Spaces:
Sleeping
Sleeping
File size: 9,489 Bytes
7277ab2 b8887b3 7277ab2 b8887b3 7277ab2 b8887b3 7277ab2 b8887b3 7277ab2 b8887b3 7277ab2 b8887b3 7277ab2 b8887b3 7277ab2 b8887b3 7277ab2 b8887b3 7277ab2 b8887b3 7277ab2 b8887b3 273af2d b8887b3 7277ab2 b8887b3 7277ab2 b8887b3 7277ab2 b8887b3 7277ab2 b8887b3 7277ab2 b8887b3 7277ab2 b8887b3 7277ab2 8dc23f4 7277ab2 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 |
"""
Gradio interface for plotting attention.
"""
import chess
import chess.pgn
import io
import gradio as gr
import os
from lczerolens import LczeroBoard, LczeroModel, Lens, InputEncoding
from demo import constants
from demo.utils import get_info
def get_model(model_name: str):
return LczeroModel.from_onnx_path(os.path.join(constants.ONNX_MODEL_DIRECTORY, model_name))
def get_gradients(model: LczeroModel, board: LczeroBoard, input_encoding: InputEncoding, target: str):
lens = Lens.from_name("gradient")
def init_target(model):
if target == "best_move":
return getattr(model, "output/policy").output.max(dim=1).values
else:
wdl_index = {"win": 0, "draw": 1, "loss": 2}[target]
return getattr(model, "output/wdl").output[:, wdl_index]
results = lens.analyse(model, board, init_target=init_target, model_kwargs={"input_encoding": input_encoding})
return results["input_grad"]
def get_board(game_pgn:str, board_fen:str):
if game_pgn:
try:
board = LczeroBoard()
pgn = io.StringIO(game_pgn)
game = chess.pgn.read_game(pgn)
for move in game.mainline_moves():
board.push(move)
except Exception as e:
print(e)
gr.Warning("Error parsing PGN, using starting position.")
board = LczeroBoard()
else:
try:
board = LczeroBoard(board_fen)
except Exception as e:
print(e)
gr.Warning("Invalid FEN, using starting position.")
board = LczeroBoard()
return board
def render_gradients(board: LczeroBoard, gradients, average_over_planes:bool, begin_average_index:int, end_average_index:int, plane_index:int):
if average_over_planes:
heatmap = gradients[0, begin_average_index:end_average_index].mean(dim=0).view(64)
else:
heatmap = gradients[0, plane_index].view(64)
board.render_heatmap(
heatmap,
save_to=f"{constants.FIGURE_DIRECTORY}/gradients.svg",
)
return f"{constants.FIGURE_DIRECTORY}/gradients_board.svg", f"{constants.FIGURE_DIRECTORY}/gradients_colorbar.svg"
def initial_load(model_name: str, board_fen: str, game_pgn: str, input_encoding: InputEncoding, target: str, average_over_planes:bool, begin_average_index:int, end_average_index:int, plane_index: int):
model = get_model(model_name)
board = get_board(game_pgn, board_fen)
gradients = get_gradients(model, board, input_encoding, target)
info = get_info(model, board)
plots = render_gradients(board, gradients, average_over_planes, begin_average_index, end_average_index, plane_index)
return model, board, gradients, info, *plots
def on_board_change(model: LczeroModel, game_pgn: str, board_fen: str, input_encoding: InputEncoding, target: str, average_over_planes:bool, begin_average_index:int, end_average_index:int, plane_index: int):
board = get_board(game_pgn, board_fen)
gradients = get_gradients(model, board, input_encoding, target)
info = get_info(model, board)
plots = render_gradients(board, gradients, average_over_planes, begin_average_index, end_average_index, plane_index)
return board, gradients, info, *plots
def on_model_change(model_name: str, board: LczeroBoard, input_encoding: InputEncoding, target: str, average_over_planes:bool, begin_average_index:int, end_average_index:int, plane_index: int):
model = get_model(model_name)
gradients = get_gradients(model, board, input_encoding, target)
info = get_info(model, board)
plots = render_gradients(board, gradients, average_over_planes, begin_average_index, end_average_index, plane_index)
return model, gradients, info, *plots
def on_input_encoding_change(model: LczeroModel, board: LczeroBoard, input_encoding: InputEncoding, target: str, average_over_planes:bool, begin_average_index:int, end_average_index:int, plane_index: int):
gradients = get_gradients(model, board, input_encoding, target)
plots = render_gradients(board, gradients, average_over_planes, begin_average_index, end_average_index, plane_index)
return gradients, *plots
def on_target_change(model: LczeroModel, board: LczeroBoard, input_encoding: InputEncoding, target: str, average_over_planes:bool, begin_average_index:int, end_average_index:int, plane_index: int):
gradients = get_gradients(model, board, input_encoding, target)
plots = render_gradients(board, gradients, average_over_planes, begin_average_index, end_average_index, plane_index)
return gradients, *plots
with gr.Blocks() as interface:
with gr.Row():
with gr.Column():
with gr.Group():
gr.Markdown(
"Specify the game PGN or FEN string that you want to analyse (PGN overrides FEN)."
)
game_pgn = gr.Textbox(
label="Game PGN",
lines=1,
value="",
)
board_fen = gr.Textbox(
label="Board FEN",
lines=1,
max_lines=1,
value=chess.STARTING_FEN,
)
input_encoding = gr.Radio(
label="Input encoding",
choices=[
("classical", InputEncoding.INPUT_CLASSICAL_112_PLANE),
("repeated", InputEncoding.INPUT_CLASSICAL_112_PLANE_REPEATED),
("no history repeated", InputEncoding.INPUT_CLASSICAL_112_PLANE_NO_HISTORY_REPEATED),
("no history zeros", InputEncoding.INPUT_CLASSICAL_112_PLANE_NO_HISTORY_ZEROS)
],
value=InputEncoding.INPUT_CLASSICAL_112_PLANE,
)
model_name = gr.Dropdown(
label="Model",
choices=constants.ONNX_MODEL_NAMES,
)
with gr.Group():
info = gr.Textbox(label="Info", lines=1, value="")
with gr.Group():
target = gr.Radio(
["win", "draw", "loss", "best_move"], label="Target",
value="win",
)
average_over_planes = gr.Checkbox(label="Average over Planes", value=False)
with gr.Accordion("Average over planes", open=False):
begin_average_index = gr.Slider(
label="Begin average index",
minimum=0,
maximum=111,
step=1,
value=0,
)
end_average_index = gr.Slider(
label="End average index",
minimum=0,
maximum=111,
step=1,
value=111,
)
plane_index = gr.Slider(
label="Plane index",
minimum=0,
maximum=111,
step=1,
value=0,
)
with gr.Column():
image_board = gr.Image(label="Board", interactive=False)
colorbar = gr.Image(label="Colorbar", interactive=False)
model = gr.State(value=None)
board = gr.State(value=None)
gradients = gr.State(value=None)
interface.load(
initial_load,
inputs=[model_name, game_pgn, board_fen, input_encoding, target, average_over_planes, begin_average_index, end_average_index, plane_index],
outputs=[model, board, gradients, info, image_board, colorbar],
concurrency_id="trace_queue"
)
game_pgn.submit(
on_board_change,
inputs=[model, game_pgn, board_fen, input_encoding, target, average_over_planes, begin_average_index, end_average_index, plane_index],
outputs=[board, gradients, info, image_board, colorbar],
concurrency_id="trace_queue"
)
board_fen.submit(
on_board_change,
inputs=[model, game_pgn, board_fen, input_encoding, target, average_over_planes, begin_average_index, end_average_index, plane_index],
outputs=[board, gradients, info, image_board, colorbar],
concurrency_id="trace_queue"
)
model_name.change(
on_model_change,
inputs=[model_name, board, input_encoding, target, average_over_planes, begin_average_index, end_average_index, plane_index],
outputs=[model, gradients, info, image_board, colorbar],
concurrency_id="trace_queue"
)
input_encoding.change(
on_input_encoding_change,
inputs=[model, board, input_encoding, target, average_over_planes, begin_average_index, end_average_index, plane_index],
outputs=[gradients, image_board, colorbar],
concurrency_id="trace_queue"
)
target.change(
on_target_change,
inputs=[model, board, input_encoding, target, average_over_planes, begin_average_index, end_average_index, plane_index],
outputs=[gradients, image_board, colorbar],
concurrency_id="trace_queue"
)
for render_arg in [average_over_planes, begin_average_index, end_average_index, plane_index]:
render_arg.change(
render_gradients,
inputs=[board, gradients, average_over_planes, begin_average_index, end_average_index, plane_index],
outputs=[image_board, colorbar],
)
|