File size: 15,990 Bytes
3133fdb |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 |
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
import logging
import unittest
from typing import Callable, Tuple
import torch
import torch.nn as nn
from fvcore.common.benchmark import benchmark
from pytorchvideo.layers.accelerator.mobile_cpu.convolutions import (
Conv3d3x3x3DwBnAct,
Conv3dPwBnAct,
)
from pytorchvideo.models.accelerator.mobile_cpu.residual_blocks import (
X3dBottleneckBlock,
)
from torch.utils.mobile_optimizer import optimize_for_mobile
TORCH_VERSION: Tuple[int, ...] = tuple(int(x) for x in torch.__version__.split(".")[:2])
if TORCH_VERSION >= (1, 11):
from torch.ao.quantization import (
convert,
DeQuantStub,
fuse_modules,
get_default_qconfig,
prepare,
QuantStub,
# quantize_fx
)
else:
from torch.quantization import (
convert,
DeQuantStub,
fuse_modules,
get_default_qconfig,
prepare,
QuantStub,
# quantize_fx
)
class TestBenchmarkEfficientBlocks(unittest.TestCase):
def setUp(self):
super().setUp()
torch.set_rng_state(torch.manual_seed(42).get_state())
def test_benchmark_conv3d_pw_bn_relu(self, num_iters: int = 20) -> None:
"""
Benchmark Conv3dPwBnAct with ReLU activation.
Note efficient block Conv3dPwBnAct is designed for mobile cpu with qnnpack
backend, and benchmarking on server with another backend (e.g., fbgemm) may
have different latency result compared to running on mobile cpu with qnnpack.
Running on x86 based server cpu with qnnpack may also have different latency as
running on mobile cpu with qnnpack, as qnnpack is optimized for
ARM based mobile cpu.
Args:
num_iters (int): number of iterations to perform benchmarking.
"""
torch.backends.quantized.engine = "qnnpack"
kwargs_list = [
{
"mode": "original",
"input_blob_size": (1, 48, 4, 40, 40),
"in_channels": 48,
"out_channels": 108,
"quantize": False,
},
{
"mode": "deployable",
"input_blob_size": (1, 48, 4, 40, 40),
"in_channels": 48,
"out_channels": 108,
"quantize": False,
},
{
"mode": "original",
"input_blob_size": (1, 48, 4, 40, 40),
"in_channels": 48,
"out_channels": 108,
"quantize": True,
},
{
"mode": "deployable",
"input_blob_size": (1, 48, 4, 40, 40),
"in_channels": 48,
"out_channels": 108,
"quantize": True,
},
{
"mode": "deployable",
"input_blob_size": (1, 48, 4, 40, 40),
"in_channels": 48,
"out_channels": 108,
"quantize": True,
"native_conv3d_op_qnnpack": True,
},
]
def _benchmark_conv3d_pw_bn_relu_forward(**kwargs) -> Callable:
assert kwargs["mode"] in ("original", "deployable"), (
"kwargs['mode'] must be either 'original' or 'deployable',"
"but got {}.".format(kwargs["mode"])
)
input_tensor = torch.randn((kwargs["input_blob_size"]))
conv_block = Conv3dPwBnAct(
kwargs["in_channels"],
kwargs["out_channels"],
use_bn=False, # assume BN has already been fused for forward
)
if kwargs["mode"] == "deployable":
native_conv3d_op_qnnpack = kwargs.get("native_conv3d_op_qnnpack", False)
conv_block.convert(
kwargs["input_blob_size"],
convert_for_quantize=kwargs["quantize"],
native_conv3d_op_qnnpack=native_conv3d_op_qnnpack,
)
conv_block.eval()
def func_to_benchmark_dummy() -> None:
return
if kwargs["quantize"] is True:
if kwargs["mode"] == "original": # manually fuse conv and relu
conv_block.kernel = fuse_modules(
conv_block.kernel, ["conv", "act.act"]
)
conv_block = nn.Sequential(
QuantStub(),
conv_block,
DeQuantStub(),
)
conv_block.qconfig = get_default_qconfig("qnnpack")
conv_block = prepare(conv_block)
try:
conv_block = convert(conv_block)
except Exception as e:
logging.info(
"benchmark_conv3d_pw_bn_relu: "
"catch exception '{}' with kwargs of {}".format(e, kwargs)
)
return func_to_benchmark_dummy
try:
traced_model = torch.jit.trace(conv_block, input_tensor, strict=False)
except Exception as e:
logging.info(
"benchmark_conv3d_pw_bn_relu: "
"catch exception '{}' with kwargs of {}".format(e, kwargs)
)
return func_to_benchmark_dummy
if kwargs["quantize"] is False:
traced_model = optimize_for_mobile(traced_model)
logging.info(f"model arch: {traced_model}")
def func_to_benchmark() -> None:
try:
_ = traced_model(input_tensor)
except Exception as e:
logging.info(
"benchmark_conv3d_pw_bn_relu: "
"catch exception '{}' with kwargs of {}".format(e, kwargs)
)
return
return func_to_benchmark
benchmark(
_benchmark_conv3d_pw_bn_relu_forward,
"benchmark_conv3d_pw_bn_relu",
kwargs_list,
num_iters=num_iters,
warmup_iters=2,
)
self.assertTrue(True)
def test_benchmark_conv3d_3x3x3_dw_bn_relu(self, num_iters: int = 20) -> None:
"""
Benchmark Conv3d3x3x3DwBnAct with ReLU activation.
Note efficient block Conv3d3x3x3DwBnAct is designed for mobile cpu with qnnpack
backend, and benchmarking on server with another backend (e.g., fbgemm) may have
different latency result compared as running on mobile cpu.
Args:
num_iters (int): number of iterations to perform benchmarking.
"""
torch.backends.quantized.engine = "qnnpack"
kwargs_list = [
{
"mode": "original",
"input_blob_size": (1, 48, 4, 40, 40),
"in_channels": 48,
"quantize": False,
},
{
"mode": "deployable",
"input_blob_size": (1, 48, 4, 40, 40),
"in_channels": 48,
"quantize": False,
},
{
"mode": "original",
"input_blob_size": (1, 48, 4, 40, 40),
"in_channels": 48,
"quantize": True,
},
{
"mode": "deployable",
"input_blob_size": (1, 48, 4, 40, 40),
"in_channels": 48,
"quantize": True,
},
{
"mode": "deployable",
"input_blob_size": (1, 48, 4, 40, 40),
"in_channels": 48,
"quantize": True,
"native_conv3d_op_qnnpack": True,
},
]
def _benchmark_conv3d_3x3x3_dw_bn_relu_forward(**kwargs) -> Callable:
assert kwargs["mode"] in ("original", "deployable"), (
"kwargs['mode'] must be either 'original' or 'deployable',"
"but got {}.".format(kwargs["mode"])
)
input_tensor = torch.randn((kwargs["input_blob_size"]))
conv_block = Conv3d3x3x3DwBnAct(
kwargs["in_channels"],
use_bn=False, # assume BN has already been fused for forward
)
def func_to_benchmark_dummy() -> None:
return
if kwargs["mode"] == "deployable":
native_conv3d_op_qnnpack = kwargs.get("native_conv3d_op_qnnpack", False)
conv_block.convert(
kwargs["input_blob_size"],
convert_for_quantize=kwargs["quantize"],
native_conv3d_op_qnnpack=native_conv3d_op_qnnpack,
)
conv_block.eval()
if kwargs["quantize"] is True:
if kwargs["mode"] == "original": # manually fuse conv and relu
conv_block.kernel = fuse_modules(
conv_block.kernel, ["conv", "act.act"]
)
conv_block = nn.Sequential(
QuantStub(),
conv_block,
DeQuantStub(),
)
conv_block.qconfig = get_default_qconfig("qnnpack")
conv_block = prepare(conv_block)
try:
conv_block = convert(conv_block)
except Exception as e:
logging.info(
"benchmark_conv3d_3x3x3_dw_bn_relu: "
"catch exception '{}' with kwargs of {}".format(e, kwargs)
)
return func_to_benchmark_dummy
try:
traced_model = torch.jit.trace(conv_block, input_tensor, strict=False)
except Exception as e:
logging.info(
"benchmark_conv3d_3x3x3_dw_bn_relu: "
"catch exception '{}' with kwargs of {}".format(e, kwargs)
)
return func_to_benchmark_dummy
if kwargs["quantize"] is False:
traced_model = optimize_for_mobile(traced_model)
logging.info(f"model arch: {traced_model}")
def func_to_benchmark() -> None:
try:
_ = traced_model(input_tensor)
except Exception as e:
logging.info(
"benchmark_conv3d_3x3x3_dw_bn_relu: "
"catch exception '{}' with kwargs of {}".format(e, kwargs)
)
return
return func_to_benchmark
benchmark(
_benchmark_conv3d_3x3x3_dw_bn_relu_forward,
"benchmark_conv3d_3x3x3_dw_bn_relu",
kwargs_list,
num_iters=num_iters,
warmup_iters=2,
)
self.assertTrue(True)
def test_benchmark_x3d_bottleneck_block(self, num_iters: int = 20) -> None:
"""
Benchmark X3dBottleneckBlock.
Note efficient block X3dBottleneckBlock is designed for mobile cpu with qnnpack
backend, and benchmarking on server/laptop may have different latency result
compared to running on mobile cpu.
Args:
num_iters (int): number of iterations to perform benchmarking.
"""
torch.backends.quantized.engine = "qnnpack"
kwargs_list = [
{
"mode": "original",
"input_blob_size": (1, 48, 4, 20, 20),
"in_channels": 48,
"mid_channels": 108,
"out_channels": 48,
"quantize": False,
},
{
"mode": "deployable",
"input_blob_size": (1, 48, 4, 20, 20),
"in_channels": 48,
"mid_channels": 108,
"out_channels": 48,
"quantize": False,
},
{
"mode": "original",
"input_blob_size": (1, 48, 4, 20, 20),
"in_channels": 48,
"mid_channels": 108,
"out_channels": 48,
"quantize": True,
},
{
"mode": "deployable",
"input_blob_size": (1, 48, 4, 20, 20),
"in_channels": 48,
"mid_channels": 108,
"out_channels": 48,
"quantize": True,
},
{
"mode": "deployable",
"input_blob_size": (1, 48, 4, 20, 20),
"in_channels": 48,
"mid_channels": 108,
"out_channels": 48,
"quantize": True,
"native_conv3d_op_qnnpack": True,
},
]
def _benchmark_x3d_bottleneck_forward(**kwargs) -> Callable:
assert kwargs["mode"] in ("original", "deployable"), (
"kwargs['mode'] must be either 'original' or 'deployable',"
"but got {}.".format(kwargs["mode"])
)
input_tensor = torch.randn((kwargs["input_blob_size"]))
conv_block = X3dBottleneckBlock(
kwargs["in_channels"],
kwargs["mid_channels"],
kwargs["out_channels"],
use_bn=(False, False, False), # Assume BN has been fused for forward
)
if kwargs["mode"] == "deployable":
native_conv3d_op_qnnpack = kwargs.get("native_conv3d_op_qnnpack", False)
conv_block.convert(
kwargs["input_blob_size"],
convert_for_quantize=kwargs["quantize"],
native_conv3d_op_qnnpack=native_conv3d_op_qnnpack,
)
conv_block.eval()
def func_to_benchmark_dummy() -> None:
return
if kwargs["quantize"] is True:
conv_block = nn.Sequential(
QuantStub(),
conv_block,
DeQuantStub(),
)
conv_block.qconfig = get_default_qconfig("qnnpack")
conv_block = prepare(conv_block)
try:
conv_block = convert(conv_block)
traced_model = torch.jit.trace(
conv_block, input_tensor, strict=False
)
except Exception as e:
logging.info(
"benchmark_x3d_bottleneck_forward: "
"catch exception '{}' with kwargs of {}".format(e, kwargs)
)
return func_to_benchmark_dummy
try:
traced_model = torch.jit.trace(conv_block, input_tensor, strict=False)
except Exception as e:
logging.info(
"benchmark_x3d_bottleneck_forward: "
"catch exception '{}' with kwargs of {}".format(e, kwargs)
)
return func_to_benchmark_dummy
if kwargs["quantize"] is False:
traced_model = optimize_for_mobile(traced_model)
logging.info(f"model arch: {traced_model}")
def func_to_benchmark() -> None:
try:
_ = traced_model(input_tensor)
except Exception as e:
logging.info(
"benchmark_x3d_bottleneck_forward: "
"catch exception '{}' with kwargs of {}".format(e, kwargs)
)
return
return func_to_benchmark
benchmark(
_benchmark_x3d_bottleneck_forward,
"benchmark_x3d_bottleneck_forward",
kwargs_list,
num_iters=num_iters,
warmup_iters=2,
)
self.assertTrue(True)
|