sparse_reward_classifier / metrics.json
imendezval's picture
Upload 2 files
1bb7603 verified
Raw
History Blame Contribute Delete
19 kB
{
"config": {
"cache_dir": "le-wm/models",
"dataset": "pusht_expert_train",
"checkpoint": "hf_pusht/weights.pt",
"output_dir": "models/probes/pusht_lewm_1M_sparse_imagined50_balanced_hn_fpr005",
"latent_cache": "models/probes/pusht_lewm_1M/latents.npz",
"imagined_cache": "models/probes/pusht_lewm_1M_sparse_imagined50_fpr005/objective_met_imagined_train.npz",
"imagined_val_cache": "models/probes/pusht_lewm_1M_sparse_imagined50_fpr005/objective_met_imagined_val.npz",
"imagined_test_cache": "models/probes/pusht_lewm_1M_sparse_imagined50_fpr005/objective_met_imagined_test.npz",
"max_samples": 1000000,
"sample_block_size": 16,
"batch_size": 512,
"encode_batch_size": 256,
"rollout_batch_size": 64,
"epochs": 50,
"patience": 8,
"lr": 0.001,
"weight_decay": 0.0001,
"ridge": 0.001,
"mlp_hidden": 256,
"mlp_depth": 2,
"device": "cuda",
"force_recache": false,
"force_imagined_recache": false,
"seed": 3072,
"probes": [
"objective_met"
],
"frameskip": 5,
"context_steps": 3,
"imagined_rollout_horizon": 20,
"include_imagined_rollouts": true,
"imagined_fraction": 0.5,
"eval_imagined_rollouts": true,
"imagined_eval_samples": 100000,
"sparse_imagined_sampling": "step_class_balanced",
"sparse_negative_sampling": "mixed",
"sparse_near_success_margin": 3.0,
"sparse_threshold_policy": "target_fpr",
"sparse_target_fpr": 0.05,
"sparse_threshold_source": "imagined_val",
"sparse_condition_on_step": false,
"sparse_hard_negative_mining": true,
"sparse_hard_negative_fraction": 0.25,
"sparse_hard_negative_epochs": 10,
"no_normalize_actions": false,
"objective_x": 256.0,
"objective_y": 256.0,
"objective_angle": 0.7853981633974483,
"objective_pos_tol": 20.0,
"objective_angle_tol": 0.3490658503988659
},
"feature_normalizer": {
"mean": [
[
-0.03391525521874428,
0.019303878769278526,
-0.0164620541036129,
0.03722295165061951,
0.02091093547642231,
0.00969670433551073,
-0.013262114487588406,
-0.0005771169089712203,
0.03022521547973156,
-0.010804723016917706,
0.03679666295647621,
0.0062456498853862286,
0.0895083025097847,
0.007039526477456093,
-0.021761879324913025,
-0.05038325861096382,
-0.007485213689506054,
-0.09354504197835922,
0.04211828485131264,
0.03040301240980625,
-0.00042473580106161535,
0.04087348282337189,
-0.03649146109819412,
0.07045347988605499,
-0.011310693807899952,
0.027283063158392906,
0.05222737789154053,
0.04991189017891884,
0.016643211245536804,
-0.006640330888330936,
0.003260034369304776,
0.0672045573592186,
-0.06359367817640305,
-0.009787686169147491,
-0.009507933631539345,
0.03875196352601051,
0.0018135369755327702,
-0.05590800940990448,
-0.002706798491999507,
0.05259759724140167,
0.020517321303486824,
0.021625978872179985,
-0.007564537227153778,
0.06991475075483322,
-0.01550955232232809,
0.033487316220998764,
-0.012956546619534492,
-0.0024987845681607723,
-0.03555292263627052,
-0.03498265892267227,
-0.008246713317930698,
0.05797659233212471,
0.0005199211882427335,
-0.08659344911575317,
-0.0030259103514254093,
-0.02483794279396534,
0.030581284314393997,
-0.08070463687181473,
-0.005035382695496082,
-0.025906288996338844,
-0.003722444176673889,
0.01232780423015356,
0.03581054136157036,
0.016551155596971512,
0.03455762565135956,
0.0232564527541399,
0.010521938093006611,
0.028451384976506233,
0.08314843475818634,
-0.031917180866003036,
0.06973885744810104,
0.017116885632276535,
0.0486535020172596,
0.028238065540790558,
0.05253681540489197,
-0.03212514892220497,
-0.008910557255148888,
0.008561389520764351,
0.05128587782382965,
0.009337232448160648,
-0.036750562489032745,
-0.03447572886943817,
0.011068688705563545,
-0.0008096068631857634,
0.04183712974190712,
0.042710136622190475,
0.12208849936723709,
-0.05603555217385292,
-0.02850712649524212,
-0.02957111969590187,
-0.03809715434908867,
0.004972326103597879,
-0.01603185012936592,
-0.018352825194597244,
0.008428851142525673,
0.04260270670056343,
-0.021392200142145157,
0.025897447019815445,
-0.004322818946093321,
0.04037991911172867,
0.04704087972640991,
-0.019779665395617485,
0.02166721224784851,
-0.010891584679484367,
0.07379323244094849,
0.017370931804180145,
0.03816729038953781,
-0.03272922337055206,
0.02131444588303566,
-0.008266504853963852,
0.03152778372168541,
0.03288315236568451,
0.0067713079042732716,
0.04723222181200981,
-0.033543363213539124,
-0.025321317836642265,
0.059431519359350204,
0.012598431669175625,
-0.031756896525621414,
-0.007516713347285986,
-0.005022983532398939,
-0.018501440063118935,
0.02159678377211094,
0.0042152684181928635,
-0.027531269937753677,
0.01141191553324461,
0.05188140645623207,
0.04684270918369293,
0.0023072012700140476,
-0.026472127065062523,
-0.024370897561311722,
0.013914326205849648,
0.04883209615945816,
0.0734897330403328,
0.013325628824532032,
0.0032136233057826757,
0.012516571208834648,
0.03837522119283676,
-0.002306437585502863,
-0.019875627011060715,
0.06812016665935516,
-0.09673334658145905,
0.05006089434027672,
-0.0334119014441967,
-0.00866574514657259,
-0.010518704541027546,
-0.06406617164611816,
-0.013224405236542225,
0.03852275758981705,
-0.05321602523326874,
-0.026803482323884964,
-0.05369143933057785,
-0.008000238798558712,
-0.028696438297629356,
0.005369434133172035,
0.033132776618003845,
-0.014477824792265892,
-0.035954587161540985,
-0.02955002523958683,
-0.001216321368701756,
0.002995032351464033,
0.06163417920470238,
0.009160122834146023,
0.03310099244117737,
-0.08168330788612366,
-0.05001722648739815,
-0.042233556509017944,
0.03416861221194267,
-0.004668304231017828,
0.02179359830915928,
-0.01448215451091528,
-0.025015246123075485,
-0.036915116012096405,
-0.008385919965803623,
-0.010830557905137539,
0.04772862419486046,
0.04292108863592148,
-0.010487840510904789,
-0.01799752749502659,
-0.05754796043038368,
-0.0037053318228572607,
0.023770539090037346,
0.07541859894990921,
0.02685737982392311,
0.004286227282136679,
0.05141935870051384,
0.06011693924665451,
-0.056893061846494675,
0.07156100124120712,
-0.003438927698880434,
0.02360234595835209,
-3.9179798477562144e-05
]
],
"std": [
[
0.9764314889907837,
1.0211230516433716,
0.9537190198898315,
0.9804485440254211,
0.9653822183609009,
0.9893568158149719,
1.0452691316604614,
0.8841189742088318,
0.9992064833641052,
0.9486748576164246,
0.963719367980957,
1.155754566192627,
1.0895198583602905,
0.9194809198379517,
1.0995614528656006,
1.0311801433563232,
1.0202665328979492,
0.9967256188392639,
0.9117249846458435,
0.8812009692192078,
1.031898021697998,
0.9735895395278931,
0.8732784986495972,
0.9987170696258545,
0.9813572764396667,
0.8988218903541565,
0.992881178855896,
0.8459854125976562,
0.999336838722229,
1.0886225700378418,
0.9308749437332153,
0.9938654899597168,
1.1240042448043823,
1.0124523639678955,
1.030287504196167,
1.0460314750671387,
0.9806774258613586,
1.0205169916152954,
1.036083459854126,
1.113065242767334,
0.9208481907844543,
0.7366371750831604,
1.0975638628005981,
1.0625592470169067,
0.9376500248908997,
0.9680666923522949,
1.0336203575134277,
1.0803191661834717,
0.9476468563079834,
1.0788174867630005,
0.8555576801300049,
0.8978824019432068,
0.9674569368362427,
1.0748887062072754,
0.8746950626373291,
1.0838682651519775,
0.9522006511688232,
1.0308846235275269,
0.9887060523033142,
0.7987900972366333,
1.0105385780334473,
0.9565116167068481,
1.061678409576416,
1.0479332208633423,
0.9825246930122375,
1.0140668153762817,
1.0726234912872314,
1.0204355716705322,
1.0739799737930298,
1.0033323764801025,
1.0935736894607544,
1.0205678939819336,
1.0440372228622437,
1.0128377676010132,
1.0341851711273193,
1.0143182277679443,
1.0131605863571167,
1.0955194234848022,
1.1120330095291138,
0.9844388365745544,
1.0389149188995361,
1.0802085399627686,
0.8403464555740356,
1.0851470232009888,
1.005518913269043,
0.9517987370491028,
1.0308743715286255,
1.0642484426498413,
0.9759883284568787,
0.9674128293991089,
0.9534948468208313,
0.977166473865509,
1.005882740020752,
0.9354233145713806,
0.9561018347740173,
1.0553257465362549,
0.9118838310241699,
1.0273126363754272,
1.1243360042572021,
0.9836583733558655,
1.0264487266540527,
1.0053898096084595,
1.0093896389007568,
0.8318060636520386,
1.0413427352905273,
0.9435185790061951,
1.0582057237625122,
0.9668142199516296,
0.9945045113563538,
0.8779057264328003,
1.0486721992492676,
1.059097170829773,
0.9552039504051208,
0.9445965886116028,
1.063852071762085,
0.8942506313323975,
0.9917181134223938,
1.0602153539657593,
1.0200163125991821,
1.0244213342666626,
0.9898468852043152,
0.9715942144393921,
1.079428791999817,
1.024674892425537,
1.0347089767456055,
0.9917073249816895,
1.1281354427337646,
1.0171256065368652,
0.9848858714103699,
0.9594829678535461,
0.8874891996383667,
0.9576436281204224,
1.0805714130401611,
1.0281935930252075,
1.0561119318008423,
1.0014591217041016,
0.9745699763298035,
1.0737097263336182,
1.0942798852920532,
0.945891797542572,
0.9173468947410583,
1.0728681087493896,
1.0073468685150146,
1.077857494354248,
1.07002592086792,
1.0006595849990845,
0.9740257263183594,
1.0437283515930176,
0.9780846238136292,
1.094173789024353,
0.9038551449775696,
1.0628260374069214,
1.1331734657287598,
0.9150418043136597,
1.0270310640335083,
1.1034488677978516,
0.9617258310317993,
1.0133018493652344,
0.9494602084159851,
1.081701636314392,
1.0030829906463623,
0.9401823282241821,
1.0448291301727295,
1.146396279335022,
1.0936627388000488,
0.9121557474136353,
0.9991093873977661,
0.9443554282188416,
0.9845660924911499,
1.05245041847229,
0.9437494277954102,
1.0455018281936646,
0.9653962254524231,
0.9539138078689575,
0.9768818616867065,
1.1123111248016357,
0.9158292412757874,
1.024539589881897,
0.9845446944236755,
0.9192908406257629,
0.8840417861938477,
1.009352445602417,
1.0474157333374023,
1.012848138809204,
0.9426905512809753,
1.0978808403015137,
1.0075842142105103,
1.0328251123428345,
1.0016602277755737,
1.0136278867721558,
0.9837449789047241,
1.0099163055419922
]
]
},
"probes": {
"objective_met": {
"class_counts": {
"train_positive": 234136,
"train_negative": 565864,
"val_positive": 9575,
"val_negative": 90425,
"test_positive": 9244,
"test_negative": 90756
},
"train_source_counts": {
"gt": 400000,
"imagined": 400000,
"imagined_sampling": "step_class_balanced",
"negative_sampling": "mixed"
},
"threshold_policy": "target_fpr",
"threshold_source": "imagined_val",
"condition_on_step": false,
"hard_negative_mining": true,
"linear": {
"threshold": 0.7027591991424558,
"positive_rate": 0.09244000166654587,
"mean_predicted_probability": 0.10874494910240173,
"roc_auc": 0.9988601481008255,
"average_precision": 0.9892172807999081,
"tp": 8225.0,
"tn": 90504.0,
"fp": 252.0,
"fn": 1019.0,
"accuracy": 0.98729,
"balanced_accuracy": 0.9434948294988474,
"precision": 0.9702725020644096,
"recall": 0.889766334919948,
"specificity": 0.9972233240777469,
"f1": 0.9282771852604255
},
"mlp": {
"threshold": 0.22969225093722448,
"positive_rate": 0.09244000166654587,
"mean_predicted_probability": 0.09291515499353409,
"roc_auc": 0.9997839354766374,
"average_precision": 0.9979031588181242,
"tp": 9125.0,
"tn": 90370.0,
"fp": 386.0,
"fn": 119.0,
"accuracy": 0.99495,
"balanced_accuracy": 0.9914368113081139,
"precision": 0.9594154137314689,
"recall": 0.9871267849415837,
"specificity": 0.9957468376746441,
"f1": 0.9730738469741401
},
"imagined_val": {
"linear": {
"threshold": 0.7027591991424558,
"positive_rate": 0.09217000007629395,
"mean_predicted_probability": 0.1610293835401535,
"roc_auc": 0.9744638244622095,
"average_precision": 0.8262833856980545,
"tp": 7992.0,
"tn": 86775.0,
"fp": 4008.0,
"fn": 1225.0,
"accuracy": 0.94767,
"balanced_accuracy": 0.911472089677066,
"precision": 0.666,
"recall": 0.8670934143430618,
"specificity": 0.9558507650110704,
"f1": 0.7533581561955037
},
"mlp": {
"threshold": 0.22969225093722448,
"positive_rate": 0.09217000007629395,
"mean_predicted_probability": 0.11417478322982788,
"roc_auc": 0.9802446668368997,
"average_precision": 0.8547125660676285,
"tp": 8642.0,
"tn": 86425.0,
"fp": 4358.0,
"fn": 575.0,
"accuracy": 0.95067,
"balanced_accuracy": 0.9448053468822426,
"precision": 0.6647692307692308,
"recall": 0.9376152761202127,
"specificity": 0.9519954176442726,
"f1": 0.7779628212629968
}
},
"imagined_test": {
"linear": {
"threshold": 0.7027591991424558,
"positive_rate": 0.09167999774217606,
"mean_predicted_probability": 0.15385475754737854,
"roc_auc": 0.9776629208313851,
"average_precision": 0.8375050401750187,
"tp": 7814.0,
"tn": 87332.0,
"fp": 3500.0,
"fn": 1354.0,
"accuracy": 0.95146,
"balanced_accuracy": 0.9068898576079776,
"precision": 0.690648753756408,
"recall": 0.8523123909249564,
"specificity": 0.9614673242909988,
"f1": 0.7630114246655602
},
"mlp": {
"threshold": 0.22969225093722448,
"positive_rate": 0.09167999774217606,
"mean_predicted_probability": 0.10984945297241211,
"roc_auc": 0.985993707415197,
"average_precision": 0.8567641919276587,
"tp": 8651.0,
"tn": 86988.0,
"fp": 3844.0,
"fn": 517.0,
"accuracy": 0.95639,
"balanced_accuracy": 0.9506441575894404,
"precision": 0.6923569427771108,
"recall": 0.943608202443281,
"specificity": 0.9576801127355998,
"f1": 0.7986890089092
}
}
}
},
"latent_cache_metadata": {
"dataset": "pusht_expert_train",
"checkpoint": "hf_pusht/weights.pt",
"cache_dir": "le-wm/models",
"max_samples": 1000000,
"sample_block_size": 16,
"split": {
"train_fraction": 0.8,
"val_fraction": 0.1,
"seed": 3072
}
},
"imagined_cache": "/home/inigo/data/git/repo/offline-rl-with-le-wm/models/probes/pusht_lewm_1M_sparse_imagined50_fpr005/objective_met_imagined_train.npz",
"imagined_val_cache": "/home/inigo/data/git/repo/offline-rl-with-le-wm/models/probes/pusht_lewm_1M_sparse_imagined50_fpr005/objective_met_imagined_val.npz",
"imagined_test_cache": "/home/inigo/data/git/repo/offline-rl-with-le-wm/models/probes/pusht_lewm_1M_sparse_imagined50_fpr005/objective_met_imagined_test.npz",
"imagined_cache_metadata": {
"dataset": "pusht_expert_train",
"checkpoint": "hf_pusht/weights.pt",
"frameskip": 5,
"context_steps": 3,
"rollout_horizon": 20,
"num_samples": 400000,
"split": "train",
"normalize_actions": true
},
"imagined_val_cache_metadata": {
"dataset": "pusht_expert_train",
"checkpoint": "hf_pusht/weights.pt",
"frameskip": 5,
"context_steps": 3,
"rollout_horizon": 20,
"num_samples": 100000,
"split": "val",
"normalize_actions": true
},
"imagined_test_cache_metadata": {
"dataset": "pusht_expert_train",
"checkpoint": "hf_pusht/weights.pt",
"frameskip": 5,
"context_steps": 3,
"rollout_horizon": 20,
"num_samples": 100000,
"split": "test",
"normalize_actions": true
}
}