Spaces:
Starting
on
L40S
Starting
on
L40S
File size: 9,692 Bytes
d7e58f0 |
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 |
# Copyright (c) OpenMMLab. All rights reserved.
import os
import platform
import numpy as np
import pytest
import torch
import torch.distributed as dist
import torch.nn as nn
if platform.system() == 'Windows':
import regex as re
else:
import re
class TestSyncBN:
def dist_init(self):
rank = int(os.environ['SLURM_PROCID'])
world_size = int(os.environ['SLURM_NTASKS'])
local_rank = int(os.environ['SLURM_LOCALID'])
node_list = str(os.environ['SLURM_NODELIST'])
node_parts = re.findall('[0-9]+', node_list)
os.environ['MASTER_ADDR'] = (f'{node_parts[1]}.{node_parts[2]}' +
f'.{node_parts[3]}.{node_parts[4]}')
os.environ['MASTER_PORT'] = '12341'
os.environ['WORLD_SIZE'] = str(world_size)
os.environ['RANK'] = str(rank)
dist.init_process_group('nccl')
torch.cuda.set_device(local_rank)
def _test_syncbn_train(self, size=1, half=False):
if 'SLURM_NTASKS' not in os.environ or int(
os.environ['SLURM_NTASKS']) != 4:
print('must run with slurm has 4 processes!\n'
'srun -p test --gres=gpu:4 -n4')
return
else:
print('Running syncbn test')
from mmcv.ops import SyncBatchNorm
assert size in (1, 2, 4)
if not dist.is_initialized():
self.dist_init()
rank = dist.get_rank()
torch.manual_seed(9)
torch.cuda.manual_seed(9)
self.x = torch.rand(16, 3, 2, 3).cuda()
self.y_bp = torch.rand(16, 3, 2, 3).cuda()
if half:
self.x = self.x.half()
self.y_bp = self.y_bp.half()
dist.broadcast(self.x, src=0)
dist.broadcast(self.y_bp, src=0)
torch.cuda.synchronize()
if size == 1:
groups = [None, None, None, None]
groups[0] = dist.new_group([0])
groups[1] = dist.new_group([1])
groups[2] = dist.new_group([2])
groups[3] = dist.new_group([3])
group = groups[rank]
elif size == 2:
groups = [None, None, None, None]
groups[0] = groups[1] = dist.new_group([0, 1])
groups[2] = groups[3] = dist.new_group([2, 3])
group = groups[rank]
elif size == 4:
group = dist.group.WORLD
syncbn = SyncBatchNorm(3, group=group).cuda()
syncbn.weight.data[0] = 0.2
syncbn.weight.data[1] = 0.5
syncbn.weight.data[2] = 0.7
syncbn.train()
bn = nn.BatchNorm2d(3).cuda()
bn.weight.data[0] = 0.2
bn.weight.data[1] = 0.5
bn.weight.data[2] = 0.7
bn.train()
sx = self.x[rank * 4:rank * 4 + 4]
sx.requires_grad_()
sy = syncbn(sx)
sy.backward(self.y_bp[rank * 4:rank * 4 + 4])
smean = syncbn.running_mean
svar = syncbn.running_var
sx_grad = sx.grad
sw_grad = syncbn.weight.grad
sb_grad = syncbn.bias.grad
if size == 1:
x = self.x[rank * 4:rank * 4 + 4]
y_bp = self.y_bp[rank * 4:rank * 4 + 4]
elif size == 2:
x = self.x[rank // 2 * 8:rank // 2 * 8 + 8]
y_bp = self.y_bp[rank // 2 * 8:rank // 2 * 8 + 8]
elif size == 4:
x = self.x
y_bp = self.y_bp
x.requires_grad_()
y = bn(x)
y.backward(y_bp)
if size == 2:
y = y[rank % 2 * 4:rank % 2 * 4 + 4]
elif size == 4:
y = y[rank * 4:rank * 4 + 4]
mean = bn.running_mean
var = bn.running_var
if size == 1:
x_grad = x.grad
w_grad = bn.weight.grad
b_grad = bn.bias.grad
elif size == 2:
x_grad = x.grad[rank % 2 * 4:rank % 2 * 4 + 4]
w_grad = bn.weight.grad / 2
b_grad = bn.bias.grad / 2
elif size == 4:
x_grad = x.grad[rank * 4:rank * 4 + 4]
w_grad = bn.weight.grad / 4
b_grad = bn.bias.grad / 4
assert np.allclose(mean.data.cpu().numpy(),
smean.data.cpu().numpy(), 1e-3)
assert np.allclose(var.data.cpu().numpy(),
svar.data.cpu().numpy(), 1e-3)
assert np.allclose(y.data.cpu().numpy(), sy.data.cpu().numpy(), 1e-3)
assert np.allclose(w_grad.data.cpu().numpy(),
sw_grad.data.cpu().numpy(), 1e-3)
assert np.allclose(b_grad.data.cpu().numpy(),
sb_grad.data.cpu().numpy(), 1e-3)
assert np.allclose(x_grad.data.cpu().numpy(),
sx_grad.data.cpu().numpy(), 1e-2)
def _test_syncbn_empty_train(self, size=1, half=False):
if 'SLURM_NTASKS' not in os.environ or int(
os.environ['SLURM_NTASKS']) != 4:
print('must run with slurm has 4 processes!\n'
'srun -p test --gres=gpu:4 -n4')
return
else:
print('Running syncbn test')
from mmcv.ops import SyncBatchNorm
assert size in (1, 2, 4)
if not dist.is_initialized():
self.dist_init()
rank = dist.get_rank()
torch.manual_seed(9)
torch.cuda.manual_seed(9)
self.x = torch.rand(0, 3, 2, 3).cuda()
self.y_bp = torch.rand(0, 3, 2, 3).cuda()
if half:
self.x = self.x.half()
self.y_bp = self.y_bp.half()
dist.broadcast(self.x, src=0)
dist.broadcast(self.y_bp, src=0)
torch.cuda.synchronize()
if size == 1:
groups = [None, None, None, None]
groups[0] = dist.new_group([0])
groups[1] = dist.new_group([1])
groups[2] = dist.new_group([2])
groups[3] = dist.new_group([3])
group = groups[rank]
elif size == 2:
groups = [None, None, None, None]
groups[0] = groups[1] = dist.new_group([0, 1])
groups[2] = groups[3] = dist.new_group([2, 3])
group = groups[rank]
elif size == 4:
group = dist.group.WORLD
syncbn = SyncBatchNorm(3, group=group, stats_mode='N').cuda()
syncbn.weight.data[0] = 0.2
syncbn.weight.data[1] = 0.5
syncbn.weight.data[2] = 0.7
syncbn.train()
bn = nn.BatchNorm2d(3).cuda()
bn.weight.data[0] = 0.2
bn.weight.data[1] = 0.5
bn.weight.data[2] = 0.7
bn.train()
sx = self.x[rank * 4:rank * 4 + 4]
sx.requires_grad_()
sy = syncbn(sx)
sy.backward(self.y_bp[rank * 4:rank * 4 + 4])
smean = syncbn.running_mean
svar = syncbn.running_var
sx_grad = sx.grad
sw_grad = syncbn.weight.grad
sb_grad = syncbn.bias.grad
if size == 1:
x = self.x[rank * 4:rank * 4 + 4]
y_bp = self.y_bp[rank * 4:rank * 4 + 4]
elif size == 2:
x = self.x[rank // 2 * 8:rank // 2 * 8 + 8]
y_bp = self.y_bp[rank // 2 * 8:rank // 2 * 8 + 8]
elif size == 4:
x = self.x
y_bp = self.y_bp
x.requires_grad_()
y = bn(x)
y.backward(y_bp)
if size == 2:
y = y[rank % 2 * 4:rank % 2 * 4 + 4]
elif size == 4:
y = y[rank * 4:rank * 4 + 4]
mean = bn.running_mean
var = bn.running_var
if size == 1:
x_grad = x.grad
w_grad = bn.weight.grad
b_grad = bn.bias.grad
elif size == 2:
x_grad = x.grad[rank % 2 * 4:rank % 2 * 4 + 4]
w_grad = bn.weight.grad / 2
b_grad = bn.bias.grad / 2
elif size == 4:
x_grad = x.grad[rank * 4:rank * 4 + 4]
w_grad = bn.weight.grad / 4
b_grad = bn.bias.grad / 4
assert np.allclose(mean.data.cpu().numpy(),
smean.data.cpu().numpy(), 1e-3)
assert np.allclose(var.data.cpu().numpy(),
svar.data.cpu().numpy(), 1e-3)
assert np.allclose(y.data.cpu().numpy(), sy.data.cpu().numpy(), 1e-3)
assert np.allclose(w_grad.data.cpu().numpy(),
sw_grad.data.cpu().numpy(), 1e-3)
assert np.allclose(b_grad.data.cpu().numpy(),
sb_grad.data.cpu().numpy(), 1e-3)
assert np.allclose(x_grad.data.cpu().numpy(),
sx_grad.data.cpu().numpy(), 1e-2)
# 'stats_mode' only allows 'default' and 'N'
with pytest.raises(AssertionError):
SyncBatchNorm(3, group=group, stats_mode='X')
def test_syncbn_1(self):
self._test_syncbn_train(size=1)
def test_syncbn_2(self):
self._test_syncbn_train(size=2)
def test_syncbn_4(self):
self._test_syncbn_train(size=4)
def test_syncbn_1_half(self):
self._test_syncbn_train(size=1, half=True)
def test_syncbn_2_half(self):
self._test_syncbn_train(size=2, half=True)
def test_syncbn_4_half(self):
self._test_syncbn_train(size=4, half=True)
def test_syncbn_empty_1(self):
self._test_syncbn_empty_train(size=1)
def test_syncbn_empty_2(self):
self._test_syncbn_empty_train(size=2)
def test_syncbn_empty_4(self):
self._test_syncbn_empty_train(size=4)
def test_syncbn_empty_1_half(self):
self._test_syncbn_empty_train(size=1, half=True)
def test_syncbn_empty_2_half(self):
self._test_syncbn_empty_train(size=2, half=True)
def test_syncbn_empty_4_half(self):
self._test_syncbn_empty_train(size=4, half=True)
|