1 files changed, 49 insertions, 0 deletions
diff --git a/experiments/conv_local_smoke.py b/experiments/conv_local_smoke.py
index 894f9d3..7d24154 100644
--- a/experiments/conv_local_smoke.py
+++ b/experiments/conv_local_smoke.py
@@ -12,7 +12,9 @@ from sdil.conv import (CIFARHierarchicalFAResNet, CIFARLocalResNet,
                        CIFARSDILResNet, ConvSDILConfig,
                        channel_subspace_apical_calibration,
                        conv_hierarchical_step, conv_local_step,
+                       hierarchical_mirror_observations,
                        hierarchical_parameter_subspace_calibration,
+                       normalized_response_mirror_update,
                        simultaneous_conv_node_perturbation,
                        vectorizer_subspace_apical_calibration)
 
@@ -756,6 +758,52 @@ def hierarchical_parameter_calibration_checks():
     }
 
 
+def normalized_response_mirror_checks():
+    """Audit local response estimation and absence of W access in the update."""
+    net = CIFARHierarchicalFAResNet(
+        depth=8, base_width=2, seed=121, dtype=torch.float64,
+        normalization="batchnorm")
+    observations = hierarchical_mirror_observations(
+        net, batch_size=16, noise_std=1.0,
+        generator=torch.Generator().manual_seed(122))
+    metrics, _ = normalized_response_mirror_update(
+        net, observations, eta=1.0)
+    pairs = list(zip(net.Q[1:], net.W[1:])) + [
+        (net.R_out, -net.W_out.t())]
+    cosines = [float(F.cosine_similarity(
+        feedback.flatten(), target.flatten(), dim=0))
+        for feedback, target in pairs]
+    norm_ratios = [float(feedback.norm() / target.norm())
+                   for feedback, target in pairs]
+    assert sum(cosines) / len(cosines) > 0.985
+    assert min(cosines) > 0.95
+    assert min(norm_ratios) > 0.90 and max(norm_ratios) < 1.10
+
+    # The update consumes observations only. Changing every forward parameter
+    # after those observations were generated must not change the Q/R update.
+    left = CIFARHierarchicalFAResNet(
+        depth=8, base_width=2, seed=123, dtype=torch.float64)
+    right = CIFARHierarchicalFAResNet(
+        depth=8, base_width=2, seed=123, dtype=torch.float64)
+    shared_observations = hierarchical_mirror_observations(
+        left, batch_size=2, generator=torch.Generator().manual_seed(124))
+    for value in right.W + [right.W_out]:
+        value.normal_(generator=torch.Generator().manual_seed(value.numel()))
+    normalized_response_mirror_update(left, shared_observations, eta=0.2)
+    normalized_response_mirror_update(right, shared_observations, eta=0.2)
+    independence_error = max(float((a - b).abs().max()) for a, b in zip(
+        left.Q[1:] + [left.R_out], right.Q[1:] + [right.R_out]))
+    assert independence_error == 0.0
+    return {
+        "mirror_estimate_mean_forward_cosine": sum(cosines) / len(cosines),
+        "mirror_estimate_min_forward_cosine": min(cosines),
+        "mirror_estimate_min_norm_ratio": min(norm_ratios),
+        "mirror_estimate_max_norm_ratio": max(norm_ratios),
+        "mirror_update_forward_parameter_independence_error": independence_error,
+        "mirror_update_rms": metrics["mirror_update_rms"],
+    }
+
+
 def apical_learning_checks():
     torch.manual_seed(11)
     net = CIFARSDILResNet(depth=8, base_width=2, seed=6)
@@ -863,6 +911,7 @@ def main():
     report.update(vectorizer_subspace_estimator_check())
     report.update(hierarchical_feedback_checks())
     report.update(hierarchical_parameter_calibration_checks())
+    report.update(normalized_response_mirror_checks())
     report.update(apical_learning_checks())
     print(report)
     print("ALL CONVOLUTIONAL LOCAL-ELIGIBILITY CHECKS PASSED")