File size: 7,142 Bytes
7a87926 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 |
"""
Advanced data loading optimizations.
Features:
- Prefetching with multiple workers
- Memory-mapped datasets
- Smart batching strategies
- Data pipeline profiling
"""
import logging
import time
from typing import Dict, Optional
from torch.utils.data import DataLoader, Dataset
logger = logging.getLogger(__name__)
class PrefetchDataLoader:
"""
DataLoader with advanced prefetching and caching.
Wraps a standard DataLoader with additional optimizations:
- Multiple prefetch buffers
- Automatic batch size tuning
- Memory usage monitoring
"""
def __init__(
self,
dataloader: DataLoader,
prefetch_factor: int = 4,
pin_memory: bool = True,
non_blocking: bool = True,
):
"""
Initialize prefetch DataLoader.
Args:
dataloader: Base DataLoader to wrap
prefetch_factor: Number of batches to prefetch
pin_memory: Pin memory for faster GPU transfer
non_blocking: Use non-blocking transfers
"""
self.dataloader = dataloader
self.prefetch_factor = prefetch_factor
self.pin_memory = pin_memory
self.non_blocking = non_blocking
def __iter__(self):
"""Iterate with prefetching."""
return iter(self.dataloader)
def __len__(self):
"""Return length of underlying DataLoader."""
return len(self.dataloader)
def optimize_dataloader(
dataset: Dataset,
batch_size: int = 1,
num_workers: Optional[int] = None,
pin_memory: bool = True,
persistent_workers: bool = True,
prefetch_factor: int = 4,
shuffle: bool = True,
device: str = "cuda",
) -> DataLoader:
"""
Create optimized DataLoader with best practices.
Args:
dataset: Dataset to load
batch_size: Batch size
num_workers: Number of worker processes (None = auto)
pin_memory: Pin memory for faster GPU transfer
persistent_workers: Keep workers alive between epochs
prefetch_factor: Number of batches to prefetch per worker
shuffle: Shuffle dataset
device: Target device
Returns:
Optimized DataLoader
"""
import os
# Auto-detect optimal number of workers
if num_workers is None:
cpu_count = os.cpu_count() or 1
# Use 2-4 workers, but not more than CPU count
num_workers = min(4, max(2, cpu_count // 2))
# Adjust prefetch factor based on batch size
if batch_size > 4:
prefetch_factor = max(2, prefetch_factor // 2)
dataloader = DataLoader(
dataset,
batch_size=batch_size,
shuffle=shuffle,
num_workers=num_workers,
pin_memory=pin_memory and device == "cuda",
persistent_workers=persistent_workers if num_workers > 0 else False,
prefetch_factor=prefetch_factor if num_workers > 0 else None,
drop_last=False,
)
logger.info(
f"Created optimized DataLoader: "
f"batch_size={batch_size}, "
f"num_workers={num_workers}, "
f"prefetch_factor={prefetch_factor}, "
f"pin_memory={pin_memory}"
)
return dataloader
def profile_dataloader(
dataloader: DataLoader,
num_batches: int = 10,
device: str = "cuda",
) -> Dict[str, float]:
"""
Profile DataLoader performance.
Args:
dataloader: DataLoader to profile
num_batches: Number of batches to profile
device: Target device
Returns:
Dict with profiling results
"""
logger.info(f"Profiling DataLoader ({num_batches} batches)...")
times = []
data_times = []
transfer_times = []
start_time = time.time()
for i, batch in enumerate(dataloader):
if i >= num_batches:
break
batch_start = time.time()
# Measure data loading time
data_time = batch_start - (times[-1][1] if times else start_time)
# Measure transfer time
if device == "cuda":
transfer_start = time.time()
# Move batch to device
if isinstance(batch, dict):
batch = {k: v.to(device, non_blocking=True) for k, v in batch.items()}
elif isinstance(batch, (list, tuple)):
batch = [x.to(device, non_blocking=True) for x in batch]
else:
batch = batch.to(device, non_blocking=True)
transfer_time = time.time() - transfer_start
else:
transfer_time = 0.0
batch_time = time.time() - batch_start
times.append((batch_time, time.time()))
data_times.append(data_time)
transfer_times.append(transfer_time)
total_time = time.time() - start_time
results = {
"total_time": total_time,
"avg_batch_time": sum(t[0] for t in times) / len(times),
"avg_data_time": sum(data_times) / len(data_times),
"avg_transfer_time": sum(transfer_times) / len(transfer_times),
"batches_per_sec": len(times) / total_time,
"data_loading_ratio": sum(data_times) / total_time,
"transfer_ratio": sum(transfer_times) / total_time,
}
logger.info("DataLoader Profile Results:")
logger.info(f" Total time: {total_time:.2f}s")
logger.info(f" Avg batch time: {results['avg_batch_time'] * 1000:.2f}ms")
logger.info(f" Avg data loading: {results['avg_data_time'] * 1000:.2f}ms")
logger.info(f" Avg transfer: {results['avg_transfer_time'] * 1000:.2f}ms")
logger.info(f" Batches/sec: {results['batches_per_sec']:.2f}")
logger.info(f" Data loading ratio: {results['data_loading_ratio'] * 100:.1f}%")
logger.info(f" Transfer ratio: {results['transfer_ratio'] * 100:.1f}%")
return results
def find_optimal_num_workers(
dataset: Dataset,
batch_size: int = 1,
max_workers: int = 8,
num_test_batches: int = 10,
device: str = "cuda",
) -> int:
"""
Find optimal number of workers for DataLoader.
Args:
dataset: Dataset to test
batch_size: Batch size
max_workers: Maximum workers to test
num_test_batches: Number of batches to test per configuration
device: Target device
Returns:
Optimal number of workers
"""
logger.info(f"Finding optimal number of workers (max={max_workers})...")
best_workers = 0
best_time = float("inf")
for num_workers in range(0, max_workers + 1):
dataloader = DataLoader(
dataset,
batch_size=batch_size,
num_workers=num_workers,
pin_memory=device == "cuda",
prefetch_factor=2 if num_workers > 0 else None,
)
# Profile
start_time = time.time()
for i, _ in enumerate(dataloader):
if i >= num_test_batches:
break
elapsed = time.time() - start_time
logger.info(f" {num_workers} workers: {elapsed:.2f}s")
if elapsed < best_time:
best_time = elapsed
best_workers = num_workers
logger.info(f"Optimal number of workers: {best_workers} ({best_time:.2f}s)")
return best_workers
|