.gitignore

@@ -1,6 +1,6 @@
 **/data/*
+**/*.zip
 **/__pycache__
 /env
 **/runs/*
-**/wandb/*
+**/wandb/*
-**/models/*

eric_wordle/ai.py

@@ -5,7 +5,7 @@ import numpy as np
 
 from stable_baselines3 import PPO, DQN
 from letter_guess import LetterGuessingEnv
-import torch
+
 
 def load_valid_words(file_path='wordle_words.txt'):
     """
@@ -23,8 +23,7 @@ def load_valid_words(file_path='wordle_words.txt'):
 
 
 class AI:
-    def __init__(self, vocab_file, model_file, num_letters=5, num_guesses=6, use_q_model=False, device="cuda"):
+    def __init__(self, vocab_file, model_file, num_letters=5, num_guesses=6, use_q_model=False):
-        self.device = device
         self.vocab_file = vocab_file
         self.num_letters = num_letters
         self.num_guesses = 6
@@ -38,28 +37,26 @@ class AI:
         self.use_q_model = use_q_model
         if use_q_model:
             # we initialize the same q env as the model train ONLY to simplify storing/calculating the gym state, not used to control the game at all
-            self.q_env = LetterGuessingEnv(load_valid_words(vocab_file))
+            self.q_env = LetterGuessingEnv(vocab_file)
             self.q_env_state, _ = self.q_env.reset()
 
             # load model
-            self.q_model = PPO.load(model_file, device=self.device)
+            self.q_model = PPO.load(model_file)
 
-        self.reset("")
+        self.reset()
 
     def solve_eval(self, results_callback):
         num_guesses = 0
         while [len(e) for e in self.domains] != [1 for _ in range(self.num_letters)]:
             num_guesses += 1
-            if self.use_q_model:
-                self.freeze_state = self.q_env.clone_state()
 
             # sample a word, this would use the q_env_state if the q_model is used
-            word = self.sample(num_guesses)
+            word = self.sample()
 
             # get emulated results
             results = results_callback(word)
             if self.use_q_model:
-                self.q_env.set_state(self.freeze_state)
+                self.q_env.set_state(self.q_env_state)
                 # step the q_env to match the guess we just made
                 for i in range(len(word)):
                     char = word[i]
@@ -73,11 +70,13 @@ class AI:
         num_guesses = 0
         while [len(e) for e in self.domains] != [1 for _ in range(self.num_letters)]:
             num_guesses += 1
-            if self.use_q_model:
+            word = self.sample()
-                self.freeze_state = self.q_env.clone_state()
 
-            # sample a word, this would use the q_env_state if the q_model is used
+            # # Always start with these two words
-            word = self.sample(num_guesses)
+            # if num_guesses == 1:
+            #     word = 'soare'
+            # elif num_guesses == 2:
+            #     word = 'culti'
 
             print('-----------------------------------------------')
             print(f'Guess #{num_guesses}/{self.num_guesses}: {word}')
@@ -97,16 +96,10 @@ class AI:
                     results.append(result)
                     break
 
-            if self.use_q_model:
-                self.q_env.set_state(self.freeze_state)
-                # step the q_env to match the guess we just made
-                for i in range(len(word)):
-                    char = word[i]
-                    action = ord(char) - ord('a')
-                    self.q_env_state, _, _, _, _ = self.q_env.step(action)
-
             self.arc_consistency(word, results)
-        return num_guesses, word
+
+        print(f'You did it! The word is {"".join([e[0] for e in self.domains])}')
+        return num_guesses
 
     def arc_consistency(self, word, results):
         self.possible_letters += [word[i] for i in range(len(word)) if results[i] == '1']
@@ -126,15 +119,14 @@ class AI:
             if results[i] == '2':
                 self.domains[i] = [word[i]]
 
-    def reset(self, target_word):
+    def reset(self):
         self.domains = [list(string.ascii_lowercase) for _ in range(self.num_letters)]
         self.possible_letters = []
 
         if self.use_q_model:
             self.q_env_state, _ = self.q_env.reset()
-            self.q_env.target_word = target_word
 
-    def sample(self, num_guesses):
+    def sample(self):
         """
         Samples a best word given the current domains
         :return:
@@ -151,15 +143,15 @@ class AI:
         for word, _ in self.best_words:
             # reset the state back to before we guessed a word
             if pattern.match(word) and False not in [e in word for e in self.possible_letters]:
-                if self.use_q_model and num_guesses == 3:
+                if self.use_q_model:
-                    self.q_env.set_state(self.freeze_state)
+                    self.q_env.set_state(self.q_env_state)
                     # Use policy to grade word
                     # get the state and action pairs
                     curr_qval = 0
 
                     for l in word:
                         action = ord(l) - ord('a')
-                        q_val, _, _ = self.q_model.policy.evaluate_actions(self.q_model.policy.obs_to_tensor(self.q_env.get_obs())[0], torch.Tensor(np.array([action])).to(self.device))
+                        q_val = self.q_model.policy.evaluate_actions(self.q_env.get_obs(), action)
                         _, _, _, _, _ = self.q_env.step(action)
                         curr_qval += q_val
 

eric_wordle/eval.py

@@ -28,26 +28,22 @@ def main(args):
     if args.n is None:
         raise Exception('Need to specify n (i.e. n = 1 for wordle, n = 4 for quordle, n = 16 for sedecordle).')
 
-    ai = AI(args.vocab_file, args.model_file, use_q_model=args.q_model, device=args.device)
+    ai = AI(args.vocab_file, args.model_file, use_q_model=args.q_model)
 
     total_guesses = 0
     wins = 0
     num_eval = args.num_eval
 
-    np.random.seed(0)
-
     for i in tqdm(range(num_eval)):
         idx = np.random.choice(range(len(ai.vocab)))
         solution = ai.vocab[idx]
-
-        ai.reset(solution)
-
         guesses, word = ai.solve_eval(results_callback=result_callback)
         if word != solution:
             total_guesses += 5
         else:
             total_guesses += guesses
             wins += 1
+        ai.reset()
 
     print(f"q_model?: {args.q_model} \t average guesses per game: {total_guesses / num_eval} \t win rate: {wins / num_eval}")
 
@@ -58,6 +54,5 @@ if __name__ == '__main__':
     parser.add_argument('--num_eval', dest="num_eval", type=int, default=1000)
     parser.add_argument('--model_file', dest="model_file", type=str, default='wordle_ppo_model')
     parser.add_argument('--q_model', dest="q_model", type=bool, default=False)
-    parser.add_argument('--device', dest="device", type=str, default="cuda")
     args = parser.parse_args()
     main(args)

eric_wordle/main.py

@@ -5,9 +5,8 @@ from ai import AI
 def main(args):
     if args.n is None:
         raise Exception('Need to specify n (i.e. n = 1 for wordle, n = 4 for quordle, n = 16 for sedecordle).')
-    print(f"using q model? {args.q_model}")
+
-    ai = AI(args.vocab_file, args.model_file, use_q_model=args.q_model, device=args.device)
+    ai = AI(args.vocab_file)
-    ai.reset("lingo")
     ai.solve()
 
 
@@ -15,8 +14,5 @@ if __name__ == '__main__':
     parser = argparse.ArgumentParser()
     parser.add_argument('--n', dest='n', type=int, default=None)
     parser.add_argument('--vocab_file', dest='vocab_file', type=str, default='wordle_words.txt')
-    parser.add_argument('--model_file', dest="model_file", type=str, default='wordle_ppo_model')
-    parser.add_argument('--q_model', dest="q_model", type=bool, default=False)
-    parser.add_argument('--device', dest="device", type=str, default="cuda")
     args = parser.parse_args()
     main(args)

inference.sh

@@ -1 +0,0 @@
-python eric_wordle/main.py --n 1 --vocab_file wordle_words.txt  --q_model True --model_file wordle_ppo_model --device cpu