Lstm model

.gitignore

@@ -142,3 +142,6 @@ savedmodel/*
 *.slp
 *.npz
 *.tfrecord
+scratch*
+Logs/
+melee/

advantage_bar.py

@@ -10,30 +10,22 @@
 import pathlib
 from model import AdvantageBarModel
 
-parser = argparse.ArgumentParser(description='AI-powered Melee in-game advantage bar')
-parser.add_argument('--train',
-                    '-t',
-                    action='store_true',
-                    help='Training mode')
-parser.add_argument('--evaluate',
-                    '-e',
-                    action='store_true',
-                    help='Evaluation mode')
-parser.add_argument('--predict',
-                    '-p',
-                    help='Prediction mode. Specify the directory where dolphin is')
-parser.add_argument('--build',
-                    '-b',
-                    action='store_true',
-                    help='Build dataset from SLP files')
-parser.add_argument('--max_split',
-                    '-m',
-                    help='Split building dataset into this many pieces')
-parser.add_argument('--split',
-                    '-s',
-                    help='Handle split number S')
+parser = argparse.ArgumentParser(description="AI-powered Melee in-game advantage bar")
+parser.add_argument("--train", "-t", action="store_true", help="Training mode")
+parser.add_argument("--evaluate", "-e", action="store_true", help="Evaluation mode")
+parser.add_argument(
+    "--predict", "-p", help="Prediction mode. Specify the directory where dolphin is"
+)
+parser.add_argument(
+    "--build", "-b", action="store_true", help="Build dataset from SLP files"
+)
+parser.add_argument(
+    "--max_split", "-m", help="Split building dataset into this many pieces"
+)
+parser.add_argument("--split", "-s", help="Handle split number S")
 args = parser.parse_args()
 
+
 def who_died(past_p1, past_p2, current_p1, current_p2):
     """Returns who died."""
     if past_p1 > current_p1:
@@ -42,11 +34,22 @@ def who_died(past_p1, past_p2, current_p1, current_p2):
         return 1
     return -1
 
+
 def _float_feature(value):
-  return tf.train.FeatureList(feature=[tf.train.Feature(float_list=tf.train.FloatList(value=[x])) for x in value])
+    return tf.train.FeatureList(
+        feature=[
+            tf.train.Feature(float_list=tf.train.FloatList(value=[x])) for x in value
+        ]
+    )
+
 
 def _int64_feature(value):
-  return tf.train.FeatureList(feature=[tf.train.Feature(int64_list=tf.train.Int64List(value=[x])) for x in value])
+    return tf.train.FeatureList(
+        feature=[
+            tf.train.Feature(int64_list=tf.train.Int64List(value=[x])) for x in value
+        ]
+    )
+
 
 if args.build:
     """Builds the tfrecord dataset
@@ -67,8 +70,10 @@ def _int64_feature(value):
     if args.max_split and args.split:
         max_split, split = int(args.max_split), int(args.split)
 
-    directory = 'training_data/'
-    num_files = len([f for f in os.listdir(directory)if os.path.isfile(os.path.join(directory, f))])
+    directory = "training_data/"
+    num_files = len(
+        [f for f in os.listdir(directory) if os.path.isfile(os.path.join(directory, f))]
+    )
     bar = progressbar.ProgressBar(maxval=num_files)
     file_index = 0
     for entry in bar(os.scandir(directory)):
@@ -92,12 +97,14 @@ def _int64_feature(value):
             "stage": [],
             "frame": [],
             "stock_winner": [],
-            "game_winner": []
+            "game_winner": [],
         }
         if entry.path.endswith(".slp") and entry.is_file():
             console = None
             try:
-                console = melee.Console(is_dolphin=False, path=entry.path, allow_old_version=True)
+                console = melee.Console(
+                    is_dolphin=False, path=entry.path, allow_old_version=True
+                )
             except Exception as ex:
                 print("Got error, skipping file", ex)
                 continue
@@ -106,7 +113,7 @@ def _int64_feature(value):
             except ubjson.decoder.DecoderException as ex:
                 print("Got error, skipping file", ex)
                 continue
-            stocks = (4,4)
+            stocks = (4, 4)
             ports = None
             # Pick a game to be part of the evaluation set 20% of the time
             is_evaluation = random.random() > 0.8
@@ -149,13 +156,19 @@ def _int64_feature(value):
                         frames["player2_stock"].append(player_two.stock)
                         frames["player2_action"].append(player_two.action.value)
                         frames["player2_character"].append(player_one.character.value)
-                        frames["stage"].append(AdvantageBarModel.stage_flatten(gamestate.stage.value))
+                        frames["stage"].append(
+                            AdvantageBarModel.stage_flatten(gamestate.stage.value)
+                        )
                         frames["frame"].append(gamestate.frame)
 
                         # Did someone lose a stock? Add the labels on
-                        died = who_died(stocks[0], stocks[1], player_one.stock, player_two.stock)
+                        died = who_died(
+                            stocks[0], stocks[1], player_one.stock, player_two.stock
+                        )
                         if died > -1:
-                            frames_needed = len(frames["frame"]) - len(frames["stock_winner"])
+                            frames_needed = len(frames["frame"]) - len(
+                                frames["stock_winner"]
+                            )
                             frames["stock_winner"].extend([died] * frames_needed)
                         stocks = (player_one.stock, player_two.stock)
 
@@ -167,37 +180,66 @@ def _int64_feature(value):
 
             filename = None
             if is_evaluation:
-                filename = "tfrecords/eval" / pathlib.Path(pathlib.Path(entry.path + ".tfrecord").name)
+                filename = "tfrecords/eval" / pathlib.Path(
+                    pathlib.Path(entry.path + ".tfrecord").name
+                )
             else:
-                filename = "tfrecords/train" /  pathlib.Path(pathlib.Path(entry.path + ".tfrecord").name)
+                filename = "tfrecords/train" / pathlib.Path(
+                    pathlib.Path(entry.path + ".tfrecord").name
+                )
             if len(frames) > 0 and game_winner > -1:
                 with tf.io.TFRecordWriter(str(filename)) as file_writer:
                     # This is all that we actually have full data for
                     data_cap = len(frames["stock_winner"])
 
                     # "Context" features are static for the whole data record. Not in the time series
-                    context_features = tf.train.Features(feature={
-                        "game_winner": tf.train.Feature(float_list=tf.train.FloatList(value=[game_winner]))
-                    })
+                    context_features = tf.train.Features(
+                        feature={
+                            "game_winner": tf.train.Feature(
+                                float_list=tf.train.FloatList(value=[game_winner])
+                            )
+                        }
+                    )
 
                     features = {
-                        "player1_character": _int64_feature(frames["player1_character"][:data_cap]),
+                        "player1_character": _int64_feature(
+                            frames["player1_character"][:data_cap]
+                        ),
                         "player1_x": _float_feature(frames["player1_x"][:data_cap]),
                         "player1_y": _float_feature(frames["player1_y"][:data_cap]),
-                        "player1_percent": _float_feature(frames["player1_percent"][:data_cap]),
-                        "player1_stock": _float_feature(frames["player1_stock"][:data_cap]),
-                        "player1_action": _int64_feature(frames["player1_action"][:data_cap]),
+                        "player1_percent": _float_feature(
+                            frames["player1_percent"][:data_cap]
+                        ),
+                        "player1_stock": _float_feature(
+                            frames["player1_stock"][:data_cap]
+                        ),
+                        "player1_action": _int64_feature(
+                            frames["player1_action"][:data_cap]
+                        ),
                         "player2_x": _float_feature(frames["player2_x"][:data_cap]),
                         "player2_y": _float_feature(frames["player2_y"][:data_cap]),
-                        "player2_percent": _float_feature(frames["player2_percent"][:data_cap]),
-                        "player2_stock": _float_feature(frames["player2_stock"][:data_cap]),
-                        "player2_action": _int64_feature(frames["player2_action"][:data_cap]),
-                        "player2_character": _int64_feature(frames["player2_character"][:data_cap]),
+                        "player2_percent": _float_feature(
+                            frames["player2_percent"][:data_cap]
+                        ),
+                        "player2_stock": _float_feature(
+                            frames["player2_stock"][:data_cap]
+                        ),
+                        "player2_action": _int64_feature(
+                            frames["player2_action"][:data_cap]
+                        ),
+                        "player2_character": _int64_feature(
+                            frames["player2_character"][:data_cap]
+                        ),
                         "stage": _int64_feature(frames["stage"][:data_cap]),
                         "frame": _float_feature(frames["frame"][:data_cap]),
-                        "stock_winner": _float_feature(frames["stock_winner"][:data_cap]),
+                        "stock_winner": _float_feature(
+                            frames["stock_winner"][:data_cap]
+                        ),
                     }
-                    newexample = tf.train.SequenceExample(feature_lists=tf.train.FeatureLists(feature_list=features), context=context_features)
+                    newexample = tf.train.SequenceExample(
+                        feature_lists=tf.train.FeatureLists(feature_list=features),
+                        context=context_features,
+                    )
                     file_writer.write(newexample.SerializeToString())
 
 if args.train:
@@ -213,8 +255,7 @@ def _int64_feature(value):
     model.load()
 
     # Start a real game
-    console = melee.Console(path=args.predict,
-                            is_dolphin=False)
+    console = melee.Console(path=args.predict, is_dolphin=False)
     # Run the console
     console.run()
 

lstm notes.txt

@@ -0,0 +1,11 @@
+If an entire game is used to predict the winner, it will no doubt learn that the person 
+closest to losing their last stock determines the winner so if a data point is a single
+game, it will bias its learning to the end of the game and get 100% accuracy, but in
+real time it will not have access to that.
+
+So for each game, it should be broken into several records.  Probably 5 seconds at a 
+time, 2 frames per second, but also right padded so that almost the entire game 
+history can be taken into account.  
+
+500 time steps per record, and using 2 frames per second, the time length for input is
+just over 4 minutes.  Matches that go longer will truncate from the beginning of the match.