File size: 10,911 Bytes
0651b21 | 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 | import torch
import torch.nn as nn
"""
updates statistic from a full data
Memory optimization note: If you encounter CUDA out of memory errors, consider setting:
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
This can help reduce memory fragmentation.
"""
class RunningMeanStd(nn.Module):
def __init__(self, insize, epsilon=1e-05, per_channel=False, norm_only=False):
super().__init__()
print("RunningMeanStd: ", insize)
self.insize = insize
self.mean_size = insize[0]
self.epsilon = epsilon
self.norm_only = norm_only
self.per_channel = per_channel
if per_channel:
if len(self.insize) == 3:
self.axis = [0, 2, 3]
if len(self.insize) == 2:
self.axis = [0, 2]
if len(self.insize) == 1:
self.axis = [0]
in_size = self.insize[0]
else:
self.axis = [0]
in_size = insize
self.register_buffer("running_mean", torch.zeros(in_size, dtype=torch.float32))
self.register_buffer("running_var", torch.ones(in_size, dtype=torch.float32))
self.register_buffer("count", torch.ones((), dtype=torch.float32))
self.frozen = False
self.frozen_partial = False
def freeze(self):
self.frozen = True
def unfreeze(self):
self.frozen = False
def freeze_partial(self, diff):
self.frozen_partial = True
self.diff = diff
def sync_across_gpus(self, accelerator):
if accelerator.num_processes <= 1:
return
# ZL: this is the right formulation but a crude sync. The correct way is to sync
# batch stats.
flat_stats = torch.cat(
[
self.running_mean.flatten(), # (D,)
self.running_var.flatten(), # (D,)
]
) # shape = (2D + C,)
# Gather stats from all processes
gathered = accelerator.gather(flat_stats[None]) # shape = (world_size, 2D + C)
# Reshape
world_size = gathered.shape[0]
D = self.running_mean.numel()
means_ = gathered[:, :D].reshape(world_size, *self.running_mean.shape).mean(dim=0)
vars_ = gathered[:, D : 2 * D].reshape(world_size, *self.running_var.shape).mean(dim=0)
# Update local stats
self.running_mean.copy_(means_)
self.running_var.copy_(vars_)
def _update_mean_var_count_from_moments(
self, mean, var, count, batch_mean, batch_var, batch_count
):
delta = batch_mean - mean
tot_count = count + batch_count
new_mean = mean + delta * batch_count / tot_count
m_a = var * count
m_b = batch_var * batch_count
M2 = m_a + m_b + delta**2 * count * batch_count / tot_count
new_var = M2 / tot_count
new_count = tot_count
return new_mean, new_var, new_count
def forward(self, input, unnorm=False):
# change shape
input_shape = input.shape
if len(input.shape) == 3:
input = input.reshape(-1, input_shape[-1])
if self.per_channel:
if len(self.insize) == 3:
# Use broadcasting instead of expand_as to avoid memory issues
current_mean = self.running_mean.view([1, self.insize[0], 1, 1])
current_var = self.running_var.view([1, self.insize[0], 1, 1])
elif len(self.insize) == 2:
current_mean = self.running_mean.view([1, self.insize[0], 1])
current_var = self.running_var.view([1, self.insize[0], 1])
elif len(self.insize) == 1:
current_mean = self.running_mean.view([1, self.insize[0]])
current_var = self.running_var.view([1, self.insize[0]])
else:
current_mean = self.running_mean
current_var = self.running_var
# get output
if unnorm:
y = torch.clamp(input, min=-5.0, max=5.0)
y = torch.sqrt(current_var + self.epsilon) * y + current_mean
else:
if self.norm_only:
y = input / torch.sqrt(current_var + self.epsilon)
else:
# Use in-place operations where possible to reduce memory usage
y = input - current_mean
y = y / torch.sqrt(current_var + self.epsilon)
y = torch.clamp(y, min=-5.0, max=5.0)
# update After normalization, so that the values used for training and testing are the same.
if self.training and not self.frozen:
mean = input.mean(self.axis) # along channel axis
var = input.var(self.axis)
new_mean, new_var, new_count = self._update_mean_var_count_from_moments(
self.running_mean, self.running_var, self.count, mean, var, input.size()[0]
)
if self.frozen_partial:
# Only update the last bit (futures)
self.running_mean[-self.diff :], self.running_var[-self.diff :], self.count = (
new_mean[-self.diff :],
new_var[-self.diff :],
new_count,
)
else:
self.running_mean, self.running_var, self.count = new_mean, new_var, new_count
if len(input_shape) == 3:
y = y.view(input_shape)
return y
class RunningMeanStdObs(nn.Module):
def __init__(self, insize, epsilon=1e-05, per_channel=False, norm_only=False):
assert isinstance(insize, dict)
super().__init__()
self.running_mean_std = nn.ModuleDict(
{k: RunningMeanStd(v, epsilon, per_channel, norm_only) for k, v in insize.items()}
)
def forward(self, input, unnorm=False):
res = {k: self.running_mean_std[k](v, unnorm) for k, v in input.items()}
return res
from collections.abc import Sequence
from copy import deepcopy
class VecNorm(nn.Module):
"""Simple running normalization for observations.
Keeps track of running mean and variance to normalize observations on-the-fly.
Args:
obs_keys: List of observation keys to normalize
decay: Decay rate for running statistics (default: 0.99)
eps: Small constant for numerical stability (default: 1e-4)
device: Device to store statistics on
"""
def __init__(
self,
obs_keys: Sequence[str] | None,
decay: float = 0.9999,
eps: float = 1e-4,
):
super().__init__()
self.obs_keys = obs_keys
self.decay = decay
self.eps = eps
# Running statistics as nn.Parameters
self.sum: dict[str, nn.Parameter] = nn.ParameterDict({})
self.ssq: dict[str, nn.Parameter] = nn.ParameterDict({})
self.cnt: dict[str, nn.Parameter] = nn.ParameterDict({})
self.mean: dict[str, nn.Parameter] = nn.ParameterDict({})
self.var: dict[str, nn.Parameter] = nn.ParameterDict({})
self.initialized = False
self.frozen = False
def init_stats(self, obs_dict: dict[str, torch.Tensor]):
"""Initialize running statistics based on observation shapes."""
if self.obs_keys is None:
self.obs_keys = list(obs_dict.keys())
for key in self.obs_keys:
if key in obs_dict:
v = obs_dict[key]
self.sum[key] = nn.Parameter(torch.zeros_like(v[0]), requires_grad=False)
self.ssq[key] = nn.Parameter(torch.zeros_like(v[0]), requires_grad=False)
self.cnt[key] = nn.Parameter(torch.zeros_like(v[0]), requires_grad=False)
self.mean[key] = nn.Parameter(torch.zeros_like(v[0]), requires_grad=False)
self.var[key] = nn.Parameter(torch.ones_like(v[0]), requires_grad=False)
self.initialized = True
def update(self, obs_dict: dict[str, torch.Tensor]):
"""Update running statistics with new observations."""
if not self.initialized:
self.init_stats(obs_dict)
if self.frozen:
return
for key in self.obs_keys:
if key not in obs_dict:
continue
x = obs_dict[key]
sum_x = x.sum(dim=0)
ssq_x = (x**2).sum(dim=0)
cnt_x = x.shape[0]
self.sum[key] = self.sum[key] * self.decay + sum_x
self.ssq[key] = self.ssq[key] * self.decay + ssq_x
self.cnt[key] = self.cnt[key] * self.decay + cnt_x
self.mean[key] = self.sum[key] / self.cnt[key]
self.var[key] = (self.ssq[key] / self.cnt[key] - self.mean[key] ** 2).clamp(
min=self.eps
)
def normalize(self, obs_dict: dict[str, torch.Tensor]) -> dict[str, torch.Tensor]:
"""Normalize observations using running statistics."""
if not self.initialized:
self.init_stats(obs_dict)
return obs_dict.copy()
normalized = obs_dict.copy()
for key in self.obs_keys:
normalized[key] = (normalized[key] - self.mean[key]) / self.var[key].sqrt()
return normalized
def denormalize(self, obs_dict: dict[str, torch.Tensor]) -> dict[str, torch.Tensor]:
"""Denormalize observations using running statistics."""
if not self.initialized:
self.init_stats(obs_dict)
return obs_dict.copy()
denormalized = obs_dict.copy()
for key in self.obs_keys:
denormalized[key] = denormalized[key] * self.var[key].sqrt() + self.mean[key]
return denormalized
def freeze(self):
"""Freeze running statistics updates."""
self.frozen = True
def unfreeze(self):
"""Unfreeze running statistics updates."""
self.frozen = False
def get_stats(self):
"""Get current running statistics."""
return {
"sum": {key: param.data.clone() for key, param in self.sum.items()},
"ssq": {key: param.data.clone() for key, param in self.ssq.items()},
"cnt": {key: param.data.clone() for key, param in self.cnt.items()},
"mean": {key: param.data.clone() for key, param in self.mean.items()},
"var": {key: param.data.clone() for key, param in self.var.items()},
}
def load_stats(self, stats):
"""Load running statistics."""
for key in stats["sum"]:
self.sum[key] = nn.Parameter(deepcopy(stats["sum"][key]), requires_grad=False)
self.ssq[key] = nn.Parameter(deepcopy(stats["ssq"][key]), requires_grad=False)
self.cnt[key] = nn.Parameter(deepcopy(stats["cnt"][key]), requires_grad=False)
self.mean[key] = nn.Parameter(deepcopy(stats["mean"][key]), requires_grad=False)
self.var[key] = nn.Parameter(deepcopy(stats["var"][key]), requires_grad=False)
self.initialized = True
|