update to chessv6 sl policy model

app.py

@@ -9,40 +9,14 @@ from transformers import DebertaV2ForSequenceClassification, AutoTokenizer, pipe
 
 token = os.environ['auth_token']
 
-tokenizer = AutoTokenizer.from_pretrained('jrahn/chessv3', use_auth_token=token)
-model = DebertaV2ForSequenceClassification.from_pretrained('jrahn/chessv4', use_auth_token=token)
+tokenizer = AutoTokenizer.from_pretrained('jrahn/chessv6', use_auth_token=token)
+model = DebertaV2ForSequenceClassification.from_pretrained('jrahn/chessv6', use_auth_token=token)
 pipe = pipeline(task="text-classification", model=model, tokenizer=tokenizer)
 
-empty_field = '0'
-board_split = ' | '
-nums = {str(n): empty_field * n for n in range(1, 9)}
-nums_rev = {v:k for k,v in reversed(nums.items())}
-
-
-def encode_fen(fen):
-    # decompress fen representation
-    # prepare for sub-word tokenization
-    fen_board, fen_rest = fen.split(' ', 1)
-    for n in nums:
-        fen_board = fen_board.replace(n, nums[n])
-    fen_board = '+' + fen_board
-    fen_board = fen_board.replace('/', ' +')
-    return board_split.join([fen_board, fen_rest])
-
-def decode_fen_repr(fen_repr):
-    fen_board, fen_rest = fen_repr.split(board_split, 1)
-    for n in nums_rev:
-        fen_board = fen_board.replace(n, nums_rev[n])
-    fen_board = fen_board.replace(' +', '/')
-    fen_board = fen_board.replace('+', '')
-    return ' '.join([fen_board, fen_rest])
-
 def predict_move(fen, top_k=3):
-    fen_prep = encode_fen(fen)
-    preds = pipe(fen_prep, top_k=top_k)
+    preds = pipe(fen, top_k=top_k)
     weights = [p['score'] for p in preds]
     p = random.choices(preds, weights=weights)[0]
-    # discard illegal moves (https://python-chess.readthedocs.io/en/latest/core.html#chess.Board.legal_moves), then select top_k
     return p['label']
 
 def btn_load(inp_fen):
@@ -76,9 +50,9 @@ def btn_play(inp_fen, inp_move, inp_notation, inp_k):
 with gr.Blocks() as block:
     gr.Markdown(
     '''
-    # Play YoloChess - Policy Network v0.4
-    110M Parameter Transformer (DeBERTaV2-base architecture)  
-    - pre-trained (MLM) from scratch on chess positions in modified FEN notation
+    # Play YoloChess - Policy Network v0.6
+    87M Parameter Transformer (DeBERTaV2-base architecture)  
+    - pre-trained (MLM) from scratch on chess positions in FEN notation
     - fine-tuned for text classification (moves) on expert games.  
     '''
     )