setting up zero-shot experiments (done and tested for MuseGen)

main.py

@@ -15,23 +15,14 @@ def main(args):
     print('Running generalized funnelling...')
 
     data = MultilingualDataset.load(args.dataset)
-    data.set_view(languages=['it', 'da'])
+    data.set_view(languages=['it', 'da', 'nl'])
     data.show_dimensions()
     lX, ly = data.training()
-
-    # Testing zero shot experiments
-    # zero_shot_setting = True
-    # if zero_shot_setting:
-    #     # _lX = {}
-    #     _ly = {}
-    #     train_langs = ['it']
-    #     for train_lang in train_langs:
-    #         # _lX[train_lang] = lX[train_lang]
-    #         _ly[train_lang] = ly[train_lang]
-    #     ly = _ly
-
     lXte, lyte = data.test()
 
+    zero_shot = True
+    zscl_train_langs = ['it']   # Todo: testing zero shot
+
     # Init multilingualIndex - mandatory when deploying Neural View Generators...
     if args.gru_embedder or args.bert_embedder:
         multilingualIndex = MultilingualIndex()
@@ -45,7 +36,8 @@ def main(args):
         embedder_list.append(posteriorEmbedder)
 
     if args.muse_embedder:
-        museEmbedder = MuseGen(muse_dir=args.muse_dir, n_jobs=args.n_jobs, zero_shot=True)
+        museEmbedder = MuseGen(muse_dir=args.muse_dir, n_jobs=args.n_jobs,
+                               zero_shot=zero_shot, train_langs=zscl_train_langs)  # Todo: testing zero shot
         embedder_list.append(museEmbedder)
 
     if args.wce_embedder:
@@ -82,6 +74,8 @@ def main(args):
     # Testing ----------------------------------------
     print('\n[Testing Generalized Funnelling]')
     time_te = time.time()
+    # Zero shot scenario -> setting first tier learners zero_shot param to False
+    gfun.set_zero_shot(val=False)
     ly_ = gfun.predict(lXte)
     l_eval = evaluate(ly_true=lyte, ly_pred=ly_)
     time_te = round(time.time() - time_te, 3)
@@ -111,7 +105,7 @@ def main(args):
                             microf1=microf1,
                             macrok=macrok,
                             microk=microk,
-                            notes=f'Train langs: {sorted(lX.keys())}')
+                            notes=f'Train langs: {sorted(zscl_train_langs)}' if zero_shot else '')
     print('Averages: MF1, mF1, MK, mK', np.round(np.mean(np.array(metrics), axis=0), 3))
 
     overall_time = round(time.time() - time_init, 3)

src/funnelling.py

@@ -48,14 +48,17 @@ class DocEmbedderList:
 
         for embedder in self.embedders:
             lZ = embedder.transform(lX)
-            for lang in langs:
+            for lang in sorted(lZ.keys()):
                 Z = lZ[lang]
                 if lZparts[lang] is None:
                     lZparts[lang] = Z
                 else:
                     lZparts[lang] += Z
         n_embedders = len(self.embedders)
-        return {lang: lZparts[lang]/n_embedders for lang in langs}  # Averaging feature spaces
+        # Zero shot experiments: removing k:v if v is None (i.e, it is a lang that will be used in zero shot setting)
+        lZparts = {k: v for k, v in lZparts.items() if v is not None}
+
+        return {lang: lZparts[lang]/n_embedders for lang in sorted(lZparts.keys())}  # Averaging feature spaces
 
     def fit_transform(self, lX, ly):
         return self.fit(lX, ly).transform(lX)
@@ -122,3 +125,9 @@ class Funnelling:
         lZ = self.first_tier.transform(lX)
         ly = self.meta.predict(lZ)
         return ly
+
+    def set_zero_shot(self, val: bool):
+        for embedder in self.first_tier.embedders:
+            embedder.embedder.set_zero_shot(val)
+        return
+

src/view_generators.py

@@ -93,13 +93,18 @@ class VanillaFunGen(ViewGen):
     def fit_transform(self, lX, ly):
         return self.fit(lX, ly).transform(lX)
 
+    def set_zero_shot(self, val: bool):
+        self.zero_shot = val
+        print('# TODO: PosteriorsGen has not been configured for zero-shot experiments')
+        return
+
 
 class MuseGen(ViewGen):
     """
     View Generator (m): generates document representation via MUSE embeddings (Fasttext multilingual word
     embeddings). Document embeddings are obtained via weighted sum of document's constituent embeddings.
     """
-    def __init__(self, muse_dir='../embeddings', zero_shot=False, n_jobs=-1):
+    def __init__(self, muse_dir='../embeddings', zero_shot=False, train_langs: list = None, n_jobs=-1):
         """
         Init the MuseGen.
         :param muse_dir: string, path to folder containing muse embeddings
@@ -111,7 +116,11 @@ class MuseGen(ViewGen):
         self.langs = None
         self.lMuse = None
         self.vectorizer = TfidfVectorizerMultilingual(sublinear_tf=True, use_idf=True)
+        # Zero shot parameters
         self.zero_shot = zero_shot
+        if train_langs is None:
+            train_langs = ['it']
+        self.train_langs = train_langs
 
     def fit(self, lX, ly):
         """
@@ -138,6 +147,7 @@ class MuseGen(ViewGen):
         """
         # Testing zero-shot experiments
         if self.zero_shot:
+            lX = self.zero_shot_experiments(lX)
             lX = {l: self.vectorizer.vectorizer[l].transform(lX[l]) for l in self.langs if lX[l] is not None}
         else:
             lX = self.vectorizer.transform(lX)
@@ -148,22 +158,23 @@ class MuseGen(ViewGen):
         return lZ
 
     def fit_transform(self, lX, ly):
-        print('## NB: Calling fit_transform!')
-        if self.zero_shot:
-            return self.fit(lX, ly).transform(self.zero_shot_experiments(lX))
         return self.fit(lX, ly).transform(lX)
 
-    def zero_shot_experiments(self, lX, train_langs: list = ['it']):
+    def zero_shot_experiments(self, lX):
-        print(f'# Zero-shot setting! Training langs will be set to: {sorted(train_langs)}')
+        print(f'# Zero-shot setting! Training langs will be set to: {sorted(self.train_langs)}')
         _lX = {}
         for lang in self.langs:
-            if lang in train_langs:
+            if lang in self.train_langs:
                 _lX[lang] = lX[lang]
             else:
                 _lX[lang] = None
         lX = _lX
         return lX
 
+    def set_zero_shot(self, val: bool):
+        self.zero_shot = val
+        return
+
 
 class WordClassGen(ViewGen):
     """
@@ -214,6 +225,11 @@ class WordClassGen(ViewGen):
     def fit_transform(self, lX, ly):
         return self.fit(lX, ly).transform(lX)
 
+    def set_zero_shot(self, val: bool):
+        self.zero_shot = val
+        print('# TODO: WordClassGen has not been configured for zero-shot experiments')
+        return
+
 
 class RecurrentGen(ViewGen):
     """
@@ -335,6 +351,11 @@ class RecurrentGen(ViewGen):
     def fit_transform(self, lX, ly):
         return self.fit(lX, ly).transform(lX)
 
+    def set_zero_shot(self, val: bool):
+        self.zero_shot = val
+        print('# TODO: RecurrentGen has not been configured for zero-shot experiments')
+        return
+
 
 class BertGen(ViewGen):
     """
@@ -405,3 +426,8 @@ class BertGen(ViewGen):
     def fit_transform(self, lX, ly):
         # we can assume that we have already indexed data for transform() since we are first calling fit()
         return self.fit(lX, ly).transform(lX)
+
+    def set_zero_shot(self, val: bool):
+        self.zero_shot = val
+        print('# TODO: BertGen has not been configured for zero-shot experiments')
+        return