minor cleanup

Author: csinva

imodelsx/embgam/embed.py

@@ -11,7 +11,7 @@ def generate_ngrams_list(
     parsing: str='',
     nlp_chunks=None,
 ):
-    """Get list of ngrams from sentence
+    """Get list of ngrams from sentence using a tokenizer
 
     Params
     ------
@@ -131,19 +131,17 @@ def embed_and_sum_function(
         sentence = example
     # seqs = sentence
 
-    if isinstance(sentence, str):
-        seqs = generate_ngrams_list(
-            sentence, ngrams=ngrams, tokenizer_ngrams=tokenizer_ngrams,
-            parsing=parsing, nlp_chunks=nlp_chunks, all_ngrams=all_ngrams,
-        )
-    elif isinstance(sentence, list):
-        raise Exception('batched mode not supported')
-        # seqs = list(map(generate_ngrams_list, sentence))
+    assert isinstance(sentence, str), 'sentence must be a string (batched mode not supported)'
+    seqs = generate_ngrams_list(
+        sentence, ngrams=ngrams, tokenizer_ngrams=tokenizer_ngrams,
+        parsing=parsing, nlp_chunks=nlp_chunks, all_ngrams=all_ngrams,
+    )
+    # seqs = list(map(generate_ngrams_list, sentence))
+
 
-    # maybe a smarter way to deal with pooling here?
     seq_len = len(seqs)
     if seq_len == 0:
-        seqs = ["dummy"]
+        seqs = ["dummy"] # will multiply embedding by 0 so doesn't matter
 
     if 'bert' in checkpoint.lower():  # has up to two keys, 'last_hidden_state', 'pooler_output'
         if not hasattr(tokenizer_embeddings, 'pad_token') or tokenizer_embeddings.pad_token is None:

imodelsx/embgam/embgam.py

@@ -168,18 +168,23 @@ def cache_linear_coefs(self, X: ArrayLike, model=None, tokenizer_embeddings=None
             print('\tNothing to update!')
             return
 
-        # compute embeddings
-        """
-        # Faster version that needs more memory
-        tokens = tokenizer(ngrams_list, padding=args.padding,
-                           truncation=True, return_tensors="pt")
-        tokens = tokens.to(device)
+        embs = self._get_embs(ngrams_list, model, tokenizer_embeddings)
+        if self.normalize_embs:
+            embs = self.normalizer.transform(embs)
 
-        output = model(**tokens) # this takes a while....
-        embs = output['pooler_output'].cpu().detach().numpy()
-        return embs
+        # save coefs
+        coef_embs = self.linear.coef_.squeeze().transpose()
+        linear_coef = embs @ coef_embs
+        self.coefs_dict_ = {
+            **coefs_dict_old,
+            **{ngrams_list[i]: linear_coef[i]
+               for i in range(len(ngrams_list))}
+        }
+        print('coefs_dict_ len', len(self.coefs_dict_))
+
+    def _get_embs(self, ngrams_list, model, tokenizer_embeddings):
+        """Get embeddings for a list of ngrams (not summed!)
         """
-        # Slower way to run things but won't run out of mem
         embs = []
         for i in tqdm(range(len(ngrams_list))):
             tokens = tokenizer_embeddings(
@@ -191,18 +196,19 @@ def cache_linear_coefs(self, X: ArrayLike, model=None, tokenizer_embeddings=None
                 emb = emb.mean(axis=1)
             embs.append(emb)
         embs = np.array(embs).squeeze()
-        if self.normalize_embs:
-            embs = self.normalizer.transform(embs)
+        return embs
 
-        # save coefs
-        coef_embs = self.linear.coef_.squeeze().transpose()
-        linear_coef = embs @ coef_embs
-        self.coefs_dict_ = {
-            **coefs_dict_old,
-            **{ngrams_list[i]: linear_coef[i]
-               for i in range(len(ngrams_list))}
-        }
-        print('coefs_dict_ len', len(self.coefs_dict_))
+        """
+        # Faster version that needs more memory
+        tokens = tokenizer(ngrams_list, padding=args.padding,
+                           truncation=True, return_tensors="pt")
+        tokens = tokens.to(device)
+
+        output = model(**tokens) # this takes a while....
+        embs = output['pooler_output'].cpu().detach().numpy()
+        return embs
+        """
+        
 
     def _get_ngrams_list(self, X):
         all_ngrams = set()
@@ -251,7 +257,7 @@ def _predict_cached(self, X, warn):
         n_unseen_ngrams = 0
         for x in X:
             pred = 0
-            seqs = imodelsx.embgam.embed.generate_ngrams_list(
+            seqs = imodelsx.embgam.embed.generate_ngraxms_list(
                 x,
                 ngrams=self.ngrams,
                 tokenizer_ngrams=self.tokenizer_ngrams,

setup.py

@@ -21,7 +21,7 @@
 
 setuptools.setup(
     name="imodelsx",
-    version="0.04",
+    version="0.05",
     author="Chandan Singh, John X. Morris",
     author_email="chansingh@microsoft.com",
     description="Library to explain a dataset in natural language.",