added attackmodel class

wrapper class for loading model and attacking on target dataset

environment.yml

@@ -8,4 +8,6 @@ dependencies:
     - flair>=0.11
     - pycld2
     - torchinfo
-    - textattack[transformers]
+    - textattack[transformers]
+    - tensorflow_hub
+    - tensorflow

src/models/attack_model.py

@@ -0,0 +1,90 @@
+import torch
+from textattack import AttackArgs, Attacker
+from textattack.attack_recipes import TextFoolerJin2019
+from textattack.datasets import HuggingFaceDataset
+from textattack.models.wrappers import HuggingFaceModelWrapper
+from transformers import AutoModelForSequenceClassification, AutoTokenizer
+
+from common.constants import MDL_TWIT_POLARITY  # huggingface model paths
+from common.constants import MDL_IMDB_POLARITY, MDL_RT_POLARITY, MDL_TWIT_IRONY
+import time
+
+
+class AttackModel:
+    """
+    Loads attack models from a given huggingface model path into a wrapper class.
+    and preps the model and dataset for attack.
+    """
+    # Models to load as static variables for easy access:
+    IMDB_POLARITY = MDL_IMDB_POLARITY
+    RT_POLARITY = MDL_RT_POLARITY # Rotten Tomatoes
+    TWIT_POLARITY = MDL_TWIT_POLARITY
+    TWIT_IRONY = MDL_TWIT_IRONY
+
+    def __init__(self, model_path=MDL_IMDB_POLARITY, use_cuda=False, 
+                 attack_recipe=TextFoolerJin2019, target_dataset='rotten_tomatoes'):
+        self.model_path = model_path
+
+        # extracting model and tokenizer from model path::
+        self.model_tokenizer = AutoTokenizer.from_pretrained(self.model_path)
+        self.model = AutoModelForSequenceClassification.from_pretrained(self.model_path)
+
+        # Move model to GPU if available:
+        self.device = torch.device("cuda") if use_cuda else torch.device("cpu")
+        self.model.to(self.device)
+
+        # wrapping model for textattack
+        self.model_wrapped = HuggingFaceModelWrapper(self.model, self.model_tokenizer)
+
+        # building attack and getting dataset
+        self.attack = attack_recipe.build(self.model_wrapped)
+        self.target_dataset = self.set_target_dataset(target_dataset)
+
+    def set_target_dataset(self, new_target: str): # changes the target dataset
+        self.target_dataset = HuggingFaceDataset( # dataset that is targeted by the attack
+            name_or_dataset=new_target,
+            subset=None,
+            split='test',
+            shuffle=False)
+        return self.target_dataset
+
+    def generate_target_examples(self, num_examples=10, log=False, 
+                     disable_stdout=True, silent=True, **kwargs):
+        """
+        This initiates the attack on the target domain by generating adversarial examples 
+        using our attack model.
+
+        Args:
+            num_examples (`int`, optional): Number of examples to generate. Defaults to 10.
+            log (`bool`, optional): Logs the examples to a file if True. Defaults to False.
+            disable_stdout (`bool`, optional): Disable displaying individual attack results to stdout. Defaults to True.
+            silent (`bool`, optional): Disable all logging (except for errors). Defaults to True.
+
+        Returns:
+            List[AttackResults]: returns a list of textattack.AttackResults containing the original and perturbed text as well as outputs
+        """
+
+        if log:
+            log_to_csv = "attacks/{}-{}-{}.csv".format(self.model_path.split("/")[-1], 
+                                                       self.target_dataset._name,
+                                                       time.strftime("%m%d_%H%M"))
+        else:
+            log_to_csv = None
+
+
+        attack_args = AttackArgs(
+            num_examples=num_examples,
+            log_to_csv=log_to_csv, 
+            disable_stdout=disable_stdout, 
+            silent=silent,
+            **kwargs
+        )
+
+        attack = Attacker(self.attack, self.target_dataset, attack_args)
+        return attack.attack_dataset()
+
+
+
+
+
+

src/models/model_tester.py

@@ -0,0 +1,5 @@
+class Tester:
+    def __init__(self, model):
+        self.model
+
+