Update lichess_sf.py

lichess_sf.py

@@ -1,4 +1,3 @@
-#%%
 import datasets
 import zstandard as zstd
 import io
@@ -40,11 +39,10 @@ class Lichess(datasets.GeneratorBasedBuilder):
         return info
 
     def _get_filepaths(self):
-        #months = ["01", "02", "03", "04", "05", "06",
-        #          "07", "08", "09", "10", "11", "12"]
-        months = ["01", "02", "03", "04", "08", "09", "10", "11", "12"]
+        months = ["01", "02", "03", "04", "05", "06",
+                  "07", "08", "09", "10", "11", "12"]
         shards = ["0", "1", "2", "3"]
-        filepaths = [ self.config.name + "/2023/" + m for m in months]
+        filepaths = ["lichess/" + self.config.name + "/2023/" + m for m in months]
 
         if self.config.name == "pgn":
             paths = []
@@ -61,48 +59,63 @@ class Lichess(datasets.GeneratorBasedBuilder):
     def _split_generators(self, dl_manager: datasets.DownloadManager):
         filepaths = self._get_filepaths()
         downloaded_files = dl_manager.download(filepaths)
+        # Hack.
+        # Hardcoded for 48 shards and 4 num_workers:
+        files = [downloaded_files[12*i:12*i+12] for i in range(4)]
         generator = datasets.SplitGenerator(name=datasets.Split.TRAIN,
-                                            gen_kwargs={'filepaths': downloaded_files})
+                                            gen_kwargs={'filepaths': files})
         return [generator]
-
+    
     def _generate_examples(self, filepaths):
-        dctx = zstd.ZstdDecompressor()
-        for filepath in filepaths:
-            n = 0
-            with open(filepath, "rb") as file:
-                with dctx.stream_reader(file) as sr:
-                    pgn = io.TextIOWrapper(sr)
+        """ Each worker receives a random set of the .zst files (the raw dataset).
+            Each worker will cycle through its set of files. They read a single game
+            from file 1, then a single game from file 2, etc. ...
+            The purpose is to create batches that contain games from a diverse mix
+            of time periods. -> Reduces distribution shift.
+        """
+        filepaths = filepaths[0]
+        files =          [open(filepath, "rb")     for filepath in filepaths]
+        dctxs =          [zstd.ZstdDecompressor()  for file in files]
+        stream_readers = [dctx.stream_reader(file) for dctx, file in zip(dctxs, files)]
+        pgns =           [io.TextIOWrapper(sr)     for sr in stream_readers]
 
-                    if self.config.name == "pgn":
-                        while True:
-                            elos = pgn.readline()
-                            game_pgn = pgn.readline()
-                            next_line = pgn.readline()
+        n = 0
+        n_files = len(files)
+        # lower bound of number of positions per .zst file to stop the loop.
+        # need something smarter at some point
+        n_positions = 2 * 10**6
+        while n <= n_files * n_positions:
+            # cycle through the different shards
+            pgn = pgns[n % n_files]
 
-                            if game_pgn:
-                                white_elo, black_elo = elos.split(" ")
-                                _id = n
-                                n += 1
-                                yield _id, {"WhiteElo": int(white_elo),
-                                            "BlackElo": int(black_elo),
-                                            "pgn": game_pgn}
-                            else:
-                                break
-                    
-                    elif self.config.name == "fen":
-                        while True:
-                            elos = pgn.readline()
-                            game = pgn.readline()
+            if self.config.name == "pgn":
+                elos = pgn.readline()
+                game_pgn = pgn.readline()
+                next_line = pgn.readline()
+
+                if game_pgn:
+                    white_elo, black_elo = elos.split(" ")
+                    _id = n
+                    n += 1
+                    yield _id, {"WhiteElo": int(white_elo),
+                                "BlackElo": int(black_elo),
+                                "pgn": game_pgn}
+                else:
+                    break
+        
+            elif self.config.name == "fen":
+                elos = pgn.readline()
+                game = pgn.readline()
 
-                            if game:
-                                white_elo, black_elo = elos.split(" ")
-                                fens, moves, scores = game.split(";")
-                                _id = n
-                                n += 1
-                                yield _id, {"WhiteElo": white_elo,
-                                            "BlackElo": black_elo,
-                                            "fens": fens,
-                                            "moves": moves,
-                                            "scores": scores}
-                            else:
-                                break
+                if game:
+                    white_elo, black_elo = elos.split(" ")
+                    fens, moves, scores = game.split(";")
+                    _id = n
+                    n += 1
+                    yield _id, {"WhiteElo": white_elo,
+                                "BlackElo": black_elo,
+                                "fens": fens,
+                                "moves": moves,
+                                "scores": scores}
+                else:
+                    break