Architectural sensitivity measures how much a model’s output shifts when its parameters are perturbed. For quantization, injected Gaussian weight noise approximates rounding perturbations; the relative output change estimates sensitivity. Architecture family alone does not reliably predict the result. section 4.1
Measured sensitivity
With Gaussian noise scaled to 1% of each weight tensor’s mean absolute value, averaged over ten trials, MobileNetV2 shifts by about 11.6%, versus 7.6% for ResNet-18 and 7.4% for BERT-base. section 4.1
Listing 4.1 Measuring architecture noise sensitivity
def measure_noise_sensitivity(model, input_tensor, noise_scale=0.01, num_trials=10):
model.eval()
with torch.no_grad():
baseline = model(input_tensor)
if hasattr(baseline, 'last_hidden_state'):
baseline = baseline.last_hidden_state #A
output_changes = []
for _ in range(num_trials):
original_weights = {}
for name, param in model.named_parameters():
original_weights[name] = param.data.clone()
noise = torch.randn_like(param) * noise_scale * param.abs().mean()
param.data.add_(noise) #B
noisy_output = model(input_tensor)
if hasattr(noisy_output, 'last_hidden_state'):
noisy_output = noisy_output.last_hidden_state
relative_change = (noisy_output - baseline).abs().mean() / (baseline.abs().mean() + 1e-8)
output_changes.append(relative_change.item())
for name, param in model.named_parameters():
param.data.copy_(original_weights[name]) #C
return sum(output_changes) / len(output_changes)
torch.manual_seed(42)
resnet = resnet18(weights="IMAGENET1K_V1")
mobilenet = mobilenet_v2(weights="IMAGENET1K_V1")
bert = BertModel.from_pretrained('bert-base-uncased')
print(f"ResNet-18: {measure_noise_sensitivity(resnet, torch.randn(1, 3, 224, 224)):.4f}")
print(f"MobileNetV2: {measure_noise_sensitivity(mobilenet, torch.randn(1, 3, 224, 224)):.4f}")
print(f"BERT-base: {measure_noise_sensitivity(bert, torch.randint(0, 30000, (1, 32))):.4f}")Inspect the listing to see how parameter noise is applied and output changes are measured. section 4.1
