test_lookup_table_op.py 16.1 KB
Newer Older
1
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
D
dzhwinter 已提交
2
#
D
dzhwinter 已提交
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
D
dzhwinter 已提交
6
#
D
dzhwinter 已提交
7
#     http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 已提交
8
#
D
dzhwinter 已提交
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
import unittest
16

17
import numpy as np
W
wanghuancoder 已提交
18 19 20 21 22 23
from eager_op_test import (
    OpTest,
    check_out_dtype,
    paddle_static_guard,
    skip_check_grad_ci,
)
L
LoneRanger 已提交
24
from op import Operator
25

26
import paddle
27
import paddle.nn.functional as F
28 29
from paddle import fluid
from paddle.fluid import Program, core, program_guard
30 31


Q
qijun 已提交
32
class TestLookupTableOp(OpTest):
33
    def setUp(self):
Q
qijun 已提交
34
        self.op_type = "lookup_table"
35
        table = np.random.random((17, 31)).astype("float64")
36
        ids = np.random.randint(0, 17, 4).astype("int64")
37 38
        ids_expand = np.expand_dims(ids, axis=1)
        self.inputs = {'W': table, 'Ids': ids_expand}
39 40
        self.outputs = {'Out': table[ids]}

Q
qijun 已提交
41
    def test_check_output(self):
42
        self.check_output(check_cinn=True)
43

Q
qijun 已提交
44
    def test_check_grad(self):
45
        self.check_grad(['W'], 'Out', no_grad_set=set('Ids'), check_cinn=True)
46 47


F
fengjiayi 已提交
48 49 50
class TestLookupTableOpWithTensorIds(OpTest):
    def setUp(self):
        self.op_type = "lookup_table"
51
        table = np.random.random((17, 31)).astype("float64")
52 53 54
        ids = np.random.randint(low=0, high=17, size=(2, 4, 5, 1)).astype(
            "int64"
        )
F
fengjiayi 已提交
55 56 57 58
        self.inputs = {'W': table, 'Ids': ids}
        self.outputs = {'Out': table[ids.flatten()].reshape((2, 4, 5, 31))}

    def test_check_output(self):
59
        self.check_output(check_cinn=True)
F
fengjiayi 已提交
60 61

    def test_check_grad(self):
62
        self.check_grad(['W'], 'Out', no_grad_set=set('Ids'), check_cinn=True)
F
fengjiayi 已提交
63 64


65 66 67
@skip_check_grad_ci(
    reason="Since paddings are not trainable and fixed in forward,"
    "the gradient of paddings makes no sense and we don't "
68 69
    "test the gradient here."
)
70 71 72 73 74
class TestLookupTableOpWithPadding(TestLookupTableOp):
    def test_check_output(self):
        ids = np.squeeze(self.inputs['Ids'])
        padding_idx = np.random.choice(ids, 1)[0]
        self.outputs['Out'][ids == padding_idx] = np.zeros(31)
75
        self.attrs = {'padding_idx': int(padding_idx)}
76
        self.check_output(check_cinn=True)
77 78


79 80 81
@skip_check_grad_ci(
    reason="Since paddings are not trainable and fixed in forward,"
    "the gradient of paddings makes no sense and we don't "
82 83
    "test the gradient here."
)
F
fengjiayi 已提交
84 85 86 87 88 89
class TestLookupTableOpWithTensorIdsAndPadding(TestLookupTableOpWithTensorIds):
    def test_check_output(self):
        ids = self.inputs['Ids']
        flatten_idx = ids.flatten()
        padding_idx = np.random.choice(flatten_idx, 1)[0]
        self.outputs['Out'][np.squeeze(ids == padding_idx)] = np.zeros(31)
90
        self.attrs = {'padding_idx': padding_idx}
91
        self.check_output(check_cinn=True)
F
fengjiayi 已提交
92

Q
qiaolongfei 已提交
93

94
class TestLookupTableWIsSelectedRows(unittest.TestCase):
F
fengjiayi 已提交
95
    def prepare_ids(self, scope, place):
Q
qiaolongfei 已提交
96 97 98
        ids_tensor = scope.var('Ids').get_tensor()
        ids_array = np.array([[0], [4], [3], [5]]).astype("int64")
        ids_tensor.set(ids_array, place)
F
fengjiayi 已提交
99
        return ids_array
Q
qiaolongfei 已提交
100

F
fengjiayi 已提交
101
    def prepare_w(self, scope, place):
Q
qiaolongfei 已提交
102 103 104 105 106 107 108 109 110
        rows = [0, 1, 2, 3, 4, 5, 6]
        row_numel = 12

        w_selected_rows = scope.var('W').get_selected_rows()
        w_selected_rows.set_height(len(rows))
        w_selected_rows.set_rows(rows)
        w_array = np.ones((len(rows), row_numel)).astype("float32")
        for i in range(len(rows)):
            w_array[i] *= i
Q
qiaolongfei 已提交
111 112
        w_tensor = w_selected_rows.get_tensor()
        w_tensor.set(w_array, place)
Q
qiaolongfei 已提交
113

F
fengjiayi 已提交
114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129
    def create_out_tensor(self, scope, place):
        return scope.var('Out').get_tensor()

    def check_result(self, ids_array, result_array):
        # all(): return True if all elements of the iterable are true (or if the iterable is empty)
        for idx, row in enumerate(ids_array):
            assert (row[0] == result_array[idx]).all()

    def check_with_place(self, place):
        scope = core.Scope()

        ids_array = self.prepare_ids(scope, place)

        self.prepare_w(scope, place)

        out_tensor = self.create_out_tensor(scope, place)
Q
qiaolongfei 已提交
130 131 132 133 134 135

        # create and run lookup_table operator
        lookup_table = Operator("lookup_table", W='W', Ids='Ids', Out='Out')
        lookup_table.run(scope, place)

        # get result from Out
Q
qiaolongfei 已提交
136
        result_array = np.array(out_tensor)
F
fengjiayi 已提交
137 138

        self.check_result(ids_array, result_array)
Q
qiaolongfei 已提交
139 140 141 142 143 144 145 146

    def test_w_is_selected_rows(self):
        places = [core.CPUPlace()]
        # currently only support CPU
        for place in places:
            self.check_with_place(place)


147 148 149
class TestLookupTableWithTensorIdsWIsSelectedRows(
    TestLookupTableWIsSelectedRows
):
F
fengjiayi 已提交
150 151
    def prepare_ids(self, scope, place):
        ids_tensor = scope.var('Ids').get_tensor()
152 153 154
        ids_array = np.random.randint(low=0, high=6, size=(2, 4, 3, 1)).astype(
            "int64"
        )
F
fengjiayi 已提交
155 156 157 158 159 160 161 162
        ids_tensor.set(ids_array, place)
        return ids_array

    def check_result(self, ids_array, result_array):
        for idx, row in np.ndenumerate(ids_array):
            assert (row == result_array[idx]).all()


163
class TestEmbedOpError(unittest.TestCase):
164
    def test_errors(self):
W
wanghuancoder 已提交
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
        with paddle_static_guard():
            with program_guard(Program(), Program()):
                input_data = np.random.randint(0, 10, (4, 1)).astype("int64")

                def test_Variable():
                    # the input type must be Variable
                    fluid.layers.embedding(input=input_data, size=(10, 64))

                self.assertRaises(TypeError, test_Variable)

                def test_input_dtype():
                    # the input dtype must be int64
                    input = paddle.static.data(
                        name='x', shape=[4, 1], dtype='float32'
                    )
                    fluid.layers.embedding(input=input, size=(10, 64))

                self.assertRaises(TypeError, test_input_dtype)

                def test_param_dtype():
                    # dtype must be float32 or float64
                    input2 = paddle.static.data(
                        name='x2', shape=[4, 1], dtype='int64'
                    )
                    fluid.layers.embedding(
                        input=input2, size=(10, 64), dtype='int64'
                    )

                self.assertRaises(TypeError, test_param_dtype)

                input3 = paddle.static.data(
                    name='x3', shape=[4, 1], dtype='int64'
197
                )
198
                fluid.layers.embedding(
W
wanghuancoder 已提交
199
                    input=input3, size=(10, 64), dtype='float16'
200
                )
201 202


203 204 205
class TestLookupTableOpInt8(OpTest):
    def setUp(self):
        self.op_type = "lookup_table"
206 207 208
        table = np.random.randint(low=-128, high=127, size=(17, 31)).astype(
            "int8"
        )
209 210 211 212 213 214
        ids = np.random.randint(0, 17, 4).astype("int64")
        ids_expand = np.expand_dims(ids, axis=1)
        self.inputs = {'W': table, 'Ids': ids_expand}
        self.outputs = {'Out': table[ids]}

    def test_check_output(self):
215
        self.check_output(check_cinn=True)
216 217

    def test_check_grad(self):
218
        # since int8 type only be used in test and inference, there is
219 220 221 222 223 224 225
        # no gradient implement, so we don't need to test it
        pass


class TestLookupTableOpWithTensorIdsInt8(OpTest):
    def setUp(self):
        self.op_type = "lookup_table"
226 227 228 229 230 231
        table = np.random.randint(low=-128, high=127, size=(17, 31)).astype(
            "int8"
        )
        ids = np.random.randint(low=0, high=17, size=(2, 4, 5, 1)).astype(
            "int64"
        )
232 233 234 235
        self.inputs = {'W': table, 'Ids': ids}
        self.outputs = {'Out': table[ids.flatten()].reshape((2, 4, 5, 31))}

    def test_check_output(self):
236
        self.check_output(check_cinn=True)
237 238

    def test_check_grad(self):
239
        # since int8 type only be used in test and inference, there is
240 241 242 243 244 245 246 247 248 249
        # no gradient implement, so we don't need to test it
        pass


class TestLookupTableOpWithPaddingInt8(TestLookupTableOpInt8):
    def test_check_output(self):
        ids = np.squeeze(self.inputs['Ids'])
        padding_idx = np.random.choice(ids, 1)[0]
        self.outputs['Out'][ids == padding_idx] = np.zeros(31)
        self.attrs = {'padding_idx': int(padding_idx)}
250
        self.check_output(check_cinn=True)
251 252 253 254 255 256 257 258

    def test_check_grad(self):
        # Since paddings are not trainable and fixed in forward, the gradient of
        # paddings makes no sense and we don't test the gradient here.
        pass


class TestLookupTableOpWithTensorIdsAndPaddingInt8(
259 260
    TestLookupTableOpWithTensorIdsInt8
):
261 262 263 264 265
    def test_check_output(self):
        ids = self.inputs['Ids']
        flatten_idx = ids.flatten()
        padding_idx = np.random.choice(flatten_idx, 1)[0]
        self.outputs['Out'][np.squeeze(ids == padding_idx)] = np.zeros(31)
266
        self.attrs = {'padding_idx': padding_idx}
267
        self.check_output(check_cinn=True)
268 269 270 271 272 273 274

    def test_check_grad(self):
        # Since paddings are not trainable and fixed in forward, the gradient of
        # paddings makes no sense and we don't test the gradient here.
        pass


275
class TestLookupTableWIsSelectedRowsInt8(unittest.TestCase):
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
    def prepare_ids(self, scope, place):
        ids_tensor = scope.var('Ids').get_tensor()
        ids_array = np.array([[0], [4], [3], [5]]).astype("int64")
        ids_tensor.set(ids_array, place)
        return ids_array

    def prepare_w(self, scope, place):
        rows = [0, 1, 2, 3, 4, 5, 6]
        row_numel = 12

        w_selected_rows = scope.var('W').get_selected_rows()
        w_selected_rows.set_height(len(rows))
        w_selected_rows.set_rows(rows)
        w_array = np.ones((len(rows), row_numel)).astype("int8")
        for i in range(len(rows)):
            w_array[i] *= i
        w_tensor = w_selected_rows.get_tensor()
        w_tensor.set(w_array, place)

    def create_out_tensor(self, scope, place):
        return scope.var('Out').get_tensor()

    def check_result(self, ids_array, result_array):
        # all(): return True if all elements of the iterable are true (or if the iterable is empty)
        for idx, row in enumerate(ids_array):
            assert (row[0] == result_array[idx]).all()

    def check_with_place(self, place):
        scope = core.Scope()

        ids_array = self.prepare_ids(scope, place)

        self.prepare_w(scope, place)

        out_tensor = self.create_out_tensor(scope, place)

        # create and run lookup_table operator
        lookup_table = Operator("lookup_table", W='W', Ids='Ids', Out='Out')
        lookup_table.run(scope, place)

        # get result from Out
        result_array = np.array(out_tensor)

        self.check_result(ids_array, result_array)

    def test_w_is_selected_rows(self):
        places = [core.CPUPlace()]
        # currently only support CPU
        for place in places:
            self.check_with_place(place)


class TestLookupTableWithTensorIdsWIsSelectedRowsInt8(
329 330
    TestLookupTableWIsSelectedRowsInt8
):
331 332
    def prepare_ids(self, scope, place):
        ids_tensor = scope.var('Ids').get_tensor()
333 334 335
        ids_array = np.random.randint(low=0, high=6, size=(2, 4, 3, 1)).astype(
            "int64"
        )
336 337 338 339 340 341 342 343
        ids_tensor.set(ids_array, place)
        return ids_array

    def check_result(self, ids_array, result_array):
        for idx, row in np.ndenumerate(ids_array):
            assert (row == result_array[idx]).all()


344 345 346 347
@skip_check_grad_ci(reason="Int16 type only be used in test and inference.")
class TestLookupTableOpInt16(OpTest):
    def setUp(self):
        self.op_type = "lookup_table"
348 349 350
        table = np.random.randint(low=-128, high=127, size=(17, 31)).astype(
            "int16"
        )
351 352 353 354 355 356
        ids = np.random.randint(0, 17, 4).astype("int64")
        ids_expand = np.expand_dims(ids, axis=1)
        self.inputs = {'W': table, 'Ids': ids_expand}
        self.outputs = {'Out': table[ids]}

    def test_check_output(self):
357
        self.check_output(check_cinn=True)
358 359 360 361 362 363


@skip_check_grad_ci(reason="Int16 type only be used in test and inference.")
class TestLookupTableOpWithTensorIdsInt16(OpTest):
    def setUp(self):
        self.op_type = "lookup_table"
364 365 366 367 368 369
        table = np.random.randint(low=-128, high=127, size=(17, 31)).astype(
            "int16"
        )
        ids = np.random.randint(low=0, high=17, size=(2, 4, 5, 1)).astype(
            "int64"
        )
370 371 372 373
        self.inputs = {'W': table, 'Ids': ids}
        self.outputs = {'Out': table[ids.flatten()].reshape((2, 4, 5, 31))}

    def test_check_output(self):
374
        self.check_output(check_cinn=True)
375 376 377 378 379 380 381 382 383


@skip_check_grad_ci(reason="Int16 type only be used in test and inference.")
class TestLookupTableOpWithPaddingInt16(TestLookupTableOpInt16):
    def test_check_output(self):
        ids = np.squeeze(self.inputs['Ids'])
        padding_idx = np.random.choice(ids, 1)[0]
        self.outputs['Out'][ids == padding_idx] = np.zeros(31)
        self.attrs = {'padding_idx': int(padding_idx)}
384
        self.check_output(check_cinn=True)
385 386 387 388


@skip_check_grad_ci(reason="Int16 type only be used in test and inference.")
class TestLookupTableOpWithTensorIdsAndPaddingInt16(
389 390
    TestLookupTableOpWithTensorIdsInt16
):
391 392 393 394 395
    def test_check_output(self):
        ids = self.inputs['Ids']
        flatten_idx = ids.flatten()
        padding_idx = np.random.choice(flatten_idx, 1)[0]
        self.outputs['Out'][np.squeeze(ids == padding_idx)] = np.zeros(31)
396
        self.attrs = {'padding_idx': padding_idx}
397
        self.check_output(check_cinn=True)
398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452


class TestLookupTableWIsSelectedRowsInt16(unittest.TestCase):
    def prepare_ids(self, scope, place):
        ids_tensor = scope.var('Ids').get_tensor()
        ids_array = np.array([[0], [4], [3], [5]]).astype("int64")
        ids_tensor.set(ids_array, place)
        return ids_array

    def prepare_w(self, scope, place):
        rows = [0, 1, 2, 3, 4, 5, 6]
        row_numel = 12

        w_selected_rows = scope.var('W').get_selected_rows()
        w_selected_rows.set_height(len(rows))
        w_selected_rows.set_rows(rows)
        w_array = np.ones((len(rows), row_numel)).astype("int16")
        for i in range(len(rows)):
            w_array[i] *= i
        w_tensor = w_selected_rows.get_tensor()
        w_tensor.set(w_array, place)

    def create_out_tensor(self, scope, place):
        return scope.var('Out').get_tensor()

    def check_result(self, ids_array, result_array):
        for idx, row in enumerate(ids_array):
            assert (row[0] == result_array[idx]).all()

    def check_with_place(self, place):
        scope = core.Scope()

        ids_array = self.prepare_ids(scope, place)

        self.prepare_w(scope, place)

        out_tensor = self.create_out_tensor(scope, place)

        # create and run lookup_table operator
        lookup_table = Operator("lookup_table", W='W', Ids='Ids', Out='Out')
        lookup_table.run(scope, place)

        # get result from Out
        result_array = np.array(out_tensor)

        self.check_result(ids_array, result_array)

    def test_w_is_selected_rows(self):
        places = [core.CPUPlace()]
        # currently only support CPU
        for place in places:
            self.check_with_place(place)


class TestLookupTableWithTensorIdsWIsSelectedRowsInt16(
453 454
    TestLookupTableWIsSelectedRowsInt16
):
455 456
    def prepare_ids(self, scope, place):
        ids_tensor = scope.var('Ids').get_tensor()
457 458 459
        ids_array = np.random.randint(low=0, high=6, size=(2, 4, 3, 1)).astype(
            "int64"
        )
460 461 462 463 464 465 466 467
        ids_tensor.set(ids_array, place)
        return ids_array

    def check_result(self, ids_array, result_array):
        for idx, row in np.ndenumerate(ids_array):
            assert (row == result_array[idx]).all()


468 469 470
class TestOutDtype(unittest.TestCase):
    def test_dtype(self):
        api_fn = F.embedding
471 472 473 474 475 476
        check_out_dtype(
            api_fn,
            in_specs=[([10, 16], 'int64'), ([100, 64],)],
            expect_dtypes=['float32', 'float64'],
            target_index=1,
        )
477 478


Q
qijun 已提交
479
if __name__ == "__main__":
480
    unittest.main()