File size: 11,297 Bytes
7a0c684 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 |
from typing import Optional, Union, Tuple, List, Dict
import numpy as np
from dataclasses import dataclass
from enum import Enum
import warnings
from .core.db_manager import HeliumDBManager
from virtual_gpu_driver.src.ai.tensor_types import Tensor, Device, DType
import hashlib
import json
from functools import lru_cache
class NormType(Enum):
"""Supported normalization types"""
BATCH = "batch"
LAYER = "layer"
GROUP = "group"
INSTANCE = "instance"
RMS = "rms"
def normalize(input: Tensor,
mean: Optional[Tensor] = None,
variance: Optional[Tensor] = None,
weight: Optional[Tensor] = None,
bias: Optional[Tensor] = None,
eps: float = 1e-5) -> Tuple[Tensor, Tensor, Tensor]:
"""
Normalizes the input using mean and variance. If mean/variance not provided,
they are computed from the input.
Args:
input: Input tensor
mean: Optional pre-computed mean
variance: Optional pre-computed variance
weight: Optional scale parameter
bias: Optional bias parameter
eps: Small constant for numerical stability
Returns:
Tuple of (normalized tensor, mean, variance)
"""
# Calculate mean and variance if not provided
if mean is None or variance is None:
# Compute stats along last dimension
axes = tuple(range(input.ndim - 1))
mean = input.mean(axis=axes, keepdims=True)
variance = input.var(axis=axes, keepdims=True)
# Normalize
denom = (variance + eps).sqrt()
normalized = (input - mean) / denom
# Apply scale and bias if provided
if weight is not None:
normalized = normalized * weight
if bias is not None:
normalized = normalized + bias
return normalized, mean, variance
@dataclass
class NormConfig:
"""Configuration for normalization layers"""
norm_type: NormType
num_features: int
eps: float = 1e-5
momentum: float = 0.1
affine: bool = True
num_groups: int = 32 # For group norm
track_running_stats: bool = True
dtype: np.dtype = np.float32
use_cache: bool = True
class NormalizationCache:
"""Cache manager for normalization computations"""
def __init__(self):
self.db = HeliumDBManager.get_instance()
self.running_means: Dict[str, np.ndarray] = {}
self.running_vars: Dict[str, np.ndarray] = {}
def _compute_key(self, x: np.ndarray, norm_type: NormType) -> str:
"""Compute cache key for input"""
hasher = hashlib.sha256()
hasher.update(x.tobytes())
hasher.update(norm_type.value.encode())
return hasher.hexdigest()
def get(self, key: str) -> Optional[Dict[str, np.ndarray]]:
"""Get cached computation"""
return self.db.get_activation(key)
def set(self, key: str, value: Dict[str, np.ndarray], metadata: Dict):
"""Cache computation"""
self.db.set_activation(key, value, metadata)
def update_running_stats(
self,
key: str,
mean: np.ndarray,
var: np.ndarray,
momentum: float
):
"""Update running statistics"""
if key in self.running_means:
self.running_means[key] = (
(1 - momentum) * self.running_means[key] +
momentum * mean
)
self.running_vars[key] = (
(1 - momentum) * self.running_vars[key] +
momentum * var
)
else:
self.running_means[key] = mean
self.running_vars[key] = var
class Normalization:
"""
Unified normalization implementation with support for:
- Multiple normalization types
- Hardware acceleration
- Mixed precision
- Computation caching
- Running statistics tracking
"""
def __init__(
self,
config: NormConfig,
driver = None
):
"""Initialize normalization layer"""
self.config = config
self.driver = driver
self.cache = NormalizationCache()
# Initialize learnable parameters if needed
if config.affine:
self.gamma = np.ones(config.num_features, dtype=config.dtype)
self.beta = np.zeros(config.num_features, dtype=config.dtype)
else:
self.gamma = None
self.beta = None
@staticmethod
@lru_cache(maxsize=128)
def _get_reshape_dims(
input_shape: Tuple[int, ...],
num_features: int,
norm_type: NormType
) -> Tuple[Tuple[int, ...], Tuple[int, ...]]:
"""Get reshaping dimensions for parameters"""
ndim = len(input_shape)
if norm_type == NormType.BATCH:
param_shape = (1, num_features) + (1,) * (ndim - 2)
reduction_axes = (0,) + tuple(range(2, ndim))
elif norm_type == NormType.LAYER:
param_shape = (1,) * (ndim - 1) + (num_features,)
reduction_axes = tuple(range(ndim - 1))
else: # GROUP, INSTANCE
param_shape = (1, num_features) + (1,) * (ndim - 2)
reduction_axes = (2,) + tuple(range(3, ndim))
return param_shape, reduction_axes
def _check_input(self, x: np.ndarray):
"""Validate input tensor"""
if x.ndim < 2:
raise ValueError(f"Expected at least 2D input, got shape {x.shape}")
if self.config.norm_type in [NormType.BATCH, NormType.GROUP]:
if x.shape[1] != self.config.num_features:
raise ValueError(
f"Expected {self.config.num_features} features, got {x.shape[1]}"
)
def _compute_stats(
self,
x: np.ndarray,
reduction_axes: Tuple[int, ...]
) -> Tuple[np.ndarray, np.ndarray]:
"""Compute mean and variance"""
if self.driver and hasattr(self.driver, 'reduce_mean'):
mean = self.driver.reduce_mean(x, axis=reduction_axes, keepdims=True)
var = self.driver.reduce_var(x, axis=reduction_axes, keepdims=True)
else:
mean = np.mean(x, axis=reduction_axes, keepdims=True)
var = np.var(x, axis=reduction_axes, keepdims=True)
return mean, var
def normalize(
self,
x: np.ndarray,
training: bool = True
) -> np.ndarray:
"""
Apply normalization to input tensor
"""
self._check_input(x)
# Get cache key and check cache
if self.config.use_cache and not training:
cache_key = self.cache._compute_key(x, self.config.norm_type)
cached = self.cache.get(cache_key)
if cached is not None:
return cached['output']
# Get reshaping dimensions
param_shape, reduction_axes = self._get_reshape_dims(
x.shape,
self.config.num_features,
self.config.norm_type
)
# Special handling for group norm
if self.config.norm_type == NormType.GROUP:
groups = self.config.num_groups
N, C = x.shape[:2]
x = x.reshape(N, groups, C // groups, *x.shape[2:])
reduction_axes = (2,) + tuple(range(3, x.ndim))
# Compute statistics
mean, var = self._compute_stats(x, reduction_axes)
# Update running statistics during training
if training and self.config.track_running_stats:
self.cache.update_running_stats(
str(id(self)),
mean,
var,
self.config.momentum
)
# Normalize
x_norm = (x - mean) / np.sqrt(var + self.config.eps)
# Reshape back if group norm
if self.config.norm_type == NormType.GROUP:
x_norm = x_norm.reshape(N, C, *x.shape[3:])
# Apply affine transform if needed
if self.config.affine:
gamma = self.gamma.reshape(param_shape)
beta = self.beta.reshape(param_shape)
out = gamma * x_norm + beta
else:
out = x_norm
# Cache result if needed
if self.config.use_cache and not training:
self.cache.set(
cache_key,
{
'output': out,
'mean': mean,
'var': var
},
{
'shape': x.shape,
'dtype': str(x.dtype),
'norm_type': self.config.norm_type.value
}
)
return out
@classmethod
def batch_norm(
cls,
x: np.ndarray,
num_features: Optional[int] = None,
**kwargs
) -> np.ndarray:
"""Convenience method for batch normalization"""
config = NormConfig(
norm_type=NormType.BATCH,
num_features=num_features or x.shape[1],
**kwargs
)
return cls(config).normalize(x)
@classmethod
def layer_norm(
cls,
x: np.ndarray,
num_features: Optional[int] = None,
**kwargs
) -> np.ndarray:
"""Convenience method for layer normalization"""
config = NormConfig(
norm_type=NormType.LAYER,
num_features=num_features or x.shape[-1],
**kwargs
)
return cls(config).normalize(x)
@classmethod
def group_norm(
cls,
x: np.ndarray,
num_features: Optional[int] = None,
num_groups: int = 32,
**kwargs
) -> np.ndarray:
"""Convenience method for group normalization"""
config = NormConfig(
norm_type=NormType.GROUP,
num_features=num_features or x.shape[1],
num_groups=num_groups,
**kwargs
)
return cls(config).normalize(x)
@classmethod
def instance_norm(
cls,
x: np.ndarray,
num_features: Optional[int] = None,
**kwargs
) -> np.ndarray:
"""Convenience method for instance normalization"""
config = NormConfig(
norm_type=NormType.INSTANCE,
num_features=num_features or x.shape[1],
**kwargs
)
return cls(config).normalize(x)
@classmethod
def rms_norm(
cls,
x: np.ndarray,
num_features: Optional[int] = None,
**kwargs
) -> np.ndarray:
"""Convenience method for RMS normalization"""
config = NormConfig(
norm_type=NormType.RMS,
num_features=num_features or x.shape[-1],
track_running_stats=False, # RMS norm doesn't use running stats
**kwargs
)
return cls(config).normalize(x)
|