test_imperative_se_resnext.py 17.0 KB
Newer Older
Y
Yan Xu 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58
# 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 contextlib
import unittest
import numpy as np
import six

import paddle
import paddle.fluid as fluid
from paddle.fluid import core
from paddle.fluid.layer_helper import LayerHelper
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, FC
from paddle.fluid.dygraph.base import to_variable
from test_imperative_base import new_program_scope

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],
        "steps": [0.1, 0.01, 0.001, 0.0001]
    },
    "batch_size": batch_size,
    "lr": 0.1,
    "total_images": 6149,
}


def optimizer_setting(params):
    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.
        #batch_size = ls["batch_size"]
        #step = int(total_images / batch_size + 1)

        #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)]
Y
Yan Xu 已提交
59
        optimizer = fluid.optimizer.SGD(learning_rate=0.01)
Y
Yan Xu 已提交
60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111

    return optimizer


class ConvBNLayer(fluid.dygraph.Layer):
    def __init__(self,
                 name_scope,
                 num_channels,
                 num_filters,
                 filter_size,
                 stride=1,
                 groups=1,
                 act=None):
        super(ConvBNLayer, self).__init__(name_scope)

        self._conv = Conv2D(
            self.full_name(),
            num_channels=num_channels,
            num_filters=num_filters,
            filter_size=filter_size,
            stride=stride,
            padding=(filter_size - 1) // 2,
            groups=groups,
            act=None,
            bias_attr=None)

        self._batch_norm = BatchNorm(self.full_name(), num_filters, act=act)

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

        return y


class SqueezeExcitation(fluid.dygraph.Layer):
    def __init__(self, name_scope, num_channels, reduction_ratio):

        super(SqueezeExcitation, self).__init__(name_scope)
        self._pool = Pool2D(
            self.full_name(), pool_size=0, pool_type='avg', global_pooling=True)
        self._squeeze = FC(
            self.full_name(),
            size=num_channels // reduction_ratio,
            param_attr=fluid.ParamAttr(
                initializer=fluid.initializer.Constant(value=0.05)),
            act='relu')
        self._excitation = FC(
            self.full_name(),
            size=num_channels,
            param_attr=fluid.ParamAttr(
                initializer=fluid.initializer.Constant(value=0.05)),
Y
Yan Xu 已提交
112
            act='sigmoid')
Y
Yan Xu 已提交
113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317

    def forward(self, input):
        y = self._pool(input)
        y = self._squeeze(y)
        y = self._excitation(y)
        y = fluid.layers.elementwise_mul(x=input, y=y, axis=0)
        return y


class BottleneckBlock(fluid.dygraph.Layer):
    def __init__(self,
                 name_scope,
                 num_channels,
                 num_filters,
                 stride,
                 cardinality,
                 reduction_ratio,
                 shortcut=True):
        super(BottleneckBlock, self).__init__(name_scope)

        self.conv0 = ConvBNLayer(
            self.full_name(),
            num_channels=num_channels,
            num_filters=num_filters,
            filter_size=1)
        self.conv1 = ConvBNLayer(
            self.full_name(),
            num_channels=num_filters,
            num_filters=num_filters,
            filter_size=3,
            stride=stride,
            groups=cardinality)
        self.conv2 = ConvBNLayer(
            self.full_name(),
            num_channels=num_filters,
            num_filters=num_filters * 4,
            filter_size=1,
            act='relu')

        self.scale = SqueezeExcitation(
            self.full_name(),
            num_channels=num_filters * 4,
            reduction_ratio=reduction_ratio)

        if not shortcut:
            self.short = ConvBNLayer(
                self.full_name(),
                num_channels=num_channels,
                num_filters=num_filters * 4,
                filter_size=1,
                stride=stride)

        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)

        y = fluid.layers.elementwise_add(x=short, y=scale)

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


class SeResNeXt(fluid.dygraph.Layer):
    def __init__(self, name_scope, layers=50, class_dim=102):
        super(SeResNeXt, self).__init__(name_scope)

        self.layers = layers
        supported_layers = [50, 101, 152]
        assert layers in supported_layers, \
            "supported layers are {} but input layer is {}".format(supported_layers, layers)

        if layers == 50:
            cardinality = 32
            reduction_ratio = 16
            depth = [3, 4, 6, 3]
            num_filters = [128, 256, 512, 1024]
            self.conv0 = ConvBNLayer(
                self.full_name(),
                num_channels=3,
                num_filters=64,
                filter_size=7,
                stride=2,
                act='relu')
            self.pool = Pool2D(
                self.full_name(),
                pool_size=3,
                pool_stride=2,
                pool_padding=1,
                pool_type='max')
        elif layers == 101:
            cardinality = 32
            reduction_ratio = 16
            depth = [3, 4, 23, 3]
            num_filters = [128, 256, 512, 1024]
            self.conv0 = ConvBNLayer(
                self.full_name(),
                num_channels=3,
                num_filters=3,
                filter_size=7,
                stride=2,
                act='relu')
            self.pool = Pool2D(
                self.full_name(),
                pool_size=3,
                pool_stride=2,
                pool_padding=1,
                pool_type='max')
        elif layers == 152:
            cardinality = 64
            reduction_ratio = 16
            depth = [3, 8, 36, 3]
            num_filters = [128, 256, 512, 1024]
            self.conv0 = ConvBNLayer(
                self.full_name(),
                num_channels=3,
                num_filters=3,
                filter_size=7,
                stride=2,
                act='relu')
            self.conv1 = ConvBNLayer(
                self.full_name(),
                num_channels=64,
                num_filters=3,
                filter_size=7,
                stride=2,
                act='relu')
            self.conv2 = ConvBNLayer(
                self.full_name(),
                num_channels=64,
                num_filters=3,
                filter_size=7,
                stride=2,
                act='relu')
            self.pool = Pool2D(
                self.full_name(),
                pool_size=3,
                pool_stride=2,
                pool_padding=1,
                pool_type='max')

        self.bottleneck_block_list = []
        num_channels = 64
        for block in range(len(depth)):
            shortcut = False
            for i in range(depth[block]):
                bottleneck_block = self.add_sublayer(
                    'bb_%d_%d' % (block, i),
                    BottleneckBlock(
                        self.full_name(),
                        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))
                num_channels = bottleneck_block._num_channels_out
                self.bottleneck_block_list.append(bottleneck_block)
                shortcut = True

        self.pool2d_avg = Pool2D(
            self.full_name(), pool_size=7, pool_type='avg', global_pooling=True)
        import math
        stdv = 1.0 / math.sqrt(2048 * 1.0)

        self.out = FC(self.full_name(),
                      size=class_dim,
                      act='softmax',
                      param_attr=fluid.param_attr.ParamAttr(
                          initializer=fluid.initializer.Uniform(-stdv, stdv)))

    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)
            y = self.conv1(inputs)
            y = self.conv2(inputs)
            y = self.pool(y)

        for bottleneck_block in self.bottleneck_block_list:
            y = bottleneck_block(y)
        y = self.pool2d_avg(y)
        y = fluid.layers.dropout(y, dropout_prob=0.2)
        y = self.out(y)
        return y


class TestImperativeResneXt(unittest.TestCase):
    def test_se_resnext_float32(self):
        seed = 90

        batch_size = train_parameters["batch_size"]
318
        batch_num = 1
Y
Yan Xu 已提交
319
        epoch_num = 1
Y
Yan Xu 已提交
320 321 322 323 324 325 326 327 328 329 330
        with fluid.dygraph.guard():
            fluid.default_startup_program().random_seed = seed
            fluid.default_main_program().random_seed = seed

            se_resnext = SeResNeXt("se_resnext")
            optimizer = optimizer_setting(train_parameters)
            np.random.seed(seed)
            import random
            random.seed = seed
            train_reader = paddle.batch(
                paddle.dataset.flowers.train(use_xmap=False),
Y
Yan Xu 已提交
331 332
                batch_size=batch_size,
                drop_last=True)
Y
Yan Xu 已提交
333 334 335

            dy_param_init_value = {}
            for param in se_resnext.parameters():
L
lujun 已提交
336
                dy_param_init_value[param.name] = param.numpy()
Y
Yan Xu 已提交
337 338 339 340 341 342 343 344 345 346 347 348 349 350 351
            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

                    dy_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)

                    img = to_variable(dy_x_data)
                    label = to_variable(y_data)
L
lujun 已提交
352
                    label.stop_gradient = True
Y
Yan Xu 已提交
353 354 355 356 357

                    out = se_resnext(img)
                    loss = fluid.layers.cross_entropy(input=out, label=label)
                    avg_loss = fluid.layers.mean(x=loss)

L
lujun 已提交
358
                    dy_out = avg_loss.numpy()
Y
Yan Xu 已提交
359 360 361 362

                    if batch_id == 0:
                        for param in se_resnext.parameters():
                            if param.name not in dy_param_init_value:
L
lujun 已提交
363 364
                                dy_param_init_value[param.name] = param.numpy()
                    avg_loss.backward()
Y
Yan Xu 已提交
365 366 367 368 369 370 371 372 373 374 375 376

                    #dy_grad_value = {}
                    #for param in se_resnext.parameters():
                    #    if param.trainable:
                    #        np_array = np.array(param._ivar._grad_ivar().value()
                    #                            .get_tensor())
                    #        dy_grad_value[param.name + core.grad_var_suffix()] = np_array

                    optimizer.minimize(avg_loss)
                    se_resnext.clear_gradients()

                    dy_param_value = {}
Y
Yan Xu 已提交
377
                    for param in se_resnext.parameters():
L
lujun 已提交
378
                        dy_param_value[param.name] = param.numpy()
Y
Yan Xu 已提交
379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394

        with new_program_scope():
            fluid.default_startup_program().random_seed = seed
            fluid.default_main_program().random_seed = seed

            exe = fluid.Executor(fluid.CPUPlace(
            ) if not core.is_compiled_with_cuda() else fluid.CUDAPlace(0))

            se_resnext = SeResNeXt("se_resnext")
            optimizer = optimizer_setting(train_parameters)

            np.random.seed(seed)
            import random
            random.seed = seed
            train_reader = paddle.batch(
                paddle.dataset.flowers.train(use_xmap=False),
Y
Yan Xu 已提交
395 396
                batch_size=batch_size,
                drop_last=True)
Y
Yan Xu 已提交
397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421

            img = fluid.layers.data(
                name='pixel', shape=[3, 224, 224], dtype='float32')
            label = fluid.layers.data(name='label', shape=[1], dtype='int64')
            out = se_resnext(img)
            loss = fluid.layers.cross_entropy(input=out, label=label)
            avg_loss = fluid.layers.mean(x=loss)
            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:
                    static_grad_name_list.append(param.name +
                                                 core.grad_var_suffix())

            out = exe.run(fluid.default_startup_program(),
                          fetch_list=static_param_name_list)

            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
422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457
            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

                    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])

                    fetch_list = [avg_loss.name]
                    fetch_list.extend(static_param_name_list)
                    fetch_list.extend(static_grad_name_list)
                    out = exe.run(
                        fluid.default_main_program(),
                        feed={"pixel": static_x_data,
                              "label": y_data},
                        fetch_list=fetch_list)

                    static_param_value = {}
                    static_grad_value = {}
                    static_out = out[0]
                    param_start_pos = 1
                    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]
                    for i in range(grad_start_pos,
                                   len(static_grad_name_list) + grad_start_pos):
                        static_grad_value[static_grad_name_list[
                            i - grad_start_pos]] = out[i]
Y
Yan Xu 已提交
458 459 460 461 462 463 464 465
        self.assertTrue(np.allclose(static_out, dy_out))

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

        for key, value in six.iteritems(static_param_init_value):
            self.assertTrue(np.allclose(value, dy_param_init_value[key]))
            self.assertTrue(np.isfinite(value.all()))
            self.assertFalse(np.isnan(value.any()))
Y
Yan Xu 已提交
466 467 468 469 470 471
        # FIXME(Yancey1989): np.array(_ivar.value().get_tensor()) leads to memory lake
        #self.assertEqual(len(dy_grad_value), len(static_grad_value))
        #for key, value in six.iteritems(static_grad_value):
        #    self.assertTrue(np.allclose(value, dy_grad_value[key]))
        #    self.assertTrue(np.isfinite(value.all()))
        #    self.assertFalse(np.isnan(value.any()))
Y
Yan Xu 已提交
472 473 474 475 476 477 478 479 480 481

        self.assertEqual(len(dy_param_value), len(static_param_value))
        for key, value in six.iteritems(static_param_value):
            self.assertTrue(np.allclose(value, dy_param_value[key]))
            self.assertTrue(np.isfinite(value.all()))
            self.assertFalse(np.isnan(value.any()))


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