bpnet-fnbm-grid-rows-v1

Per-(window, strand) profile measures on valid and test for the best-val-loss checkpoint and three reference predictors

Dataset Info

• Rows: 12312864
• Columns: 24

Columns

Column	Type	Description
row_id	Value('large_string')	run:split:predictor:npz row:strand
run	Value('large_string')	Run directory name: -init_lr-interaction_feature_l2_penalty-seed
model	Value('large_string')	Model class of the run (BPNetFactorizedNBM). The predictor column says which predictor a row scores.
control	Value('bool')	Whether the run's model used the PATCH-CAP control track (always true in this grid)
init_lr	Value('float64')	Initial learning rate of the run (grid factor)
interaction_feature_l2_penalty	Value('float64')	model.kwargs.interaction_feature_l2_penalty of the run (grid factor)
seed	Value('int64')	Run seed
split	Value('large_string')	valid (chr2/3/4, 29,277 windows) or test (chr1/8/9, 27,727 windows)
predictor	Value('large_string')	model = the run's best-val-loss checkpoint; uniform, mean_profile, control = reference predictors fit on train
row	Value('int64')	Row index of the window in the split's npz
strand	Value('int64')	0 = plus, 1 = minus
source	Value('large_string')	TF whose peak summit the window is centred on (Oct4, Sox2, Nanog, Klf4)
chrom	Value('large_string')	Chromosome of the window (mm10)
start	Value('int64')	Window start coordinate (mm10)
n_reads	Value('float64')	Observed Sox2 ChIP-nexus reads on this strand of the window
profile_mnll	Value('float64')	Multinomial NLL of this strand's counts under softmax(profile logits) over its 1000 positions
profile_ce	Value('float64')	NLL per read in nats (-sum y log p / n); uniform = ln 1000 = 6.9078
profile_jsd	Value('float64')	Jensen-Shannon divergence (natural log, range [0, 0.693]) between predicted and observed profile
profile_pearson	Value('float64')	Pearson r along the 1000 positions between predicted probabilities and counts
profile_spearman	Value('float64')	Spearman r along positions, ties in torch's unstable argsort order (bpnet-lite's call)
profile_jsd_smooth	Value('float64')	profile_jsd against observed counts smoothed with a Gaussian (sigma 7, width 81)
profile_pearson_smooth	Value('float64')	profile_pearson against the same smoothed counts
pred_log_count	Value('float64')	Predicted log(1 + total reads over both strands) for the window (same on both strand rows)
true_log_count	Value('float64')	Observed log(1 + total reads over both strands) for the window

Generation Parameters

{
  "script_name": "loop.py / scripts/bpnet_eval.py (factorized-model-iterations, branch bpnet-canary, commit dce5aaa); jobs/bpnet/fnbm_grid.sbatch",
  "model": "BPNetFactorizedNBM",
  "description": "Per-(window, strand) profile measures on valid and test for the best-val-loss checkpoint and three reference predictors",
  "hyperparameters": {
    "num_epochs": 50,
    "early_stopping_patience": 10,
    "lr_decay_patience": 5,
    "lr_decay_factor": 0.5,
    "min_lr": 1e-05,
    "count_weight": 1.0,
    "gradient_clip": 1.0,
    "batch_size": 64,
    "rc_augment": true,
    "grid": {
      "init_lr": [
        0.001,
        0.0001,
        5e-05
      ],
      "interaction_feature_l2_penalty": [
        0.0001,
        0.001,
        0.002
      ],
      "seed": [
        0,
        1,
        2
      ]
    }
  },
  "input_datasets": [],
  "experiment_name": "bpnet-fnbm-grid",
  "job_id": "torch:19801051",
  "cluster": "torch",
  "artifact_status": "final",
  "canary": false
}

Usage

from datasets import load_dataset

dataset = load_dataset("arushram/bpnet-fnbm-grid-rows-v1", split="train")
print(f"Loaded {len(dataset)} rows")

---