hey

console_play.py

@@ -21,11 +21,12 @@ def console_print(str, pycharm=True):
 
 
 def play_unconstrained():
-    #generator = CTRLGenerator()
-    generator = WebGenerator(CRED_FILE)
-    prompt = get_story_start("classic")
+    generator = CTRLGenerator()
+    #generator = WebGenerator(CRED_FILE)
+    prompt = get_story_start("haunted")
     story_manager = UnconstrainedStoryManager(generator, prompt)
 
+    print("\n")
     console_print(str(story_manager.story))
     while (True):
         action = ""
@@ -38,6 +39,7 @@ def play_unconstrained():
 
 
 def play_constrained():
+    print("\n")
     #generator = WebGenerator(CRED_FILE)
     generator = CTRLGenerator()
     story_start = "haunted"

generator/ctrl/ctrl_generator.py

@@ -10,6 +10,8 @@ from tensorflow.python.ops import math_ops
 from tensorflow.python.ops import embedding_ops
 import fastBPE
 from story.utils import *
+import warnings
+warnings.filterwarnings("ignore")
 
 pos_action_starts = ["You attack", "You tell", "You use", "You go"]
 
@@ -20,7 +22,7 @@ def loss(labels, logits):
 
 class CTRLGenerator():
 
-    def __init__(self, control_code="Horror Text: ", generate_num=64, temperature=0.3):
+    def __init__(self, control_code="Horror Text: ", generate_num=64, temperature=0.5):
 
         self.generate_num=generate_num
         model_dir = "generator/ctrl/model/seqlen256_v1.ckpt/"
@@ -147,9 +149,9 @@ class CTRLGenerator():
 
         prompt = self.control_code + prompt
 
-        # print("******************************")
-        # print(" DEBUG:: Prompt to generate by is \n", prompt)
-        # print("******************************")
+        print("******************************")
+        print(" DEBUG:: Prompt to generate by is \n", prompt)
+        print("******************************")
 
         prompt_length = len(prompt)
 
@@ -162,6 +164,8 @@ class CTRLGenerator():
         padded_text = text + [0] * (total_text_len - len(text))
         tokens_generated = np.tile(padded_text, (1, 1))
         result = ""
+        max_new_lines = 5
+        num_new_lines = 0
         for token in range(len(text) - 1, total_text_len - 1):
 
             if first_token:
@@ -205,14 +209,18 @@ class CTRLGenerator():
                     prompt_logits[_token][generated_token] /= self.penalty
 
             # disallow some tokens
-            forbidden_tokens = ['<unk>', 'Sco@@', '\n']
+            forbidden_tokens = ['<unk>', 'Sco@@']
+
+            if num_new_lines >= max_new_lines:
+                forbidden_tokens.append('\n')
+
             for forbidden_token in forbidden_tokens:
                 prompt_logits[_token][self.word2idx[forbidden_token]] = -1e8
 
             # Make sure only a possible verb is chosen.
             if first_token:
                 for word in get_possible_verbs():
-                    prompt_logits[_token][self.word2idx[word]] += 10
+                    prompt_logits[_token][self.word2idx[word]] += 5
 
             # compute probabilities from logits
             prompt_probs = np.exp(prompt_logits[_token])
@@ -257,9 +265,12 @@ class CTRLGenerator():
             # like which one was chosen,
             # what the top25 were,
             # here is your opportunity.
-            #print('chosen:', repr(self.idx2word[idx]))
+            print('chosen:', repr(self.idx2word[idx]))
             # print('top25 alternatives:', pruned_list[:25])
 
+            if self.idx2word[idx] == "\n":
+                num_new_lines += 1
+
             # assign the token for generation
             tokens_generated[0][token + 1] = idx
 

story/story_manager.py

@@ -66,7 +66,6 @@ class UnconstrainedStoryManager(StoryManager):
         return result
 
     def generate_result(self, action):
-        print("Story context is ", self.story_context())
         block = self.generator.generate(self.story_context() + action)
         block = cut_trailing_sentence(block)
         block = story_replace(block)

story/utils.py

@@ -17,7 +17,7 @@ def get_action_verbs(key):
 
     return data_loaded["action_verbs"][key]
 
-
+# TODO add capital words to remove words
 def remove_profanity(text):
     remove_words = ["fuck", "Fuck", "shit", "rape", "bastard", "bitch"]
     for word in remove_words:
@@ -87,7 +87,8 @@ first_to_second_mappings = [
     ("we ","you "),
     ("We ","You "),
     (" mine"," yours"),
-    (" me", " you"),
+    (" me ", " you "),
+    (" me.", " you."),
     (" us ", " you "),
     (" us.", " you."),
     (" our", " your")