You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.

test_batchnorm.py 14 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396
  1. # -*- coding: utf-8 -*-
  2. # MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
  3. #
  4. # Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
  5. #
  6. # Unless required by applicable law or agreed to in writing,
  7. # software distributed under the License is distributed on an
  8. # "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  9. import functools
  10. import platform
  11. import numpy as np
  12. import pytest
  13. import megengine as mge
  14. import megengine.distributed as dist
  15. from megengine import Tensor, jit
  16. from megengine.autodiff.grad_manager import GradManager
  17. from megengine.core._trace_option import use_symbolic_shape
  18. from megengine.module import BatchNorm1d, BatchNorm2d, SyncBatchNorm
  19. _assert_allclose = functools.partial(np.testing.assert_allclose, atol=5e-6, rtol=5e-6)
  20. @pytest.mark.require_ngpu(2)
  21. @pytest.mark.isolated_distributed
  22. def test_syncbn():
  23. nr_chan = 8
  24. data_shape = (3, nr_chan, 4, 16)
  25. momentum = 0.9
  26. eps = 1e-5
  27. running_mean = np.zeros((1, nr_chan, 1, 1), dtype=np.float32)
  28. running_var = np.ones((1, nr_chan, 1, 1), dtype=np.float32)
  29. steps = 4
  30. nr_ranks = 2
  31. server = dist.Server()
  32. port = server.py_server_port
  33. @dist.launcher(n_gpus=2)
  34. def worker(data, yv_expect, running_mean, running_var):
  35. rank = dist.get_rank()
  36. bn = SyncBatchNorm(nr_chan, momentum=momentum, eps=eps)
  37. for i in range(steps):
  38. yv = bn(Tensor(data[rank][i]))
  39. _assert_allclose(yv.numpy(), yv_expect[rank])
  40. _assert_allclose(bn.running_mean.numpy(), running_mean)
  41. _assert_allclose(bn.running_var.numpy(), running_var)
  42. xv = []
  43. for i in range(steps):
  44. xv.append(np.random.normal(loc=2.3, size=data_shape).astype(np.float32))
  45. xv_transposed = np.transpose(xv[i], [0, 2, 3, 1]).reshape(
  46. (data_shape[0] * data_shape[2] * data_shape[3], nr_chan)
  47. )
  48. mean = np.mean(xv_transposed, axis=0).reshape(1, nr_chan, 1, 1)
  49. var_biased = np.var(xv_transposed, axis=0).reshape((1, nr_chan, 1, 1))
  50. sd = np.sqrt(var_biased + eps)
  51. var_unbiased = np.var(xv_transposed, axis=0, ddof=1).reshape((1, nr_chan, 1, 1))
  52. running_mean = running_mean * momentum + mean * (1 - momentum)
  53. running_var = running_var * momentum + var_unbiased * (1 - momentum)
  54. yv_expect = (xv[i] - mean) / sd
  55. data = []
  56. for i in range(nr_ranks):
  57. data.append([])
  58. for j in range(steps):
  59. data[i].append(xv[j][:, :, :, i * 8 : i * 8 + 8])
  60. yv_expect = [yv_expect[:, :, :, i * 8 : i * 8 + 8] for i in range(nr_ranks)]
  61. worker(data, yv_expect, running_mean, running_var)
  62. def test_batchnorm():
  63. nr_chan = 8
  64. data_shape = (3, nr_chan, 4)
  65. momentum = 0.9
  66. bn = BatchNorm1d(nr_chan, momentum=momentum)
  67. running_mean = np.zeros((1, nr_chan, 1), dtype=np.float32)
  68. running_var = np.ones((1, nr_chan, 1), dtype=np.float32)
  69. for i in range(3):
  70. xv = np.random.normal(loc=2.3, size=data_shape).astype(np.float32)
  71. mean = np.mean(np.mean(xv, axis=0, keepdims=True), axis=2, keepdims=True)
  72. xv_transposed = np.transpose(xv, [0, 2, 1]).reshape(
  73. (data_shape[0] * data_shape[2], nr_chan)
  74. )
  75. var_biased = np.var(xv_transposed, axis=0).reshape((1, nr_chan, 1))
  76. sd = np.sqrt(var_biased + bn.eps)
  77. var_unbiased = np.var(xv_transposed, axis=0, ddof=1).reshape((1, nr_chan, 1))
  78. running_mean = running_mean * momentum + mean * (1 - momentum)
  79. running_var = running_var * momentum + var_unbiased * (1 - momentum)
  80. yv = bn(Tensor(xv))
  81. yv_expect = (xv - mean) / sd
  82. _assert_allclose(yv.numpy(), yv_expect)
  83. _assert_allclose(bn.running_mean.numpy().reshape(-1), running_mean.reshape(-1))
  84. _assert_allclose(bn.running_var.numpy().reshape(-1), running_var.reshape(-1))
  85. # test set 'training' flag to False
  86. mean_backup = bn.running_mean.numpy()
  87. var_backup = bn.running_var.numpy()
  88. bn.training = False
  89. xv = np.random.normal(loc=2.3, size=data_shape).astype(np.float32)
  90. data = Tensor(xv)
  91. yv1 = bn(data)
  92. yv2 = bn(data)
  93. np.testing.assert_equal(yv1.numpy(), yv2.numpy())
  94. np.testing.assert_equal(mean_backup, bn.running_mean.numpy())
  95. np.testing.assert_equal(var_backup, bn.running_var.numpy())
  96. yv_expect = (xv - running_mean) / np.sqrt(running_var + bn.eps)
  97. _assert_allclose(yv1.numpy(), yv_expect)
  98. def test_syncbn1d():
  99. nr_chan = 8
  100. data_shape = (3, nr_chan, 4)
  101. momentum = 0.9
  102. bn = SyncBatchNorm(nr_chan, momentum=momentum)
  103. running_mean = np.zeros((1, nr_chan, 1), dtype=np.float32)
  104. running_var = np.ones((1, nr_chan, 1), dtype=np.float32)
  105. for i in range(3):
  106. xv = np.random.normal(loc=2.3, size=data_shape).astype(np.float32)
  107. mean = np.mean(np.mean(xv, axis=0, keepdims=True), axis=2, keepdims=True)
  108. xv_transposed = np.transpose(xv, [0, 2, 1]).reshape(
  109. (data_shape[0] * data_shape[2], nr_chan)
  110. )
  111. var_biased = np.var(xv_transposed, axis=0).reshape((1, nr_chan, 1))
  112. sd = np.sqrt(var_biased + bn.eps)
  113. var_unbiased = np.var(xv_transposed, axis=0, ddof=1).reshape((1, nr_chan, 1))
  114. running_mean = running_mean * momentum + mean * (1 - momentum)
  115. running_var = running_var * momentum + var_unbiased * (1 - momentum)
  116. yv = bn(Tensor(xv))
  117. yv_expect = (xv - mean) / sd
  118. _assert_allclose(yv.numpy(), yv_expect)
  119. _assert_allclose(bn.running_mean.numpy().reshape(-1), running_mean.reshape(-1))
  120. _assert_allclose(bn.running_var.numpy().reshape(-1), running_var.reshape(-1))
  121. # test set 'training' flag to False
  122. mean_backup = bn.running_mean.numpy()
  123. var_backup = bn.running_var.numpy()
  124. bn.training = False
  125. xv = np.random.normal(loc=2.3, size=data_shape).astype(np.float32)
  126. data = Tensor(xv)
  127. yv1 = bn(data)
  128. yv2 = bn(data)
  129. np.testing.assert_equal(yv1.numpy(), yv2.numpy())
  130. np.testing.assert_equal(mean_backup, bn.running_mean.numpy())
  131. np.testing.assert_equal(var_backup, bn.running_var.numpy())
  132. yv_expect = (xv - running_mean) / np.sqrt(running_var + bn.eps)
  133. _assert_allclose(yv1.numpy(), yv_expect)
  134. def test_batchnorm2d():
  135. nr_chan = 8
  136. data_shape = (3, nr_chan, 16, 16)
  137. momentum = 0.9
  138. bn = BatchNorm2d(nr_chan, momentum=momentum)
  139. running_mean = np.zeros((1, nr_chan, 1, 1), dtype=np.float32)
  140. running_var = np.ones((1, nr_chan, 1, 1), dtype=np.float32)
  141. for i in range(3):
  142. xv = np.random.normal(loc=2.3, size=data_shape).astype(np.float32)
  143. xv_transposed = np.transpose(xv, [0, 2, 3, 1]).reshape(
  144. (data_shape[0] * data_shape[2] * data_shape[3], nr_chan)
  145. )
  146. mean = np.mean(xv_transposed, axis=0).reshape(1, nr_chan, 1, 1)
  147. var_biased = np.var(xv_transposed, axis=0).reshape((1, nr_chan, 1, 1))
  148. sd = np.sqrt(var_biased + bn.eps)
  149. var_unbiased = np.var(xv_transposed, axis=0, ddof=1).reshape((1, nr_chan, 1, 1))
  150. running_mean = running_mean * momentum + mean * (1 - momentum)
  151. running_var = running_var * momentum + var_unbiased * (1 - momentum)
  152. yv = bn(Tensor(xv))
  153. yv_expect = (xv - mean) / sd
  154. _assert_allclose(yv.numpy(), yv_expect)
  155. _assert_allclose(bn.running_mean.numpy(), running_mean)
  156. _assert_allclose(bn.running_var.numpy(), running_var)
  157. # test set 'training' flag to False
  158. mean_backup = bn.running_mean.numpy()
  159. var_backup = bn.running_var.numpy()
  160. bn.training = False
  161. xv = np.random.normal(loc=2.3, size=data_shape).astype(np.float32)
  162. data = Tensor(xv)
  163. yv1 = bn(data)
  164. yv2 = bn(data)
  165. np.testing.assert_equal(yv1.numpy(), yv2.numpy())
  166. np.testing.assert_equal(mean_backup, bn.running_mean.numpy())
  167. np.testing.assert_equal(var_backup, bn.running_var.numpy())
  168. yv_expect = (xv - running_mean) / np.sqrt(running_var + bn.eps)
  169. _assert_allclose(yv1.numpy(), yv_expect)
  170. def test_syncbn2d():
  171. nr_chan = 8
  172. data_shape = (3, nr_chan, 16, 16)
  173. momentum = 0.9
  174. bn = SyncBatchNorm(nr_chan, momentum=momentum)
  175. running_mean = np.zeros((1, nr_chan, 1, 1), dtype=np.float32)
  176. running_var = np.ones((1, nr_chan, 1, 1), dtype=np.float32)
  177. for i in range(3):
  178. xv = np.random.normal(loc=2.3, size=data_shape).astype(np.float32)
  179. xv_transposed = np.transpose(xv, [0, 2, 3, 1]).reshape(
  180. (data_shape[0] * data_shape[2] * data_shape[3], nr_chan)
  181. )
  182. mean = np.mean(xv_transposed, axis=0).reshape(1, nr_chan, 1, 1)
  183. var_biased = np.var(xv_transposed, axis=0).reshape((1, nr_chan, 1, 1))
  184. sd = np.sqrt(var_biased + bn.eps)
  185. var_unbiased = np.var(xv_transposed, axis=0, ddof=1).reshape((1, nr_chan, 1, 1))
  186. running_mean = running_mean * momentum + mean * (1 - momentum)
  187. running_var = running_var * momentum + var_unbiased * (1 - momentum)
  188. yv = bn(Tensor(xv))
  189. yv_expect = (xv - mean) / sd
  190. _assert_allclose(yv.numpy(), yv_expect)
  191. _assert_allclose(bn.running_mean.numpy(), running_mean)
  192. _assert_allclose(bn.running_var.numpy(), running_var)
  193. # test set 'training' flag to False
  194. mean_backup = bn.running_mean.numpy()
  195. var_backup = bn.running_var.numpy()
  196. bn.training = False
  197. xv = np.random.normal(loc=2.3, size=data_shape).astype(np.float32)
  198. data = Tensor(xv)
  199. yv1 = bn(data)
  200. yv2 = bn(data)
  201. np.testing.assert_equal(yv1.numpy(), yv2.numpy())
  202. np.testing.assert_equal(mean_backup, bn.running_mean.numpy())
  203. np.testing.assert_equal(var_backup, bn.running_var.numpy())
  204. yv_expect = (xv - running_mean) / np.sqrt(running_var + bn.eps)
  205. _assert_allclose(yv1.numpy(), yv_expect)
  206. def test_batchnorm_no_stats():
  207. nr_chan = 8
  208. data_shape = (3, nr_chan, 4)
  209. bn = BatchNorm1d(8, track_running_stats=False)
  210. for i in range(4):
  211. if i == 2:
  212. bn.training = False
  213. xv = np.random.normal(loc=2.3, size=data_shape).astype(np.float32)
  214. mean = np.mean(np.mean(xv, axis=0, keepdims=True), axis=2, keepdims=True)
  215. var = np.var(
  216. np.transpose(xv, [0, 2, 1]).reshape(
  217. (data_shape[0] * data_shape[2], nr_chan)
  218. ),
  219. axis=0,
  220. ).reshape((1, nr_chan, 1))
  221. sd = np.sqrt(var + bn.eps)
  222. yv = bn(Tensor(xv))
  223. yv_expect = (xv - mean) / sd
  224. _assert_allclose(yv.numpy(), yv_expect)
  225. def test_syncbn_no_stats():
  226. nr_chan = 8
  227. data_shape = (3, nr_chan, 4)
  228. bn = SyncBatchNorm(8, track_running_stats=False)
  229. for i in range(4):
  230. if i == 2:
  231. bn.training = False
  232. xv = np.random.normal(loc=2.3, size=data_shape).astype(np.float32)
  233. mean = np.mean(np.mean(xv, axis=0, keepdims=True), axis=2, keepdims=True)
  234. var = np.var(
  235. np.transpose(xv, [0, 2, 1]).reshape(
  236. (data_shape[0] * data_shape[2], nr_chan)
  237. ),
  238. axis=0,
  239. ).reshape((1, nr_chan, 1))
  240. sd = np.sqrt(var + bn.eps)
  241. yv = bn(Tensor(xv))
  242. yv_expect = (xv - mean) / sd
  243. _assert_allclose(yv.numpy(), yv_expect)
  244. def test_batchnorm2d_no_stats():
  245. nr_chan = 8
  246. data_shape = (3, nr_chan, 16, 16)
  247. bn = BatchNorm2d(8, track_running_stats=False)
  248. for i in range(4):
  249. if i == 2:
  250. bn.training = False
  251. xv = np.random.normal(loc=2.3, size=data_shape).astype(np.float32)
  252. xv_transposed = np.transpose(xv, [0, 2, 3, 1]).reshape(
  253. (data_shape[0] * data_shape[2] * data_shape[3], nr_chan)
  254. )
  255. mean = np.mean(xv_transposed, axis=0).reshape(1, nr_chan, 1, 1)
  256. var = np.var(xv_transposed, axis=0).reshape((1, nr_chan, 1, 1))
  257. sd = np.sqrt(var + bn.eps)
  258. yv = bn(Tensor(xv))
  259. yv_expect = (xv - mean) / sd
  260. _assert_allclose(yv.numpy(), yv_expect)
  261. def test_syncbn2d_no_stats():
  262. nr_chan = 8
  263. data_shape = (3, nr_chan, 16, 16)
  264. bn = SyncBatchNorm(8, track_running_stats=False)
  265. for i in range(4):
  266. if i == 2:
  267. bn.training = False
  268. xv = np.random.normal(loc=2.3, size=data_shape).astype(np.float32)
  269. xv_transposed = np.transpose(xv, [0, 2, 3, 1]).reshape(
  270. (data_shape[0] * data_shape[2] * data_shape[3], nr_chan)
  271. )
  272. mean = np.mean(xv_transposed, axis=0).reshape(1, nr_chan, 1, 1)
  273. var = np.var(xv_transposed, axis=0).reshape((1, nr_chan, 1, 1))
  274. sd = np.sqrt(var + bn.eps)
  275. yv = bn(Tensor(xv))
  276. yv_expect = (xv - mean) / sd
  277. _assert_allclose(yv.numpy(), yv_expect)
  278. def test_syncbn2d_grad():
  279. nr_chan = 8
  280. data_shape = (3, nr_chan, 16, 16)
  281. syncbn = SyncBatchNorm(8, track_running_stats=False)
  282. bn = BatchNorm2d(8, track_running_stats=False)
  283. for i in range(4):
  284. if i == 2:
  285. syncbn.training = False
  286. bn.training = False
  287. inp = Tensor(np.random.normal(loc=2.3, size=data_shape).astype(np.float32))
  288. diff = Tensor(np.random.normal(size=data_shape).astype(np.float32))
  289. with GradManager().attach(inp) as gm:
  290. oup = syncbn(inp)
  291. gm.backward(oup, diff)
  292. grad = inp.grad
  293. inp.grad = None
  294. with GradManager().attach(inp) as gm:
  295. oup_expect = bn(inp)
  296. gm.backward(oup_expect, diff)
  297. grad_expect = inp.grad
  298. inp.grad = None
  299. _assert_allclose(oup.numpy(), oup_expect.numpy())
  300. _assert_allclose(grad.numpy(), grad_expect.numpy())
  301. @pytest.mark.parametrize("dim", [1, 2])
  302. @pytest.mark.parametrize("is_symbolic", [None, False, True])
  303. def test_batchnorm_empty_tensor(dim, is_symbolic):
  304. if dim == 1:
  305. m = BatchNorm1d(4, affine=True)
  306. inp = mge.tensor(np.random.randn(0, 4, 0).astype("float32"))
  307. elif dim == 2:
  308. m = BatchNorm2d(4, affine=True)
  309. inp = mge.tensor(np.random.randn(0, 4, 0, 0).astype("float32"))
  310. else:
  311. raise NotImplementedError
  312. m.train()
  313. def fn(inp):
  314. return m(inp)
  315. if is_symbolic is not None:
  316. fn = jit.trace(symbolic=is_symbolic)(fn)
  317. for _ in range(3):
  318. out = fn(inp)
  319. np.testing.assert_equal(out.numpy(), inp)
  320. if is_symbolic is None:
  321. break

MegEngine 安装包中集成了使用 GPU 运行代码所需的 CUDA 环境,不用区分 CPU 和 GPU 版。 如果想要运行 GPU 程序,请确保机器本身配有 GPU 硬件设备并安装好驱动。 如果你想体验在云端 GPU 算力平台进行深度学习开发的感觉,欢迎访问 MegStudio 平台