Files
drc-ners-nlp/core/utils/data_loader.py
T

169 lines
6.1 KiB
Python

import gc
import logging
from pathlib import Path
from typing import Optional, Union, Iterator, Dict
import pandas as pd
from core.config.pipeline_config import PipelineConfig
OPTIMIZED_DTYPES = {
# Numeric columns with appropriate bit-width
"year": "Int16", # Years fit in 16-bit integer
"words": "Int8", # Word counts typically < 128
"length": "Int16", # Name lengths fit in 16-bit
"annotated": "Int8", # Binary flag (0/1)
"ner_tagged": "Int8", # Binary flag (0/1)
# Categorical columns (memory efficient for repeated values)
"sex": "category",
"province": "category",
"region": "category",
"identified_category": "category",
"transformation_type": "category",
# String columns with proper string dtype
"name": "string",
"probable_native": "string",
"probable_surname": "string",
"identified_name": "string",
"identified_surname": "string",
"ner_entities": "string",
}
class DataLoader:
"""Reusable data loading utilities"""
def __init__(self, config: PipelineConfig, custom_dtypes: Optional[Dict] = None):
self.config = config
self.dtypes = {**OPTIMIZED_DTYPES, **(custom_dtypes or {})}
def load_csv_chunked(
self, filepath: Union[str, Path], chunk_size: Optional[int] = None
) -> Iterator[pd.DataFrame]:
"""Load CSV file in chunks for memory efficiency"""
chunk_size = chunk_size or self.config.processing.chunk_size
encodings = self.config.processing.encoding_options
filepath = Path(filepath)
for encoding in encodings:
try:
logging.info(f"Reading {filepath} with encoding: {encoding}")
# Read with optimal dtypes
chunk_iter = pd.read_csv(
filepath,
encoding=encoding,
chunksize=chunk_size,
on_bad_lines="skip",
dtype=self.dtypes,
)
for i, chunk in enumerate(chunk_iter):
logging.debug(f"Processing optimized chunk {i + 1}")
yield chunk
logging.info(f"Successfully read {filepath} with encoding: {encoding}")
return
except Exception as e:
logging.warning(f"Failed with encoding {encoding}: {e}")
continue
raise ValueError(f"Unable to decode {filepath} with any encoding: {encodings}")
def load_csv_complete(self, filepath: Union[str, Path]) -> pd.DataFrame:
"""Load complete CSV with memory optimization"""
chunks = []
for chunk in self.load_csv_chunked(filepath):
chunks.append(chunk)
if not chunks:
return pd.DataFrame()
logging.info(f"Concatenating {len(chunks)} optimized chunks")
df = pd.concat(chunks, ignore_index=True, copy=False)
# Cleanup chunks from memory
del chunks
gc.collect()
# Apply dataset size limiting if configured
if self.config.data.max_dataset_size is not None:
df = self._limit_dataset_size(df)
return df
def _limit_dataset_size(self, df: pd.DataFrame) -> pd.DataFrame:
"""Limit dataset size with optional sex balancing"""
max_size = self.config.data.max_dataset_size
if max_size is None or len(df) <= max_size:
return df
if self.config.data.balance_by_sex and "sex" in df.columns:
return self._balanced_sample(df, max_size)
else:
# Simple random sampling
return df.sample(n=max_size, random_state=self.config.data.random_seed)
def _balanced_sample(self, df: pd.DataFrame, max_size: int) -> pd.DataFrame:
"""Sample data with balanced sex distribution"""
# Get unique sex values
sex_values = df["sex"].dropna().unique()
if len(sex_values) == 0:
logging.warning(f"No valid values found in sex column 'sex', using random sampling")
return df.sample(n=max_size, random_state=self.config.data.random_seed)
# Calculate samples per sex category
samples_per_sex = max_size // len(sex_values)
remaining_samples = max_size % len(sex_values)
balanced_samples = []
for i, sex in enumerate(sex_values):
# Use boolean indexing instead of creating temporary DataFrames
sex_mask = df["sex"] == sex
sex_indices = df[sex_mask].index
# Distribute remaining samples to first categories
current_samples = samples_per_sex + (1 if i < remaining_samples else 0)
current_samples = min(current_samples, len(sex_indices))
if current_samples > 0:
# Sample indices instead of DataFrame
sampled_indices = pd.Series(sex_indices).sample(
n=current_samples, random_state=self.config.data.random_seed + i
)
balanced_samples.extend(sampled_indices.tolist())
logging.info(f"Sampled {current_samples} records for sex '{sex}'")
if not balanced_samples:
logging.warning("No balanced samples could be created, using random sampling")
return df.sample(n=max_size, random_state=self.config.data.random_seed)
# Create result using iloc with indices (no copying until final step)
result = df.iloc[balanced_samples].copy()
# Shuffle the final result
result = result.sample(frac=1, random_state=self.config.data.random_seed).reset_index(
drop=True
)
logging.info(f"Created balanced dataset with {len(result)} records from {len(df)} total")
return result
@classmethod
def save_csv(
cls, df: pd.DataFrame, filepath: Union[str, Path], create_dirs: bool = True
) -> None:
"""Save DataFrame to CSV with proper handling"""
filepath = Path(filepath)
if create_dirs:
filepath.parent.mkdir(parents=True, exist_ok=True)
df.to_csv(filepath, index=False, encoding="utf-8")
logging.info(f"Saved {len(df)} rows to {filepath}")