Spaces:
Runtime error
Runtime error
File size: 9,067 Bytes
2d90381 4dd5bdd 2d90381 4dd5bdd 2d90381 44be985 2d90381 44be985 2d90381 4dd5bdd 2d90381 4dd5bdd 44be985 4dd5bdd 44be985 4dd5bdd 44be985 4dd5bdd 2501733 4dd5bdd 2501733 4dd5bdd 2501733 4dd5bdd 2501733 4dd5bdd 2d90381 4dd5bdd 2d90381 4dd5bdd 2d90381 4dd5bdd 2d90381 4dd5bdd 2d90381 4dd5bdd 2d90381 4dd5bdd 2d90381 4dd5bdd 2d90381 4dd5bdd 2d90381 4dd5bdd 2d90381 4dd5bdd 2d90381 4dd5bdd 2d90381 4dd5bdd 2501733 4dd5bdd |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 |
import streamlit as st
st.title('Numerai Example Script')
# content below adapted from
# https://github.com/numerai/example-scripts/blob/master/example_model.py
#
import pandas as pd
from lightgbm import LGBMRegressor
import gc
import json
from pathlib import Path
import os
from numerapi import NumerAPI
from utils import (
save_model,
load_model,
neutralize,
get_biggest_change_features,
validation_metrics,
ERA_COL,
DATA_TYPE_COL,
TARGET_COL,
EXAMPLE_PREDS_COL
)
IS_RUNNING_IN_HUGGING_FACE = os.environ.get('HF_ENDPOINT') is not None
napi = NumerAPI()
current_round = napi.get_current_round()
# Tournament data changes every week so we specify the round in their name. Training
# and validation data only change periodically, so no need to download
# them every time.
Path("./v4").mkdir(parents=False, exist_ok=True)
@st.cache
def get_dataset_path():
if IS_RUNNING_IN_HUGGING_FACE:
from datasets import load_dataset_builder
ds_builder = load_dataset_builder("Numerati/numerai-datasets")
return ds_builder.cache_dir
else:
return "./v4"
@st.cache
def download_dataset():
print('download_dataset')
if IS_RUNNING_IN_HUGGING_FACE:
napi.download_dataset("v4/train.parquet")
napi.download_dataset("v4/validation.parquet")
napi.download_dataset("v4/validation_example_preds.parquet")
napi.download_dataset("v4/features.json")
napi.download_dataset("v4/live.parquet", f"v4/live_{current_round}.parquet")
print('done download_dataset')
@st.cache
def load_dataset(feature_set: str):
dataset_path = get_dataset_path()
print(f'load_dataset with feature_set {feature_set} and path {dataset_path}')
# read the feature metadata and get a feature set (or all the features)
with open(f"{dataset_path}/features.json", "r") as f:
feature_metadata = json.load(f)
# features = list(feature_metadata["feature_stats"].keys()) # get all the features
# features = feature_metadata["feature_sets"]["small"] # get the small
# feature set
features = feature_metadata["feature_sets"][feature_set] # get the medium feature set
# read in just those features along with era and target columns
read_columns = features + [ERA_COL, DATA_TYPE_COL, TARGET_COL]
# note: sometimes when trying to read the downloaded data you get an error about invalid magic parquet bytes...
# if so, delete the file and rerun the napi.download_dataset to fix the
# corrupted file
training_data = pd.read_parquet(f'{dataset_path}/train.parquet',
columns=read_columns)
validation_data = pd.read_parquet(f'{dataset_path}/validation.parquet',
columns=read_columns)
live_data = pd.read_parquet(f'v4/live_{current_round}.parquet',
columns=read_columns)
# pare down the number of eras to every 4th era
# every_4th_era = training_data[ERA_COL].unique()[::4]
# training_data = training_data[training_data[ERA_COL].isin(every_4th_era)]
# getting the per era correlation of each feature vs the target
all_feature_corrs = training_data.groupby(ERA_COL).apply(
lambda era: era[features].corrwith(era[TARGET_COL])
)
# find the riskiest features by comparing their correlation vs
# the target in each half of training data; we'll use these later
riskiest_features = get_biggest_change_features(all_feature_corrs, 50)
# "garbage collection" (gc) gets rid of unused data and frees up memory
gc.collect()
print('done with feature_set', feature_set)
return training_data, validation_data, live_data, features, riskiest_features
feature_set = st.selectbox(
'Which feature set should be used?',
('small', 'medium', 'fncv3_features', 'v2_equivalent_features', 'v3_equivalent_features'))
data_load_state = st.text('Loading data...')
download_dataset()
training_data, validation_data, live_data, features, riskiest_features = load_dataset(feature_set)
data_load_state.text('Loading data...done!')
st.subheader('Raw data')
st.write(training_data.head())
st.subheader('Model Configuration')
n_estimators = st.slider('n_estimators', 100, 10000, 2000)
learning_rate = st.slider('learning_rate', 0.0001, 0.1, 0.01)
max_depth = st.slider('max_depth', 2, 20, 5)
params = {"n_estimators": n_estimators,
"learning_rate": learning_rate,
"max_depth": max_depth,
"num_leaves": 2 ** 5,
"colsample_bytree": 0.1
}
model_name = f"model_target"
@st.cache
def get_model_and_fit(model_name, *params):
print('get_model_and_fit')
model = load_model(model_name)
if not model:
with st.spinner('Wait model training...'):
print(f"model not found, creating new one")
model = LGBMRegressor(**params)
# train on all of train and save the model so we don't have to
# train next time
model.fit(training_data.filter(like='feature_', axis='columns'),
training_data[TARGET_COL])
print(f"saving new model: {model_name}")
save_model(model, model_name)
st.success('Done model training!')
gc.collect()
print('done get_model_and_fit')
has_model_preds = False
@st.cache
def get_model_preds(model_name, *params):
print('get_model_preds')
model = load_model(model_name)
has_model_preds = False
nans_per_col = live_data[live_data["data_type"]
== "live"][features].isna().sum()
# check for nans and fill nans
if nans_per_col.any():
total_rows = len(live_data[live_data["data_type"] == "live"])
print(f"Number of nans per column this week: {nans_per_col[nans_per_col > 0]}")
print(f"out of {total_rows} total rows")
print(f"filling nans with 0.5")
live_data.loc[:, features] = live_data.loc[:, features].fillna(0.5)
else:
print("No nans in the features this week!")
# double check the feature that the model expects vs what is available to prevent our
# pipeline from failing if Numerai adds more data and we don't have time
# to retrain!
model_expected_features = model.booster_.feature_name()
if set(model_expected_features) != set(features):
print(f"New features are available! Might want to retrain model {model_name}.")
validation_data.loc[:, f"preds_{model_name}"] = model.predict(
validation_data.loc[:, model_expected_features])
live_data.loc[:, f"preds_{model_name}"] = model.predict(
live_data.loc[:, model_expected_features])
gc.collect()
# neutralize our predictions to the riskiest features
validation_data[f"preds_{model_name}_neutral_riskiest_50"] = neutralize(
df=validation_data,
columns=[f"preds_{model_name}"],
neutralizers=riskiest_features,
proportion=1.0,
normalize=True,
era_col=ERA_COL
)
live_data[f"preds_{model_name}_neutral_riskiest_50"] = neutralize(
df=live_data,
columns=[f"preds_{model_name}"],
neutralizers=riskiest_features,
proportion=1.0,
normalize=True,
era_col=ERA_COL
)
model_to_submit = f"preds_{model_name}_neutral_riskiest_50"
# rename best model to "prediction" and rank from 0 to 1 to meet upload
# requirements
validation_data["prediction"] = validation_data[model_to_submit].rank(pct=True)
live_data["prediction"] = live_data[model_to_submit].rank(pct=True)
validation_prediction_fname = f"validation_predictions_{current_round}.csv"
validation_data["prediction"].to_csv(validation_prediction_fname)
live_data["prediction"].to_csv(f"live_predictions_{current_round}.csv")
validation_preds = pd.read_parquet(f'{get_dataset_path()}/validation_example_preds.parquet')
validation_data[EXAMPLE_PREDS_COL] = validation_preds["prediction"]
# get some stats about each of our models to compare...
# fast_mode=True so that we skip some of the stats that are slower to calculate
print('start validation_metrics')
validation_stats = validation_metrics(validation_data, [model_to_submit, f"preds_{model_name}"], example_col=EXAMPLE_PREDS_COL, fast_mode=True, target_col=TARGET_COL)
st.markdown(validation_stats[["mean", "sharpe"]].to_markdown())
# st.write(f'''
# Done! Next steps:
# 1. Go to numer.ai/tournament (make sure you have an account)
# 2. Submit validation_predictions_{current_round}.csv to the diagnostics tool
# 3. Submit tournament_predictions_{current_round}.csv to the "Upload Predictions" button
# ''')
has_model_preds = True
st.button('Start model training', on_click=get_model_and_fit, args=[model_name, params])
st.button('Start model evaluation', on_click=get_model_preds, args=[model_name, params])
if has_model_preds:
st.download_button('Validation data for diagnostics tool', validation_data["prediction"], validation_prediction_fname)
|