 0, 1, 0, 0),
-    (127, 128, 0, 1, 0), (128, 127, 1, 0, 0),
-    (255, 255, 0, 0, 1), (255, 0, 1, 0, 0), (0, 255, 0, 1, 0),
-]
-for a, b, exp_gt, exp_lt, exp_eq in cmp_tests:
-    if gt(a, b) != exp_gt: failures.append(f'gt({a},{b})')
-    if lt(a, b) != exp_lt: failures.append(f'lt({a},{b})')
-    if eq(a, b) != exp_eq: failures.append(f'eq({a},{b})')
-print('  24 tests')
-
-print('\n[9] POPCOUNT SINGLE BITS + EXTREMES')
-w_pop = model['pattern_recognition.popcount.weight']
-b_pop = model['pattern_recognition.popcount.bias']
-for i in range(8):
-    val = 1 << i
-    bits = int_to_bits(val)
-    pc = int((bits @ w_pop + b_pop).item())
-    if pc != 1: failures.append(f'popcount(1<<{i})')
-if int((int_to_bits(0) @ w_pop + b_pop).item()) != 0: failures.append('popcount(0)')
-if int((int_to_bits(255) @ w_pop + b_pop).item()) != 8: failures.append('popcount(255)')
-print('  10 tests')
-
-print('\n[10] DISTRIBUTIVITY: A & (B | C) = (A & B) | (A & C)')
-for a, b, c in [(255, 15, 240), (170, 85, 51), (0, 255, 0)]:
-    lhs = and_8bit(a, or_8bit(b, c))
-    rhs = or_8bit(and_8bit(a, b), and_8bit(a, c))
-    if lhs != rhs: failures.append(f'distrib: {a},{b},{c}')
-print('  3 tests')
-
-print('\n' + '=' * 70)
-if failures:
-    print(f'FAILURES: {len(failures)}')
-    for f in failures[:20]:
-        print(f'  {f}')
-else:
-    print('ALL 127 SKEPTICAL TESTS PASSED')
-print('=' * 70)