Speed up startup, massively decrease ram usage

By only instanciating the `counterfitted_GLOVE_embedding` when
necessary, the startup time gets cut by two thirds, while the ram usage
decreases by at least two gigabytes.

textattack/constraints/semantics/sentence_encoders/thought_vector.py

@@ -20,8 +20,10 @@ class ThoughtVector(SentenceEncoder):
     """
 
     def __init__(
-        self, embedding=WordEmbedding.counterfitted_GLOVE_embedding(), **kwargs
+        self, embedding=None, **kwargs
     ):
+        if embedding is None:
+            embedding = WordEmbedding.counterfitted_GLOVE_embedding()
         if not isinstance(embedding, AbstractWordEmbedding):
             raise ValueError(
                 "`embedding` object must be of type `textattack.shared.AbstractWordEmbedding`."

textattack/constraints/semantics/word_embedding_distance.py

@@ -24,14 +24,16 @@ class WordEmbeddingDistance(Constraint):
 
     def __init__(
         self,
-        embedding=WordEmbedding.counterfitted_GLOVE_embedding(),
+        embedding=None,
         include_unknown_words=True,
         min_cos_sim=None,
         max_mse_dist=None,
         cased=False,
         compare_against_original=True,
     ):
         super().__init__(compare_against_original)
+        if embedding is None:
+            embedding = WordEmbedding.counterfitted_GLOVE_embedding()
         self.include_unknown_words = include_unknown_words
         self.cased = cased
 

textattack/transformations/word_swaps/word_swap_embedding.py

@@ -31,10 +31,12 @@ class WordSwapEmbedding(WordSwap):
     def __init__(
         self,
         max_candidates=15,
-        embedding=WordEmbedding.counterfitted_GLOVE_embedding(),
+        embedding=None,
         **kwargs
     ):
         super().__init__(**kwargs)
+        if embedding is None:
+            embedding = WordEmbedding.counterfitted_GLOVE_embedding()
         self.max_candidates = max_candidates
         if not isinstance(embedding, AbstractWordEmbedding):
             raise ValueError(