| { | |
| "ops": [ | |
| "sin", | |
| "cos", | |
| "log", | |
| "sqrt", | |
| "square", | |
| "cube", | |
| "quad" | |
| ], | |
| "d": 256, | |
| "config": { | |
| "run_name": "compose_v8", | |
| "data": "data/compose_v3", | |
| "latents": "latents_v14.npz", | |
| "ops": [ | |
| "sin", | |
| "cos", | |
| "log", | |
| "sqrt", | |
| "square", | |
| "cube", | |
| "quad" | |
| ], | |
| "seed": 0, | |
| "model": { | |
| "hidden": 768, | |
| "depth": 3, | |
| "mult": 2, | |
| "dropout": 0.1, | |
| "skip_init": "identity", | |
| "head": "factored", | |
| "structure": "exact", | |
| "quad_from_square": true | |
| }, | |
| "loss": { | |
| "w_cos": 0.0, | |
| "w_nce": 0.1, | |
| "tau": 0.05, | |
| "w_gain": 1.0 | |
| }, | |
| "train": { | |
| "steps": 40000, | |
| "batch_size": 1024, | |
| "lr": 5e-05, | |
| "weight_decay": 0.05, | |
| "warmup": 500, | |
| "min_lr_frac": 0.05, | |
| "schedule": "wsd", | |
| "decay_frac": 0.0, | |
| "grad_clip": 1.0, | |
| "eval_every": 2000, | |
| "eval_n": 3000, | |
| "patience": 0, | |
| "ema": 0.0, | |
| "max_lift_frac": 0.25, | |
| "clean_exclude": "data/v4", | |
| "aug_sign": true | |
| } | |
| }, | |
| "best": { | |
| "sin": { | |
| "val_top1": 0.7093333333333334, | |
| "step": 34000 | |
| }, | |
| "cos": { | |
| "val_top1": 0.649, | |
| "step": 32000 | |
| }, | |
| "log": { | |
| "val_top1": 0.6013834846519671, | |
| "step": 38000 | |
| }, | |
| "sqrt": { | |
| "val_top1": 0.5649109792284867, | |
| "step": 40000 | |
| }, | |
| "square": { | |
| "val_top1": 0.9226666666666666, | |
| "step": 40000 | |
| }, | |
| "cube": { | |
| "val_top1": 0.8953333333333333, | |
| "step": 40000 | |
| }, | |
| "quad": { | |
| "val_top1": 0.8016666666666666, | |
| "step": 38000 | |
| } | |
| }, | |
| "component": "compose", | |
| "description": "7 composition operators (sin cos log sqrt square cube quad)", | |
| "source_checkpoint": "runs/compose_v8/opbank.pt", | |
| "domain": [ | |
| -3.141592653589793, | |
| 3.141592653589793 | |
| ], | |
| "format": "func-jepa-1" | |
| } |