Assignment #1 and #2 and #3

discern/discern_tabular.py

@@ -4,6 +4,8 @@
 from discern import util
 from discern.discern_base import DisCERN
 from sklearn.preprocessing import MinMaxScaler
+import torch
+from transformers import T5Tokenizer, T5ForConditionalGeneration
 
 class DisCERNTabular(DisCERN):
     """
@@ -102,9 +104,19 @@ def find_cf(self, test_instance, test_label, desired_class='opposite', **kwargs)
                 if abs(val_x - val_nun) <= self.threshold:
                     None
                 else:
-                    x_adapted[indices[now_index]] = nun_data[indices[now_index]]
-                    changes +=1
-                    amounts += abs(val_x - val_nun)
+                    if self.feature_names[indices[now_index]]=="age" and nun_data[indices[now_index]] > x_adapted[indices[now_index]]:
+                        None
+                    elif self.feature_names[indices[now_index]] in [' Black', ' Other', ' White']:
+                        None
+                    elif self.feature_names[indices[now_index]]==' Self-emp-inc' and nun_data[indices[now_index]]==" Local-gov":
+                        None
+                    elif self.feature_names[indices[now_index]]==' Self-emp-inc' and nun_data[indices[now_index]]==" State-gov":
+                        None
+                    else:
+                        x_adapted[indices[now_index]] = nun_data[indices[now_index]]
+                        changes +=1
+                        amounts += abs(val_x - val_nun)
+
             new_class = self.model.predict([x_adapted])[0]
             # print('new_class: '+str(new_class))
             now_index += 1
@@ -120,5 +132,48 @@ def find_cf(self, test_instance, test_label, desired_class='opposite', **kwargs)
                 break
         return x_adapted, sparsity, proximity
 
-    def show_cf(self, test_instance, test_label, cf, cf_label, **kwargs):
-        None
+    def show_cf(self, test_instance, cf, **kwargs):
+
+        PATH = "../discern/NLG_model/model.pt"
+        if torch.cuda.is_available():
+            dev = torch.device("cuda:0") 
+        else:
+            dev = torch.device("cpu")
+
+        tokenizer = T5Tokenizer.from_pretrained('t5-base')
+        model = T5ForConditionalGeneration.from_pretrained('t5-base', return_dict=True)
+        model.eval()
+        model.to(dev)
+        model.load_state_dict(torch.load(PATH))
+
+        l_test = []
+        for i in test_instance:
+            l_test.append(str(i))
+            l_test.append(str(test_instance[i]))
+        test_instance = '|'.join(l_test)
+        input_ids = tokenizer.encode(test_instance, return_tensors="pt")
+        input_ids=input_ids.to(dev)
+        outputs = model.generate(input_ids,
+            do_sample=True, 
+            max_length=50, 
+            top_k=50, 
+            top_p=0.95)
+        out_test = tokenizer.decode(outputs[0])
+
+        l_cf = []
+        for i in cf:
+            l_cf.append(str(i))
+            l_cf.append(str(cf[i]))
+        cf = '|'.join(l_cf)
+        input_ids = tokenizer.encode(cf, return_tensors="pt")
+        input_ids=input_ids.to(dev)
+        outputs = model.generate(input_ids,
+            do_sample=True, 
+            max_length=50, 
+            top_k=50, 
+            top_p=0.95)
+        out_cf = tokenizer.decode(outputs[0])
+
+        print(f"Instance: {out_test}")
+        print(f"Counterfactual: {out_cf}")
+

requirements.txt

@@ -1,6 +1,7 @@
 numpy
 pandas
 scikit-learn
-heapq
 tqdm
 lime
+transformers==4.11.2
+sentencepiece==0.1.96