File size: 2,744 Bytes
4c27541
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
{
  "deltas_sha256": "a0ae6ba35bb8f09b279c87be84a840015629844b909eb3731c72af6515ab9db5",
  "inputs_sha256": {
    "experience.npz": "aaffa2749869d2b90e3256d2fd6f9cec38ad3c4e516ca0583c5826038a27d6a1",
    "forward.py": "7f68b0e5a81ed5e1dc9d3a6f5a240d86a7f1298bfcb65d3be4acb5bf70f42e4f",
    "gated_bn_clip2_auto_v2.py": "de248188aa467283ab2dc2d7d95edaf1b8862df07072dd032f50fc0a63b89d25",
    "layout.json": "65d9eb530c12e1879f4cd9b7d74d75b991980711e9c64c428e6536a6c1b8cd0f",
    "params.npz": "55ea3ab8d3aeb06d90548455298c4e7d2a33050f62b92ba007f56b4c90a45666"
  },
  "program_file": "gated_bn_clip2_auto_v2.py",
  "program_sha256": "de248188aa467283ab2dc2d7d95edaf1b8862df07072dd032f50fc0a63b89d25",
  "run": {
    "duration_s": 0.685,
    "exit_code": 0,
    "result": {
      "hypothesis": "The new condition is a blur: conv features of new images are shrunk (std ~0.6-0.75x) and shifted, so the head collapses predictions onto cat/bird/ship. New and clean images are linearly separable in the input of the penultimate fc layer (mean-difference direction). Repurposing dead penultimate units as a gated linear correction that applies, only to new-condition inputs, a BN-style re-standardization of the penultimate-layer input (propagated through the linearized head) restores class evidence on new images while leaving clean images untouched.",
      "hypothesis_type": "str",
      "keys": [
        "fc1.weight",
        "fc1.bias",
        "fc2.weight"
      ],
      "predicted_effect": "New-condition accuracy rises substantially (evidence oracle: +15-25 points); clean predictions are almost unchanged (only clean images whose domain score leaks past the gate threshold can change), so score_old ~ 0 and score_new clearly positive.",
      "predicted_effect_type": "str",
      "status": "ok",
      "update_s": 0.39759106099882047
    },
    "status": "ok",
    "stderr": "",
    "stdout": "dead=59 rank=9 gap=7.596 kappa=9.97 cfac=2.72 gmax=1.000 g_eff=1.000\nfc1.weight rel frob 0.476\nfc2.weight rel frob 0.4541\nchanged frac 0.0987\n"
  },
  "sandbox": {
    "bwrap": "/usr/bin/bwrap",
    "driver_sha256": "9da8657cd189d50361b1903005565c7b7a73114c9b99ebddf90de10b003135d7",
    "file_size_limit_bytes": 268435456,
    "max_output_chars": 30000,
    "memory_bytes": 8589934592,
    "python": "/home/boom/BoomLi/WeightRSI/.venv-sandbox/bin/python",
    "threads": 4,
    "tmpfs_bytes": 1073741824
  },
  "shown": {
    "duration_s": 0.685,
    "program_stderr": "",
    "program_stdout": "dead=59 rank=9 gap=7.596 kappa=9.97 cfac=2.72 gmax=1.000 g_eff=1.000\nfc1.weight rel frob 0.476\nfc2.weight rel frob 0.4541\nchanged frac 0.0987\n"
  },
  "tensors": [
    "fc1.weight",
    "fc1.bias",
    "fc2.weight"
  ],
  "wall_s": 300.0
}