Test Marginal Distribution

Setting

config = {
    "ignore": ["order_item"],
    "train": {
        "epochs": 500,
        "lr": 0.01,
        "samples": None,
        "naive_dequantization": True,
    },
    "finetune": {
        "activate": True,
        "samples": 1000,
        "epochs": 50,
        "lr": 0.01,
    },
    "eval": {
        "activate": True,
        "samples": 1000,
    },
    "partitions": 5,
}

model_config = EinetConfig(
    num_channels=1,
    num_leaves=5,
    depth=2,
    num_sums=5,
    num_repetitions=5,
    num_classes=1,
    leaf_type=Normal,
    layer_type='einsum',
    dropout=0.0
)

Training

Generating evaluation queries...
Preprocessing data...
Using naive dequantization...
Training model...

Evaluating model...
GMQ: 1.3147785798696092;
50%: 1.331358507169842;
90%: 1.3920799390966372;
95%: 1.418013446751778;
MAX: 1.6900461766239159
Generating finetune queries...
Finetuning model...

Evaluating model...
GMQ: 1.0494095226517741;
50%: 1.0367809122052642;
90%: 1.0985178470366797;
95%: 1.1413706383551367;
MAX: 1.4810541052894972

Results

Column 0

Column 1

Column 2

Column 3

Column 5

Column 6