some shape fixing stuff for lang -> info plane measurement

src/examples/colors/ib.py

@@ -17,11 +17,16 @@ def language_to_encoder(language: Language) -> np.ndarray:
         where element (r, w) is the probability of word w given referent r
     """
     universe = language.universe
-    encoder = np.array(
+    # (|referents|, |words|): p(r | w)
+    prob_chip_given_expression = np.array(
         [
             [expression.meaning[referent] for expression in language.expressions]
             for referent in universe.referents
         ]
     )
+    # (|referents|, 1): p(r)
+    # prior = np.array(universe.prior)[:, None]
+    # TODO: get encoder from prob_chip_given_expression (i.e. prob_expression_given_chip)
+    encoder = prob_chip_given_expression
     print(encoder)
     return encoder

src/examples/colors/scripts/measure_natural_languages.py

@@ -9,6 +9,7 @@
 
 if __name__ == "__main__":
 
+    """
     with open("colors/outputs/natural_languages.yaml", "r") as f:
         languages = load(f, Loader=Loader)
 
@@ -21,14 +22,15 @@
     encoder = language_to_encoder(language)
     print(encoder)
     print(encoder.sum(axis=0))
+    print(encoder.sum(axis=1))
 
     print(
         ib_encoder_to_point(
             np.array(color_universe.prior), meaning_distributions, encoder
         )
     )
-    """
 
+    """
     prior = np.array(color_universe.prior)
     information_plane = pd.DataFrame.from_records(
         [
@@ -43,3 +45,4 @@
     information_plane.to_csv(
         "colors/outputs/natural_language_information_plane.csv", index=False
     )
+    """

src/examples/colors/scripts/read_color_universe.py

@@ -1,19 +1,11 @@
 import pandas as pd
-import pickle
-
-from ultk.language.semantics import Universe
 
 
 if __name__ == "__main__":
     referents = pd.read_csv("colors/data/cnum-vhcm-lab-new.txt", delimiter="\t")
     referents.sort_values(by="#cnum", inplace=True)
     # add a name column, as required by ULTK
     referents["name"] = referents["#cnum"]
+    # rename columns for access as python properties
     referents.rename(columns={"L*": "L", "a*": "a", "b*": "b"}, inplace=True)
     referents.to_csv("colors/outputs/color_universe.csv", index=False)
-    """
-    color_universe = Universe.from_dataframe(referents)
-
-    with open("colors/outputs/color_universe.pkl", "wb") as f:
-        pickle.dump(color_universe, f)
-    """

src/ultk/effcomm/rate_distortion.py

@@ -122,39 +122,35 @@ def ib_encoder_to_point(
     # IB complexity = info rate of encoder = I(meanings; words)
     complexity = information_cond(prior, encoder)
 
+    # (|meanings|, 1)
+    prior = prior[:, None]
+
     # IB accuracy/informativity = I(words; world states)
-    pMW = encoder * prior[:, None]
+    pMW = encoder * prior
     pWU = pMW.T @ meaning_dists
     accuracy = mutual_info(pWU)
 
     # expected distortion
     I_mu = information_cond(prior, meaning_dists)
     distortion = I_mu - accuracy
 
-    # TODO: debug the below!
-
     # TODO: the above is only IB optimal; should we look at the emergent listener accuracy? To do that we'll need to compute kl divergence
     # and then do I(M;U) - distortion to get the accuracy.
 
     # pu_w = decoder @ meaning_dists
     # dist_mat = ib_kl(meaning_dists, pu_w,) # getting infs; I confirmed that this because there exists an x s.t. p(x) > 0 but q(x) = 0. Ask Noga what to do here. Add a little epsilon?
     # distortion = np.sum( prior * ( encoder @ decoder ) * dist_mat )
 
-    """
     decoder_smoothed = decoder + 1e-20
     decoder_smoothed /= decoder_smoothed.sum(axis=1, keepdims=True)
-    print(decoder_smoothed.shape)
     pu_w = decoder_smoothed @ meaning_dists
-    print(pu_w.shape)
     dist_mat = ib_kl(
         meaning_dists,
         pu_w,
     )
-    print(dist_mat.shape) # NOTE: this is the wrong shape
-    distortion = np.sum(prior * (encoder @ decoder) * dist_mat)
+    distortion = np.sum(prior * encoder * dist_mat)
     # but this measure of distortion is almost an order magnitude higher than bayesian decoder
     # breakpoint()
-    """
 
     return (complexity, accuracy, distortion)