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.

observer.py 16 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447
  1. # MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
  2. #
  3. # Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
  4. #
  5. # Unless required by applicable law or agreed to in writing,
  6. # software distributed under the License is distributed on an
  7. # "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  8. import math
  9. from abc import abstractmethod
  10. from enum import Enum
  11. import numpy as np
  12. from .. import functional as F
  13. from .._internal.dtype import _metadata_dict, get_quantized_dtype
  14. from ..core import Buffer, Function, tensor
  15. from ..jit import sideeffect
  16. from ..module import Module
  17. class Round(Function):
  18. def forward(self, x):
  19. return x.round()
  20. def backward(self, output_grads):
  21. return output_grads
  22. class Observer(Module):
  23. r"""
  24. A base class for Observer Module.
  25. :param dtype: a string indicating to collect scale and zero_point of which dtype
  26. :param narrow_range: Whether the absolute value of ``qmin`` is the same as ``qmax``,
  27. instead of 1 greater. Usually True for weight and False for activation.
  28. """
  29. def __init__(self, dtype: str, narrow_range: bool = False):
  30. super().__init__()
  31. if dtype not in _metadata_dict.keys():
  32. raise ValueError(
  33. "unknown dtype: {}, only support {}".format(
  34. dtype, _metadata_dict.keys()
  35. )
  36. )
  37. self.dtype = dtype
  38. self.narrow_range = narrow_range
  39. self.qmin = (
  40. -_metadata_dict[dtype].qmax if narrow_range else _metadata_dict[dtype].qmin
  41. )
  42. self.qmax = _metadata_dict[dtype].qmax
  43. self.enabled = True
  44. def get_dtype(self):
  45. q_dict = self.get_qparams()
  46. numpy_scale = None if "scale" not in q_dict else q_dict["scale"].numpy()[0]
  47. numpy_zero_point = (
  48. None if "zero_point" not in q_dict else q_dict["zero_point"].numpy()[0]
  49. )
  50. return get_quantized_dtype(self.dtype, numpy_scale, numpy_zero_point)
  51. def enable(self):
  52. self.enabled = True
  53. def disable(self):
  54. self.enabled = False
  55. def train(self, mode: bool = True, recursive: bool = True) -> None:
  56. super().train(mode, recursive)
  57. if mode:
  58. self.enable()
  59. else:
  60. self.disable()
  61. @abstractmethod
  62. def forward(self, x):
  63. pass
  64. @abstractmethod
  65. def get_qparams(self, **kwargs):
  66. pass
  67. class ObserverMode(Enum):
  68. SYMMERTIC = 1
  69. ASYMMERTIC = 2
  70. def create_observer_dict(mode):
  71. if mode == ObserverMode.SYMMERTIC:
  72. return {
  73. "mode": ObserverMode.SYMMERTIC,
  74. "scale": None,
  75. }
  76. else:
  77. return {
  78. "mode": ObserverMode.ASYMMERTIC,
  79. "scale": None,
  80. "zero_point": None,
  81. }
  82. class MinMaxObserver(Observer):
  83. def __init__(
  84. self,
  85. mode=ObserverMode.SYMMERTIC,
  86. eps=0.00001,
  87. dtype="qint8",
  88. narrow_range: bool = False,
  89. ):
  90. super().__init__(dtype, narrow_range)
  91. self.mode = mode
  92. self.min_val = Buffer(np.finfo(np.float32).max, dtype=np.float32)
  93. self.max_val = Buffer(np.finfo(np.float32).min, dtype=np.float32)
  94. self.scale_limit = eps
  95. def _calculate_qparams(self, inp_min_val, inp_max_val):
  96. min_val = F.minimum(0.0, inp_min_val)
  97. max_val = F.maximum(0.0, inp_max_val)
  98. q_dict = create_observer_dict(self.mode)
  99. q_dict["min_val"] = inp_min_val
  100. q_dict["max_val"] = inp_max_val
  101. if self.mode == ObserverMode.SYMMERTIC:
  102. symmetric_max_vals = F.maximum(-min_val, max_val)
  103. # use maximun to avoid scale too small at the begin
  104. q_dict["scale"] = F.maximum(
  105. symmetric_max_vals / ((self.qmax - self.qmin) / 2), self.scale_limit
  106. )
  107. # zero_point = self.zero_point
  108. else:
  109. # use maximun to avoid scale too small at the begin
  110. q_dict["scale"] = F.maximum(
  111. (max_val - min_val) / (self.qmax - self.qmin), self.scale_limit,
  112. )
  113. # caculate zero_point
  114. q_dict["zero_point"] = self.qmin - Round()((min_val / q_dict["scale"]))
  115. return q_dict
  116. def get_qparams(self):
  117. return self._calculate_qparams(self.min_val, self.max_val)
  118. def forward(self, x_orig):
  119. if self.enabled:
  120. # stop gradient
  121. x = F.zero_grad(x_orig)
  122. # find max and min
  123. F.add_update(
  124. self.min_val,
  125. F.minimum(self.min_val, x.min()),
  126. alpha=0.0,
  127. beta=1.0,
  128. bias=0.0,
  129. )
  130. F.add_update(
  131. self.max_val,
  132. F.maximum(self.max_val, x.max()),
  133. alpha=0.0,
  134. beta=1.0,
  135. bias=0.0,
  136. )
  137. return x_orig
  138. class ExponentialMovingAverageObserver(MinMaxObserver):
  139. def __init__(
  140. self,
  141. momentum=0.9,
  142. mode=ObserverMode.SYMMERTIC,
  143. eps=0.00001,
  144. dtype="qint8",
  145. narrow_range: bool = False,
  146. ):
  147. super().__init__(mode, eps, dtype, narrow_range)
  148. self.momentum = Buffer(momentum)
  149. self.runtime_momentum = Buffer(0.0)
  150. def set_momentum(self, momentum):
  151. self.momentum.set_value(momentum)
  152. def forward(self, x_orig):
  153. if self.enabled:
  154. # stop gradient
  155. x = F.zero_grad(x_orig)
  156. # Exponential Moving Average
  157. tmp_min = (
  158. self.min_val * self.runtime_momentum
  159. + (1 - self.runtime_momentum) * x.min()
  160. )
  161. tmp_max = (
  162. self.max_val * self.runtime_momentum
  163. + (1 - self.runtime_momentum) * x.max()
  164. )
  165. F.add_update(self.min_val, tmp_min, alpha=0.0, beta=1.0, bias=0.0)
  166. F.add_update(self.max_val, tmp_max, alpha=0.0, beta=1.0, bias=0.0)
  167. F.add_update(
  168. self.runtime_momentum, self.momentum, alpha=0.0, beta=1.0, bias=0.0
  169. )
  170. return x_orig
  171. class HistogramObserver(MinMaxObserver):
  172. def __init__(
  173. self,
  174. bins=2048,
  175. upsample_rate=128,
  176. mode=ObserverMode.SYMMERTIC,
  177. eps=0.00001,
  178. dtype="qint8",
  179. narrow_range: bool = False,
  180. ):
  181. super().__init__(mode, eps, dtype, narrow_range)
  182. self.bins = bins
  183. self.upsample_rate = upsample_rate
  184. self.dst_nbins = _metadata_dict[dtype].qmax - _metadata_dict[dtype].qmin + 1
  185. self.histogram = Buffer([-1] + [0.0] * (bins - 1))
  186. def _non_linear_param_search(self):
  187. r"""Non-linear parameter search.
  188. An approximation for L2 error minimization for selecting min/max.
  189. By selecting new min/max, we filter out outliers in input distribution.
  190. """
  191. np_min_val = self.min_val.numpy()[0]
  192. np_max_val = self.max_val.numpy()[0]
  193. np_histogram = self.histogram.numpy()
  194. assert len(np_histogram) == self.bins, "bins mistmatch"
  195. bin_width = (np_max_val - np_min_val) / self.bins
  196. def _get_norm(delta_begin, delta_end, density, norm_type):
  197. r"""
  198. Compute the norm of the values uniformaly distributed between
  199. delta_begin and delta_end.
  200. norm = density * (integral_{begin, end} x^2)
  201. = density * (end^3 - begin^3) / 3
  202. """
  203. assert norm_type == "L2", "Only L2 norms are currently supported"
  204. norm = 0.0
  205. if norm_type == "L2":
  206. norm = (
  207. delta_end * delta_end * delta_end
  208. - delta_begin * delta_begin * delta_begin
  209. ) / 3
  210. return density * norm
  211. def _compute_quantization_error(next_start_bin, next_end_bin, norm_type):
  212. r"""
  213. Compute the quantization error if we use start_bin to end_bin as the
  214. min and max to do the quantization.
  215. """
  216. norm = 0.0
  217. dst_bin_width = (
  218. bin_width * (next_end_bin - next_start_bin + 1) / self.dst_nbins
  219. )
  220. if dst_bin_width == 0.0:
  221. return 0.0
  222. for src_bin in range(self.bins):
  223. # distances from the beginning of first dst_bin to the beginning and
  224. # end of src_bin
  225. src_bin_begin = (src_bin - next_start_bin) * bin_width
  226. src_bin_end = src_bin_begin + bin_width
  227. # which dst_bins the beginning and end of src_bin belong to?
  228. dst_bin_of_begin = min(
  229. self.dst_nbins - 1,
  230. max(0.0, math.floor(src_bin_begin / dst_bin_width)),
  231. )
  232. dst_bin_of_end = min(
  233. self.dst_nbins - 1,
  234. max(0.0, math.floor(src_bin_end / dst_bin_width)),
  235. )
  236. dst_bin_of_begin_center = (
  237. dst_bin_of_begin * dst_bin_width + dst_bin_width / 2
  238. )
  239. density = np_histogram[src_bin] / bin_width
  240. if dst_bin_of_begin == dst_bin_of_end:
  241. # if src_bin is entirely within 1 dst_bin
  242. delta_begin = src_bin_begin - dst_bin_of_begin_center
  243. delta_end = src_bin_end - dst_bin_of_begin_center
  244. norm = norm + _get_norm(delta_begin, delta_end, density, norm_type)
  245. else:
  246. delta_begin = src_bin_begin - dst_bin_of_begin_center
  247. delta_end = dst_bin_width / 2
  248. norm = norm + _get_norm(delta_begin, delta_end, density, norm_type)
  249. norm = norm + (dst_bin_of_end - dst_bin_of_begin - 1) * _get_norm(
  250. -dst_bin_width / 2, dst_bin_width / 2, density, norm_type
  251. )
  252. dst_bin_of_end_center = (
  253. dst_bin_of_end * dst_bin_width + dst_bin_width / 2
  254. )
  255. delta_begin = -dst_bin_width / 2
  256. delta_end = src_bin_end - dst_bin_of_end_center
  257. norm = norm + _get_norm(delta_begin, delta_end, density, norm_type)
  258. return norm
  259. # cumulative sum
  260. total = sum(np_histogram)
  261. cSum = np.cumsum(np_histogram, axis=0)
  262. stepsize = 1e-5 # granularity
  263. alpha = 0.0 # lower bound
  264. beta = 1.0 # upper bound
  265. start_bin = 0
  266. end_bin = self.bins - 1
  267. norm_min = float("inf")
  268. while alpha < beta:
  269. # Find the next step
  270. next_alpha = alpha + stepsize
  271. next_beta = beta - stepsize
  272. # find the left and right bins between the quantile bounds
  273. l = start_bin
  274. r = end_bin
  275. while l < end_bin and cSum[l] < next_alpha * total:
  276. l = l + 1
  277. while r > start_bin and cSum[r] > next_beta * total:
  278. r = r - 1
  279. # decide the next move
  280. next_start_bin = start_bin
  281. next_end_bin = end_bin
  282. if (l - start_bin) > (end_bin - r):
  283. # move the start bin
  284. next_start_bin = l
  285. alpha = next_alpha
  286. else:
  287. # move the end bin
  288. next_end_bin = r
  289. beta = next_beta
  290. if next_start_bin == start_bin and next_end_bin == end_bin:
  291. continue
  292. # calculate the quantization error using next_start_bin and next_end_bin
  293. norm = _compute_quantization_error(next_start_bin, next_end_bin, "L2")
  294. if norm > norm_min:
  295. break
  296. norm_min = norm
  297. start_bin = next_start_bin
  298. end_bin = next_end_bin
  299. new_min = self.min_val + bin_width * start_bin
  300. new_max = self.min_val + bin_width * (end_bin + 1)
  301. return new_min, new_max
  302. def get_qparams(self):
  303. new_min, new_max = self._non_linear_param_search()
  304. return self._calculate_qparams(new_min, new_max)
  305. def _combine_histograms(
  306. self, orig_hist, new_hist, upsample_rate, downsample_rate, start_idx, Nbins
  307. ):
  308. # First up-sample the histogram with new data by a factor of L
  309. # This creates an approximate probability density thats piecwise constant
  310. upsampled_histogram = new_hist.repeat(upsample_rate)
  311. # Now insert the upsampled histogram into the output
  312. # histogram, which is initialized with zeros.
  313. # The offset at which the histogram is introduced is determined
  314. # by the start index as the output histogram can cover a wider range
  315. histogram_with_output_range = np.zeros((Nbins * downsample_rate))
  316. histogram_with_output_range[
  317. start_idx : Nbins * upsample_rate + start_idx
  318. ] = upsampled_histogram
  319. # Compute integral histogram, double precision is needed to ensure
  320. # that there are no overflows
  321. integral_histogram = np.cumsum(histogram_with_output_range, 0)[
  322. downsample_rate - 1 :: downsample_rate
  323. ]
  324. # Finally perform interpolation
  325. shifted_integral_histogram = np.zeros((Nbins))
  326. shifted_integral_histogram[1:Nbins] = integral_histogram[0:-1]
  327. interpolated_histogram = (
  328. integral_histogram - shifted_integral_histogram
  329. ) / upsample_rate
  330. orig_hist = orig_hist + interpolated_histogram
  331. return orig_hist
  332. def _adjust_min_max(self, combined_min, combined_max, upsample_rate):
  333. # We ensure that:
  334. # (combined_max - combined_min)/(downsample_rate*Nbins) = (max - min)/(upsample_rate*Nbins)
  335. # This allows us to have a common grid of resolution s, where we can align
  336. # the input histogram
  337. # start_idx maps min_val to the histogram bin index.
  338. np_min_val = self.min_val.numpy()[0]
  339. np_max_val = self.max_val.numpy()[0]
  340. hist_bin_width = (np_max_val - np_min_val) / (self.bins * upsample_rate)
  341. downsample_rate = int(
  342. np.ceil((combined_max - combined_min) / (self.bins * hist_bin_width))
  343. )
  344. e = downsample_rate * (self.bins * hist_bin_width) - (
  345. combined_max - combined_min
  346. )
  347. combined_max = combined_max + e / 2
  348. combined_min = combined_min - e / 2
  349. start_idx = int(np.round((np_min_val - combined_min) / hist_bin_width))
  350. return combined_min, combined_max, downsample_rate, start_idx
  351. @sideeffect
  352. def sideeffect_forward(self, x_orig):
  353. x = x_orig.numpy()
  354. min_val = self.min_val.numpy()[0]
  355. max_val = self.max_val.numpy()[0]
  356. histogram = self.histogram.numpy()
  357. new_min = x.min()
  358. new_max = x.max()
  359. if histogram[0] == -1:
  360. new_histogram, _ = np.histogram(x, self.bins, (new_min, new_max))
  361. else:
  362. new_min = min(new_min, min_val)
  363. new_max = max(new_max, max_val)
  364. # combine the existing histogram and new histogram into 1 histogram
  365. # We do this by first upsampling the histogram to a dense grid
  366. # and then downsampling the histogram efficiently
  367. (new_min, new_max, downsample_rate, start_idx,) = self._adjust_min_max(
  368. new_min, new_max, self.upsample_rate
  369. )
  370. new_histogram, _ = np.histogram(x, self.bins, (new_min, new_max))
  371. new_histogram = new_histogram.astype(np.float64)
  372. if new_min == min_val and new_max == max_val:
  373. new_histogram += histogram
  374. else:
  375. new_histogram = self._combine_histograms(
  376. new_histogram,
  377. histogram,
  378. self.upsample_rate,
  379. downsample_rate,
  380. start_idx,
  381. self.bins,
  382. )
  383. self.histogram.set_value(new_histogram)
  384. self.min_val.set_value(new_min)
  385. self.max_val.set_value(new_max)
  386. def forward(self, x_orig):
  387. self.sideeffect_forward(x_orig)
  388. return x_orig

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