File size: 6,520 Bytes
27411b0 |
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 |
from collections import defaultdict
from dataclasses import dataclass
from typing import Literal, Type
import pandas as pd
from typing import List
import gradio as gr
import numpy as np
import pandas as pd
from viewer.literals import REFERENCE_RUNS, TASK_CONSISTENCY_BUTTON_CLOSE_LABEL, TASK_CONSISTENCY_BUTTON_LABEL
@dataclass
class PlotOptions:
smoothing: int
interpolate: bool
pct: bool
merge_seeds: str
@dataclass(frozen=True)
class TaskInfo:
# Source file from which the task was fetched
filename: str
name: str
metrics: dict[str, float]
hashes: dict[str, str]
@dataclass(frozen=True)
class RunInfo:
name: str
seed: int
step: int
tasks: list[TaskInfo]
@property
def full_name(self):
return f"{self.name}-seed-{self.seed}" if not self.name.endswith("-") else self.name
RunData = list[RunInfo]
def get_run_name_seed(run_name):
if "-seed-" not in run_name:
return run_name, 42
run_name, seed = run_name.split("-seed-")
return run_name, int(seed)
def select_runs(df: pd.DataFrame, runs_to_include: list[str] | None = None, runs_to_exclude: list[str] | None = None):
conditions = pd.Series(True, index=df.index)
if runs_to_include:
conditions_include = [(df['runname'] == get_run_name_seed(run)[0]) & (df['seed'] == get_run_name_seed(run)[1]) for run in runs_to_include]
conditions = pd.concat(conditions_include, axis=1).any(axis=1)
if runs_to_exclude:
conditions_exclude = [(df['runname'] == get_run_name_seed(run)[0]) & (df['seed'] == get_run_name_seed(run)[1]) for run in runs_to_exclude]
conditions = ~pd.concat(conditions_exclude, axis=1).any(axis=1)
return df[conditions]
BASELINE_GROUPING_MODE = Literal["Mean", "Median", "Min", "Max"]
def get_groupped_score(df: pd.DataFrame, runs: list[str], groupping_mode: BASELINE_GROUPING_MODE):
if len(runs) == 0:
return pd.DataFrame(columns=df.columns)
tasks_or_agg = [col for col in df.columns if is_task_column(col) or is_aggregate_column(col)]
res = select_runs(df, runs_to_include=runs)
if groupping_mode == "Mean":
return res.groupby("steps")[tasks_or_agg].mean().reset_index()
elif groupping_mode == "Median":
return res.groupby("steps")[tasks_or_agg].median().reset_index()
elif groupping_mode == "Min":
return res.groupby("steps")[tasks_or_agg].min().reset_index()
elif groupping_mode == "Max":
return res.groupby("steps")[tasks_or_agg].max().reset_index()
def check_task_hash_consistency(run_data: RunData, check_task_consistency_button):
if not run_data or check_task_consistency_button == TASK_CONSISTENCY_BUTTON_CLOSE_LABEL:
return gr.update(value={}, visible=False), gr.update(value=TASK_CONSISTENCY_BUTTON_LABEL)
# Ignore the continuation tokens, as they vary with generative tasks
hash_keys = ["hash_examples", "hash_full_prompts"]
task_hashes = defaultdict(lambda: defaultdict(list))
for run in run_data:
for task_info in run.tasks:
hashes = task_info.hashes
hash_values = tuple(hashes.get(k) for k in hash_keys)
task_hashes[task_info.name][hash_values].append({
"name": run.name,
"step": run.step,
"filename": task_info.filename
})
conflicts = {}
for task, hash_groups in task_hashes.items():
if len(hash_groups) > 1:
conflicts[task] = [
{
"runs": runs,
"hashes": dict(zip(hash_keys, hash_values))
}
for hash_values, runs in hash_groups.items()
]
return gr.Json(value={"conflicts": conflicts}, visible=True), gr.Button(value=TASK_CONSISTENCY_BUTTON_CLOSE_LABEL)
def create_df_from_run_data(run_data: RunData):
df = pd.DataFrame([
{
"runname": run.name,
"seed": run.seed,
"steps": run.step,
"agg_score_micro": 0,
**{
f"{task_info.name}/{metric}": value
for task_info in run.tasks
for metric, value in task_info.metrics.items()
}
} for run in run_data
])
df = df.fillna(0)
return df
def is_task_column(column: str):
return "/" in column
def is_aggregate_column(column: str):
return column.startswith("agg_score")
def is_baseline_run(run: str):
return any(run.startswith(prefix) for prefix in ["random", "dummy", "baseline"])
def is_reference_run(run: str):
return any([ref_run + "-" in run for ref_run in REFERENCE_RUNS])
def z_score_normalize(df: pd.DataFrame, normalization_runs: List[str], columns: List[str], variability_window: int = 1) -> pd.DataFrame:
# without 2 runs we can't estimate the std
if len(normalization_runs) <= 1:
return df
normalization_df = select_runs(df, runs_to_include=normalization_runs)
# Group by steps and calculate mean and std for all columns at once
grouped = normalization_df.groupby('steps')[columns]
means = grouped.mean()
stds = grouped.std()
# Ensure we don't divide by zero
stds = stds.replace(0, 1)
# fetch values at the highest step
last_means = means.loc[means.index.max()]
# fetch and average the last N steps defined by the window size
last_window_stds = stds.sort_index(ascending=False).head(variability_window).mean()
df[columns] = (df[columns].sub(last_means[columns], axis=1)
.div(last_window_stds[columns], axis=1))
return df
def rescale_scores(df: pd.DataFrame, normalization_runs: List[str], columns: List[str]) -> pd.DataFrame:
baseline = get_groupped_score(df, normalization_runs, "Mean")
# Prepare baseline values and df for vectorized operation
baseline = baseline.set_index("steps").reindex(df["steps"].unique()).interpolate().reset_index()
rescaled_cols = baseline.columns[~((baseline <= 0.0).all() | (baseline == 1.0).all())]
rescaled_cols = rescaled_cols[(rescaled_cols != 'steps') & rescaled_cols.isin(columns)]
df_with_baseline = df.merge(baseline[list(rescaled_cols) + ['steps']], on=["steps"], how="left", suffixes=("", "_baseline")).fillna(0)
df[rescaled_cols] = df[rescaled_cols].sub(df_with_baseline[rescaled_cols + '_baseline'].values)
df[rescaled_cols] = df[rescaled_cols].div(1 - df_with_baseline[rescaled_cols + '_baseline'].values)
return df |