From c6af287a6177cda3863e6f25f2bdab8abd611ad8 Mon Sep 17 00:00:00 2001 From: YurenHao0426 Date: Wed, 22 Jul 2026 06:18:36 -0500 Subject: oral-a: measure training-state apical alignment --- sdil/conv.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/sdil/conv.py b/sdil/conv.py index 40def21..0436525 100644 --- a/sdil/conv.py +++ b/sdil/conv.py @@ -753,7 +753,7 @@ def conv_alignment_report(net, x, y, config): parameters = net.W + net.gamma + net.beta + [net.W_out, net.b_out] for parameter in parameters: parameter.requires_grad_(True) - forward = net.forward(x) + forward = net.forward(x, training=True, update_stats=False) gradients = torch.autograd.grad( F.cross_entropy(forward["logits"], y), forward["hiddens"]) batch = x.shape[0] @@ -771,6 +771,7 @@ def conv_alignment_report(net, x, y, config): return float(F.cosine_similarity(left, right, dim=1).mean()) report = { + "normalization_state": "training_batch_stats_without_running_update", "teaching_negative_gradient_cosine": [ cosine(left, right) for left, right in zip(teaching, negative_gradients)], "raw_negative_gradient_cosine": [ -- cgit v1.2.3