Finetune DM0 got very terrifying loss

[lucasjinreal]

the loss like:

{'loss': 8.3281, 'grad_norm': 322.77156161507077, 'learning_rate': 0.0, 'epoch': 0.0}                                                                                             | 19/92 [00:03<00:11,  6.29it/s]
{'loss': 8.5, 'grad_norm': 273.2512527507693, 'learning_rate': 5.0000000000000004e-08, 'epoch': 0.0}
{'loss': 7.9219, 'grad_norm': 246.70768633306642, 'learning_rate': 1.0000000000000001e-07, 'epoch': 0.0}
{'loss': 6.9141, 'grad_norm': 207.61659902097261, 'learning_rate': 1.5000000000000002e-07, 'epoch': 0.0}
{'loss': 6.5938, 'grad_norm': 238.2175499387362, 'learning_rate': 2.0000000000000002e-07, 'epoch': 0.0}
{'loss': 7.0312, 'grad_norm': 215.7555159500225, 'learning_rate': 2.5000000000000004e-07, 'epoch': 0.0}
{'loss': 9.0625, 'grad_norm': 292.1837063871723, 'learning_rate': 3.0000000000000004e-07, 'epoch': 0.0}
{'loss': 7.8125, 'grad_norm': 249.52821941121618, 'learning_rate': 3.5000000000000004e-07, 'epoch': 0.0}
{'loss': 7.9609, 'grad_norm': 281.87461514835957, 'learning_rate': 4.0000000000000003e-07, 'epoch': 0.0}
{'loss': 8.2422, 'grad_norm': 220.2848272714045, 'learning_rate': 4.5e-07, 'epoch': 0.0}
{'loss': 8.8281, 'grad_norm': 309.4026145868916, 'learning_rate': 5.000000000000001e-07, 'epoch': 0.0}
{'loss': 8.375, 'grad_norm': 246.82614914951165, 'learning_rate': 5.5e-07, 'epoch': 0.0}
{'loss': 8.4844, 'grad_norm': 267.56563534299346, 'learning_rate': 6.000000000000001e-07, 'epoch': 0.0}
{'loss': 7.75, 'grad_norm': 254.66540924254167, 'learning_rate': 6.5e-07, 'epoch': 0.0}
{'loss': 8.5703, 'grad_norm': 236.04995774767977, 'learning_rate': 7.000000000000001e-07, 'epoch': 0.0}
{'loss': 7.6055, 'grad_norm': 244.71480797209944, 'learning_rate': 7.5e-07, 'epoch': 0.0}
{'loss': 7.4844, 'grad_norm': 239.66509686659916, 'learning_rate': 8.000000000000001e-07, 'epoch': 0.0}
{'loss': 5.6172, 'grad_norm': 196.99852716316857, 'learning_rate': 8.500000000000001e-07, 'epoch': 0.0}
{'loss': 6.5312, 'grad_norm': 209.76915835265947, 'learning_rate': 9e-07, 'epoch': 0.0}
{'loss': 7.3906, 'grad_norm': 209.8876629848969, 'learning_rate': 9.5e-07, 'epoch': 0.0}
{'loss': 5.9844, 'grad_norm': 197.52531679833206, 'learning_rate': 1.0000000000000002e-06, 'epoch': 0.0}
{'loss': 4.9141, 'grad_norm': 176.5006773319435, 'learning_rate': 1.0500000000000001e-06, 'epoch': 0.0}
{'loss': 4.6719, 'grad_norm': 159.39841595633183, 'learning_rate': 1.1e-06, 'epoch': 0.0}
{'loss': 5.5391, 'grad_norm': 246.40031142756555, 'learning_rate': 1.15e-06, 'epoch': 0.0}
{'loss': 6.8672, 'grad_norm': 313.6043602156962, 'learning_rate': 1.2000000000000002e-06, 'epoch': 0.0}
{'loss': 4.918, 'grad_norm': 170.7606607499666, 'learning_rate': 1.25e-06, 'epoch': 0.0}
{'loss': 4.2773, 'grad_norm': 184.55600488249698, 'learning_rate': 1.3e-06, 'epoch': 0.0}
{'loss': 4.6211, 'grad_norm': 373.14208295733414, 'learning_rate': 1.35e-06, 'epoch': 0.0}
{'loss': 4.4883, 'grad_norm': 346.7172125262856, 'learning_rate': 1.4000000000000001e-06, 'epoch': 0.0}
{'loss': 3.9102, 'grad_norm': 270.9716257650011, 'learning_rate': 1.45e-06, 'epoch': 0.0}

it's that normal? It seems not like a VLA fashioned loss.

[TalentBoy2333]

Thank you for using Dexbotic.

This loss does look rather unusual. Can you show your experiment norm file in dexbotic/norm_assets. I think it might be related to the norm calculation.

[lucasjinreal]

Am using a 20 dim, 10 for left and another 10 for right, am using same data finetuned pi0.5 got no problem, unlikely a dataset issue.
here is the norm status i have:

{
  "norm_stats": {
    "default": {
      "min": -1,
      "max": 1
    },
    "action": {
      "min": [
        -0.22909217740297316,
        -0.34187245479784906,
        -0.16764292054474353,
        -6.24746151266098,
        -0.8278781593590975,
        -6.242435348510742,
        -1.3552099466323641e-05,
        -0.22253079099953174,
        -0.25445848705172536,
        -0.1836433614462614,
        -6.260026783800125,
        -0.7498763909459114,
        -6.237409295749664,
        -0.05973217776987003,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0
      ],
      "max": [
        0.20785048677027224,
        0.5076031101797707,
        0.20924269043803212,
        6.247461463928223,
        0.963147219443321,
        6.242366218185426,
        0.06026176686733961,
        0.24575681301802388,
        0.24196554211378096,
        0.18176415565013893,
        6.255000637054444,
        0.8937823958758266,
        6.239904945421218,
        0.05963652784122096,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0
      ],
      "mean": [
        -0.0003853877433054956,
        0.0007306670896923745,
        -0.0011364989214451323,
        0.03654705501546644,
        0.006134213136216658,
        0.0034879869032055483,
        0.03629060535569406,
        -0.0012605499199800437,
        -0.000600332485169393,
        -0.0014568852311355226,
        0.017656159329983774,
        0.006814231340934181,
        0.04113114739203431,
        -0.00024956181893120245,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0
      ],
      "std": [
        0.0565540551815235,
        0.07096283268827523,
        0.05149424025522513,
        2.1483310518032352,
        0.1763593971855789,
        2.022556809609136,
        0.026246241017056537,
        0.05208878941591823,
        0.04618576588626645,
        0.04012263513461467,
        2.0749658421718733,
        0.15095649269647488,
        1.6071311164231308,
        0.020451321335037315,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0
      ]
    },
    "state": {
      "min": [
        0.034795538062602285,
        -0.36924759843945504,
        0.0717046156220138,
        -3.1372819442749025,
        -0.20151048215031625,
        -3.1385384731292723,
        -1.3552099466323641e-05,
        0.03081692423820495,
        -0.15925544939637185,
        0.10152572415173054,
        -3.1385384731292723,
        -0.11308792293071745,
        -3.1372819442749025,
        3.603599891066552e-05,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0
      ],
      "max": [
        0.5302837542824448,
        0.41212332770824434,
        0.4778286756694317,
        3.140336215639114,
        1.5355551875308158,
        3.1403362226009364,
        0.06026176686733961,
        0.5252191450089216,
        0.3542859651446343,
        0.4252568843513727,
        3.140336215639114,
        1.5433564427375792,
        3.1403362226009364,
        0.05991809597976098,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0
      ],
      "mean": [
        0.25205901768571076,
        -0.0662840930865076,
        0.20094749684081395,
        -1.257679132223258,
        0.40007935753853524,
        1.2122270765050698,
        0.03637827677756685,
        0.23957563157661038,
        0.08987595533075236,
        0.21091035993157026,
        1.3981852355059594,
        0.521069339878782,
        -1.6587256649618192,
        0.03616423602859954,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0
      ],
      "std": [
        0.09389227527453549,
        0.10809575278633178,
        0.0654208821031667,
        2.563685689653422,
        0.3026987514295694,
        2.4309380655718944,
        0.026255269304874113,
        0.09101097241799283,
        0.07506062744505591,
        0.053717267846462455,
        2.4758535214172688,
        0.25443014586113055,
        2.203187749639332,
        0.02604700591525368,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0
      ]
    }
  }
}

[TalentBoy2333]

However, judging from the non-zero bits, it appears that this norm file only uses 14 dimensions. Are you sure you provided the correct norm file?