test_rnn_nets.py 12.5 KB
Newer Older
F
Feiyu Chan 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# 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
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# 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.

import paddle
paddle.set_default_dtype("float64")
from paddle.fluid.layers import sequence_mask

import numpy as np
import unittest

from convert import convert_params_for_net
from rnn_numpy import SimpleRNN, LSTM, GRU

25 26
bidirectional_list = ["bidirectional", "bidirect"]

F
Feiyu Chan 已提交
27 28 29 30 31 32

class TestSimpleRNN(unittest.TestCase):
    def __init__(self, time_major=True, direction="forward", place="cpu"):
        super(TestSimpleRNN, self).__init__("runTest")
        self.time_major = time_major
        self.direction = direction
33
        self.num_directions = 2 if direction in bidirectional_list else 1
34
        self.place = place
F
Feiyu Chan 已提交
35 36

    def setUp(self):
37 38 39 40
        # Since `set_device` is global, set `set_device` in `setUp` rather than
        # `__init__` to avoid using an error device set by another test case.
        place = paddle.set_device(self.place)
        paddle.disable_static(place)
F
Feiyu Chan 已提交
41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59
        rnn1 = SimpleRNN(
            16, 32, 2, time_major=self.time_major, direction=self.direction)
        rnn2 = paddle.nn.SimpleRNN(
            16, 32, 2, time_major=self.time_major, direction=self.direction)
        convert_params_for_net(rnn1, rnn2)

        self.rnn1 = rnn1
        self.rnn2 = rnn2

    def test_with_initial_state(self):
        rnn1 = self.rnn1
        rnn2 = self.rnn2

        x = np.random.randn(12, 4, 16)
        if not self.time_major:
            x = np.transpose(x, [1, 0, 2])
        prev_h = np.random.randn(2 * self.num_directions, 4, 32)

        y1, h1 = rnn1(x, prev_h)
60
        y2, h2 = rnn2(paddle.to_tensor(x), paddle.to_tensor(prev_h))
F
Feiyu Chan 已提交
61 62 63 64 65 66 67 68 69 70 71 72
        np.testing.assert_allclose(y1, y2.numpy(), atol=1e-8, rtol=1e-5)
        np.testing.assert_allclose(h1, h2.numpy(), atol=1e-8, rtol=1e-5)

    def test_with_zero_state(self):
        rnn1 = self.rnn1
        rnn2 = self.rnn2

        x = np.random.randn(12, 4, 16)
        if not self.time_major:
            x = np.transpose(x, [1, 0, 2])

        y1, h1 = rnn1(x)
73
        y2, h2 = rnn2(paddle.to_tensor(x))
F
Feiyu Chan 已提交
74 75 76 77 78 79 80 81 82 83 84 85 86 87
        np.testing.assert_allclose(y1, y2.numpy(), atol=1e-8, rtol=1e-5)
        np.testing.assert_allclose(h1, h2.numpy(), atol=1e-8, rtol=1e-5)

    def test_with_input_lengths(self):
        rnn1 = self.rnn1
        rnn2 = self.rnn2

        x = np.random.randn(12, 4, 16)
        if not self.time_major:
            x = np.transpose(x, [1, 0, 2])
        sequence_length = np.array([12, 10, 9, 8], dtype=np.int64)

        y1, h1 = rnn1(x, sequence_length=sequence_length)

88
        seq_len = paddle.to_tensor(sequence_length)
F
Feiyu Chan 已提交
89 90 91
        mask = sequence_mask(seq_len, dtype=paddle.get_default_dtype())
        if self.time_major:
            mask = paddle.transpose(mask, [1, 0])
92
        y2, h2 = rnn2(paddle.to_tensor(x), sequence_length=seq_len)
93 94
        mask = paddle.unsqueeze(mask, -1)
        y2 = paddle.multiply(y2, mask)
F
Feiyu Chan 已提交
95 96 97 98

        np.testing.assert_allclose(y1, y2.numpy(), atol=1e-8, rtol=1e-5)
        np.testing.assert_allclose(h1, h2.numpy(), atol=1e-8, rtol=1e-5)

G
Guo Sheng 已提交
99 100 101
    def test_predict(self):
        predict_test_util(self.place, "SimpleRNN")

F
Feiyu Chan 已提交
102 103 104 105
    def runTest(self):
        self.test_with_initial_state()
        self.test_with_zero_state()
        self.test_with_input_lengths()
G
Guo Sheng 已提交
106
        self.test_predict()
F
Feiyu Chan 已提交
107 108 109 110 111 112 113


class TestGRU(unittest.TestCase):
    def __init__(self, time_major=True, direction="forward", place="cpu"):
        super(TestGRU, self).__init__("runTest")
        self.time_major = time_major
        self.direction = direction
114
        self.num_directions = 2 if direction in bidirectional_list else 1
115
        self.place = place
F
Feiyu Chan 已提交
116 117

    def setUp(self):
118 119 120 121
        # Since `set_device` is global, set `set_device` in `setUp` rather than
        # `__init__` to avoid using an error device set by another test case.
        place = paddle.set_device(self.place)
        paddle.disable_static(place)
F
Feiyu Chan 已提交
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
        rnn1 = GRU(16,
                   32,
                   2,
                   time_major=self.time_major,
                   direction=self.direction)
        rnn2 = paddle.nn.GRU(16,
                             32,
                             2,
                             time_major=self.time_major,
                             direction=self.direction)
        convert_params_for_net(rnn1, rnn2)

        self.rnn1 = rnn1
        self.rnn2 = rnn2

    def test_with_initial_state(self):
        rnn1 = self.rnn1
        rnn2 = self.rnn2

        x = np.random.randn(12, 4, 16)
        if not self.time_major:
            x = np.transpose(x, [1, 0, 2])
        prev_h = np.random.randn(2 * self.num_directions, 4, 32)

        y1, h1 = rnn1(x, prev_h)
147
        y2, h2 = rnn2(paddle.to_tensor(x), paddle.to_tensor(prev_h))
F
Feiyu Chan 已提交
148 149 150 151 152 153 154 155 156 157 158 159
        np.testing.assert_allclose(y1, y2.numpy(), atol=1e-8, rtol=1e-5)
        np.testing.assert_allclose(h1, h2.numpy(), atol=1e-8, rtol=1e-5)

    def test_with_zero_state(self):
        rnn1 = self.rnn1
        rnn2 = self.rnn2

        x = np.random.randn(12, 4, 16)
        if not self.time_major:
            x = np.transpose(x, [1, 0, 2])

        y1, h1 = rnn1(x)
160
        y2, h2 = rnn2(paddle.to_tensor(x))
F
Feiyu Chan 已提交
161 162 163 164 165 166 167 168 169 170 171 172 173 174
        np.testing.assert_allclose(y1, y2.numpy(), atol=1e-8, rtol=1e-5)
        np.testing.assert_allclose(h1, h2.numpy(), atol=1e-8, rtol=1e-5)

    def test_with_input_lengths(self):
        rnn1 = self.rnn1
        rnn2 = self.rnn2

        x = np.random.randn(12, 4, 16)
        if not self.time_major:
            x = np.transpose(x, [1, 0, 2])
        sequence_length = np.array([12, 10, 9, 8], dtype=np.int64)

        y1, h1 = rnn1(x, sequence_length=sequence_length)

175
        seq_len = paddle.to_tensor(sequence_length)
F
Feiyu Chan 已提交
176 177 178
        mask = sequence_mask(seq_len, dtype=paddle.get_default_dtype())
        if self.time_major:
            mask = paddle.transpose(mask, [1, 0])
179
        y2, h2 = rnn2(paddle.to_tensor(x), sequence_length=seq_len)
180 181
        mask = paddle.unsqueeze(mask, -1)
        y2 = paddle.multiply(y2, mask)
F
Feiyu Chan 已提交
182 183 184 185

        np.testing.assert_allclose(y1, y2.numpy(), atol=1e-8, rtol=1e-5)
        np.testing.assert_allclose(h1, h2.numpy(), atol=1e-8, rtol=1e-5)

G
Guo Sheng 已提交
186 187 188
    def test_predict(self):
        predict_test_util(self.place, "GRU")

F
Feiyu Chan 已提交
189 190 191 192
    def runTest(self):
        self.test_with_initial_state()
        self.test_with_zero_state()
        self.test_with_input_lengths()
G
Guo Sheng 已提交
193
        self.test_predict()
F
Feiyu Chan 已提交
194 195 196 197 198 199 200


class TestLSTM(unittest.TestCase):
    def __init__(self, time_major=True, direction="forward", place="cpu"):
        super(TestLSTM, self).__init__("runTest")
        self.time_major = time_major
        self.direction = direction
201
        self.num_directions = 2 if direction in bidirectional_list else 1
202
        self.place = place
F
Feiyu Chan 已提交
203 204

    def setUp(self):
205 206 207 208
        # Since `set_device` is global, set `set_device` in `setUp` rather than
        # `__init__` to avoid using an error device set by another test case.
        place = paddle.set_device(self.place)
        paddle.disable_static(place)
F
Feiyu Chan 已提交
209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229
        rnn1 = LSTM(
            16, 32, 2, time_major=self.time_major, direction=self.direction)
        rnn2 = paddle.nn.LSTM(
            16, 32, 2, time_major=self.time_major, direction=self.direction)
        convert_params_for_net(rnn1, rnn2)

        self.rnn1 = rnn1
        self.rnn2 = rnn2

    def test_with_initial_state(self):
        rnn1 = self.rnn1
        rnn2 = self.rnn2

        x = np.random.randn(12, 4, 16)
        if not self.time_major:
            x = np.transpose(x, [1, 0, 2])
        prev_h = np.random.randn(2 * self.num_directions, 4, 32)
        prev_c = np.random.randn(2 * self.num_directions, 4, 32)

        y1, (h1, c1) = rnn1(x, (prev_h, prev_c))
        y2, (h2, c2) = rnn2(
230 231
            paddle.to_tensor(x),
            (paddle.to_tensor(prev_h), paddle.to_tensor(prev_c)))
F
Feiyu Chan 已提交
232 233 234 235 236 237 238 239 240 241 242 243 244
        np.testing.assert_allclose(y1, y2.numpy(), atol=1e-8, rtol=1e-5)
        np.testing.assert_allclose(h1, h2.numpy(), atol=1e-8, rtol=1e-5)
        np.testing.assert_allclose(c1, c2.numpy(), atol=1e-8, rtol=1e-5)

    def test_with_zero_state(self):
        rnn1 = self.rnn1
        rnn2 = self.rnn2

        x = np.random.randn(12, 4, 16)
        if not self.time_major:
            x = np.transpose(x, [1, 0, 2])

        y1, (h1, c1) = rnn1(x)
245
        y2, (h2, c2) = rnn2(paddle.to_tensor(x))
F
Feiyu Chan 已提交
246 247 248 249 250 251 252 253 254 255 256 257 258 259 260
        np.testing.assert_allclose(y1, y2.numpy(), atol=1e-8, rtol=1e-5)
        np.testing.assert_allclose(h1, h2.numpy(), atol=1e-8, rtol=1e-5)
        np.testing.assert_allclose(c1, c2.numpy(), atol=1e-8, rtol=1e-5)

    def test_with_input_lengths(self):
        rnn1 = self.rnn1
        rnn2 = self.rnn2

        x = np.random.randn(12, 4, 16)
        if not self.time_major:
            x = np.transpose(x, [1, 0, 2])
        sequence_length = np.array([12, 10, 9, 8], dtype=np.int64)

        y1, (h1, c1) = rnn1(x, sequence_length=sequence_length)

261
        seq_len = paddle.to_tensor(sequence_length)
F
Feiyu Chan 已提交
262 263 264
        mask = sequence_mask(seq_len, dtype=paddle.get_default_dtype())
        if self.time_major:
            mask = paddle.transpose(mask, [1, 0])
265
        y2, (h2, c2) = rnn2(paddle.to_tensor(x), sequence_length=seq_len)
266 267
        mask = paddle.unsqueeze(mask, -1)
        y2 = paddle.multiply(y2, mask)
F
Feiyu Chan 已提交
268 269 270 271 272

        np.testing.assert_allclose(y1, y2.numpy(), atol=1e-8, rtol=1e-5)
        np.testing.assert_allclose(h1, h2.numpy(), atol=1e-8, rtol=1e-5)
        np.testing.assert_allclose(c1, c2.numpy(), atol=1e-8, rtol=1e-5)

273
    def test_predict(self):
G
Guo Sheng 已提交
274
        predict_test_util(self.place, "LSTM")
275
        predict_test_util(self.place, "LSTM", False)
276

F
Feiyu Chan 已提交
277 278 279 280
    def runTest(self):
        self.test_with_initial_state()
        self.test_with_zero_state()
        self.test_with_input_lengths()
281
        self.test_predict()
F
Feiyu Chan 已提交
282 283


284
def predict_test_util(place, mode, stop_gradient=True):
G
Guo Sheng 已提交
285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301
    place = paddle.set_device(place)
    paddle.seed(123)
    np.random.seed(123)

    class Net(paddle.nn.Layer):
        def __init__(self):
            super(Net, self).__init__()
            self.rnn = getattr(paddle.nn, mode)(16,
                                                32,
                                                2,
                                                direction="bidirectional",
                                                dropout=0.1)

        def forward(self, input):
            return self.rnn(input)

    x = paddle.randn((4, 10, 16))
302
    x.stop_gradient = stop_gradient
G
Guo Sheng 已提交
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
    seq_len = paddle.to_tensor(np.array([10, 6, 8, 5]))
    mask = sequence_mask(seq_len, maxlen=10, dtype=x.dtype)
    mask = paddle.unsqueeze(mask, [2])
    rnn = Net()
    y, _ = rnn(x)
    y = y * mask
    loss = paddle.mean(y)
    loss.backward()
    optimizer = paddle.optimizer.Adam(
        learning_rate=0.1, parameters=rnn.parameters())
    optimizer.step()
    rnn.eval()
    y, _ = rnn(x)
    # `jit.to_static` would include a train_program, eval mode might cause
    # some errors currently, such as dropout grad op gets `is_test == True`.
    rnn.train()

    rnn = paddle.jit.to_static(
        rnn, [paddle.static.InputSpec(
            shape=[None, None, 16], dtype=x.dtype)])
    paddle.jit.save(rnn, "./inference/%s_infer" % mode)

    paddle.enable_static()

    new_scope = paddle.static.Scope()
    with paddle.static.scope_guard(new_scope):
        exe = paddle.static.Executor(place)
        [inference_program, feed_target_names,
         fetch_targets] = paddle.static.load_inference_model(
332
             "./inference/%s_infer" % mode, exe)
G
Guo Sheng 已提交
333 334 335 336 337 338 339 340
        results = exe.run(inference_program,
                          feed={feed_target_names[0]: x.numpy()},
                          fetch_list=fetch_targets)
        np.testing.assert_equal(
            y.numpy(), results[0])  # eval results equal predict results
    paddle.disable_static()


F
Feiyu Chan 已提交
341 342 343 344
def load_tests(loader, tests, pattern):
    suite = unittest.TestSuite()
    devices = ["cpu", "gpu"] if paddle.fluid.is_compiled_with_cuda() \
        else ["cpu"]
345
    for direction in ["forward", "bidirectional", "bidirect"]:
F
Feiyu Chan 已提交
346 347 348 349 350
        for time_major in [True, False]:
            for device in devices:
                for test_class in [TestSimpleRNN, TestLSTM, TestGRU]:
                    suite.addTest(test_class(time_major, direction, device))
    return suite
351

352

353 354
if __name__ == '__main__':
    unittest.main()