Adapt all metrics + fix unit test

Author: EtienneCmb

hoi/core/tests/test_combinatory.py

@@ -1,33 +1,79 @@
 import pytest
+
 import numpy as np
+import jax.numpy as jnp
+
 from math import comb as ccomb
 from hoi.core.combinatory import combinations, _combinations
 from collections.abc import Iterable
 
 
 class TestCombinatory(object):
+    @pytest.mark.parametrize("target", [[], [20], [20, 21]])
     @pytest.mark.parametrize(
         "n", [np.random.randint(5, 10) for _ in range(10)]
     )
     @pytest.mark.parametrize(
         "k", [np.random.randint(5, 10) for _ in range(10)]
     )
     @pytest.mark.parametrize("order", [True, False])
-    def test_single_combinations(self, n, k, order):
-        c = list(_combinations(n, k, order))
+    def test_single_combinations(self, n, k, order, target):
+        c = list(_combinations(n, k, order, target))
+
+        # test that the number of combinations is correct
         assert len(c) == ccomb(n, k)
-        pass
 
-    @pytest.mark.parametrize("n", [np.random.randint(5, 10) for _ in range(2)])
-    @pytest.mark.parametrize(
-        "min", [np.random.randint(1, 10) for _ in range(2)]
-    )
+        # check the order
+        if order:
+            assert all([o == k + len(target) for o in c])
+        else:
+            assert all([len(o) == k + len(target) for o in c])
+
+        # check that targets are included
+        if len(target) and not order:
+            assert all([all([m in o for m in target]) for o in c])
+
+    @pytest.mark.parametrize("fill", [-1, -10])
+    @pytest.mark.parametrize("target", [None, [20], [20, 21]])
+    @pytest.mark.parametrize("order", [True, False])
+    @pytest.mark.parametrize("astype", ["numpy", "jax", "iterator"])
     @pytest.mark.parametrize(
         "max", [_ for _ in range(2)]
     )  # addition to minimum size
-    @pytest.mark.parametrize("astype", ["numpy", "jax", "iterator"])
-    @pytest.mark.parametrize("order_val", [True, False])
-    def test_combinations(self, n, min, max, astype, order_val):
-        combs = combinations(n, min, min + max, astype, order_val)
-        assert isinstance(combs, Iterable)
-        pass
+    @pytest.mark.parametrize(
+        "min", [np.random.randint(1, 10) for _ in range(2)]
+    )
+    @pytest.mark.parametrize("n", [np.random.randint(5, 10) for _ in range(2)])
+    def test_combinations(self, n, min, max, astype, order, target, fill):
+        # get combinations
+        combs = combinations(
+            n,
+            min,
+            maxsize=min + max,
+            astype=astype,
+            order=order,
+            target=target,
+            fill_value=fill,
+        )
+
+        # check the number of multiplets
+        n_mults = 0
+        for c in range(min, min + max + 1):
+            n_mults += ccomb(n, c)
+        if astype in ["jax", "numpy"]:
+            assert combs.shape[0] == n_mults
+        elif astype == "iterator":
+            assert len([c for c in combs]) == n_mults
+
+        # check type
+        if astype == "numpy":
+            assert isinstance(combs, np.ndarray)
+        elif astype == "jax":
+            assert isinstance(combs, jnp.ndarray)
+        elif astype == "iterator":
+            assert isinstance(combs, Iterable)
+
+
+if __name__ == "__main__":
+    # TestCombinatory().test_single_combinations(5, 3, False, [21, 22])
+    TestCombinatory().test_combinations(10, 2, 3, "iterator", False, None, -1)

hoi/metrics/base_hoi.py

@@ -216,8 +216,6 @@ def compute_entropies(
         pbar.close()
 
         self._entropies = h_x
-        self._multiplets = h_idx
-        self._order = order
 
         return h_x, h_idx, order
 

hoi/metrics/gradient_oinfo.py

@@ -80,20 +80,17 @@ def fit(self, minsize=2, maxsize=None, method="gcmi", **kwargs):
             The NumPy array containing values of higher-rder interactions of
             shape (n_multiplets, n_variables)
         """
-        # ____________________________ TASK-FREE ______________________________
-        hoi_tf = self._oinf_tf.fit(
+        kw_oinfo = dict(
             minsize=minsize, maxsize=maxsize, method=method, **kwargs
         )
 
+        # ____________________________ TASK-FREE ______________________________
+        hoi_tf = self._oinf_tf.fit(**kw_oinfo)
+
         self._multiplets = self._oinf_tf._multiplets
 
         # __________________________ TASK-RELATED _____________________________
-        hoi_tr = self._oinf_tr.fit(
-            minsize=minsize,
-            maxsize=maxsize,
-            method=method,
-            **kwargs
-        )
+        hoi_tr = self._oinf_tr.fit(**kw_oinfo)
 
         return hoi_tr - hoi_tf
 
@@ -118,4 +115,4 @@ def fit(self, minsize=2, maxsize=None, method="gcmi", **kwargs):
     model = GradientOinfo(x, y=y)
     hoi = model.fit(minsize=2, maxsize=None, method="gcmi")
 
-    print(get_nbest_mult(hoi, model=model, minsize=3, maxsize=3, n_best=3))
+    print(get_nbest_mult(hoi, model=model, minsize=3, maxsize=3, n_best=3))

hoi/metrics/info_tot.py

@@ -1,4 +1,3 @@
-from math import comb as ccomb
 from functools import partial
 
 import numpy as np
@@ -7,7 +6,6 @@
 import jax.numpy as jnp
 
 from hoi.metrics.base_hoi import HOIEstimator
-from hoi.core.combinatory import combinations
 from hoi.core.entropies import prepare_for_entropy
 from hoi.core.mi import get_mi, compute_mi_comb
 from hoi.utils.progressbar import get_pbar

hoi/metrics/infotopo.py

@@ -1,3 +1,5 @@
+import warnings
+
 import numpy as np
 
 import jax
@@ -64,12 +66,6 @@ class InfoTopo(HOIEstimator):
     x : array_like
         Standard NumPy arrays of shape (n_samples, n_features) or
         (n_samples, n_features, n_variables)
-    y : array_like
-        The feature of shape (n_samples,) for estimating task-related O-info.
-    multiplets : list | None
-        List of multiplets to compute. Should be a list of multiplets, for
-        example [(0, 1, 2), (2, 7, 8, 9)]. By default, all multiplets are
-        going to be computed.
 
     References
     ----------
@@ -83,9 +79,14 @@ class InfoTopo(HOIEstimator):
     _negative = "synergy"
     _symmetric = True
 
-    def __init__(self, x, y=None, multiplets=None, verbose=None):
+    def __init__(self, x, y=None, verbose=None):
+        # for infotopo, the multiplets are set to None because this metric
+        # first require to compute entropies and then associate them to form
+        # the MI. Same for the target y.
+        if y is not None:
+            warnings.warn("For InfoTopo, y input is going to be ignored.")
         HOIEstimator.__init__(
-            self, x=x, y=y, multiplets=multiplets, verbose=verbose
+            self, x=x, y=None, multiplets=None, verbose=verbose
         )
 
     def fit(self, minsize=1, maxsize=None, method="gcmi", **kwargs):
@@ -120,20 +121,20 @@ def fit(self, minsize=1, maxsize=None, method="gcmi", **kwargs):
         # ____________________________ ENTROPIES ______________________________
 
         minsize, maxsize = self._check_minmax(minsize, maxsize)
-        h_x, h_idx, order = self.compute_entropies(
+        h_x, h_idx, _ = self.compute_entropies(
             minsize=1, maxsize=maxsize, method=method, **kwargs
         )
-        n_mult = h_x.shape[0]
 
         # _______________________________ HOI _________________________________
 
         # compute order and multiply entropies
+        order = (h_idx >= 0).sum(1)
         h_x_sgn = jnp.multiply(((-1.0) ** (order.reshape(-1, 1) - 1)), h_x)
 
         # subselection of multiplets
         mults, _ = self.get_combinations(minsize, maxsize=maxsize)
         h_idx_2 = jnp.where(mults == -1, -2, mults)
-        n_mult = h_idx_2.shape[0]
+        n_mult = mults.shape[0]
 
         # progress-bar definition
         pbar = scan_tqdm(n_mult, message="Mutual information")
@@ -150,28 +151,23 @@ def fit(self, minsize=1, maxsize=None, method="gcmi", **kwargs):
 
 if __name__ == "__main__":
     import matplotlib.pyplot as plt
-    from hoi.utils import landscape, get_nbest_mult
-    from matplotlib.colors import LogNorm
+    from hoi.utils import get_nbest_mult
+    from hoi.plot import plot_landscape
 
     plt.style.use("ggplot")
 
     x = np.random.rand(200, 7)
     y_red = np.random.rand(x.shape[0])
 
-    # redundancy: (1, 2, 6) + (7, 8)
-    x[:, 1] += y_red
-    x[:, 2] += y_red
-    x[:, 6] += y_red
+    # redundancy: (1, 2, 6)
+    x[:, 2] += x[:, 1]
+    x[:, 6] += x[:, 2]
     # synergy:    (0, 3, 5) + (7, 8)
-    y_syn = x[:, 0] + x[:, 3] + x[:, 5]
-    # bivariate target
-    y = np.c_[y_red, y_syn]
+    x[:, 0] = x[:, 0] + x[:, 3] + x[:, 5]
 
-    model = InfoTopo(x, y=y)
-    hoi = model.fit(maxsize=None, method="gcmi")
+    model = InfoTopo(x)
+    hoi = model.fit(minsize=3, maxsize=5, method="gcmi")
     print(get_nbest_mult(hoi, model=model, minsize=3, maxsize=3, n_best=3))
 
-    lscp = landscape(hoi.squeeze(), model.order, output="xarray")
-    lscp.plot(x="order", y="bins", cmap="jet", norm=LogNorm())
-    plt.axvline(model.undersampling, linestyle="--", color="k")
+    plot_landscape(hoi, model=model)
     plt.show()

hoi/metrics/red_mmi.py

@@ -1,4 +1,3 @@
-from math import comb as ccomb
 from functools import partial
 
 import numpy as np
@@ -7,7 +6,6 @@
 import jax.numpy as jnp
 
 from hoi.metrics.base_hoi import HOIEstimator
-from hoi.core.combinatory import combinations
 from hoi.core.entropies import prepare_for_entropy
 from hoi.core.mi import get_mi, compute_mi_comb
 from hoi.utils.progressbar import get_pbar

hoi/metrics/rsi.py

@@ -1,4 +1,3 @@
-from math import comb as ccomb
 from functools import partial
 
 import numpy as np
@@ -7,7 +6,6 @@
 import jax.numpy as jnp
 
 from hoi.metrics.base_hoi import HOIEstimator
-from hoi.core.combinatory import combinations
 from hoi.core.entropies import prepare_for_entropy
 from hoi.core.mi import get_mi, compute_mi_comb
 from hoi.utils.progressbar import get_pbar

hoi/metrics/syn_mmi.py

@@ -1,4 +1,3 @@
-from math import comb as ccomb
 from functools import partial
 
 import numpy as np
@@ -7,7 +6,6 @@
 import jax.numpy as jnp
 
 from hoi.metrics.base_hoi import HOIEstimator
-from hoi.core.combinatory import combinations
 from hoi.core.entropies import prepare_for_entropy
 from hoi.core.mi import get_mi, compute_mi_comb
 from hoi.utils.progressbar import get_pbar