Source code for onnx_diagnostic.helpers.log_helper

import enum
import io
import pprint
import re
import warnings
from typing import Any, Callable, Dict, List, Optional, Sequence, Tuple, Union
import numpy as np
import pandas
from pandas.api.types import is_numeric_dtype, is_datetime64_any_dtype
from .helper import string_sig
from ._log_helper import (
    BUCKET_SCALES,
    breaking_last_point,
    apply_excel_style,
    align_dataframe_with,
    open_dataframe,
    enumerate_csv_files,
)


[docs] class CubeViewDef: """ Defines how to compute a view. :param key_index: keys to put in the row index :param values: values to show :param ignore_unique: ignore keys with a unique value :param order: to reorder key in columns index :param key_agg: aggregate according to these columns before creating the view :param agg_args: see :meth:`pandas.core.groupby.DataFrameGroupBy.agg`, it can be also a callable to return a different aggregation method depending on the column name :param agg_kwargs: see :meth:`pandas.core.groupby.DataFrameGroupBy.agg` :param agg_multi: aggregation over multiple columns :param ignore_columns: ignore the following columns if known to overload the view :param keep_columns_in_index: keeps the columns even if there is only one unique value :param dropna: drops rows with nan if not relevant :param transpose: transpose :param f_highlight: to highlights some values :param name: name of the view, used mostly to debug :param plots: adds plot to the Excel sheet :param no_index: remove the index (but keeps the columns) Some examples of views. First example is an aggregated view for many metrics. .. code-block:: python cube = CubeLogs(...) CubeViewDef( key_index=cube._filter_column(fs, cube.keys_time), values=cube._filter_column( ["TIME_ITER", "speedup", "time_latency.*", "onnx_n_nodes"], cube.values, ), ignore_unique=True, key_agg=["model_name", "task", "model_task", "suite"], agg_args=lambda column_name: "sum" if column_name.startswith("n_") else "mean", agg_multi={"speedup_weighted": mean_weight, "speedup_geo": mean_geo}, name="agg-all", plots=True, ) Next one focuses on a couple of metrics. .. code-block:: python cube = CubeLogs(...) CubeViewDef( key_index=cube._filter_column(fs, cube.keys_time), values=cube._filter_column(["speedup"], cube.values), ignore_unique=True, keep_columns_in_index=["suite"], name="speedup", ) """
[docs] class HighLightKind(enum.IntEnum): "Codes to highlight values." NONE = 0 RED = 1 GREEN = 2
def __init__( self, key_index: Sequence[str], values: Sequence[str], ignore_unique: bool = True, order: Optional[Sequence[str]] = None, key_agg: Optional[Sequence[str]] = None, agg_args: Union[Sequence[Any], Callable[[str], Any]] = ("sum",), agg_kwargs: Optional[Dict[str, Any]] = None, agg_multi: Optional[ Dict[str, Callable[[pandas.core.groupby.DataFrameGroupBy], pandas.Series]] ] = None, ignore_columns: Optional[Sequence[str]] = None, keep_columns_in_index: Optional[Sequence[str]] = None, dropna: bool = True, transpose: bool = False, f_highlight: Optional[Callable[[Any], "CubeViewDef.HighLightKind"]] = None, name: Optional[str] = None, no_index: bool = False, plots: bool = False, ): self.key_index = key_index self.values = values self.ignore_unique = ignore_unique self.order = order self.key_agg = key_agg self.agg_args = agg_args self.agg_kwargs = agg_kwargs self.agg_multi = agg_multi self.dropna = dropna self.ignore_columns = ignore_columns self.keep_columns_in_index = keep_columns_in_index self.f_highlight = f_highlight self.transpose = transpose self.name = name self.no_index = no_index self.plots = plots def __repr__(self) -> str: "usual" return string_sig(self) # type: ignore[arg-type]
[docs] class CubePlot: """ Creates a plot. :param df: dataframe :param kind: kind of graph to plot, bar, barh, line :param split: draw a graph per line in the dataframe :param timeseries: this assumes the time is one level of the columns, this argument indices the level name It defines a graph. Usually *bar* or *barh* is used to compare experiments for every metric, a subplot by metric. .. code-block:: python CubePlot(df, kind="barh", orientation="row", split=True) *line* is usually used to plot timeseries showing the evolution of metrics over time. .. code-block:: python CubePlot( df, kind="line", orientation="row", split=True, timeseries="time", ) """ KINDS = {"bar", "barh", "line"}
[docs] @classmethod def group_columns( cls, columns: List[str], sep: str = "/", depth: int = 2 ) -> List[List[str]]: """Groups columns to have nice display.""" res: Dict[str, List[str]] = {} for c in columns: p = c.split("/") k = "/".join(p[:depth]) if k not in res: res[k] = [] res[k].append(c) new_res: Dict[str, List[str]] = {} for k, v in res.items(): if len(v) >= 3: new_res[k] = v else: if "0" not in new_res: new_res["0"] = [] new_res["0"].extend(v) groups: List[List[str]] = [sorted(v) for k, v in sorted(new_res.items())] if depth <= 1: return groups new_groups: List[List[str]] = [] for v in groups: if len(v) >= 6: new_groups.extend(cls.group_columns(v, depth=1, sep=sep)) else: new_groups.append(v) return new_groups
def __init__( self, df: pandas.DataFrame, kind: str = "bar", orientation="col", split: bool = True, timeseries: Optional[str] = None, ): assert ( not timeseries or timeseries in df.columns.names ), f"Level {timeseries!r} is not part of the columns levels {df.columns.names}" assert ( kind in self.__class__.KINDS ), f"Unexpected kind={kind!r} not in {self.__class__.KINDS}" assert split, f"split={split} not implemented" assert ( not timeseries or orientation == "row" ), f"orientation={orientation!r} must be 'row' for timeseries" self.df = df.copy() self.kind = kind self.orientation = orientation self.split = split self.timeseries = timeseries if timeseries: if isinstance(self.df.columns, pandas.MultiIndex): index_time = list(self.df.columns.names).index(self.timeseries) def _drop(t, i=index_time): return (*t[:i], *t[i + 1 :]) self.df.columns = pandas.MultiIndex.from_tuples( [("/".join(map(str, _drop(i))), i[index_time]) for i in self.df.columns], names=["metric", timeseries], ) else: if isinstance(self.df.columns, pandas.MultiIndex): self.df.columns = ["/".join(map(str, i)) for i in self.df.columns] if isinstance(self.df.index, pandas.MultiIndex): self.df.index = ["/".join(map(str, i)) for i in self.df.index] def __repr__(self) -> str: "usual" return string_sig(self) # type: ignore[arg-type]
[docs] def to_images( self, verbose: int = 0, merge: bool = True, title_suffix: Optional[str] = None ) -> List[bytes]: """ Converts data into plots and images. :param verbose: verbosity :param merge: returns all graphs in a single image (True) or an image for every graph (False) :param title_suffix: prefix for the title of every graph :return: list of binary images (format PNG) """ if self.kind in ("barh", "bar"): return self._to_images_bar(verbose=verbose, merge=merge, title_suffix=title_suffix) if self.kind == "line": return self._to_images_line( verbose=verbose, merge=merge, title_suffix=title_suffix ) raise AssertionError(f"self.kind={self.kind!r} not implemented")
@classmethod def _make_loop(cls, ensemble, verbose): if verbose: from tqdm import tqdm loop = tqdm(ensemble) else: loop = ensemble return loop def _to_images_bar( self, verbose: int = 0, merge: bool = True, title_suffix: Optional[str] = None ) -> List[bytes]: assert merge, f"merge={merge} not implemented yet" import matplotlib.pyplot as plt df = self.df.T if self.orientation == "row" else self.df title_suffix = f"\n{title_suffix}" if title_suffix else "" n_cols = 3 nn = df.shape[1] // n_cols nn += int(df.shape[1] % n_cols != 0) fig, axs = plt.subplots(nn, n_cols, figsize=(6 * n_cols, nn * df.shape[0] / 5)) pos = 0 imgs = [] for c in self._make_loop(df.columns, verbose): ax = axs[pos // n_cols, pos % n_cols] ( df[c].plot.barh(title=f"{c}{title_suffix}", ax=ax) if self.kind == "barh" else df[c].plot.bar(title=f"{c}{title_suffix}", ax=ax) ) ax.tick_params(axis="both", which="major", labelsize=8) ax.grid(True) pos += 1 # noqa: SIM113 fig.tight_layout() imgdata = io.BytesIO() fig.savefig(imgdata, format="png") imgs.append(imgdata.getvalue()) plt.close() return imgs def _to_images_line( self, verbose: int = 0, merge: bool = True, title_suffix: Optional[str] = None ) -> List[bytes]: assert merge, f"merge={merge} not implemented yet" assert ( self.orientation == "row" ), f"self.orientation={self.orientation!r} not implemented for this kind of graph." def rotate_align(ax, angle=15, align="right"): for label in ax.get_xticklabels(): label.set_rotation(angle) label.set_horizontalalignment(align) ax.tick_params(axis="both", which="major", labelsize=8) ax.grid(True) ax.legend() ax.tick_params(labelleft=True) return ax import matplotlib.pyplot as plt df = self.df.T confs = list(df.unstack(self.timeseries).index) groups = self.group_columns(confs) n_cols = len(groups) title_suffix = f"\n{title_suffix}" if title_suffix else "" fig, axs = plt.subplots( df.shape[1], n_cols, figsize=(5 * n_cols, max(len(g) for g in groups) * df.shape[1] / 2), sharex=True, sharey="row" if n_cols > 1 else False, ) imgs = [] row = 0 for c in self._make_loop(df.columns, verbose): dfc = df[[c]] dfc = dfc.unstack(self.timeseries).T.droplevel(0) if n_cols == 1: dfc.plot(title=f"{c}{title_suffix}", ax=axs[row], linewidth=3) axs[row].grid(True) rotate_align(axs[row]) else: x = list(range(dfc.shape[0])) ticks = list(dfc.index) for ii, group in enumerate(groups): ddd = dfc.loc[:, group].copy() axs[row, ii].set_xticks(x) axs[row, ii].set_xticklabels(ticks) # This is very slow # ddd.plot(ax=axs[row, ii],linewidth=3) for jj in range(ddd.shape[1]): axs[row, ii].plot(x, ddd.iloc[:, jj], lw=3, label=ddd.columns[jj]) axs[row, ii].set_title(f"{c}{title_suffix}") rotate_align(axs[row, ii]) row += 1 # noqa: SIM113 fig.tight_layout() imgdata = io.BytesIO() fig.savefig(imgdata, format="png") imgs.append(imgdata.getvalue()) plt.close() return imgs
[docs] class CubeLogs: """ Processes logs coming from experiments. A cube is basically a database with certain columns playing specific roles. * time: only one column, it is not mandatory but it is recommended to have one * keys: they are somehow coordinates, they cannot be aggregated, they are not numbers, more like categories, `(time, *keys)` identifies an element of the database in an unique way, there cannot be more than one row sharing the same key and time values * values: they are not necessary numerical, but if they are, they can be aggregated Every other columns is ignored. More columns can be added by using formulas. :param data: the raw data :param time: the time column :param keys: the keys, can include regular expressions :param values: the values, can include regular expressions :param ignored: ignores some column, acts as negative regular expressions for the other two :param recent: if more than one rows share the same keys, the cube only keeps the most recent one :param formulas: columns to add, defined with formulas :param fill_missing: a dictionary, defines values replacing missing one for some columns :param keep_last_date: overwrites all the times with the most recent one, it makes things easier for timeseries """ def __init__( self, data: Any, time: str = "date", keys: Sequence[str] = ("version_.*", "model_.*"), values: Sequence[str] = ("time_.*", "disc_.*"), ignored: Sequence[str] = (), recent: bool = False, formulas: Optional[ Union[ Sequence[str], Dict[str, Union[str, Callable[[pandas.DataFrame], pandas.Series]]], ] ] = None, fill_missing: Optional[Sequence[Tuple[str, Any]]] = None, keep_last_date: bool = False, ): self._data = data self._time = time self._keys = keys self._values = values self._ignored = ignored self.recent = recent self._formulas = formulas self.fill_missing = fill_missing self.keep_last_date = keep_last_date
[docs] def clone( self, data: Optional[pandas.DataFrame] = None, keys: Optional[Sequence[str]] = None ) -> "CubeLogs": """ Makes a copy of the dataframe. It copies the processed data not the original one. """ cube = self.__class__( data if data is not None else self.data.copy(), time=self.time, keys=keys or self.keys_no_time, values=self.values, ) cube.load() return cube
[docs] def post_load_process_piece( self, df: pandas.DataFrame, unique: bool = False ) -> pandas.DataFrame: """ Postprocesses a piece when a cube is made of multiple pieces before it gets merged. """ if not self.fill_missing: return df missing = dict(self.fill_missing) for k, v in missing.items(): if k not in df.columns: df[k] = v return df
[docs] def load(self, verbose: int = 0): """Loads and preprocesses the data. Returns self.""" if isinstance(self._data, pandas.DataFrame): if verbose: print(f"[CubeLogs.load] load from dataframe, shape={self._data.shape}") self.data = self.post_load_process_piece(self._data, unique=True) if verbose: print(f"[CubeLogs.load] after postprocessing shape={self.data.shape}") elif isinstance(self._data, list) and all(isinstance(r, dict) for r in self._data): if verbose: print(f"[CubeLogs.load] load from list of dicts, n={len(self._data)}") self.data = pandas.DataFrame(self.post_load_process_piece(self._data, unique=True)) if verbose: print(f"[CubeLogs.load] after postprocessing shape={self.data.shape}") elif isinstance(self._data, list) and all( isinstance(r, pandas.DataFrame) for r in self._data ): if verbose: print(f"[CubeLogs.load] load from list of DataFrame, n={len(self._data)}") self.data = pandas.concat( [self.post_load_process_piece(c) for c in self._data], axis=0 ) if verbose: print(f"[CubeLogs.load] after postprocessing shape={self.data.shape}") elif isinstance(self._data, list): if verbose: print("[CubeLogs.load] load from list of Cubes") cubes = [] for item in enumerate_csv_files(self._data, verbose=verbose): df = open_dataframe(item) cube = CubeLogs( df, time=self._time, keys=self._keys, values=self._values, ignored=self._ignored, recent=self.recent, ) cube.load() cubes.append(self.post_load_process_piece(cube.data)) self.data = pandas.concat(cubes, axis=0) if verbose: print(f"[CubeLogs.load] after postprocessing shape={self.data.shape}") else: raise NotImplementedError( f"Not implemented with the provided data (type={type(self._data)})" ) assert all(isinstance(c, str) for c in self.data.columns), ( f"The class only supports string as column names " f"but found {[c for c in self.data.columns if not isinstance(c, str)]}" ) if verbose: print(f"[CubeLogs.load] loaded with shape={self.data.shape}") self._initialize_columns() if verbose: print(f"[CubeLogs.load] time={self.time}") print(f"[CubeLogs.load] keys={self.keys_no_time}") print(f"[CubeLogs.load] values={self.values}") print(f"[CubeLogs.load] ignored={self.ignored}") print(f"[CubeLogs.load] ignored_values={self.ignored_values}") print(f"[CubeLogs.load] ignored_keys={self.ignored_keys}") assert self.keys_no_time, f"No keys found with {self._keys} from {self.data.columns}" assert self.values, f"No values found with {self._values} from {self.data.columns}" assert not ( set(self.keys_no_time) & set(self.values) ), f"Columns {set(self.keys_no_time) & set(self.values)} cannot be keys and values" assert not ( set(self.keys_no_time) & set(self.ignored) ), f"Columns {set(self.keys_no_time) & set(self.ignored)} cannot be keys and ignored" assert not ( set(self.values) & set(self.ignored) ), f"Columns {set(self.keys_no_time) & set(self.ignored)} cannot be values and ignored" assert ( self.time not in self.keys_no_time and self.time not in self.values and self.time not in self.ignored ), ( f"Column {self.time!r} is also a key, a value or ignored, " f"keys={sorted(self.keys_no_time)}, values={sorted(self.values)}, " f"ignored={sorted(self.ignored)}" ) self._columns = [self.time, *self.keys_no_time, *self.values, *self.ignored] self.dropped = [c for c in self.data.columns if c not in set(self.columns)] self.data = self.data[self.columns] if verbose: print(f"[CubeLogs.load] dropped={self.dropped}") print(f"[CubeLogs.load] data.shape={self.data.shape}") if verbose: print(f"[CubeLogs.load] removed columns, shape={self.data.shape}") self._preprocess() if verbose: print(f"[CubeLogs.load] preprocess, shape={self.data.shape}") if self.recent: print(f"[CubeLogs.load] keep most recent data.shape={self.data.shape}") # Let's apply the formulas if self._formulas: forms = ( {k: k for k in self._formulas} if not isinstance(self._formulas, dict) else self._formulas ) cols = set(self.values) for k, ff in forms.items(): f = self._process_formula(ff) if k in cols or f is None: if verbose: print(f"[CubeLogs.load] skip formula {k!r}") else: if verbose: print(f"[CubeLogs.load] apply formula {k!r}") self.data[k] = f(self.data) self.values.append(k) cols.add(k) self.values_for_key = {k: set(self.data[k].dropna()) for k in self.keys_time} for k in self.keys_no_time: if self.data[k].isna().max(): self.values_for_key[k].add(np.nan) self.keys_with_nans = [ c for c in self.keys_time if self.data[c].isna().astype(int).sum() > 0 ] if verbose: print(f"[CubeLogs.load] convert column {self.time!r} into date") if self.keys_with_nans: print(f"[CubeLogs.load] keys_with_nans={self.keys_with_nans}") self.data[self.time] = pandas.to_datetime(self.data[self.time]) if self.keep_last_date: times = self.data[self.time].dropna() mi, mx = times.min(), times.max() if mi != mx: print(f"[CubeLogs.load] setting all dates in column {self.time} to {mx!r}") self.data.loc[~self.data[self.time].isna(), self.time] = mx self.values_for_key[self.time] = {mx} if self.data[self.time].isna().max(): self.values_for_key[self.time].add(np.nan) if verbose: print(f"[CubeLogs.load] done, shape={self.shape}") return self
def _process_formula( self, formula: Union[str, Callable[[pandas.DataFrame], pandas.Series]] ) -> Callable[[pandas.DataFrame], pandas.Series]: assert callable(formula), f"formula={formula!r} is not supported." return formula @property def shape(self) -> Tuple[int, int]: "Returns the shape." assert hasattr(self, "data"), "Method load was not called" return self.data.shape @property def columns(self) -> Sequence[str]: "Returns the columns." assert hasattr(self, "data"), "Method load was not called" return self.data.columns def _preprocess(self): last = self.values[0] gr = self.data[[*self.keys_time, last]].groupby(self.keys_time, dropna=False).count() gr = gr[gr[last] > 1] if self.recent: cp = self.data.copy() assert ( "__index__" not in cp.columns ), f"'__index__' should not be a column in {cp.columns}" cp["__index__"] = np.arange(cp.shape[0]) gr = ( cp[[*self.keys_time, "__index__"]] .groupby(self.keys_no_time, as_index=False, dropna=False) .max() ) assert gr.shape[0] > 0, ( f"Something went wrong after the groupby.\n" f"{cp[[*self.keys, self.time, '__index__']].head().T}" ) filtered = pandas.merge(cp, gr, on=["__index__", *self.keys_time]) assert filtered.shape[0] <= self.data.shape[0], ( f"Keeping the latest row brings more row {filtered.shape} " f"(initial is {self.data.shape})." ) self.data = filtered.drop("__index__", axis=1) else: assert gr.shape[0] == 0, f"There are duplicated rows:\n{gr}" @classmethod def _filter_column(cls, filters, columns, can_be_empty=False): assert list(columns), "columns is empty" set_cols = set() for f in filters: if set(f) & {'"', "^", ".", "*", "+", "{", "}"}: reg = re.compile(f) cols = [c for c in columns if reg.search(c)] elif f in columns: # No regular expression. cols = [f] else: continue set_cols |= set(cols) assert ( can_be_empty or set_cols ), f"Filters {filters} returns an empty set from {columns}" return sorted(set_cols) def _initialize_columns(self): keys = self._filter_column(self._keys, self.data.columns) self.values = self._filter_column(self._values, self.data.columns) self.ignored = self._filter_column(self._ignored, self.data.columns, True) assert ( self._time in self.data.columns ), f"Column {self._time} not found in {pprint.pformat(sorted(self.data.columns))}" ignored_keys = set(self.ignored) & set(keys) ignored_values = set(self.ignored) & set(self.values) self.keys_no_time = [c for c in keys if c not in ignored_keys] self.values = [c for c in self.values if c not in ignored_values] self.ignored_keys = sorted(ignored_keys) self.ignored_values = sorted(ignored_values) self.time = self._time self.keys_time = [self.time, *[c for c in keys if c not in ignored_keys]] def __str__(self) -> str: "usual" return str(self.data) if hasattr(self, "data") else str(self._data)
[docs] def make_view_def(self, name: str) -> Optional[CubeViewDef]: """ Returns a view definition. :param name: name of a value :return: a CubeViewDef or None if name does not make sense """ assert name in self.values, f"{name!r} is not one of the values {self.values}" keys = sorted(self.keys_no_time) index = len(keys) // 2 + (len(keys) % 2) return CubeViewDef(key_index=keys[:index], values=[name], name=name)
[docs] def view( self, view_def: Union[str, CubeViewDef], return_view_def: bool = False, verbose: int = 0, ) -> Union[pandas.DataFrame, Tuple[pandas.DataFrame, CubeViewDef]]: """ Returns a dataframe, a pivot view. `key_index` determines the index, the other key columns determines the columns. If `ignore_unique` is True, every columns with a unique value is removed. :param view_def: view definition :param return_view_def: returns the view as well :param verbose: verbosity level :return: dataframe """ if isinstance(view_def, str): # We automatically create a view for a metric view_def_ = self.make_view_def(view_def) assert view_def_ is not None, f"Unable to create a view from {view_def!r}" view_def = view_def_ assert isinstance( view_def, CubeViewDef ), f"view_def should be a CubeViewDef, got {type(view_def)}: {view_def!r} instead" if verbose: print(f"[CubeLogs.view] -- start view {view_def.name!r}: {view_def}") key_agg = ( self._filter_column(view_def.key_agg, self.keys_time) if view_def.key_agg else [] ) set_key_agg = set(key_agg) assert set_key_agg <= set(self.keys_time), ( f"view_def.name={view_def.name!r}, " f"non existing keys in key_agg {set_key_agg - set(self.keys_time)}", f"keys={sorted(self.keys_time)}", ) values = self._filter_column(view_def.values, self.values) assert set(values) <= set(self.values), ( f"view_def.name={view_def.name!r}, " f"non existing columns in values {set(values) - set(self.values)}, " f"values={sorted(self.values)}" ) # aggregation if key_agg: final_stack = True key_index = [ c for c in self._filter_column(view_def.key_index, self.keys_time) if c not in set_key_agg ] keys_no_agg = [c for c in self.keys_time if c not in set_key_agg] if verbose: print(f"[CubeLogs.view] aggregation of {set_key_agg}") print(f"[CubeLogs.view] groupby {keys_no_agg}") data_red = self.data[[*keys_no_agg, *values]] assert set(key_index) <= set(data_red.columns), ( f"view_def.name={view_def.name!r}, " f"nnable to find {set(key_index) - set(data_red.columns)}, " f"key_agg={key_agg}, keys_no_agg={keys_no_agg},\n--\n" f"selected={pprint.pformat(sorted(data_red.columns))},\n--\n" f"keys={pprint.pformat(sorted(self.keys_time))}" ) grouped_data = data_red.groupby(keys_no_agg, as_index=True, dropna=False) if callable(view_def.agg_args): agg_kwargs = view_def.agg_kwargs or {} agg_args = ({c: view_def.agg_args(c) for c in values},) else: agg_args = view_def.agg_args # type: ignore[assignment] agg_kwargs = view_def.agg_kwargs or {} data = grouped_data.agg(*agg_args, **agg_kwargs) if view_def.agg_multi: append = [] for k, f in view_def.agg_multi.items(): cv = grouped_data.apply(f, include_groups=False) append.append(cv.to_frame(k)) data = pandas.concat([data, *append], axis=1) set_all_keys = set(keys_no_agg) values = list(data.columns) data = data.reset_index(drop=False) else: key_index = self._filter_column(view_def.key_index, self.keys_time) if verbose: print(f"[CubeLogs.view] no aggregation, index={key_index}") data = self.data[[*self.keys_time, *values]] set_all_keys = set(self.keys_time) final_stack = False assert set(key_index) <= set_all_keys, ( f"view_def.name={view_def.name!r}, " f"Non existing keys in key_index {set(key_index) - set_all_keys}" ) # remove unnecessary column set_key_columns = { c for c in self.keys_time if c not in key_index and c not in set(key_agg) } key_index0 = key_index if view_def.ignore_unique: unique = { k for k, v in self.values_for_key.items() if k in set_all_keys and len(v) <= 1 } keep_anyway = ( set(view_def.keep_columns_in_index) if view_def.keep_columns_in_index else set() ) key_index = [k for k in key_index if k not in unique or k in keep_anyway] key_columns = [k for k in set_key_columns if k not in unique or k in keep_anyway] if verbose: print(f"[CubeLogs.view] unique={unique}, keep_anyway={keep_anyway}") print( f"[CubeLogs.view] columns with unique values " f"{set(key_index0) - set(key_index)}" ) else: if verbose: print("[CubeLogs.view] keep all columns") key_columns = sorted(set_key_columns) unique = set() _md = lambda s: {k: v for k, v in self.values_for_key.items() if k in s} # noqa: E731 all_cols = set(key_columns) | set(key_index) | set(key_agg) | unique assert all_cols == set(self.keys_time), ( f"view_def.name={view_def.name!r}, " f"key_columns + key_index + key_agg + unique != keys, left=" f"{set(self.keys_time) - all_cols}, " f"unique={unique}, index={set(key_index)}, columns={set(key_columns)}, " f"agg={set(key_agg)}, keys={set(self.keys_time)}, values={values}" ) # reorder if view_def.order: subset = self._filter_column(view_def.order, all_cols | {self.time}) corder = [o for o in view_def.order if o in subset] assert set(corder) <= set_key_columns, ( f"view_def.name={view_def.name!r}, " f"non existing columns from order in key_columns " f"{set(corder) - set_key_columns}" ) key_columns = [ *[o for o in corder if o in key_columns], *[c for c in key_columns if c not in view_def.order], ] else: corder = None if view_def.dropna: data, key_index, key_columns, values = self._dropna( # type: ignore[assignment] data, key_index, key_columns, values, keep_columns_in_index=view_def.keep_columns_in_index, ) if view_def.ignore_columns: if verbose: print(f"[CubeLogs.view] ignore_columns {view_def.ignore_columns}") data = data.drop(view_def.ignore_columns, axis=1) seti = set(view_def.ignore_columns) if view_def.keep_columns_in_index: seti -= set(view_def.keep_columns_in_index) key_index = [c for c in key_index if c not in seti] key_columns = [c for c in key_columns if c not in seti] values = [c for c in values if c not in seti] # final verification if verbose: print(f"[CubeLogs.view] key_index={key_index}") print(f"[CubeLogs.view] key_columns={key_columns}") g = data[[*key_index, *key_columns]].copy() g["count"] = 1 r = g.groupby([*key_index, *key_columns], dropna=False).sum() not_unique = r[r["count"] > 1] assert not_unique.shape[0] == 0, ( f"view_def.name={view_def.name!r}, " f"unable to run the pivot with index={sorted(key_index)}, " f"key={sorted(key_columns)}, key_agg={key_agg}, values={sorted(values)}, " f"columns={sorted(data.columns)}, ignored={view_def.ignore_columns}, " f"not unique={set(data.columns) - unique}" f"\n--\n{not_unique.head()}" ) # pivot if verbose: print(f"[CubeLogs.view] values={values}") if key_index: piv = data.pivot(index=key_index[::-1], columns=key_columns, values=values) else: # pivot does return the same rank with it is empty. # Let's add arficially one data = data.copy() data["ALL"] = "ALL" piv = data.pivot(index=["ALL"], columns=key_columns, values=values) if isinstance(piv, pandas.Series): piv = piv.to_frame(name="series") names = list(piv.columns.names) assert ( "METRICS" not in names ), f"Not implemented when a level METRICS already exists {names!r}" names[0] = "METRICS" piv.columns = piv.columns.set_names(names) if final_stack: piv = piv.stack("METRICS", future_stack=True) if view_def.transpose: piv = piv.T if isinstance(piv, pandas.Series): piv = piv.to_frame("VALUE") piv.sort_index(inplace=True) if isinstance(piv.columns, pandas.MultiIndex): if corder: # reorder the levels for the columns with the view definition new_corder = [c for c in corder if c in piv.columns.names] new_names = [ *[c for c in piv.columns.names if c not in new_corder], *new_corder, ] piv.columns = piv.columns.reorder_levels(new_names) elif self.time in piv.columns.names: # put time at the end new_names = list(piv.columns.names) ind = new_names.index(self.time) if ind < len(new_names) - 1: del new_names[ind] new_names.append(self.time) piv.columns = piv.columns.reorder_levels(new_names) if view_def.no_index: piv = piv.reset_index(drop=False) else: piv.sort_index(inplace=True, axis=1) # final step, force columns with numerical values to be float for c in list(piv.columns): s = piv[c] if not pandas.api.types.is_object_dtype(s): continue try: sf = s.astype(float) except (ValueError, TypeError): continue piv[c] = sf if verbose: print(f"[CubeLogs.view] levels {piv.index.names}, {piv.columns.names}") print(f"[CubeLogs.view] -- done view {view_def.name!r}") return (piv, view_def) if return_view_def else piv
def _dropna( self, data: pandas.DataFrame, key_index: Sequence[str], key_columns: Sequence[str], values: Sequence[str], keep_columns_in_index: Optional[Sequence[str]] = None, ) -> Tuple[pandas.DataFrame, Sequence[str], Sequence[str], Sequence[str]]: set_keep_columns_in_index = ( set(keep_columns_in_index) if keep_columns_in_index else set() ) v = data[values] new_data = data[~v.isnull().all(1)] if data.shape == new_data.shape: return data, key_index, key_columns, values new_data = new_data.copy() new_key_index = [] for c in key_index: if c in set_keep_columns_in_index: new_key_index.append(c) continue v = new_data[c] sv = set(v.dropna()) if len(sv) > 1 or (v.isna().max() and len(sv) > 0): new_key_index.append(c) new_key_columns = [] for c in key_columns: if c in set_keep_columns_in_index: new_key_columns.append(c) continue v = new_data[c] sv = set(v.dropna()) if len(sv) > 1 or (v.isna().max() and len(sv) > 0): new_key_columns.append(c) for c in set(key_index) | set(key_columns): s = new_data[c] if s.isna().max(): if pandas.api.types.is_numeric_dtype( s ) and not pandas.api.types.is_object_dtype(s): min_v = s.dropna().min() assert ( min_v >= 0 ), f"Unable to replace nan values in column {c!r}, min_v={min_v}" new_data[c] = s.fillna(-1) else: new_data[c] = s.fillna("NAN") return new_data, new_key_index, new_key_columns, values
[docs] def describe(self) -> pandas.DataFrame: """Basic description of all variables.""" rows = [] for name in self.data.columns: values = self.data[name] dtype = values.dtype nonan = values.dropna() obs = dict( name=name, dtype=str(dtype), missing=len(values) - len(nonan), kind=( "time" if name == self.time else ( "keys" if name in self.keys_no_time else ( "values" if name in self.values else ("ignored" if name in self.ignored else "unused") ) ) ), ) if len(nonan) > 0: obs.update(dict(count=len(nonan))) if is_numeric_dtype(nonan) and not pandas.api.types.is_object_dtype(nonan): obs.update( dict( min=nonan.min(), max=nonan.max(), mean=nonan.mean(), sum=nonan.sum(), n_values=len(set(nonan)), ) ) elif obs["kind"] == "time": unique = set(nonan) obs["n_values"] = len(unique) o = dict( min=str(nonan.min()), max=str(nonan.max()), n_values=len(set(nonan)), ) o["values"] = f"{o['min']} - {o['max']}" obs.update(o) else: unique = set(nonan) obs["n_values"] = len(unique) if len(unique) < 20: obs["values"] = ",".join(map(str, sorted(unique))) rows.append(obs) return pandas.DataFrame(rows).set_index("name")
[docs] def to_excel( self, output: str, views: Union[Sequence[str], Dict[str, Union[str, CubeViewDef]]], main: Optional[str] = "main", raw: Optional[str] = "raw", verbose: int = 0, csv: Optional[Sequence[str]] = None, time_mask: bool = False, sbs: Optional[Dict[str, Dict[str, Any]]] = None, ): """ Creates an excel file with a list of views. :param output: output file to create :param views: sequence or dictionary of views to append :param main: add a page with statitcs on all variables :param raw: add a page with the raw data :param csv: views to dump as csv files (same name as outputs + view naw) :param verbose: verbosity :param time_mask: color the background of the cells if one of the value for the last date is unexpected, assuming they should remain stale :param sbs: configurations to compare side-by-side, this adds two tabs, one gathering raw data about the two configurations, the other one is aggregated by metrics """ if verbose: print(f"[CubeLogs.to_excel] create Excel file {output}, shape={self.shape}") time_mask &= len(self.data[self.time].unique()) > 2 cube_time = self.cube_time(fill_other_dates=True) if time_mask else None views = {k: k for k in views} if not isinstance(views, dict) else views f_highlights = {} plots = [] with pandas.ExcelWriter(output, engine="openpyxl") as writer: if main: assert main not in views, f"{main!r} is duplicated in views {sorted(views)}" df = self.describe().sort_values("name") if verbose: print(f"[CubeLogs.to_excel] add sheet {main!r} with shape {df.shape}") df.to_excel(writer, sheet_name=main, freeze_panes=(1, 1)) time_mask_view: Dict[str, pandas.DataFrame] = {} for name, view in views.items(): if view is None: continue df, tview = self.view(view, return_view_def=True, verbose=max(verbose - 1, 0)) if cube_time is not None: cube_mask = cube_time.view(view) aligned = align_dataframe_with(cube_mask, df) if aligned is not None: assert aligned.shape == df.shape, ( f"Shape mismatch between the view {df.shape} and the mask " f"{time_mask_view[name].shape}" ) time_mask_view[name] = aligned if verbose: print( f"[CubeLogs.to_excel] compute mask for view {name!r} " f"with shape {aligned.shape}" ) if tview is None: continue memory = df.memory_usage(deep=True).sum() if verbose: print( f"[CubeLogs.to_excel] add sheet {name!r} with shape " f"{df.shape} ({memory} bytes), index={df.index.names}, " f"columns={df.columns.names}" ) if self.time in df.columns.names: # Let's convert the time into str fr = df.columns.to_frame() if is_datetime64_any_dtype(fr[self.time]): dt = fr[self.time] has_time = (dt != dt.dt.normalize()).any() sdt = dt.apply( lambda t, has_time=has_time: t.strftime( "%Y-%m-%dT%H-%M-%S" if has_time else "%Y-%m-%d" ) ) fr[self.time] = sdt df.columns = pandas.MultiIndex.from_frame(fr) if csv and name in csv: name_csv = f"{output}.{name}.csv" if verbose: print(f"[CubeLogs.to_excel] saving sheet {name!r} in {name_csv!r}") df.reset_index(drop=False).to_csv(f"{output}.{name}.csv", index=False) if memory > 2**22: msg = ( f"[CubeLogs.to_excel] skipping {name!r}, " f"too big for excel with {memory} bytes" ) if verbose: print(msg) else: warnings.warn(msg, category=RuntimeWarning, stacklevel=0) else: df.to_excel( writer, sheet_name=name, freeze_panes=(df.columns.nlevels + df.index.nlevels, df.index.nlevels), ) f_highlights[name] = tview.f_highlight if tview.plots: plots.append( CubePlot( df, kind="line", orientation="row", split=True, timeseries=self.time, ) if self.time in df.columns.names else CubePlot(df, kind="barh", orientation="row", split=True) ) if raw: assert main not in views, f"{main!r} is duplicated in views {sorted(views)}" # Too long. # self._apply_excel_style(raw, writer, self.data) if csv and "raw" in csv: df.reset_index(drop=False).to_csv(f"{output}.raw.csv", index=False) memory = df.memory_usage(deep=True).sum() if memory > 2**22: msg = ( f"[CubeLogs.to_excel] skipping 'raw', " f"too big for excel with {memory} bytes" ) if verbose: print(msg) else: warnings.warn(msg, category=RuntimeWarning, stacklevel=0) else: if verbose: print(f"[CubeLogs.to_excel] add sheet 'raw' with shape {self.shape}") self.data.to_excel( writer, sheet_name="raw", freeze_panes=(1, 1), index=True ) if sbs: if verbose: for k, v in sbs.items(): print(f"[CubeLogs.to_excel] sbs {k}: {v}") name = "∧".join(sbs) sbs_raw, sbs_agg = self.sbs(sbs) if verbose: print(f"[CubeLogs.to_excel] add sheet {name!r} with shape {sbs_raw.shape}") print( f"[CubeLogs.to_excel] add sheet '{name}-AGG' " f"with shape {sbs_agg.shape}" ) sbs_raw = sbs_raw.reset_index(drop=False) sbs_raw.to_excel( writer, sheet_name=name, freeze_panes=( sbs_raw.columns.nlevels + sbs_raw.index.nlevels, sbs_raw.index.nlevels, ), ) sbs_agg.to_excel( writer, sheet_name=f"{name}-AGG", freeze_panes=( sbs_agg.columns.nlevels + sbs_agg.index.nlevels, sbs_agg.index.nlevels, ), ) if plots: from openpyxl.drawing.image import Image if verbose: print(f"[CubeLogs.to_excel] plots {len(plots)} plots") sheet = writer.book.create_sheet("plots") pos = 0 empty_row = 1 times = self.data[self.time].dropna() mini, maxi = times.min(), times.max() title_suffix = (str(mini) if mini == maxi else f"{mini}-{maxi}").replace( " 00:00:00", "" ) for plot in plots: imgs = plot.to_images( verbose=verbose, merge=True, title_suffix=title_suffix ) for img in imgs: y = (pos // 2) * 16 loc = f"A{y}" if pos % 2 == 0 else f"M{y}" sheet.add_image(Image(io.BytesIO(img)), loc) if verbose: no = f"{output}.png" print(f"[CubeLogs.to_excel] dump graphs into {no!r}") with open(no, "wb") as f: f.write(img) pos += 1 empty_row += len(plots) + 2 if verbose: print(f"[CubeLogs.to_excel] applies style to {output!r}") apply_excel_style( writer, f_highlights, time_mask_view=time_mask_view, verbose=verbose # type: ignore[arg-type] ) if verbose: print(f"[CubeLogs.to_excel] done with {len(views)} views")
[docs] def cube_time(self, fill_other_dates: bool = False, threshold: float = 1.2) -> "CubeLogs": """ Aggregates the data over time to detect changes on the last value. If *fill_other_dates* is True, all dates are kept, but values are filled with 0. *threshold* determines the bandwidth within the values are expected, should be a factor of the standard deviation. """ unique_time = self.data[self.time].unique() assert len(unique_time) > 2, f"Not enough dates to proceed: unique_time={unique_time}" gr = self.data[[*self.keys_no_time, *self.values]].groupby( self.keys_no_time, dropna=False ) dgr = gr.agg( lambda series, th=threshold: int(breaking_last_point(series, threshold=th)[0]) ) tm = unique_time.max() assert dgr.shape[0] > 0, ( f"Unexpected output shape={dgr.shape}, unique_time={unique_time}, " f"data.shape={self.data.shape}" ) dgr[self.time] = tm if fill_other_dates: other_df = [] other_dates = [t for t in unique_time if t != tm] for t in other_dates: df = dgr.copy() df[self.time] = t for c in df.columns: if c != self.time: df[c] = 0 other_df.append(df) dgr = pandas.concat([dgr, *other_df], axis=0) assert dgr.shape[0] > 0, ( f"Unexpected output shape={dgr.shape}, unique_time={unique_time}, " f"data.shape={self.data.shape}, " f"other_df shapes={[df.shape for df in other_df]}" ) return self.clone(data=dgr.reset_index(drop=False))
[docs] def sbs( self, configs: Dict[str, Dict[str, Any]], column_name: str = "CONF" ) -> Tuple[pandas.DataFrame, pandas.DataFrame]: """ Creates a side-by-side for two configurations. Every configuration a dictionary column:value which filters in the rows to keep in order to compute the side by side. Every configuration is given a name (the key in configs), it is added in column column_name. :param configs: example ``dict(CFA=dict(exporter="E1", opt="O"), CFB=dict(exporter="E2", opt="O"))`` :param column_name: column to add with the name of the configuration :return: data and aggregated date """ assert ( len(configs) >= 2 ), f"A side by side needs at least two configs but configs={configs}" set_keys_time = set(self.keys_time) columns_index = None data_list = [] for name_conf, conf in configs.items(): if columns_index is None: columns_index = list(conf.keys()) assert set(columns_index) <= set_keys_time, ( f"Configuration {conf} includes columns outside the keys " f"{', '.join(sorted(set_keys_time))}" ) else: assert set(columns_index) == set(conf), ( f"Every conf should share the same keys but conf={conf} " f"is different from {set(columns_index)}" ) data = self.data for k, v in conf.items(): data = data[data[k] == v] assert data.shape[0] > 0, f"No rows found for conf={conf}" assert ( column_name not in data.columns ), f"column_name={column_name!r} is already in {data.columns}" data = data.copy() data[column_name] = name_conf data_list.append(data) new_data = pandas.concat(data_list, axis=0) cube = self.clone(new_data, keys=[*self.keys_no_time, column_name]) key_index = set(self.keys_time) - {*columns_index, column_name} # type: ignore[misc] view = CubeViewDef( key_index=set(key_index), # type: ignore[arg-type] name="sbs", values=cube.values, keep_columns_in_index=[self.time], ) view_res = cube.view(view) assert isinstance(view_res, pandas.DataFrame), "not needed but mypy complains" # add metrics index_column_name = list(view_res.columns.names).index(column_name) index_metrics = list(view_res.columns.names).index("METRICS") def _mkc(m, s): c = ["" for c in view_res.columns.names] c[index_column_name] = s c[index_metrics] = m return tuple(c) list_configs = list(configs.items()) mean_columns = [ c for c in view_res.columns if pandas.api.types.is_numeric_dtype(view_res[c]) and not pandas.api.types.is_object_dtype(view_res[c]) ] assert mean_columns, f"No numerical columns in {view_res.dtypes}" view_res = view_res[mean_columns].copy() metrics = sorted(set(c[index_metrics] for c in view_res.columns)) assert metrics, ( f"No numerical metrics detected in " f"view_res.columns.names={view_res.columns.names}, " f"columns={view_res.dtypes}" ) sum_columns = [] columns_to_add = [] for i in range(len(list_configs)): for j in range(i + 1, len(list_configs)): for m in metrics: iname, ci = list_configs[i] jname, cj = list_configs[j] ci = ci.copy() cj = cj.copy() ci["METRICS"] = m cj["METRICS"] = m ci["CONF"] = iname cj["CONF"] = jname ci_name = tuple(ci[n] for n in view_res.columns.names) cj_name = tuple(cj[n] for n in view_res.columns.names) assert ci_name in view_res.columns or cj_name in view_res.columns, ( f"Unable to find column {ci_name} or {cj_name} " f"in columns {view_res.columns}, metrics={metrics}" ) if ci_name not in view_res.columns or cj_name not in view_res.columns: # One config does not have such metric. continue si = view_res[ci_name] sj = view_res[cj_name] sinan = si.isna() sjnan = sj.isna() n1 = iname n2 = jname nas = pandas.DataFrame( { _mkc(m, f"∅{n1}∧∅{n2}"): (sinan & sjnan).astype(int), _mkc(m, f"∅{n1}{n2}"): (sinan & ~sjnan).astype(int), _mkc(m, f"{n1}∧∅{n2}"): (~sinan & sjnan).astype(int), _mkc(m, f"{n1}{n2}"): (~sinan & ~sjnan).astype(int), _mkc(m, f"{n1}<{n2}"): (si < sj).astype(int), _mkc(m, f"{n1}=={n2}"): (si == sj).astype(int), _mkc(m, f"{n1}>{n2}"): (si > sj).astype(int), } ) nas.columns.names = view_res.columns.names columns_to_add.append(nas) sum_columns.extend(nas.columns) view_res = pandas.concat([view_res, *columns_to_add], axis=1) res = view_res.stack("METRICS", future_stack=True) # type: ignore[union-attr] res = res.reorder_levels( [res.index.nlevels - 1, *list(range(res.index.nlevels - 1))] ).sort_index() # aggregated metrics aggs = { **{k: "mean" for k in mean_columns}, # noqa: C420 **{k: "sum" for k in sum_columns}, # noqa: C420 } flat = view_res.groupby(self.time).agg(aggs) flat = flat.stack("METRICS", future_stack=True) return res, flat
[docs] class CubeLogsPerformance(CubeLogs): """ Processes logs coming from experiments. """ def __init__( self, data: Any, time: str = "DATE", keys: Sequence[str] = ( "^version_.*", "^model_.*", "device", "opt_patterns", "suite", "memory_peak", "machine", "exporter", "dynamic", "rtopt", "dtype", "device", "architecture", ), values: Sequence[str] = ( "^time_.*", "^disc.*", "^ERR_.*", "CMD", "^ITER", "^onnx_.*", "^op_onnx_.*", "^peak_gpu_.*", ), ignored: Sequence[str] = ("version_python",), recent: bool = True, formulas: Optional[ Union[ Sequence[str], Dict[str, Union[str, Callable[[pandas.DataFrame], pandas.Series]]], ] ] = ( "speedup", "bucket[speedup]", "ERR1", "n_models", "n_model_eager", "n_model_running", "n_model_acc01", "n_model_acc001", "n_model_dynamic", "n_model_pass", "n_model_faster", "n_model_faster2x", "n_model_faster3x", "n_model_faster4x", "n_node_attention", "n_node_control_flow", "n_node_scatter", "n_node_function", "n_node_initializer", "n_node_initializer_small", "n_node_constant", "n_node_shape", "n_node_expand", "onnx_n_nodes_no_cst", "peak_gpu_torch", "peak_gpu_nvidia", "time_export_unbiased", ), fill_missing: Optional[Sequence[Tuple[str, Any]]] = (("model_attn_impl", "eager"),), keep_last_date: bool = False, ): super().__init__( data=data, time=time, keys=keys, values=values, ignored=ignored, recent=recent, formulas=formulas, fill_missing=fill_missing, keep_last_date=keep_last_date, )
[docs] def clone( self, data: Optional[pandas.DataFrame] = None, keys: Optional[Sequence[str]] = None ) -> "CubeLogs": """ Makes a copy of the dataframe. It copies the processed data not the original one. keys can be changed as well. """ cube = self.__class__( data if data is not None else self.data.copy(), time=self.time, keys=keys or self.keys_no_time, values=self.values, recent=False, ) cube.load() return cube
def _process_formula( self, formula: Union[str, Callable[[pandas.DataFrame], pandas.Series]] ) -> Callable[[pandas.DataFrame], pandas.Series]: """ Processes a formula, converting it into a function. :param formula: a formula string :return: a function """ if callable(formula): return formula assert isinstance( formula, str ), f"Unexpected type for formula {type(formula)}: {formula!r}" def gdf(df, cname, default_value=np.nan): if cname in df.columns: return df[cname] return pandas.Series(default_value, index=df.index) def ghas_value(df, cname): if cname not in df.columns: return pandas.Series(np.nan, index=df.index) isna = df[cname].isna() return pandas.Series(np.where(isna, np.nan, 1.0), index=df.index) def gpreserve(df, cname, series): if cname not in df.columns: return pandas.Series(np.nan, index=df.index) isna = df[cname].isna() return pandas.Series(np.where(isna, np.nan, series), index=df.index).astype(float) if formula == "speedup": columns = set(self._filter_column(["^time_.*"], self.data.columns)) assert "time_latency" in columns and "time_latency_eager" in columns, ( f"Unable to apply formula {formula!r}, with columns\n" f"{pprint.pformat(sorted(columns))}" ) return lambda df: df["time_latency_eager"] / df["time_latency"] if formula == "bucket[speedup]": columns = set(self._filter_column(["^time_.*", "speedup"], self.data.columns)) assert "speedup" in columns, ( f"Unable to apply formula {formula!r}, with columns\n" f"{pprint.pformat(sorted(columns))}" ) # return lambda df: df["time_latency_eager"] / df["time_latency"] return lambda df: pandas.cut( df["speedup"], bins=BUCKET_SCALES, right=False, duplicates="raise" ) if formula == "ERR1": columns = set(self._filter_column(["^ERR_.*"], self.data.columns)) if not columns: return lambda df: np.nan def first_err(df: pandas.DataFrame) -> pandas.Series: ordered = [ c for c in [ "ERR_timeout", "ERR_load", "ERR_feeds", "ERR_warmup_eager", "ERR_export", "ERR_ort", "ERR_warmup", # "ERR_std", # "ERR_crash", # "ERR_stdout", ] if c in df.columns ] res = None for c in ordered: if res is None: res = df[c].fillna("") else: res = pandas.Series(np.where(res != "", res, df[c].fillna(""))) return res return first_err if formula.startswith("n_"): lambdas = dict( n_models=lambda df: ghas_value(df, "model_name"), n_model_eager=lambda df: ghas_value(df, "time_latency_eager"), n_model_running=lambda df: ghas_value(df, "time_latency"), n_model_acc01=lambda df: gpreserve( df, "discrepancies_abs", (gdf(df, "discrepancies_abs") <= 0.1) ), n_model_acc001=lambda df: gpreserve( df, "discrepancies_abs", gdf(df, "discrepancies_abs") <= 0.01 ), n_model_dynamic=lambda df: gpreserve( df, "discrepancies_dynamic_abs", (gdf(df, "discrepancies_dynamic_abs") <= 0.1), ), n_model_pass=lambda df: gpreserve( df, "time_latency", (gdf(df, "discrepancies_abs", np.inf) < 0.1) & (gdf(df, "time_latency_eager") > gdf(df, "time_latency", np.inf) * 0.98), ), n_model_faster=lambda df: gpreserve( df, "time_latency", gdf(df, "time_latency_eager") > gdf(df, "time_latency", np.inf) * 0.98, ), n_model_faster2x=lambda df: gpreserve( df, "time_latency", gdf(df, "time_latency_eager") > gdf(df, "time_latency", np.inf) * 1.98, ), n_model_faster3x=lambda df: gpreserve( df, "time_latency", gdf(df, "time_latency_eager") > gdf(df, "time_latency", np.inf) * 2.98, ), n_model_faster4x=lambda df: gpreserve( df, "time_latency", gdf(df, "time_latency_eager") > gdf(df, "time_latency", np.inf) * 3.98, ), n_node_attention=lambda df: gpreserve( df, "op_onnx_com.microsoft_Attention", gdf(df, "op_onnx_com.microsoft_Attention") + gdf(df, "op_onnx_com.microsoft_MultiHeadAttention"), ), n_node_control_flow=lambda df: gpreserve( df, "op_onnx__If", ( gdf(df, "op_onnx__If", 0) + gdf(df, "op_onnx__Scan", 0) + gdf(df, "op_onnx__Loop", 0) ), ), n_node_scatter=lambda df: gpreserve( df, "op_onnx__ScatterND", gdf(df, "op_onnx__ScatterND", 0) + gdf(df, "op_onnx__ScatterElements", 0), ), n_node_function=lambda df: gpreserve( df, "onnx_n_functions", gdf(df, "onnx_n_functions") ), n_node_initializer_small=lambda df: gpreserve( df, "op_onnx_initializer_small", gdf(df, "op_onnx_initializer_small") ), n_node_initializer=lambda df: gpreserve( df, "onnx_n_initializer", gdf(df, "onnx_n_initializer") ), n_node_constant=lambda df: gpreserve( df, "op_onnx__Constant", gdf(df, "op_onnx__Constant") ), n_node_shape=lambda df: gpreserve( df, "op_onnx__Shape", gdf(df, "op_onnx__Shape") ), n_node_expand=lambda df: gpreserve( df, "op_onnx__Expand", gdf(df, "op_onnx__Expand") ), ) assert ( formula in lambdas ), f"Unexpected formula={formula!r}, should be in {sorted(lambdas)}" return lambdas[formula] if formula == "onnx_n_nodes_no_cst": return lambda df: gdf(df, "onnx_n_nodes", 0) - gdf( df, "op_onnx__Constant", 0 ).fillna(0) if formula == "peak_gpu_torch": return lambda df: gdf(df, "mema_gpu_5_after_export") - gdf(df, "mema_gpu_4_reset") if formula == "peak_gpu_nvidia": return ( lambda df: (gdf(df, "memory_gpu0_peak") - gdf(df, "memory_gpu0_begin")) * 2**20 ) if formula == "time_export_unbiased": def unbiased_export(df): if "time_warmup_first_iteration" not in df.columns: return pandas.Series(np.nan, index=df.index) return pandas.Series( np.where( df["exporter"] == "inductor", df["time_warmup_first_iteration"] + df["time_export_success"], df["time_export_success"], ), index=df.index, ) return lambda df: gpreserve(df, "time_warmup_first_iteration", unbiased_export(df)) raise ValueError( f"Unexpected formula {formula!r}, available columns are\n" f"{pprint.pformat(sorted(self.data.columns))}" )
[docs] def view( self, view_def: Optional[Union[str, CubeViewDef]], return_view_def: bool = False, verbose: int = 0, ) -> Union[ Optional[pandas.DataFrame], Tuple[Optional[pandas.DataFrame], Optional[CubeViewDef]] ]: """ Returns a dataframe, a pivot view. If view_def is a string, it is replaced by a prefined view. :param view_def: view definition or a string :param return_view_def: returns the view definition as well :param verbose: verbosity level :return: dataframe or a couple (dataframe, view definition), both of them can be one if view_def cannot be interpreted """ assert view_def is not None, "view_def is None, this is not allowed." if isinstance(view_def, str): view_def = self.make_view_def(view_def) if view_def is None: return (None, None) if return_view_def else None return super().view(view_def, return_view_def=return_view_def, verbose=verbose)
[docs] def make_view_def(self, name: str) -> Optional[CubeViewDef]: """ Returns a view definition. :param name: name of the view :return: a CubeViewDef or None if name does not make sense Available views: * **agg-suite:** aggregation per suite * **disc:** discrepancies * **speedup:** speedup * **bucket_speedup:** speedup in buckets * **time:** latency * **time_export:** time to export * **counts:** status, running, faster, has control flow, ... * **err:** important errors * **cmd:** command lines * **raw-short:** raw data without all the unused columns """ fs = ["suite", "model_suite", "task", "model_name", "model_task"] index_cols = self._filter_column(fs, self.keys_time) assert index_cols, ( f"No index columns found for {fs!r} in " f"{pprint.pformat(sorted(self.keys_time))}" ) index_cols = [c for c in fs if c in set(index_cols)] f_speedup = lambda x: ( # noqa: E731 CubeViewDef.HighLightKind.NONE if not isinstance(x, (float, int)) else ( CubeViewDef.HighLightKind.RED if x < 0.9 else ( CubeViewDef.HighLightKind.GREEN if x > 1.1 else CubeViewDef.HighLightKind.NONE ) ) ) f_disc = lambda x: ( # noqa: E731 CubeViewDef.HighLightKind.NONE if not isinstance(x, (float, int)) else ( CubeViewDef.HighLightKind.RED if x > 0.1 else ( CubeViewDef.HighLightKind.GREEN if x < 0.01 else CubeViewDef.HighLightKind.NONE ) ) ) f_bucket = lambda x: ( # noqa: E731 CubeViewDef.HighLightKind.NONE if not isinstance(x, str) else ( CubeViewDef.HighLightKind.RED if x in {"[-inf, 0.8)", "[0.8, 0.9)", "[0.9, 0.95)"} else ( CubeViewDef.HighLightKind.NONE if x in {"[0.95, 0.98)", "[0.98, 1.02)", "[1.02, 1.05)"} else ( CubeViewDef.HighLightKind.GREEN if "[" in x else CubeViewDef.HighLightKind.NONE ) ) ) ) def mean_weight(gr): weight = gr["time_latency_eager"] x = gr["speedup"] if x.shape[0] == 0: return np.nan div = weight.sum() if div > 0: return (x * weight).sum() / div return np.nan def mean_geo(gr): x = gr["speedup"] return np.exp(np.log(x.dropna()).mean()) order = ["model_attn_impl", "exporter", "opt_patterns", "DATE"] implemented_views = { "agg-suite": lambda: CubeViewDef( key_index=index_cols, values=self._filter_column( [ "TIME_ITER", "speedup", "time_latency", "time_latency_eager", "time_export_success", "time_export_unbiased", "^n_.*", "target_opset", "onnx_filesize", "onnx_weight_size_torch", "onnx_weight_size_proto", "onnx_n_nodes", "onnx_n_nodes_no_cst", "op_onnx__Constant", "peak_gpu_torch", "peak_gpu_nvidia", ], self.values, ), ignore_unique=True, key_agg=["model_name", "task", "model_task"], agg_args=lambda column_name: "sum" if column_name.startswith("n_") else "mean", agg_multi={"speedup_weighted": mean_weight, "speedup_geo": mean_geo}, keep_columns_in_index=["suite"], name="agg-suite", order=order, ), "agg-all": lambda: CubeViewDef( key_index=index_cols, values=self._filter_column( [ "TIME_ITER", "speedup", "time_latency", "time_latency_eager", "time_export_success", "time_export_unbiased", "^n_.*", "target_opset", "onnx_filesize", "onnx_weight_size_torch", "onnx_weight_size_proto", "onnx_n_nodes", "onnx_n_nodes_no_cst", "peak_gpu_torch", "peak_gpu_nvidia", ], self.values, ), ignore_unique=True, key_agg=["model_name", "task", "model_task", "suite"], agg_args=lambda column_name: "sum" if column_name.startswith("n_") else "mean", agg_multi={"speedup_weighted": mean_weight, "speedup_geo": mean_geo}, name="agg-all", order=order, plots=True, ), "disc": lambda: CubeViewDef( key_index=index_cols, values=self._filter_column(["discrepancies_abs"], self.values), ignore_unique=True, keep_columns_in_index=["suite"], f_highlight=f_disc, name="disc", order=order, ), "speedup": lambda: CubeViewDef( key_index=index_cols, values=self._filter_column(["speedup"], self.values), ignore_unique=True, keep_columns_in_index=["suite"], f_highlight=f_speedup, name="speedup", order=order, ), "counts": lambda: CubeViewDef( key_index=index_cols, values=self._filter_column(["^n_.*"], self.values), ignore_unique=True, keep_columns_in_index=["suite"], name="counts", order=order, ), "peak-gpu": lambda: CubeViewDef( key_index=index_cols, values=self._filter_column(["^peak_gpu_.*"], self.values), ignore_unique=True, keep_columns_in_index=["suite"], name="peak-gpu", order=order, ), "time": lambda: CubeViewDef( key_index=index_cols, values=self._filter_column( ["time_latency", "time_latency_eager"], self.values ), ignore_unique=True, keep_columns_in_index=["suite"], name="time", order=order, ), "time_export": lambda: CubeViewDef( key_index=index_cols, values=self._filter_column(["time_export_unbiased"], self.values), ignore_unique=True, keep_columns_in_index=["suite"], name="time_export", order=order, ), "err": lambda: CubeViewDef( key_index=index_cols, values=self._filter_column( ["ERR1", "ERR_timeout", "ERR_export", "ERR_crash"], self.values ), ignore_unique=True, keep_columns_in_index=["suite"], name="err", order=order, ), "bucket-speedup": lambda: CubeViewDef( key_index=index_cols, values=self._filter_column(["bucket[speedup]"], self.values), ignore_unique=True, keep_columns_in_index=["suite"], name="bucket-speedup", f_highlight=f_bucket, order=order, ), "onnx": lambda: CubeViewDef( key_index=index_cols, values=self._filter_column( [ "onnx_filesize", "onnx_n_nodes", "onnx_n_nodes_no_cst", "onnx_weight_size_proto", "onnx_weight_size_torch", "op_onnx_initializer_small", ], self.values, ), ignore_unique=True, keep_columns_in_index=["suite"], name="onnx", order=order, ), "raw-short": lambda: CubeViewDef( key_index=self.keys_time, values=[c for c in self.values if c not in {"ERR_std", "ERR_stdout"}], ignore_unique=False, keep_columns_in_index=["suite"], name="raw-short", no_index=True, ), } cmd_col = self._filter_column(["CMD"], self.values, can_be_empty=True) if cmd_col: implemented_views["cmd"] = lambda: CubeViewDef( key_index=index_cols, values=cmd_col, ignore_unique=True, keep_columns_in_index=["suite"], name="cmd", order=order, ) assert name in implemented_views or name in {"cmd"}, ( f"Unknown view {name!r}, expected a name in {sorted(implemented_views)}," f"\n--\nkeys={pprint.pformat(sorted(self.keys_time))}, " f"\n--\nvalues={pprint.pformat(sorted(self.values))}" ) if name not in implemented_views: return None return implemented_views[name]()
[docs] def post_load_process_piece( self, df: pandas.DataFrame, unique: bool = False ) -> pandas.DataFrame: df = super().post_load_process_piece(df, unique=unique) if unique: return df cols = self._filter_column(self._keys, df) res = None for c in cols: if df[c].isna().any(): # Missing values for keys are not supposed to happen. uniq = set(df[c].dropna()) if len(uniq) == 1: if res is None: res = df.copy() res[c] = res[c].fillna(uniq.pop()) return df if res is None else res