utils.py 12.7 KB
Newer Older
1
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
2
#
3 4 5
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
6
#
7
#     http://www.apache.org/licenses/LICENSE-2.0
8
#
9 10 11 12 13 14
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

15 16 17 18 19 20 21 22
import typing
import enum
import sys
import re
import inspect
import functools
import contextlib
import collections
23 24
import numpy as np
import paddle
25
from paddle.autograd.functional import _as_tensors
26 27


28 29 30
##########################################################
# Finite Difference Utils
##########################################################
31 32 33 34 35 36 37 38
def _product(t):
    if isinstance(t, int):
        return t
    else:
        return np.product(t)


def _get_item(t, idx):
39 40 41
    assert isinstance(
        t,
        paddle.fluid.framework.Variable), "The first argument t must be Tensor."
42 43 44 45 46 47 48
    assert isinstance(idx,
                      int), "The second argument idx must be an int number."
    flat_t = paddle.reshape(t, [-1])
    return flat_t.__getitem__(idx)


def _set_item(t, idx, value):
49 50 51
    assert isinstance(
        t,
        paddle.fluid.framework.Variable), "The first argument t must be Tensor."
52 53 54 55 56 57 58 59
    assert isinstance(idx,
                      int), "The second argument idx must be an int number."
    flat_t = paddle.reshape(t, [-1])
    flat_t.__setitem__(idx, value)
    return paddle.reshape(flat_t, t.shape)


def _compute_numerical_jacobian(func, xs, delta, np_dtype):
60 61
    xs = list(_as_tensors(xs))
    ys = list(_as_tensors(func(*xs)))
62 63 64 65 66 67
    fin_size = len(xs)
    fout_size = len(ys)
    jacobian = list([] for _ in range(fout_size))
    for i in range(fout_size):
        jac_i = list([] for _ in range(fin_size))
        for j in range(fin_size):
68 69
            jac_i[j] = np.zeros((_product(ys[i].shape), _product(xs[j].shape)),
                                dtype=np_dtype)
70 71 72 73 74 75 76
        jacobian[i] = jac_i

    for j in range(fin_size):
        for q in range(_product(xs[j].shape)):
            orig = _get_item(xs[j], q)
            x_pos = orig + delta
            xs[j] = _set_item(xs[j], q, x_pos)
77
            ys_pos = _as_tensors(func(*xs))
78 79 80

            x_neg = orig - delta
            xs[j] = _set_item(xs[j], q, x_neg)
81
            ys_neg = _as_tensors(func(*xs))
82 83 84 85 86 87 88 89 90 91 92 93

            xs[j] = _set_item(xs[j], q, orig)

            for i in range(fout_size):
                for p in range(_product(ys[i].shape)):
                    y_pos = _get_item(ys_pos[i], p)
                    y_neg = _get_item(ys_neg[i], p)
                    jacobian[i][j][p][q] = (y_pos - y_neg) / delta / 2.
    return jacobian


def _compute_numerical_hessian(func, xs, delta, np_dtype):
94 95
    xs = list(_as_tensors(xs))
    ys = list(_as_tensors(func(*xs)))
96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111
    fin_size = len(xs)
    hessian = list([] for _ in range(fin_size))
    for i in range(fin_size):
        hessian_i = list([] for _ in range(fin_size))
        for j in range(fin_size):
            hessian_i[j] = np.zeros(
                (_product(xs[i].shape), _product(xs[j].shape)), dtype=np_dtype)
        hessian[i] = hessian_i

    for i in range(fin_size):
        for p in range(_product(xs[i].shape)):
            for j in range(fin_size):
                for q in range(_product(xs[j].shape)):
                    orig = _get_item(xs[j], q)
                    x_pos = orig + delta
                    xs[j] = _set_item(xs[j], q, x_pos)
112 113
                    jacobian_pos = _compute_numerical_jacobian(
                        func, xs, delta, np_dtype)
114 115
                    x_neg = orig - delta
                    xs[j] = _set_item(xs[j], q, x_neg)
116 117
                    jacobian_neg = _compute_numerical_jacobian(
                        func, xs, delta, np_dtype)
118 119
                    xs[j] = _set_item(xs[j], q, orig)
                    hessian[i][j][p][q] = (
120 121
                        jacobian_pos[0][i][0][p] -
                        jacobian_neg[0][i][0][p]) / delta / 2.
122
    return hessian
L
levi131 已提交
123 124


125 126 127 128 129 130 131 132 133 134 135 136 137
def concat_to_matrix(xs, is_batched=False):
    """Concats a tuple of tuple of Jacobian/Hessian matrix into one matrix"""
    rows = []
    for i in range(len(xs)):
        rows.append(np.concatenate([x for x in xs[i]], -1))
    return np.concatenate(rows, 1) if is_batched else np.concatenate(rows, 0)


def _compute_numerical_batch_jacobian(func,
                                      xs,
                                      delta,
                                      np_dtype,
                                      merge_batch=True):
138
    no_batch_jacobian = _compute_numerical_jacobian(func, xs, delta, np_dtype)
139 140
    xs = list(_as_tensors(xs))
    ys = list(_as_tensors(func(*xs)))
141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157
    fin_size = len(xs)
    fout_size = len(ys)
    bs = xs[0].shape[0]
    bat_jac = []
    for i in range(fout_size):
        batch_jac_i = []
        for j in range(fin_size):
            jac = no_batch_jacobian[i][j]
            jac_shape = jac.shape
            out_size = jac_shape[0] // bs
            in_size = jac_shape[1] // bs
            jac = np.reshape(jac, (bs, out_size, bs, in_size))
            batch_jac_i_j = np.zeros(shape=(out_size, bs, in_size))
            for p in range(out_size):
                for b in range(bs):
                    for q in range(in_size):
                        batch_jac_i_j[p][b][q] = jac[b][p][b][q]
158 159
            if merge_batch:
                batch_jac_i_j = np.reshape(batch_jac_i_j, (out_size, -1))
160 161 162 163 164 165 166
            batch_jac_i.append(batch_jac_i_j)
        bat_jac.append(batch_jac_i)

    return bat_jac


def _compute_numerical_batch_hessian(func, xs, delta, np_dtype):
167
    xs = list(_as_tensors(xs))
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
    batch_size = xs[0].shape[0]
    fin_size = len(xs)
    hessian = []
    for b in range(batch_size):
        x_l = []
        for j in range(fin_size):
            x_l.append(paddle.reshape(xs[j][b], shape=[1, -1]))
        hes_b = _compute_numerical_hessian(func, x_l, delta, np_dtype)
        if fin_size == 1:
            hessian.append(hes_b[0][0])
        else:
            hessian.append(hes_b)

    hessian_res = []
    for index in range(fin_size):
        x_reshape = paddle.reshape(xs[index], shape=[batch_size, -1])
        for index_ in range(fin_size):
            for i in range(x_reshape.shape[1]):
                tmp = []
                for j in range(batch_size):
                    if fin_size == 1:
                        tmp.extend(hessian[j][i])
                    else:
                        tmp.extend(hessian[j][i][index_][index])
                hessian_res.append(tmp)
        if fin_size == 1:
            return hessian_res

    hessian_result = []
    mid = len(hessian_res) // 2
    for i in range(mid):
        hessian_result.append(
200
            np.stack((hessian_res[i], hessian_res[mid + i]), axis=0))
201 202 203
    return hessian_result


L
levi131 已提交
204
def _compute_numerical_vjp(func, xs, v, delta, np_dtype):
205
    xs = _as_tensors(xs)
L
levi131 已提交
206
    jacobian = np.array(_compute_numerical_jacobian(func, xs, delta, np_dtype))
207 208
    if v is None:
        v = [paddle.ones_like(x) for x in xs]
L
levi131 已提交
209 210 211 212 213 214 215 216 217 218 219
    flat_v = np.array([v_el.numpy().reshape(-1) for v_el in v])
    vjp = [np.zeros((_product(x.shape)), dtype=np_dtype) for x in xs]
    for j in range(len(xs)):
        for q in range(_product(xs[j].shape)):
            vjp[j][q] = np.sum(jacobian[:, j, :, q].reshape(flat_v.shape) *
                               flat_v)
    vjp = [vjp[j].reshape(xs[j].shape) for j in range(len(xs))]
    return vjp


def _compute_numerical_vhp(func, xs, v, delta, np_dtype):
220
    xs = list(_as_tensors(xs))
L
levi131 已提交
221 222 223 224 225 226 227 228 229
    hessian = np.array(_compute_numerical_hessian(func, xs, delta, np_dtype))
    flat_v = np.array([v_el.numpy().reshape(-1) for v_el in v])
    vhp = [np.zeros((_product(x.shape)), dtype=np_dtype) for x in xs]
    for j in range(len(xs)):
        for q in range(_product(xs[j].shape)):
            vhp[j][q] = np.sum(hessian[:, j, :, q].reshape(flat_v.shape) *
                               flat_v)
    vhp = [vhp[j].reshape(xs[j].shape) for j in range(len(xs))]
    return vhp
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


##########################################################
# TestCases of different function.
##########################################################
def reduce(x):
    return paddle.sum(x)


def reduce_dim(x):
    return paddle.sum(x, axis=0)


def matmul(x, y):
    return paddle.matmul(x, y)


def mul(x, y):
    return x * y


def pow(x, y):
    return paddle.pow(x, y)


def o2(x, y):
    return paddle.multiply(x, y), paddle.matmul(x, y.t())


def unuse(x, y):
    return paddle.sum(x)


def nested(x):
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
    def inner(y):
        return x * y

    return inner


def square(x):
    return x * x


##########################################################
# Parameterized Test Utils.
##########################################################

TEST_CASE_NAME = 'suffix'


def place(devices, key='place'):
    """A Decorator for a class which will make the class running on different 
    devices .

    Args:
        devices (Sequence[Paddle.CUDAPlace|Paddle.CPUPlace]): Device list.
        key (str, optional): Defaults to 'place'.
    """

    def decorate(cls):
        module = sys.modules[cls.__module__].__dict__
        raw_classes = {
            k: v
            for k, v in module.items() if k.startswith(cls.__name__)
        }

        for raw_name, raw_cls in raw_classes.items():
            for d in devices:
                test_cls = dict(raw_cls.__dict__)
                test_cls.update({key: d})
                new_name = raw_name + '.' + d.__class__.__name__
                module[new_name] = type(new_name, (raw_cls, ), test_cls)
            del module[raw_name]
        return cls

    return decorate


def parameterize(fields, values=None):
    """Decorator for a unittest class which make the class running on different 
    test cases.

    Args:
        fields (Sequence): The feild name sequence of test cases.
        values (Sequence, optional): The test cases sequence. Defaults to None.

    """
    fields = [fields] if isinstance(fields, str) else fields
    params = [dict(zip(fields, vals)) for vals in values]

    def decorate(cls):
        test_cls_module = sys.modules[cls.__module__].__dict__
        for i, values in enumerate(params):
            test_cls = dict(cls.__dict__)
            values = {
                k: staticmethod(v) if callable(v) else v
                for k, v in values.items()
            }
            test_cls.update(values)
            name = cls.__name__ + str(i)
            name = name + '.' + \
                values.get('suffix') if values.get('suffix') else name

            test_cls_module[name] = type(name, (cls, ), test_cls)

        for m in list(cls.__dict__):
            if m.startswith("test"):
                delattr(cls, m)
        return cls

    return decorate


##########################################################
# Utils for transpose different Jacobian/Hessian matrix format.
##########################################################

# B is batch size, N is row size, M is column size.
MatrixFormat = enum.Enum('MatrixFormat', ('NBM', 'BNM', 'NMB', 'NM'))


def _np_transpose_matrix_format(src, src_format, des_format):
    """Transpose Jacobian/Hessian matrix format."""
    supported_format = (MatrixFormat.NBM, MatrixFormat.BNM, MatrixFormat.NMB)
    if src_format not in supported_format or des_format not in supported_format:
        raise ValueError(
            f"Supported Jacobian format is {supported_format}, but got src: {src_format}, des: {des_format}"
        )

    src_axis = {c: i for i, c in enumerate(src_format.name)}
    dst_axis = tuple(src_axis[c] for c in des_format.name)

    return np.transpose(src, dst_axis)


def _np_concat_matrix_sequence(src, src_format=MatrixFormat.NM):
    """Convert a sequence of sequence of Jacobian/Hessian matrix into one huge 
    matrix."""

    def concat_col(xs):
        if src_format in (MatrixFormat.NBM, MatrixFormat.BNM, MatrixFormat.NM):
            return np.concatenate(xs, axis=-1)
        else:
            return np.concatenate(xs, axis=1)

    def concat_row(xs):
        if src_format in (MatrixFormat.NBM, MatrixFormat.NM, MatrixFormat.NMB):
            return np.concatenate(xs, axis=0)
        else:
            return np.concatenate(xs, axis=1)

    supported_format = (MatrixFormat.NBM, MatrixFormat.BNM, MatrixFormat.NMB,
                        MatrixFormat.NM)
    if src_format not in supported_format:
        raise ValueError(
            f"Supported Jacobian format is {supported_format}, but got {src_format}"
        )
    if not isinstance(src, typing.Sequence):
        return src
    if not isinstance(src[0], typing.Sequence):
        src = [src]
    return concat_row(tuple(concat_col(xs) for xs in src))