Merge pull request #203 from thunlp/dev

Dev

OpenAttack/attackers/fd.py

@@ -70,7 +70,7 @@ def __call__(self, clsf, x_orig, target=None):
                 if iter_cnt > 5 * len(sent):    # Failed to find a substitute word
                     return None
                 try:
-                    reps = list(map(lambda x:x[0], self.substitute(sent[idx], pos=None, threshold=self.config["threshold"])))
+                    reps = list(map(lambda x:x[0], self.substitute(sent[idx], None, threshold=self.config["threshold"])))
                 except WordNotInDictionaryException:
                     continue
                 reps = list(filter(lambda x: x in self.wordid, reps))

OpenAttack/attackers/hotflip.py

@@ -109,7 +109,7 @@ def get_neighbours(self, word, POS, num):
             sub_words = list(
                 map(
                     lambda x: x[0],
-                    self.config["substitute"](word, pos=POS, threshold=threshold)[:num],
+                    self.config["substitute"](word, POS, threshold=threshold)[:num],
                 )
             )
             neighbours = []

OpenAttack/attackers/textbugger.py

@@ -157,7 +157,7 @@ def generateBugs(self, word, glove_vectors, sub_w_enabled=False, typo_enabled=Fa
 
     def bug_sub_W(self, word):
         try:
-            res = self.counterfit.__call__(word, pos=None, threshold=0.5)
+            res = self.counterfit(word, None, threshold=0.5)
             if len(res) == 0:
                 return word
             return res[0][0]

OpenAttack/attackers/textfooler.py

@@ -43,6 +43,7 @@
     "import_score_threshold": -1.,
     "sim_score_threshold": 0.5, 
     "sim_score_window": 15, 
+    "threshold": 0.5,
     "synonym_num": 50,
     "processor": DefaultTextProcessor(),
     "substitute": None,
@@ -190,12 +191,12 @@ def __call__(self, clsf, x_orig, target=None):
 
 
     def get_neighbours(self, word, pos):
-        threshold = 0.5
+        threshold = self.config["threshold"]
         try:
             return list(
                 map(
                     lambda x: x[0],
-                    self.config["substitute"](word, pos=pos, threshold=threshold)[1 : self.config["synonym_num"] + 1],
+                    self.config["substitute"](word, pos, threshold=threshold)[1 : self.config["synonym_num"] + 1],
                 )
             )
         except WordNotInDictionaryException:

OpenAttack/substitutes/chinese_hownet.py

@@ -21,14 +21,16 @@ def __init__(self):
         self.hownet_dict = DataManager.load("AttackAssist.HowNet")
         self.zh_word_list = self.hownet_dict.get_zh_words()
 
-    def __call__(self, word, pos_tag, threshold=None):
-        if pos_tag[:2] == "JJ":
+    def __call__(self, word, pos, threshold=None):
+        if pos is  None:
+            pp = None
+        elif pos[:2] == "JJ":
             pp = "adj"
-        elif pos_tag[:2] == "VB":
+        elif pos[:2] == "VB":
             pp = "verb"
-        elif pos_tag[:2] == "NN":
+        elif pos[:2] == "NN":
             pp = "noun"
-        elif pos_tag[:2] == "RB":
+        elif pos[:2] == "RB":
             pp = "adv"
         else:
             pp = None

OpenAttack/substitutes/chinese_wordnet.py

@@ -27,7 +27,9 @@ def __init__(self):
 
     def __call__(self, word, pos_tag, threshold=None):
         from nltk.corpus import wordnet as wn
-        if pos_tag[:2] == "JJ":
+        if pos_tag is None:
+            pp = None
+        elif pos_tag[:2] == "JJ":
             pp = "adj"
         elif pos_tag[:2] == "VB":
             pp = "verb"

OpenAttack/substitutes/hownet.py

@@ -25,7 +25,9 @@ def __init__(self):
 
     def __call__(self, word, pos_tag, threshold=None):
         pp = "noun"
-        if pos_tag[:2] == "JJ":
+        if pos_tag is None:
+            pp = None
+        elif pos_tag[:2] == "JJ":
             pp = "adj"
         elif pos_tag[:2] == "VB":
             pp = "verb"

examples/workflow.py

@@ -17,7 +17,7 @@ def main():
     # The load operation returns a PytorchClassifier that can be further used for Attacker and AttackEval.
 
     dataset = datasets.load_dataset("sst", split="train[:20]").map(function=dataset_mapping)
-    # Dataset.SST.sample is a list of 1k sentences sampled from test dataset of Dataset.SST.
+    # We load sst dataset using `datasets` package, and map the fields.
 
     attacker = OpenAttack.attackers.GeneticAttacker()
     # After this step, we’ve initialized a GeneticAttacker and uses the default configuration during attack process.

slow_tests/attackers_chinese.py

@@ -11,11 +11,11 @@ def get_attackers_on_chinese(dataset, clsf):
     chinese_substitute = OpenAttack.substitutes.ChineseHowNetSubstitute()
     #Attackers that current support Chinese: SememePSO TextFooler PWWS Genetic FD TextBugger
     attackers = [
-        OpenAttack.attackers.FDAttacker(word2id=clsf.word2id, embedding=clsf.embedding, processor=chinese_processor, substitute=chinese_substitute),
+        OpenAttack.attackers.FDAttacker(word2id=clsf.word2id, embedding=clsf.embedding, token_unk="[UNK]", processor=chinese_processor, substitute=chinese_substitute),
         OpenAttack.attackers.TextBuggerAttacker(processor=chinese_processor),
-        OpenAttack.attackers.TextFoolerAttacker(processor=chinese_processor, substitute=chinese_substitute),
-        OpenAttack.attackers.GeneticAttacker(processor=chinese_processor, substitute=chinese_substitute, skip_words=["的", "了", "着"]),
-        OpenAttack.attackers.PWWSAttacker(processor=chinese_processor, substitute=chinese_substitute),
+        OpenAttack.attackers.TextFoolerAttacker(processor=chinese_processor, substitute=chinese_substitute, threshold=10),
+        OpenAttack.attackers.GeneticAttacker(processor=chinese_processor, substitute=chinese_substitute, skip_words=["的", "了", "着"], neighbour_threshold=10),
+        OpenAttack.attackers.PWWSAttacker(processor=chinese_processor, substitute=chinese_substitute, threshold=10),
         OpenAttack.attackers.PSOAttacker(processor=chinese_processor, substitute=chinese_substitute)
     ]
     return attackers

slow_tests/test_chinese_default.py

@@ -29,7 +29,7 @@ def main():
         try:
             st = time.perf_counter()
             print(
-                OpenAttack.attack_evals.ChineseAttackEval(attacker, time_clsf, progress_bar=True).eval(dataset, visualize=True),
+                OpenAttack.attack_evals.ChineseAttackEval(attacker, time_clsf).eval(dataset, progress_bar=True),
                 time_clsf.total_time,
                 time.perf_counter() - st
             )