socr / spadl /statsbomb.py
scfive's picture
Upload 203 files
d6ea71e verified
"""StatsBomb event stream data to SPADL converter."""
import warnings
from typing import Any, Optional, cast
import numpy as np
import numpy.typing as npt
import pandas as pd # type: ignore
from pandera.typing import DataFrame
from . import config as spadlconfig
from .base import _add_dribbles, _fix_clearances, _fix_direction_of_play
from .schema import SPADLSchema
def convert_to_actions(
events: pd.DataFrame,
home_team_id: int,
xy_fidelity_version: Optional[int] = None,
shot_fidelity_version: Optional[int] = None,
) -> DataFrame[SPADLSchema]:
"""
Convert StatsBomb events to SPADL actions.
Parameters
----------
events : pd.DataFrame
DataFrame containing StatsBomb events from a single game.
home_team_id : int
ID of the home team in the corresponding game.
xy_fidelity_version : int, optional
Whether low or high fidelity coordinates are used in the event data.
If not specified, the fidelity version is inferred from the data.
shot_fidelity_version : int, optional
Whether low or high fidelity coordinates are used in the event data
for shots. If not specified, the fidelity version is inferred from the
data.
Returns
-------
actions : pd.DataFrame
DataFrame with corresponding SPADL actions.
"""
actions = pd.DataFrame()
# Determine xy_fidelity_version and shot_fidelity_version
infered_xy_fidelity_version, infered_shot_fidelity_version = _infer_xy_fidelity_versions(
events
)
if xy_fidelity_version is None:
xy_fidelity_version = infered_xy_fidelity_version
warnings.warn(
f"Inferred xy_fidelity_version={infered_xy_fidelity_version}."
+ " If this is incorrect, please specify the correct version"
+ " using the xy_fidelity_version argument"
)
else:
assert xy_fidelity_version in [1, 2], "xy_fidelity_version must be 1 or 2"
if shot_fidelity_version is None:
if xy_fidelity_version == 2:
shot_fidelity_version = 2
else:
shot_fidelity_version = infered_shot_fidelity_version
warnings.warn(
f"Inferred shot_fidelity_version={infered_shot_fidelity_version}."
+ " If this is incorrect, please specify the correct version"
+ " using the shot_fidelity_version argument"
)
else:
assert shot_fidelity_version in [1, 2], "shot_fidelity_version must be 1 or 2"
events = events.copy()
events = _insert_interception_passes(events)
events["extra"].fillna({}, inplace=True)
actions["game_id"] = events.game_id
actions["original_event_id"] = events.event_id
actions["period_id"] = events.period_id
actions["time_seconds"] = pd.to_timedelta(events.timestamp).dt.total_seconds()
actions["team_id"] = events.team_id
actions["player_id"] = events.player_id
# split (end)location column into x and y columns
end_location = events[["location", "extra"]].apply(_get_end_location, axis=1)
# convert StatsBomb coordinates to spadl coordinates
actions.loc[events.type_name == "Shot", ["start_x", "start_y"]] = _convert_locations(
events.loc[events.type_name == "Shot", "location"],
shot_fidelity_version,
)
actions.loc[events.type_name != "Shot", ["start_x", "start_y"]] = _convert_locations(
events.loc[events.type_name != "Shot", "location"],
shot_fidelity_version,
)
actions.loc[events.type_name == "Shot", ["end_x", "end_y"]] = _convert_locations(
end_location.loc[events.type_name == "Shot"],
shot_fidelity_version,
)
actions.loc[events.type_name != "Shot", ["end_x", "end_y"]] = _convert_locations(
end_location.loc[events.type_name != "Shot"],
shot_fidelity_version,
)
actions[["type_id", "result_id", "bodypart_id"]] = events[["type_name", "extra"]].apply(
_parse_event, axis=1, result_type="expand"
)
actions = (
actions[actions.type_id != spadlconfig.actiontypes.index("non_action")]
.sort_values(["game_id", "period_id", "time_seconds"], kind="mergesort")
.reset_index(drop=True)
)
actions = _fix_direction_of_play(actions, home_team_id)
actions = _fix_clearances(actions)
actions["action_id"] = range(len(actions))
actions = _add_dribbles(actions)
return cast(DataFrame[SPADLSchema], actions)
Location = tuple[float, float]
def _insert_interception_passes(df_events: pd.DataFrame) -> pd.DataFrame:
"""Insert interception actions before passes.
This function converts passes that are also interceptions (type 64) in the
StatsBomb event data into two separate events, first an interception and
then a pass.
Parameters
----------
df_events : pd.DataFrame
StatsBomb event dataframe
Returns
-------
pd.DataFrame
StatsBomb event dataframe in which passes that were also denoted as
interceptions in the StatsBomb notation are transformed into two events.
"""
def is_interception_pass(x: dict) -> bool: # type: ignore
return x.get("extra", {}).get("pass", {}).get("type", {}).get("name") == "Interception"
df_events_interceptions = df_events[df_events.apply(is_interception_pass, axis=1)].copy()
if not df_events_interceptions.empty:
df_events_interceptions["type_name"] = "Interception"
df_events_interceptions["extra"] = [
{"interception": {"outcome": {"id": 16, "name": "Success In Play"}}}
] * len(df_events_interceptions)
df_events = pd.concat([df_events_interceptions, df_events], ignore_index=True)
df_events = df_events.sort_values(["timestamp"], kind="mergesort")
df_events = df_events.reset_index(drop=True)
return df_events
def _infer_xy_fidelity_versions(events: pd.DataFrame) -> tuple[int, int]:
"""Find out if x and y are integers disguised as floats."""
mask_shot = events.type_name == "Shot"
mask_other = events.type_name != "Shot"
locations = events.location.apply(pd.Series)
mask_valid_location = locations.notna().any(axis=1)
high_fidelity_shots = (locations.loc[mask_valid_location & mask_shot] % 1 != 0).any(axis=None)
high_fidelity_other = (locations.loc[mask_valid_location & mask_other] % 1 != 0).any(axis=None)
xy_fidelity_version = 2 if high_fidelity_other else 1
shot_fidelity_version = 2 if high_fidelity_shots else xy_fidelity_version
return shot_fidelity_version, xy_fidelity_version
def _convert_locations(locations: pd.Series, fidelity_version: int) -> npt.NDArray[np.float32]:
"""Convert StatsBomb locations to spadl coordinates.
StatsBomb coordinates are cell-based, using a 120x80 grid, so 1,1 is the
top-left square 'yard' of the field (in landscape), even though 0,0 is the
true coordinate of the corner flag.
Some matches have metadata like "xy_fidelity_version" : "2", which means
the grid has higher granularity. In this case 0.1,0.1 is the top left
cell.
"""
# [1, 120] x [1, 80]
# +-----+------+
# | 1,1 | 2, 1 |
# +-----+------+
# | 1,2 | 2,2 |
# +-----+------+
cell_side = 0.1 if fidelity_version == 2 else 1.0
cell_relative_center = cell_side / 2
coordinates = np.empty((len(locations), 2), dtype=float)
for i, loc in enumerate(locations):
if isinstance(loc, list) and len(loc) == 2:
coordinates[i, 0] = (loc[0] - cell_relative_center) / 120 * spadlconfig.field_length
coordinates[i, 1] = (
spadlconfig.field_width
- (loc[1] - cell_relative_center) / 80 * spadlconfig.field_width
)
elif isinstance(loc, list) and len(loc) == 3:
# A coordinate in the goal frame, only used for the end location of
# Shot events. The y-coordinates and z-coordinates are always detailed
# to a tenth of a yard.
coordinates[i, 0] = (loc[0] - cell_relative_center) / 120 * spadlconfig.field_length
coordinates[i, 1] = (
spadlconfig.field_width - (loc[1] - 0.05) / 80 * spadlconfig.field_width
)
coordinates[:, 0] = np.clip(coordinates[:, 0], 0, spadlconfig.field_length)
coordinates[:, 1] = np.clip(coordinates[:, 1], 0, spadlconfig.field_width)
return coordinates
def _get_end_location(q: tuple[Location, dict[str, Any]]) -> Location:
start_location, extra = q
for event in ["pass", "shot", "carry"]:
if event in extra and "end_location" in extra[event]:
return extra[event]["end_location"]
return start_location
def _parse_event(q: tuple[str, dict[str, Any]]) -> tuple[int, int, int]:
t, x = q
events = {
"Pass": _parse_pass_event,
"Dribble": _parse_dribble_event,
"Carry": _parse_carry_event,
"Foul Committed": _parse_foul_event,
"Duel": _parse_duel_event,
"Interception": _parse_interception_event,
"Shot": _parse_shot_event,
"Own Goal Against": _parse_own_goal_event,
"Goal Keeper": _parse_goalkeeper_event,
"Clearance": _parse_clearance_event,
"Miscontrol": _parse_miscontrol_event,
}
parser = events.get(t, _parse_event_as_non_action)
a, r, b = parser(x)
actiontype = spadlconfig.actiontypes.index(a)
result = spadlconfig.results.index(r)
bodypart = spadlconfig.bodyparts.index(b)
return actiontype, result, bodypart
def _parse_event_as_non_action(_extra: dict[str, Any]) -> tuple[str, str, str]:
a = "non_action"
r = "success"
b = "foot"
return a, r, b
def _parse_pass_event(extra: dict[str, Any]) -> tuple[str, str, str]: # noqa: C901
a = "pass" # default
b = "foot" # default
p = extra.get("pass", {})
ptype = p.get("type", {}).get("name")
height = p.get("height", {}).get("name")
cross = p.get("cross")
if ptype == "Free Kick":
if height == "High Pass" or cross:
a = "freekick_crossed"
else:
a = "freekick_short"
elif ptype == "Corner":
if height == "High Pass" or cross:
a = "corner_crossed"
else:
a = "corner_short"
elif ptype == "Goal Kick":
a = "goalkick"
elif ptype == "Throw-in":
a = "throw_in"
b = "other"
elif cross:
a = "cross"
else:
a = "pass"
pass_outcome = extra.get("pass", {}).get("outcome", {}).get("name")
if pass_outcome in ["Incomplete", "Out"]:
r = "fail"
elif pass_outcome == "Pass Offside":
r = "offside"
elif pass_outcome in ["Injury Clearance", "Unknown"]:
# discard passes that are not part of the play
a = "non_action"
r = "success"
else:
r = "success"
bp = extra.get("pass", {}).get("body_part", {}).get("name")
if bp is not None:
if "Head" in bp:
b = "head"
elif bp == "Left Foot":
b = "foot_left"
elif bp == "Right Foot":
b = "foot_right"
elif "Foot" in bp or bp == "Drop Kick":
b = "foot"
else:
b = "other"
return a, r, b
def _parse_dribble_event(extra: dict[str, Any]) -> tuple[str, str, str]:
a = "take_on"
dribble_outcome = extra.get("dribble", {}).get("outcome", {}).get("name")
if dribble_outcome == "Incomplete":
r = "fail"
elif dribble_outcome == "Complete":
r = "success"
else:
r = "success"
b = "foot"
return a, r, b
def _parse_carry_event(_extra: dict[str, Any]) -> tuple[str, str, str]:
a = "dribble"
r = "success"
b = "foot"
return a, r, b
def _parse_foul_event(extra: dict[str, Any]) -> tuple[str, str, str]:
a = "foul"
foul_card = extra.get("foul_committed", {}).get("card", {}).get("name", "")
if "Yellow" in foul_card:
r = "yellow_card"
elif "Red" in foul_card:
r = "red_card"
else:
r = "fail"
b = "foot"
return a, r, b
def _parse_duel_event(extra: dict[str, Any]) -> tuple[str, str, str]:
if extra.get("duel", {}).get("type", {}).get("name") == "Tackle":
a = "tackle"
duel_outcome = extra.get("duel", {}).get("outcome", {}).get("name")
if duel_outcome in ["Lost In Play", "Lost Out"]:
r = "fail"
elif duel_outcome in ["Success in Play", "Won"]:
r = "success"
else:
r = "success"
b = "foot"
return a, r, b
return _parse_event_as_non_action(extra)
def _parse_interception_event(extra: dict[str, Any]) -> tuple[str, str, str]:
a = "interception"
interception_outcome = extra.get("interception", {}).get("outcome", {}).get("name")
if interception_outcome in ["Lost In Play", "Lost Out"]:
r = "fail"
elif interception_outcome == "Won":
r = "success"
else:
r = "success"
b = "foot"
return a, r, b
def _parse_shot_event(extra: dict[str, Any]) -> tuple[str, str, str]:
extra_type = extra.get("shot", {}).get("type", {}).get("name")
if extra_type == "Free Kick":
a = "shot_freekick"
elif extra_type == "Penalty":
a = "shot_penalty"
else:
a = "shot"
shot_outcome = extra.get("shot", {}).get("outcome", {}).get("name")
if shot_outcome == "Goal":
r = "success"
elif shot_outcome in ["Blocked", "Off T", "Post", "Saved", "Wayward"]:
r = "fail"
else:
r = "fail"
bp = extra.get("shot", {}).get("body_part", {}).get("name")
if bp is None:
b = "foot"
elif "Head" in bp:
b = "head"
elif bp == "Left Foot":
b = "foot_left"
elif bp == "Right Foot":
b = "foot_right"
elif "Foot" in bp:
b = "foot"
else:
b = "other"
return a, r, b
def _parse_own_goal_event(_extra: dict[str, Any]) -> tuple[str, str, str]:
a = "bad_touch"
r = "owngoal"
b = "foot"
return a, r, b
def _parse_goalkeeper_event(extra: dict[str, Any]) -> tuple[str, str, str]: # noqa: C901
extra_type = extra.get("goalkeeper", {}).get("type", {}).get("name")
if extra_type == "Shot Saved":
a = "keeper_save"
elif extra_type in ("Collected", "Keeper Sweeper"):
a = "keeper_claim"
elif extra_type == "Punch":
a = "keeper_punch"
else:
a = "non_action"
goalkeeper_outcome = extra.get("goalkeeper", {}).get("outcome", {}).get("name", "x")
if goalkeeper_outcome in [
"Claim",
"Clear",
"Collected Twice",
"In Play Safe",
"Success",
"Touched Out",
]:
r = "success"
elif goalkeeper_outcome in ["In Play Danger", "No Touch"]:
r = "fail"
else:
r = "success"
bp = extra.get("goalkeeper", {}).get("body_part", {}).get("name")
if bp is None:
b = "other"
elif "Head" in bp:
b = "head"
elif bp == "Left Foot":
b = "foot_left"
elif bp == "Right Foot":
b = "foot_right"
elif "Foot" in bp or bp == "Drop Kick":
b = "foot"
else:
b = "other"
return a, r, b
def _parse_clearance_event(extra: dict[str, Any]) -> tuple[str, str, str]:
a = "clearance"
r = "success"
bp = extra.get("clearance", {}).get("body_part", {}).get("name")
if bp is None:
b = "foot"
elif "Head" in bp:
b = "head"
elif bp == "Left Foot":
b = "foot_left"
elif bp == "Right Foot":
b = "foot_right"
elif "Foot" in bp:
b = "foot"
else:
b = "other"
return a, r, b
def _parse_miscontrol_event(_extra: dict[str, Any]) -> tuple[str, str, str]:
a = "bad_touch"
r = "fail"
b = "foot"
return a, r, b