test_imperative_se_resnext.py 19.0 KB
Newer Older
Y
Yan Xu 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
# Copyright (c) 2018 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 unittest
16

Y
Yan Xu 已提交
17
import numpy as np
18
from test_imperative_base import new_program_scope
Y
Yan Xu 已提交
19 20 21 22

import paddle
import paddle.fluid as fluid
from paddle.fluid import core
23
from paddle.fluid.dygraph.nn import BatchNorm
H
hong 已提交
24
from paddle.fluid.framework import _test_eager_guard
25
from paddle.fluid.layer_helper import LayerHelper
Y
Yan Xu 已提交
26

27 28 29
if fluid.is_compiled_with_cuda():
    fluid.set_flags({'FLAGS_cudnn_deterministic': True})

Y
Yan Xu 已提交
30 31 32 33 34 35 36 37 38
batch_size = 8
train_parameters = {
    "input_size": [3, 224, 224],
    "input_mean": [0.485, 0.456, 0.406],
    "input_std": [0.229, 0.224, 0.225],
    "learning_strategy": {
        "name": "piecewise_decay",
        "batch_size": batch_size,
        "epochs": [30, 60, 90],
39
        "steps": [0.1, 0.01, 0.001, 0.0001],
Y
Yan Xu 已提交
40 41 42 43 44 45 46
    },
    "batch_size": batch_size,
    "lr": 0.1,
    "total_images": 6149,
}


47
def optimizer_setting(params, parameter_list=None):
Y
Yan Xu 已提交
48 49 50 51 52 53 54
    ls = params["learning_strategy"]
    if ls["name"] == "piecewise_decay":
        if "total_images" not in params:
            total_images = 6149
        else:
            total_images = params["total_images"]
        # TODO(Yancey1989): using lr decay if it is ready.
55 56
        # batch_size = ls["batch_size"]
        # step = int(total_images / batch_size + 1)
Y
Yan Xu 已提交
57

58 59 60
        # bd = [step * e for e in ls["epochs"]]
        # base_lr = params["lr"]
        # lr = [base_lr * (0.1**i) for i in range(len(bd) + 1)]
J
Jiabin Yang 已提交
61
        if fluid._non_static_mode():
62 63 64
            optimizer = fluid.optimizer.SGD(
                learning_rate=0.01, parameter_list=parameter_list
            )
65 66
        else:
            optimizer = fluid.optimizer.SGD(learning_rate=0.01)
Y
Yan Xu 已提交
67 68 69 70 71

    return optimizer


class ConvBNLayer(fluid.dygraph.Layer):
72 73 74 75 76 77 78 79 80
    def __init__(
        self,
        num_channels,
        num_filters,
        filter_size,
        stride=1,
        groups=1,
        act=None,
    ):
81
        super().__init__()
Y
Yan Xu 已提交
82

83 84 85 86
        self._conv = paddle.nn.Conv2D(
            in_channels=num_channels,
            out_channels=num_filters,
            kernel_size=filter_size,
87 88 89 90 91
            stride=stride,
            padding=(filter_size - 1) // 2,
            groups=groups,
            bias_attr=None,
        )
Y
Yan Xu 已提交
92

93
        self._batch_norm = BatchNorm(num_filters, act=act)
Y
Yan Xu 已提交
94 95 96 97 98 99 100 101 102

    def forward(self, inputs):
        y = self._conv(inputs)
        y = self._batch_norm(y)

        return y


class SqueezeExcitation(fluid.dygraph.Layer):
103
    def __init__(self, num_channels, reduction_ratio):
Y
Yan Xu 已提交
104

105
        super().__init__()
106
        self._num_channels = num_channels
W
wangzhen38 已提交
107
        self._pool = paddle.nn.AdaptiveAvgPool2D(1)
108
        self._squeeze = paddle.nn.Linear(
109 110
            num_channels,
            num_channels // reduction_ratio,
111 112
            weight_attr=paddle.ParamAttr(
                initializer=paddle.nn.initializer.Constant(value=0.05)
113 114
            ),
        )
115 116
        self.act_1 = paddle.nn.ReLU()
        self._excitation = paddle.nn.Linear(
117 118
            num_channels // reduction_ratio,
            num_channels,
119 120
            weight_attr=paddle.ParamAttr(
                initializer=paddle.nn.initializer.Constant(value=0.05)
121 122
            ),
        )
Y
Yan Xu 已提交
123

124 125
        self.act_2 = paddle.nn.Softmax()

Y
Yan Xu 已提交
126 127
    def forward(self, input):
        y = self._pool(input)
128
        y = paddle.reshape(y, shape=[-1, self._num_channels])
Y
Yan Xu 已提交
129
        y = self._squeeze(y)
130
        y = self.act_1(y)
Y
Yan Xu 已提交
131
        y = self._excitation(y)
132
        y = self.act_2(y)
Y
Yan Xu 已提交
133 134 135 136 137
        y = fluid.layers.elementwise_mul(x=input, y=y, axis=0)
        return y


class BottleneckBlock(fluid.dygraph.Layer):
138 139 140 141 142 143 144 145 146
    def __init__(
        self,
        num_channels,
        num_filters,
        stride,
        cardinality,
        reduction_ratio,
        shortcut=True,
    ):
147
        super().__init__()
Y
Yan Xu 已提交
148

149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168
        self.conv0 = ConvBNLayer(
            num_channels=num_channels, num_filters=num_filters, filter_size=1
        )
        self.conv1 = ConvBNLayer(
            num_channels=num_filters,
            num_filters=num_filters,
            filter_size=3,
            stride=stride,
            groups=cardinality,
        )
        self.conv2 = ConvBNLayer(
            num_channels=num_filters,
            num_filters=num_filters * 4,
            filter_size=1,
            act='relu',
        )

        self.scale = SqueezeExcitation(
            num_channels=num_filters * 4, reduction_ratio=reduction_ratio
        )
Y
Yan Xu 已提交
169 170

        if not shortcut:
171 172 173 174 175 176
            self.short = ConvBNLayer(
                num_channels=num_channels,
                num_filters=num_filters * 4,
                filter_size=1,
                stride=stride,
            )
Y
Yan Xu 已提交
177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192

        self.shortcut = shortcut

        self._num_channels_out = num_filters * 4

    def forward(self, inputs):
        y = self.conv0(inputs)
        conv1 = self.conv1(y)
        conv2 = self.conv2(conv1)
        scale = self.scale(conv2)

        if self.shortcut:
            short = inputs
        else:
            short = self.short(inputs)

193
        y = paddle.add(x=short, y=scale)
Y
Yan Xu 已提交
194 195 196 197 198 199 200

        layer_helper = LayerHelper(self.full_name(), act='relu')
        y = layer_helper.append_activation(y)
        return y


class SeResNeXt(fluid.dygraph.Layer):
201
    def __init__(self, layers=50, class_dim=102):
202
        super().__init__()
Y
Yan Xu 已提交
203 204 205

        self.layers = layers
        supported_layers = [50, 101, 152]
206 207 208 209 210
        assert (
            layers in supported_layers
        ), "supported layers are {} but input layer is {}".format(
            supported_layers, layers
        )
Y
Yan Xu 已提交
211 212 213 214 215 216

        if layers == 50:
            cardinality = 32
            reduction_ratio = 16
            depth = [3, 4, 6, 3]
            num_filters = [128, 256, 512, 1024]
217 218 219 220 221 222 223
            self.conv0 = ConvBNLayer(
                num_channels=3,
                num_filters=64,
                filter_size=7,
                stride=2,
                act='relu',
            )
224
            self.pool = paddle.nn.MaxPool2D(kernel_size=3, stride=2, padding=1)
Y
Yan Xu 已提交
225 226 227 228 229
        elif layers == 101:
            cardinality = 32
            reduction_ratio = 16
            depth = [3, 4, 23, 3]
            num_filters = [128, 256, 512, 1024]
230 231 232 233 234 235 236
            self.conv0 = ConvBNLayer(
                num_channels=3,
                num_filters=64,
                filter_size=7,
                stride=2,
                act='relu',
            )
237
            self.pool = paddle.nn.MaxPool2D(kernel_size=3, stride=2, padding=1)
Y
Yan Xu 已提交
238 239 240 241 242
        elif layers == 152:
            cardinality = 64
            reduction_ratio = 16
            depth = [3, 8, 36, 3]
            num_filters = [128, 256, 512, 1024]
243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263
            self.conv0 = ConvBNLayer(
                num_channels=3,
                num_filters=64,
                filter_size=3,
                stride=2,
                act='relu',
            )
            self.conv1 = ConvBNLayer(
                num_channels=64,
                num_filters=64,
                filter_size=3,
                stride=2,
                act='relu',
            )
            self.conv2 = ConvBNLayer(
                num_channels=64,
                num_filters=128,
                filter_size=3,
                stride=1,
                act='relu',
            )
264
            self.pool = paddle.nn.MaxPool2D(kernel_size=3, stride=2, padding=1)
Y
Yan Xu 已提交
265 266 267

        self.bottleneck_block_list = []
        num_channels = 64
268 269
        if layers == 152:
            num_channels = 128
Y
Yan Xu 已提交
270 271 272 273 274
        for block in range(len(depth)):
            shortcut = False
            for i in range(depth[block]):
                bottleneck_block = self.add_sublayer(
                    'bb_%d_%d' % (block, i),
275 276 277 278 279 280 281 282 283
                    BottleneckBlock(
                        num_channels=num_channels,
                        num_filters=num_filters[block],
                        stride=2 if i == 0 and block != 0 else 1,
                        cardinality=cardinality,
                        reduction_ratio=reduction_ratio,
                        shortcut=shortcut,
                    ),
                )
Y
Yan Xu 已提交
284 285 286
                num_channels = bottleneck_block._num_channels_out
                self.bottleneck_block_list.append(bottleneck_block)
                shortcut = True
W
wangzhen38 已提交
287
        self.pool2d_avg = paddle.nn.AdaptiveAvgPool2D(1)
Y
Yan Xu 已提交
288
        import math
289

Y
Yan Xu 已提交
290 291
        stdv = 1.0 / math.sqrt(2048 * 1.0)

292 293
        self.pool2d_avg_output = num_filters[-1] * 4 * 1 * 1

294
        self.out = paddle.nn.Linear(
295 296
            self.pool2d_avg_output,
            class_dim,
297 298
            weight_attr=paddle.ParamAttr(
                initializer=paddle.nn.initializer.Uniform(-stdv, stdv)
299 300
            ),
        )
301
        self.out_act = paddle.nn.Softmax()
Y
Yan Xu 已提交
302 303 304 305 306 307 308

    def forward(self, inputs):
        if self.layers == 50 or self.layers == 101:
            y = self.conv0(inputs)
            y = self.pool(y)
        elif self.layers == 152:
            y = self.conv0(inputs)
309 310
            y = self.conv1(y)
            y = self.conv2(y)
Y
Yan Xu 已提交
311 312 313 314 315
            y = self.pool(y)

        for bottleneck_block in self.bottleneck_block_list:
            y = bottleneck_block(y)
        y = self.pool2d_avg(y)
316
        y = paddle.reshape(y, shape=[-1, self.pool2d_avg_output])
Y
Yan Xu 已提交
317
        y = self.out(y)
318
        return self.out_act(y)
Y
Yan Xu 已提交
319 320 321


class TestImperativeResneXt(unittest.TestCase):
322 323 324 325 326 327 328 329 330
    def reader_decorator(self, reader):
        def _reader_imple():
            for item in reader():
                doc = np.array(item[0]).reshape(3, 224, 224)
                label = np.array(item[1]).astype('int64').reshape(1)
                yield doc, label

        return _reader_imple

Y
Yan Xu 已提交
331 332 333 334
    def test_se_resnext_float32(self):
        seed = 90

        batch_size = train_parameters["batch_size"]
335
        batch_num = 1
Y
Yan Xu 已提交
336
        epoch_num = 1
H
hong 已提交
337 338

        def run_dygraph():
C
cnn 已提交
339
            paddle.seed(seed)
L
Leo Chen 已提交
340
            paddle.framework.random._manual_program_seed(seed)
Y
Yan Xu 已提交
341

342 343
            se_resnext = SeResNeXt()
            optimizer = optimizer_setting(
344 345
                train_parameters, parameter_list=se_resnext.parameters()
            )
Y
Yan Xu 已提交
346
            np.random.seed(seed)
347 348 349

            batch_py_reader = fluid.io.PyReader(capacity=1)
            batch_py_reader.decorate_sample_list_generator(
350 351 352 353 354 355 356 357 358
                paddle.batch(
                    self.reader_decorator(
                        paddle.dataset.flowers.train(use_xmap=False)
                    ),
                    batch_size=batch_size,
                    drop_last=True,
                ),
                places=fluid.CPUPlace(),
            )
Y
Yan Xu 已提交
359 360 361

            dy_param_init_value = {}
            for param in se_resnext.parameters():
L
lujun 已提交
362
                dy_param_init_value[param.name] = param.numpy()
Y
Yan Xu 已提交
363
            for epoch_id in range(epoch_num):
364
                for batch_id, data in enumerate(batch_py_reader()):
Y
Yan Xu 已提交
365 366 367 368

                    if batch_id >= batch_num and batch_num != -1:
                        break

369 370 371
                    img = data[0]
                    label = data[1]
                    label.stop_gradient = True
L
lujun 已提交
372
                    label.stop_gradient = True
Y
Yan Xu 已提交
373 374

                    out = se_resnext(img)
375
                    softmax_out = paddle.nn.functional.softmax(out)
376 377 378 379 380
                    loss = paddle.nn.functional.cross_entropy(
                        input=softmax_out,
                        label=label,
                        reduction='none',
                        use_softmax=False,
381
                    )
382
                    avg_loss = paddle.mean(x=loss)
Y
Yan Xu 已提交
383

L
lujun 已提交
384
                    dy_out = avg_loss.numpy()
Y
Yan Xu 已提交
385 386 387 388

                    if batch_id == 0:
                        for param in se_resnext.parameters():
                            if param.name not in dy_param_init_value:
L
lujun 已提交
389 390
                                dy_param_init_value[param.name] = param.numpy()
                    avg_loss.backward()
Y
Yan Xu 已提交
391

392 393 394
                    dy_grad_value = {}
                    for param in se_resnext.parameters():
                        if param.trainable:
395
                            np_array = np.array(
396 397 398 399 400
                                param._grad_ivar().value().get_tensor()
                            )
                            dy_grad_value[
                                param.name + core.grad_var_suffix()
                            ] = np_array
Y
Yan Xu 已提交
401 402 403 404 405

                    optimizer.minimize(avg_loss)
                    se_resnext.clear_gradients()

                    dy_param_value = {}
Y
Yan Xu 已提交
406
                    for param in se_resnext.parameters():
L
lujun 已提交
407
                        dy_param_value[param.name] = param.numpy()
Y
Yan Xu 已提交
408

409 410 411 412 413 414
                    return (
                        dy_out,
                        dy_param_init_value,
                        dy_param_value,
                        dy_grad_value,
                    )
H
hong 已提交
415 416

        with fluid.dygraph.guard():
417 418 419 420 421 422
            (
                dy_out,
                dy_param_init_value,
                dy_param_value,
                dy_grad_value,
            ) = run_dygraph()
H
hong 已提交
423 424 425

        with fluid.dygraph.guard():
            with _test_eager_guard():
426 427 428 429 430 431
                (
                    eager_out,
                    eager_param_init_value,
                    eager_param_value,
                    eager_grad_value,
                ) = run_dygraph()
H
hong 已提交
432

Y
Yan Xu 已提交
433
        with new_program_scope():
C
cnn 已提交
434
            paddle.seed(seed)
L
Leo Chen 已提交
435
            paddle.framework.random._manual_program_seed(seed)
Y
Yan Xu 已提交
436

437 438 439 440 441
            exe = fluid.Executor(
                fluid.CPUPlace()
                if not core.is_compiled_with_cuda()
                else fluid.CUDAPlace(0)
            )
Y
Yan Xu 已提交
442

443
            se_resnext = SeResNeXt()
Y
Yan Xu 已提交
444 445 446 447 448
            optimizer = optimizer_setting(train_parameters)

            np.random.seed(seed)
            train_reader = paddle.batch(
                paddle.dataset.flowers.train(use_xmap=False),
Y
Yan Xu 已提交
449
                batch_size=batch_size,
450 451
                drop_last=True,
            )
Y
Yan Xu 已提交
452

453 454 455
            img = fluid.layers.data(
                name='pixel', shape=[3, 224, 224], dtype='float32'
            )
Y
Yan Xu 已提交
456 457
            label = fluid.layers.data(name='label', shape=[1], dtype='int64')
            out = se_resnext(img)
458
            softmax_out = paddle.nn.function.softmax(out)
459 460 461 462 463 464
            loss = paddle.nn.functional.cross_entropy(
                input=softmax_out,
                label=label,
                reduction='none',
                use_softmax=False,
            )
465
            avg_loss = paddle.mean(x=loss)
Y
Yan Xu 已提交
466 467 468 469 470 471 472 473 474 475
            optimizer.minimize(avg_loss)

            # initialize params and fetch them
            static_param_init_value = {}
            static_param_name_list = []
            static_grad_name_list = []
            for param in se_resnext.parameters():
                static_param_name_list.append(param.name)
            for param in se_resnext.parameters():
                if param.trainable:
476 477 478
                    static_grad_name_list.append(
                        param.name + core.grad_var_suffix()
                    )
Y
Yan Xu 已提交
479

480 481 482 483
            out = exe.run(
                fluid.default_startup_program(),
                fetch_list=static_param_name_list,
            )
Y
Yan Xu 已提交
484 485 486

            for i in range(len(static_param_name_list)):
                static_param_init_value[static_param_name_list[i]] = out[i]
Y
Yan Xu 已提交
487 488 489 490 491
            for epoch_id in range(epoch_num):
                for batch_id, data in enumerate(train_reader()):
                    if batch_id >= batch_num and batch_num != -1:
                        break

492 493 494 495 496 497 498 499
                    static_x_data = np.array(
                        [x[0].reshape(3, 224, 224) for x in data]
                    ).astype('float32')
                    y_data = (
                        np.array([x[1] for x in data])
                        .astype('int64')
                        .reshape([batch_size, 1])
                    )
Y
Yan Xu 已提交
500 501 502 503

                    fetch_list = [avg_loss.name]
                    fetch_list.extend(static_param_name_list)
                    fetch_list.extend(static_grad_name_list)
504 505 506 507 508
                    out = exe.run(
                        fluid.default_main_program(),
                        feed={"pixel": static_x_data, "label": y_data},
                        fetch_list=fetch_list,
                    )
Y
Yan Xu 已提交
509 510 511 512 513

                    static_param_value = {}
                    static_grad_value = {}
                    static_out = out[0]
                    param_start_pos = 1
514 515 516 517 518 519 520 521 522 523
                    grad_start_pos = (
                        len(static_param_name_list) + param_start_pos
                    )
                    for i in range(
                        param_start_pos,
                        len(static_param_name_list) + param_start_pos,
                    ):
                        static_param_value[
                            static_param_name_list[i - param_start_pos]
                        ] = out[i]
Y
Yan Xu 已提交
524
                    for i in range(
525 526 527 528 529 530
                        grad_start_pos,
                        len(static_grad_name_list) + grad_start_pos,
                    ):
                        static_grad_value[
                            static_grad_name_list[i - grad_start_pos]
                        ] = out[i]
531

532
        np.testing.assert_allclose(static_out, dy_out, rtol=1e-05)
Y
Yan Xu 已提交
533 534 535

        self.assertEqual(len(dy_param_init_value), len(static_param_init_value))

536
        for key, value in static_param_init_value.items():
537 538 539
            np.testing.assert_allclose(
                value, dy_param_init_value[key], rtol=1e-05
            )
Y
Yan Xu 已提交
540 541
            self.assertTrue(np.isfinite(value.all()))
            self.assertFalse(np.isnan(value.any()))
542 543 544

        self.assertEqual(len(dy_grad_value), len(static_grad_value))

545
        for key, value in static_grad_value.items():
546
            np.testing.assert_allclose(value, dy_grad_value[key], rtol=1e-05)
547 548
            self.assertTrue(np.isfinite(value.all()))
            self.assertFalse(np.isnan(value.any()))
Y
Yan Xu 已提交
549 550

        self.assertEqual(len(dy_param_value), len(static_param_value))
551
        for key, value in static_param_value.items():
552
            np.testing.assert_allclose(value, dy_param_value[key], rtol=1e-05)
Y
Yan Xu 已提交
553 554 555
            self.assertTrue(np.isfinite(value.all()))
            self.assertFalse(np.isnan(value.any()))

H
hong 已提交
556
        # check eager
557
        np.testing.assert_allclose(static_out, eager_out, rtol=1e-05)
H
hong 已提交
558

559 560 561
        self.assertEqual(
            len(eager_param_init_value), len(static_param_init_value)
        )
H
hong 已提交
562

563
        for key, value in static_param_init_value.items():
564 565 566
            np.testing.assert_allclose(
                value, eager_param_init_value[key], rtol=1e-05
            )
H
hong 已提交
567 568 569

        self.assertEqual(len(eager_grad_value), len(static_grad_value))

570
        for key, value in static_grad_value.items():
571
            np.testing.assert_allclose(value, eager_grad_value[key], rtol=1e-05)
H
hong 已提交
572 573

        self.assertEqual(len(eager_param_value), len(static_param_value))
574
        for key, value in static_param_value.items():
575 576 577
            np.testing.assert_allclose(
                value, eager_param_value[key], rtol=1e-05
            )
H
hong 已提交
578

Y
Yan Xu 已提交
579 580

if __name__ == '__main__':
581
    paddle.enable_static()
Y
Yan Xu 已提交
582
    unittest.main()