update multimodal games

games/matchit/instancegenerator.py

@@ -1,8 +1,24 @@
-import sys
+import pandas as pd
 from clemgame.clemgame import GameInstanceGenerator
 
-GAME_NAME = "matchit"
+GAME_NAME: str = "matchit"
+# n instances to be generated
+N: int = 10 # max = len(similar_images.csv) = 161, if not using other image pairs
+# paths to image pair tables
+PATH_DIFF: str = "games/matchit/resources/image_pairs/different_images.csv"
+PATH_SIM: str = "games/matchit/resources/image_pairs/similar_images.csv"
 
+#how many questions can each player ask?
+DEC_TURN: int = 3
+# should the players be informed about the number of questions they can ask?
+INFO_NUM_QUESTIONS: bool = False
+
+SEED: int = 42
+
+# Flags that have to be at the beginning of each response; are also specified in the prompts
+FLAGS: dict[str, str] = {"description": "DESCRIPTION:", "question": "QUESTION:", "answer": "ANSWER:", "decision": "DECISION:"}
+SOL_SAME: str = "same image"
+SOL_DIFF: str = "different images"
 
 class MatchItInstanceGenerator(GameInstanceGenerator):
     def __init__(self, game_name):
@@ -11,27 +27,55 @@ def __init__(self, game_name):
 
     def on_generate(self): 
 
-        prompt_a = self.load_template('resources/initial_prompts/player_a_prompt.template')
-        prompt_b = self.load_template('resources/initial_prompts/player_b_prompt.template')
+        differents = pd.read_csv(PATH_DIFF)
+        diffs = differents.sample(n = N, random_state = SEED)
+
+        similars = pd.read_csv(PATH_SIM)
+        #similars = similars[similars.clipscore >= THRESHOLD_SIM ]
+        sims = similars.sample(n = N, random_state= SEED)[["url1", "url2"]]
+
+        # same images get sampled from the same df as different image, just doubling url1
+        sames = differents[~differents.url1.isin(diffs.url1)]
+        sams = sames.sample(n = N, random_state= SEED)[["url1"]]
+        sams["url2"] = sams[["url1"]]
+
+        initial_prompt = self.load_template('resources/initial_prompts/initial_prompt.template').replace("$FLAG$", FLAGS["description"])
+
+        desc_intro = self.load_template('resources/initial_prompts/description_introduction.template')
 
-        experiments = {"same_image": (self.load_csv("resources/image_pairs/same_image_10_test.csv"), "same image"), "similar_image": (self.load_csv("resources/image_pairs/similar_image_10_test.csv"), "different image"), "different_image": (self.load_csv("resources/image_pairs/different_image_10_test.csv"), "different image")}
+        sentence_num_questions = self.load_template('resources/initial_prompts/info_num_questions.template').replace("$DEC_TURN$", str(DEC_TURN))
 
-        print(experiments)
+        if INFO_NUM_QUESTIONS:
+            initial_prompt = initial_prompt.replace("$NUM_QUESTIONS$", sentence_num_questions)
+        else:
+            initial_prompt = initial_prompt.replace("$NUM_QUESTIONS$", "")
+
+
+        q_reprompt = self.load_template('resources/initial_prompts/q_reprompt.template').replace("$FLAG$", FLAGS["question"])
+        d_reprompt = self.load_template('resources/initial_prompts/d_reprompt.template').replace("$SOL_SAME$", SOL_SAME).replace("$SOL_DIFF$", SOL_DIFF).replace("$FLAG$", FLAGS["decision"])
+        a_request = self.load_template('resources/initial_prompts/a_request.template').replace("$FLAG$", FLAGS["answer"])
+
+
+        experiments = {"same_image": (sams, SOL_SAME), 
+                       "similar_image": (sims, SOL_DIFF), 
+                       "different_image": (diffs, SOL_DIFF)}
 
-        max_turns = 10
 
         for exp_name in experiments.keys(): 
             experiment =  self.add_experiment(exp_name)
             game_id = 0
-            experiment["prompt_a"] = prompt_a  
-            experiment["prompt_b"] = prompt_b
+            experiment["initial_prompt"] = initial_prompt
+            experiment["q_reprompt"] = q_reprompt
+            experiment["d_reprompt"] = d_reprompt
+            experiment["a_request"] = a_request
+            experiment["desc_intro"] = desc_intro
+            experiment["flags"] = FLAGS
             experiment["solution"] = experiments[exp_name][1]
 
 
-            for inst in experiments[exp_name][0]:
-                game_id = game_id
+            for index, row in experiments[exp_name][0].iterrows():
                 instance = self.add_game_instance(experiment, game_id)
-                image_a, image_b = inst[0].strip(), inst[1].strip()
+                image_a, image_b = row["url1"], row["url2"]
                 if image_a.startswith("http"):
                     instance["image_a"] = image_a
                 else:
@@ -41,37 +85,10 @@ def on_generate(self):
                 else:
                     instance["image_b"] = "games/matchit/resources/images/" + image_b
 
-                max_turns = max_turns
+                instance["decision_turn"] = DEC_TURN
 
                 game_id += 1
 
 
-
-
-#{game_id: , image_a, image_b, solution}, max_turns, prompt_a, prompt_b
-
-
-
-        # wir haben idealerweise 3 oder 4 Listen, in denen schon die Bildpaare pro Schwierigkeit festgelegt sind (mit Dateinamen). Später kann man da vielleicht auch noch draus machen, dass diese Listen hier generiert und die Bilder heruntergeladen werden, falls sie noch nicht existieren
-
-        #Ziel: 
-        # json file with key "experiments":[] followed by list of ... experiments:
-            # each experiment has: name, list of instances:   
-                # instances: id, pic1, pic2, 
-
-
-
-            # Schwierigkeit: leicht
-            # game instances:
-                # game id: int
-                # image_player_A: str (.jpg)
-                # image_player_B: str (.jpg)
-                # solution: str oder int oder so (same image/1, diff image/0)
-            # Schwierigkeit: mittel
-                # game_id: int
-                # ... etc.
-            # hart:
-
-
 if __name__ == "__main__":
     MatchItInstanceGenerator(GAME_NAME).generate()