Add clean_sparsity_persample.py: per-sample gradient stats

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

1 files changed, 78 insertions, 0 deletions

diff --git a/experiments/clean_sparsity_persample.py b/experiments/clean_sparsity_persample.py
new file mode 100644
index 0000000..e0f7e4d
--- /dev/null
+++ b/experiments/clean_sparsity_persample.py
@@ -0,0 +1,78 @@
+"""
+Per-sample gradient stats for CIFAR. One method+seed per invocation.
+Outputs CSV: each row = one sample × one layer.
+Usage: python clean_sparsity_persample.py --method bp --seed 42 --gpu 0
+"""
+import os, sys, csv, argparse, numpy as np, torch, torch.nn.functional as F
+from torch.utils.data import DataLoader
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+from models.residual_mlp import ResidualMLP
+import torchvision, torchvision.transforms as transforms
+
+def main():
+    p = argparse.ArgumentParser()
+    p.add_argument('--method', type=str, required=True)
+    p.add_argument('--seed', type=int, required=True)
+    p.add_argument('--gpu', type=int, default=0)
+    p.add_argument('--output_dir', type=str, default='results/confirmatory/persample')
+    args = p.parse_args()
+    os.makedirs(args.output_dir, exist_ok=True)
+    device = torch.device(f'cuda:{args.gpu}')
+
+    tv = transforms.Compose([transforms.ToTensor(),
+        transforms.Normalize((0.4914,0.4822,0.4465),(0.2470,0.2435,0.2616))])
+    tel = DataLoader(torchvision.datasets.CIFAR10('./data', False, download=True, transform=tv),
+                     256, False, num_workers=0)
+    for x, y in tel: x = x.view(x.size(0),-1).to(device); y = y.to(device); break
+    batch = x.size(0)
+
+    L, d = 4, 256
+    model = ResidualMLP(3072, d, 10, L).to(device)
+    model.load_state_dict(torch.load(
+        f'results/confirmatory/checkpoints_A2/{args.method}_s{args.seed}.pt',
+        map_location=device), strict=True)
+    model.eval()
+
+    h0 = model.embed(x.detach())
+    hs = [h0.clone().requires_grad_(True)]
+    for b in model.blocks: hs.append(hs[-1] + b(hs[-1]))
+    lo = model.out_head(model.out_ln(hs[-1]))
+    loss = F.cross_entropy(lo, y)
+    gs = torch.autograd.grad(loss, hs)
+
+    rows = []
+    for l in range(L):
+        g = gs[l].detach()  # (batch, d)
+        n2 = g.norm(dim=-1)
+        ninf = g.abs().max(dim=-1).values
+        n4 = (g.abs()**4).sum(-1)**(1/4)
+        n1 = g.abs().sum(-1)
+        r_inf = ninf / n2.clamp(min=1e-30)
+        pr = (n2**4 / (n4**4).clamp(min=1e-60)) / d
+        hoyer_num = (n1 / (n2 * d**0.5).clamp(min=1e-30))**2
+        gsq = g**2; te = gsq.sum(-1, keepdim=True).clamp(min=1e-60)
+        ssq, _ = gsq.sort(dim=-1, descending=True); cs = ssq.cumsum(-1)
+        k1 = max(1, int(d*0.01))-1; k5 = max(1, int(d*0.05))-1
+        topk1 = (cs[:, k1:k1+1] / te).squeeze(-1)
+        topk5 = (cs[:, k5:k5+1] / te).squeeze(-1)
+
+        for i in range(batch):
+            rows.append({
+                'method': args.method, 'seed': args.seed, 'layer': l, 'sample_id': i,
+                'grad_norm': n2[i].item(),
+                'log10_grad_norm': np.log10(max(n2[i].item(), 1e-30)),
+                'r_inf': r_inf[i].item(),
+                'pr': pr[i].item(),
+                'hoyer': hoyer_num[i].item(),
+                'topk1': topk1[i].item(),
+                'topk5': topk5[i].item(),
+            })
+
+    out = os.path.join(args.output_dir, f'{args.method}_s{args.seed}.csv')
+    with open(out, 'w', newline='') as f:
+        w = csv.DictWriter(f, fieldnames=['method','seed','layer','sample_id','grad_norm','log10_grad_norm','r_inf','pr','hoyer','topk1','topk5'])
+        w.writeheader(); w.writerows(rows)
+    print(f"[{args.method} s={args.seed}] {len(rows)} rows -> {out}", flush=True)
+
+if __name__ == '__main__':
+    main()