Merge pull request #7701 from freqtrade/add-single-precision-freqai

add option to force single precision

config_examples/config_full.example.json

@@ -204,6 +204,7 @@
     "strategy_path": "user_data/strategies/",
     "recursive_strategy_search": false,
     "add_config_files": [],
+    "reduce_df_footprint": false,
     "dataformat_ohlcv": "json",
     "dataformat_trades": "jsongz"
 }

docs/configuration.md

@@ -253,6 +253,7 @@ Mandatory parameters are marked as **Required**, which means that they are requi
 | `add_config_files` | Additional config files. These files will be loaded and merged with the current config file. The files are resolved relative to the initial file.<br> *Defaults to `[]`*. <br> **Datatype:** List of strings
 | `dataformat_ohlcv` | Data format to use to store historical candle (OHLCV) data. <br> *Defaults to `json`*. <br> **Datatype:** String
 | `dataformat_trades` | Data format to use to store historical trades data. <br> *Defaults to `jsongz`*. <br> **Datatype:** String
+| `reduce_df_footprint` | Recast all numeric columns to float32/int32, with the objective of reducing ram/disk usage (and decreasing train/inference timing in FreqAI). (Currently only affects FreqAI use-cases) <br> **Datatype:** Boolean. <br> Default: `False`.
 
 ### Parameters in the strategy
 

docs/freqai-parameter-table.md

@@ -50,3 +50,4 @@ Mandatory parameters are marked as **Required** and have to be set in one of the
 |  |  **Extraneous parameters**
 | `keras` | If the selected model makes use of Keras (typical for Tensorflow-based prediction models), this flag needs to be activated so that the model save/loading follows Keras standards. <br> **Datatype:** Boolean. <br> Default: `False`.
 | `conv_width` | The width of a convolutional neural network input tensor. This replaces the need for shifting candles (`include_shifted_candles`) by feeding in historical data points as the second dimension of the tensor. Technically, this parameter can also be used for regressors, but it only adds computational overhead and does not change the model training/prediction. <br> **Datatype:** Integer. <br> Default: `2`.
+| `reduce_df_footprint` | Recast all numeric columns to float32/int32, with the objective of reducing ram/disk usage and decreasing train/inference timing. This parameter is set in the main level of the Freqtrade configuration file (not inside FreqAI). <br> **Datatype:** Boolean. <br> Default: `False`.

freqtrade/constants.py

@@ -159,6 +159,7 @@ CONF_SCHEMA = {
         'ignore_buying_expired_candle_after': {'type': 'number'},
         'trading_mode': {'type': 'string', 'enum': TRADING_MODES},
         'margin_mode': {'type': 'string', 'enum': MARGIN_MODES},
+        'reduce_df_footprint': {'type': 'boolean', 'default': False},
         'liquidation_buffer': {'type': 'number', 'minimum': 0.0, 'maximum': 0.99},
         'backtest_breakdown': {
             'type': 'array',

freqtrade/data/converter.py

@@ -7,6 +7,7 @@ from datetime import datetime, timezone
 from operator import itemgetter
 from typing import Dict, List
 
+import numpy as np
 import pandas as pd
 from pandas import DataFrame, to_datetime
 
@@ -313,3 +314,29 @@ def convert_ohlcv_format(
                 if erase and convert_from != convert_to:
                     logger.info(f"Deleting source data for {pair} / {timeframe}")
                     src.ohlcv_purge(pair=pair, timeframe=timeframe, candle_type=candle_type)
+
+
+def reduce_dataframe_footprint(df: DataFrame) -> DataFrame:
+    """
+    Ensure all values are float32 in the incoming dataframe.
+    :param df: Dataframe to be converted to float/int 32s
+    :return: Dataframe converted to float/int 32s
+    """
+
+    logger.debug(f"Memory usage of dataframe is "
+                 f"{df.memory_usage().sum() / 1024**2:.2f} MB")
+
+    df_dtypes = df.dtypes
+    for column, dtype in df_dtypes.items():
+        if column in ['open', 'high', 'low', 'close', 'volume']:
+            continue
+        if dtype == np.float64:
+            df_dtypes[column] = np.float32
+        elif dtype == np.int64:
+            df_dtypes[column] = np.int32
+    df = df.astype(df_dtypes)
+
+    logger.debug(f"Memory usage after optimization is: "
+                 f"{df.memory_usage().sum() / 1024**2:.2f} MB")
+
+    return df

freqtrade/freqai/data_kitchen.py

@@ -19,6 +19,7 @@ from sklearn.neighbors import NearestNeighbors
 
 from freqtrade.configuration import TimeRange
 from freqtrade.constants import Config
+from freqtrade.data.converter import reduce_dataframe_footprint
 from freqtrade.exceptions import OperationalException
 from freqtrade.exchange import timeframe_to_seconds
 from freqtrade.strategy.interface import IStrategy
@@ -1275,6 +1276,9 @@ class FreqaiDataKitchen:
 
         dataframe = self.remove_special_chars_from_feature_names(dataframe)
 
+        if self.config.get('reduce_df_footprint', False):
+            dataframe = reduce_dataframe_footprint(dataframe)
+
         return dataframe
 
     def fit_labels(self) -> None:

tests/data/test_converter.py

@@ -3,18 +3,19 @@ import logging
 from pathlib import Path
 from shutil import copyfile
 
+import numpy as np
 import pytest
 
 from freqtrade.configuration.timerange import TimeRange
 from freqtrade.data.converter import (convert_ohlcv_format, convert_trades_format,
                                       ohlcv_fill_up_missing_data, ohlcv_to_dataframe,
-                                      trades_dict_to_list, trades_remove_duplicates,
-                                      trades_to_ohlcv, trim_dataframe)
+                                      reduce_dataframe_footprint, trades_dict_to_list,
+                                      trades_remove_duplicates, trades_to_ohlcv, trim_dataframe)
 from freqtrade.data.history import (get_timerange, load_data, load_pair_history,
                                     validate_backtest_data)
 from freqtrade.data.history.idatahandler import IDataHandler
 from freqtrade.enums import CandleType
-from tests.conftest import log_has, log_has_re
+from tests.conftest import generate_test_data, log_has, log_has_re
 from tests.data.test_history import _clean_test_file
 
 
@@ -344,3 +345,33 @@ def test_convert_ohlcv_format(default_conf, testdatadir, tmpdir, file_base, cand
         assert file.exists()
     for file in (files_new):
         assert not file.exists()
+
+
+def test_reduce_dataframe_footprint():
+    data = generate_test_data('15m', 40)
+
+    data['open_copy'] = data['open']
+    data['close_copy'] = data['close']
+    data['close_copy'] = data['close']
+
+    assert data['open'].dtype == np.float64
+    assert data['open_copy'].dtype == np.float64
+    assert data['close_copy'].dtype == np.float64
+
+    df2 = reduce_dataframe_footprint(data)
+
+    # Does not modify original dataframe
+    assert data['open'].dtype == np.float64
+    assert data['open_copy'].dtype == np.float64
+    assert data['close_copy'].dtype == np.float64
+
+    # skips ohlcv columns
+    assert df2['open'].dtype == np.float64
+    assert df2['high'].dtype == np.float64
+    assert df2['low'].dtype == np.float64
+    assert df2['close'].dtype == np.float64
+    assert df2['volume'].dtype == np.float64
+
+    # Changes dtype of returned dataframe
+    assert df2['open_copy'].dtype == np.float32
+    assert df2['close_copy'].dtype == np.float32

tests/freqai/test_freqai_interface.py

@@ -27,13 +27,13 @@ def is_mac() -> bool:
     return "Darwin" in machine
 
 
-@pytest.mark.parametrize('model, pca, dbscan', [
-    ('LightGBMRegressor', True, False),
-    ('XGBoostRegressor', False, True),
-    ('XGBoostRFRegressor', False, False),
-    ('CatboostRegressor', False, False),
+@pytest.mark.parametrize('model, pca, dbscan, float32', [
+    ('LightGBMRegressor', True, False, True),
+    ('XGBoostRegressor', False, True, False),
+    ('XGBoostRFRegressor', False, False, False),
+    ('CatboostRegressor', False, False, False),
     ])
-def test_extract_data_and_train_model_Standard(mocker, freqai_conf, model, pca, dbscan):
+def test_extract_data_and_train_model_Standard(mocker, freqai_conf, model, pca, dbscan, float32):
     if is_arm() and model == 'CatboostRegressor':
         pytest.skip("CatBoost is not supported on ARM")
 
@@ -43,6 +43,7 @@ def test_extract_data_and_train_model_Standard(mocker, freqai_conf, model, pca,
     freqai_conf.update({"strategy": "freqai_test_strat"})
     freqai_conf['freqai']['feature_parameters'].update({"principal_component_analysis": pca})
     freqai_conf['freqai']['feature_parameters'].update({"use_DBSCAN_to_remove_outliers": dbscan})
+    freqai_conf.update({"reduce_df_footprint": float32})
 
     strategy = get_patched_freqai_strategy(mocker, freqai_conf)
     exchange = get_patched_exchange(mocker, freqai_conf)