Simple Workflow¶

Below, we provide experiments with the dummy (synthetic data) and the molecular data (musk).

1. Examples with Dummy Data¶

1.1 Load the dummy data¶

In [ ]:

from sawmil.data import generate_dummy_bags
import numpy as np
rng = np.random.default_rng(0)

ds = generate_dummy_bags(
    n_pos=100, n_neg=100, inst_per_bag=(5, 15), d=2,
    pos_centers=((+2,+1), (+4,+3)),
    neg_centers=((-1.5,-1.0), (-3.0,+0.5)),
    pos_scales=((2.0, 0.6), (1.2, 0.8)),
    neg_scales=((1.5, 0.5), (2.5, 0.9)),
    pos_intra_rate=(0.25, 0.85),
    ensure_pos_in_every_pos_bag=True,
    neg_pos_noise_rate=(0.00, 0.05),
    pos_neg_noise_rate=(0.00, 0.20),
    outlier_rate=0.1,
    outlier_scale=8.0,
    random_state=42,
)

# Quick sanity:
X_pos, pos_idx = ds.positive_instances()
X_neg, neg_idx = ds.negative_instances()
print("Number of bags:", len(ds.bags))

from sawmil.data import generate_dummy_bags import numpy as np rng = np.random.default_rng(0) ds = generate_dummy_bags( n_pos=100, n_neg=100, inst_per_bag=(5, 15), d=2, pos_centers=((+2,+1), (+4,+3)), neg_centers=((-1.5,-1.0), (-3.0,+0.5)), pos_scales=((2.0, 0.6), (1.2, 0.8)), neg_scales=((1.5, 0.5), (2.5, 0.9)), pos_intra_rate=(0.25, 0.85), ensure_pos_in_every_pos_bag=True, neg_pos_noise_rate=(0.00, 0.05), pos_neg_noise_rate=(0.00, 0.20), outlier_rate=0.1, outlier_scale=8.0, random_state=42, ) # Quick sanity: X_pos, pos_idx = ds.positive_instances() X_neg, neg_idx = ds.negative_instances() print("Number of bags:", len(ds.bags))

Number of bags: 200

1.2. Fit the model¶

In [ ]:

from sawmil.kernels import get_kernel, Linear
k = get_kernel("linear") # base (single-instance kernel)
k1 = Linear() # equivalent k == k1
# if you want to use kernels for bags outside of the models
from sawmil.bag_kernels import make_bag_kernel
bag_k = make_bag_kernel(k, normalizer="none", p=1.0, use_intra_labels=False, fast_linear=True) # bag kernel
# Otherwise, it is handled inside each model

from sawmil.kernels import get_kernel, Linear k = get_kernel("linear") # base (single-instance kernel) k1 = Linear() # equivalent k == k1 # if you want to use kernels for bags outside of the models from sawmil.bag_kernels import make_bag_kernel bag_k = make_bag_kernel(k, normalizer="none", p=1.0, use_intra_labels=False, fast_linear=True) # bag kernel # Otherwise, it is handled inside each model

1.2.1 Fit NSK with the Linear Kernel¶

In [ ]:

from sawmil import NSK

k = get_kernel("linear", normalizer="average")
clf = NSK(C=1, kernel=k, 
          # bag kernel settings
          normalizer='average',
          p=1.0,
          fast_linear=True,
          # solver settings
          scale_C=True, 
          tol=1e-8, 
          verbose=False, 
          solver='osqp').fit(ds, None)
print("Train acc:", clf.score(ds, ds.y))
# clf.predict(ds), clf.decision_function(ds)

from sawmil import NSK k = get_kernel("linear", normalizer="average") clf = NSK(C=1, kernel=k, # bag kernel settings normalizer='average', p=1.0, fast_linear=True, # solver settings scale_C=True, tol=1e-8, verbose=False, solver='osqp').fit(ds, None) print("Train acc:", clf.score(ds, ds.y)) # clf.predict(ds), clf.decision_function(ds)

Linear kernel has no parameters to fit.

Train acc: 0.895

1.2.2 Fit NSK with the RBF Kernel¶

In [ ]:

k = get_kernel("rbf", gamma=0.8)
clf = NSK(C=10, kernel=k, scale_C=True, tol=1e-8, verbose=False, solver='osqp').fit(ds, None)
print("Train acc:", clf.score(ds, ds.y))

k = get_kernel("rbf", gamma=0.8) clf = NSK(C=10, kernel=k, scale_C=True, tol=1e-8, verbose=False, solver='osqp').fit(ds, None) print("Train acc:", clf.score(ds, ds.y))

Train acc: 0.955

1.2.3 Fit NSK with Combined Kernels¶

In [ ]:

from sawmil.kernels import Product, Polynomial, Linear, RBF, Sum, Scale

k = Sum(Linear(), 
        Scale(0.5, 
              Product(Polynomial(degree=2), RBF(gamma=1.0))))
clf = NSK(C=100, kernel=k, 
          # params to create a bag kernel
          normalizer="none",
          # svm params
          scale_C=True, 
          tol=1e-8, verbose=False, solver='gurobi').fit(ds, None)
print("Train acc:", clf.score(ds, ds.y))

from sawmil.kernels import Product, Polynomial, Linear, RBF, Sum, Scale k = Sum(Linear(), Scale(0.5, Product(Polynomial(degree=2), RBF(gamma=1.0)))) clf = NSK(C=100, kernel=k, # params to create a bag kernel normalizer="none", # svm params scale_C=True, tol=1e-8, verbose=False, solver='gurobi').fit(ds, None) print("Train acc:", clf.score(ds, ds.y))

Linear kernel has no parameters to fit.

Set parameter Username
Academic license - for non-commercial use only - expires 2026-08-04
Train acc: 1.0

1.2.3 Fit sMIL with the Linear Kernel¶

In [ ]:

from sawmil import sMIL
from sawmil.kernels import Linear

from sawmil import sMIL from sawmil.kernels import Linear

In [ ]:

k  = Linear()
clf = sMIL(C=10, kernel=k, 
           # params to create a bag kernel
           normalizer="none",
           # svm params
           scale_C=True, 
           tol=1e-8, verbose=False, solver='osqp').fit(ds, None)

k = Linear() clf = sMIL(C=10, kernel=k, # params to create a bag kernel normalizer="none", # svm params scale_C=True, tol=1e-8, verbose=False, solver='osqp').fit(ds, None)

Linear kernel has no parameters to fit.

In [ ]:

y = np.array([b.y for b in ds.bags])
# yhat = clf.predict(ds)
print("Train acc:", clf.score(ds, y))

y = np.array([b.y for b in ds.bags]) # yhat = clf.predict(ds) print("Train acc:", clf.score(ds, y))

Train acc: 0.89

1.2.4. Fit sAwMIL with the Linear kernel¶

In [ ]:

from sawmil import sAwMIL
from sawmil.kernels import get_kernel

from sawmil import sAwMIL from sawmil.kernels import get_kernel

In [ ]:

k = get_kernel('linear')
clf = sAwMIL(C=0.1, kernel=k,
             solver="gurobi", eta=0.95) # here eta is high, since all items in the bag are relevant
clf.fit(ds)
print("Train acc:", clf.score(ds, ds.y))
clf.predict(ds)

k = get_kernel('linear') clf = sAwMIL(C=0.1, kernel=k, solver="gurobi", eta=0.95) # here eta is high, since all items in the bag are relevant clf.fit(ds) print("Train acc:", clf.score(ds, ds.y)) clf.predict(ds)

Linear kernel has no parameters to fit.
Linear kernel has no parameters to fit.

Train acc: 0.805

Out[ ]:

array([1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
       1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
       1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
       1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
       1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.,
       1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 0., 0.,
       0., 0., 1., 1., 0., 1., 1., 0., 0., 0., 0., 0., 1., 0., 0., 1., 1.,
       0., 0., 1., 1., 1., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0., 0., 1.,
       1., 1., 1., 1., 1., 0., 0., 1., 0., 0., 1., 1., 0., 1., 1., 0., 0.,
       0., 0., 0., 1., 1., 1., 0., 1., 0., 0., 0., 0., 1., 1., 1., 0., 1.,
       1., 1., 0., 0., 1., 0., 1., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
       0., 1., 0., 0., 0., 1., 0., 0., 1., 0., 0., 1., 0.])

2. Experiments with the molecular data¶

2.1 Load the musk data¶

In [ ]:

from sawmil.data import load_musk_bags
train_ds, test_ds, scaler = load_musk_bags(standardize=True, test_size=0.3, random_state=42)

from sawmil.data import load_musk_bags train_ds, test_ds, scaler = load_musk_bags(standardize=True, test_size=0.3, random_state=42)

2.2 Fit models¶

2.2.1 Fit NSK with Combined Kernel¶

In [ ]:

from sawmil import NSK
from sawmil.kernels import Linear, Sum, RBF
from sklearn.metrics import matthews_corrcoef as mcc


k = Sum(Linear(),RBF(gamma=0.5))

clf = NSK(C=10, kernel=k, 
          normalizer="none",
          scale_C=True, tol=1e-8, verbose=False, solver='osqp').fit(train_ds, None)

y = ds.y
yhat = clf.predict(test_ds)
print('Matthews Correlation Coefficient:', mcc(y, yhat))

from sawmil import NSK from sawmil.kernels import Linear, Sum, RBF from sklearn.metrics import matthews_corrcoef as mcc k = Sum(Linear(),RBF(gamma=0.5)) clf = NSK(C=10, kernel=k, normalizer="none", scale_C=True, tol=1e-8, verbose=False, solver='osqp').fit(train_ds, None) y = ds.y yhat = clf.predict(test_ds) print('Matthews Correlation Coefficient:', mcc(y, yhat))

Linear kernel has no parameters to fit.

Matthews Correlation Coefficient: 0.6882604778571242

2.2.2 Fit sMIL with Combined Kernel¶

In [ ]:

from sawmil.kernels import Linear
from sawmil import sMIL

k = Sum(Linear(),RBF(gamma=0.5))

clf = sMIL(C=10, kernel=k, scale_C=True, tol=1e-8, verbose=True, solver='osqp').fit(train_ds, None)

from sawmil.kernels import Linear from sawmil import sMIL k = Sum(Linear(),RBF(gamma=0.5)) clf = sMIL(C=10, kernel=k, scale_C=True, tol=1e-8, verbose=True, solver='osqp').fit(train_ds, None)

Linear kernel has no parameters to fit.

-----------------------------------------------------------------
           OSQP v1.0.0  -  Operator Splitting QP Solver
              (c) The OSQP Developer Team
-----------------------------------------------------------------
problem:  variables n = 2199, constraints m = 2200
          nnz(P) + nnz(A) = 2423298
settings: algebra = Built-in,
          OSQPInt = 4 bytes, OSQPFloat = 8 bytes,
          linear system solver = QDLDL v0.1.8,
          eps_abs = 1.0e-08, eps_rel = 1.0e-08,
          eps_prim_inf = 1.0e-04, eps_dual_inf = 1.0e-04,
          rho = 1.00e-01 (adaptive: 50 iterations),
          sigma = 1.00e-06, alpha = 1.60, max_iter = 20000
          check_termination: on (interval 25, duality gap: on),
          time_limit: 1.00e+10 sec,
          scaling: on (10 iterations), scaled_termination: off
          warm starting: on, polishing: on, 
iter   objective    prim res   dual res   gap        rel kkt    rho         time
   1  -2.5694e+01   2.17e+00   2.12e+00  -1.93e+01   2.17e+00   1.00e-01    1.03e+00s
 200  -5.1152e+00   2.65e-04   8.27e-04  -4.07e-02   8.27e-04   1.00e-01    1.94e+00s
 400  -5.0805e+00   4.98e-05   1.40e-04  -8.78e-03   1.40e-04   1.00e-01    2.87e+00s
 600  -5.0729e+00   6.05e-06   1.62e-05  -1.26e-03   1.62e-05   1.00e-01    3.76e+00s
 800  -5.0718e+00   9.51e-07   1.06e-05  -2.23e-04   1.06e-05   1.00e-01    4.67e+00s
1000  -5.0717e+00   1.79e-07   5.62e-06  -4.12e-05   5.62e-06   1.00e-01    5.57e+00s
1200  -5.0716e+00   4.16e-08   2.96e-06  -7.74e-06   2.96e-06   1.00e-01    6.48e+00s
1400  -5.0716e+00   1.04e-08   1.62e-06  -1.48e-06   1.62e-06   1.00e-01    7.44e+00s
1600  -5.0716e+00   3.18e-09   8.88e-07  -2.93e-07   8.88e-07   1.00e-01    8.41e+00s
1800  -5.0716e+00   1.40e-09   4.85e-07  -6.18e-08   4.85e-07   1.00e-01    9.41e+00s
2000  -5.0716e+00   7.31e-10   2.65e-07  -1.49e-08   2.65e-07   1.00e-01    1.04e+01s
2200  -5.0716e+00   3.89e-10   1.45e-07  -4.44e-09   1.45e-07   1.00e-01    1.13e+01s
2400  -5.0716e+00   2.08e-10   7.99e-08  -1.64e-09   7.99e-08   1.00e-01    1.23e+01s
2475  -5.0716e+00   1.65e-10   6.39e-08  -1.17e-09   6.39e-08   1.00e-01    1.27e+01s
plsh  -5.0716e+00   3.84e-15   2.10e-13   7.42e-14   2.10e-13   --------    1.37e+01s

status:               solved
solution polishing:   successful
number of iterations: 2475
optimal objective:    -5.0716
dual objective:       -5.0716
duality gap:          7.4186e-14
primal-dual integral: 2.0957e+01
run time:             1.37e+01s
optimal rho estimate: 2.65e-02

In [ ]:

y = ds.y
yhat = clf.predict(test_ds)
print('Matthews Correlation Coefficient:', mcc(y, yhat))

y = ds.y yhat = clf.predict(test_ds) print('Matthews Correlation Coefficient:', mcc(y, yhat))

Matthews Correlation Coefficient: 0.6369615602528665

In [ ]: