test_model.py 20.2 KB
Newer Older
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.

from __future__ import division
from __future__ import print_function

import unittest

import os
import numpy as np
import shutil
import tempfile

L
Leo Chen 已提交
25
import paddle
26
from paddle import fluid
27
from paddle import to_tensor
28
from paddle.nn import Conv2d, Pool2D, Linear, ReLU, Sequential, Softmax
29

30 31
from paddle import Model
from paddle.static import InputSpec
32
from paddle.nn.layer.loss import CrossEntropyLoss
33
from paddle.metric import Accuracy
34 35 36 37
from paddle.vision.datasets import MNIST
from paddle.vision.models import LeNet
from paddle.io import DistributedBatchSampler
from paddle.hapi.model import prepare_distributed_context
38 39
from paddle.fluid.dygraph.jit import declarative
from paddle.fluid.dygraph.dygraph_to_static.program_translator import ProgramTranslator
40 41


42
class LeNetDygraph(paddle.nn.Layer):
43
    def __init__(self, num_classes=10, classifier_activation=None):
44 45 46
        super(LeNetDygraph, self).__init__()
        self.num_classes = num_classes
        self.features = Sequential(
47
            Conv2d(
48
                1, 6, 3, stride=1, padding=1),
L
LielinJiang 已提交
49
            ReLU(),
50
            Pool2D(2, 'max', 2),
51
            Conv2d(
52
                6, 16, 5, stride=1, padding=0),
L
LielinJiang 已提交
53
            ReLU(),
54 55 56 57
            Pool2D(2, 'max', 2))

        if num_classes > 0:
            self.fc = Sequential(
58 59
                Linear(400, 120), Linear(120, 84), Linear(84, 10),
                Softmax())  #Todo: accept any activation
60 61 62 63 64 65 66 67 68 69

    def forward(self, inputs):
        x = self.features(inputs)

        if self.num_classes > 0:
            x = fluid.layers.flatten(x, 1)
            x = self.fc(x)
        return x


70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85
class LeNetDeclarative(fluid.dygraph.Layer):
    def __init__(self, num_classes=10, classifier_activation=None):
        super(LeNetDeclarative, self).__init__()
        self.num_classes = num_classes
        self.features = Sequential(
            Conv2d(
                1, 6, 3, stride=1, padding=1),
            ReLU(),
            Pool2D(2, 'max', 2),
            Conv2d(
                6, 16, 5, stride=1, padding=0),
            ReLU(),
            Pool2D(2, 'max', 2))

        if num_classes > 0:
            self.fc = Sequential(
86 87
                Linear(400, 120), Linear(120, 84), Linear(84, 10),
                Softmax())  #Todo: accept any activation
88 89 90 91 92 93 94 95 96 97 98

    @declarative
    def forward(self, inputs):
        x = self.features(inputs)

        if self.num_classes > 0:
            x = fluid.layers.flatten(x, 1)
            x = self.fc(x)
        return x


99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130
class MnistDataset(MNIST):
    def __init__(self, mode, return_label=True, sample_num=None):
        super(MnistDataset, self).__init__(mode=mode)
        self.return_label = return_label
        if sample_num:
            self.images = self.images[:sample_num]
            self.labels = self.labels[:sample_num]

    def __getitem__(self, idx):
        img, label = self.images[idx], self.labels[idx]
        img = np.reshape(img, [1, 28, 28])
        if self.return_label:
            return img, np.array(self.labels[idx]).astype('int64')
        return img,

    def __len__(self):
        return len(self.images)


def compute_acc(pred, label):
    pred = np.argmax(pred, -1)
    label = np.array(label)
    correct = pred[:, np.newaxis] == label
    return np.sum(correct) / correct.shape[0]


def dynamic_train(model, dataloader):
    optim = fluid.optimizer.Adam(
        learning_rate=0.001, parameter_list=model.parameters())
    model.train()
    for inputs, labels in dataloader:
        outputs = model(inputs)
131
        loss = CrossEntropyLoss(reduction="sum")(outputs, labels)
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
        avg_loss = fluid.layers.reduce_sum(loss)
        avg_loss.backward()
        optim.minimize(avg_loss)
        model.clear_gradients()


def dynamic_evaluate(model, dataloader):
    with fluid.dygraph.no_grad():
        model.eval()
        cnt = 0
        for inputs, labels in dataloader:
            outputs = model(inputs)

            cnt += (np.argmax(outputs.numpy(), -1)[:, np.newaxis] ==
                    labels.numpy()).astype('int').sum()

    return cnt / len(dataloader.dataset)


@unittest.skipIf(not fluid.is_compiled_with_cuda(),
                 'CPU testing is not supported')
class TestModel(unittest.TestCase):
    @classmethod
    def setUpClass(cls):
        if not fluid.is_compiled_with_cuda():
            self.skipTest('module not tested when ONLY_CPU compling')
158
        cls.device = paddle.set_device('gpu')
159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174
        fluid.enable_dygraph(cls.device)

        sp_num = 1280
        cls.train_dataset = MnistDataset(mode='train', sample_num=sp_num)
        cls.val_dataset = MnistDataset(mode='test', sample_num=sp_num)
        cls.test_dataset = MnistDataset(
            mode='test', return_label=False, sample_num=sp_num)

        cls.train_loader = fluid.io.DataLoader(
            cls.train_dataset, places=cls.device, batch_size=64)
        cls.val_loader = fluid.io.DataLoader(
            cls.val_dataset, places=cls.device, batch_size=64)
        cls.test_loader = fluid.io.DataLoader(
            cls.test_dataset, places=cls.device, batch_size=64)

        seed = 333
L
Leo Chen 已提交
175 176
        paddle.manual_seed(seed)
        paddle.framework.random._manual_program_seed(seed)
177 178 179 180 181 182 183

        dy_lenet = LeNetDygraph()
        cls.init_param = dy_lenet.state_dict()
        dynamic_train(dy_lenet, cls.train_loader)

        cls.acc1 = dynamic_evaluate(dy_lenet, cls.val_loader)

184 185
        cls.inputs = [InputSpec([-1, 1, 28, 28], 'float32', 'image')]
        cls.labels = [InputSpec([None, 1], 'int64', 'label')]
186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202

        cls.save_dir = tempfile.mkdtemp()
        cls.weight_path = os.path.join(cls.save_dir, 'lenet')
        fluid.dygraph.save_dygraph(dy_lenet.state_dict(), cls.weight_path)

        fluid.disable_dygraph()

    @classmethod
    def tearDownClass(cls):
        shutil.rmtree(cls.save_dir)

    def test_fit_dygraph(self):
        self.fit(True)

    def test_fit_static(self):
        self.fit(False)

203 204 205 206 207 208
    def test_fit_dynamic_with_rank(self):
        self.fit(True, 2, 0)

    def test_fit_static_with_rank(self):
        self.fit(False, 2, 0)

209 210 211 212 213 214 215 216 217 218 219 220 221 222 223
    def test_evaluate_dygraph(self):
        self.evaluate(True)

    def test_evaluate_static(self):
        self.evaluate(False)

    def test_predict_dygraph(self):
        self.predict(True)

    def test_predict_static(self):
        self.predict(False)

    def test_prepare_context(self):
        prepare_distributed_context()

224
    def fit(self, dynamic, num_replicas=None, rank=None):
225 226
        fluid.enable_dygraph(self.device) if dynamic else None
        seed = 333
L
Leo Chen 已提交
227 228
        paddle.manual_seed(seed)
        paddle.framework.random._manual_program_seed(seed)
229

230
        net = LeNet(classifier_activation=None)
231
        optim_new = fluid.optimizer.Adam(
232 233
            learning_rate=0.001, parameter_list=net.parameters())
        model = Model(net, inputs=self.inputs, labels=self.labels)
234 235
        model.prepare(
            optim_new,
236
            loss=CrossEntropyLoss(reduction="sum"),
237
            metrics=Accuracy())
238 239 240 241 242 243
        model.fit(self.train_dataset, batch_size=64, shuffle=False)

        result = model.evaluate(self.val_dataset, batch_size=64)
        np.testing.assert_allclose(result['acc'], self.acc1)

        train_sampler = DistributedBatchSampler(
244 245 246 247 248
            self.train_dataset,
            batch_size=64,
            shuffle=False,
            num_replicas=num_replicas,
            rank=rank)
249
        val_sampler = DistributedBatchSampler(
250 251 252 253 254
            self.val_dataset,
            batch_size=64,
            shuffle=False,
            num_replicas=num_replicas,
            rank=rank)
255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272

        train_loader = fluid.io.DataLoader(
            self.train_dataset,
            batch_sampler=train_sampler,
            places=self.device,
            return_list=True)

        val_loader = fluid.io.DataLoader(
            self.val_dataset,
            batch_sampler=val_sampler,
            places=self.device,
            return_list=True)

        model.fit(train_loader, val_loader)
        fluid.disable_dygraph() if dynamic else None

    def evaluate(self, dynamic):
        fluid.enable_dygraph(self.device) if dynamic else None
273 274
        model = Model(LeNet(), self.inputs, self.labels)
        model.prepare(metrics=Accuracy())
275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293
        model.load(self.weight_path)
        result = model.evaluate(self.val_dataset, batch_size=64)
        np.testing.assert_allclose(result['acc'], self.acc1)

        sampler = DistributedBatchSampler(
            self.val_dataset, batch_size=64, shuffle=False)

        val_loader = fluid.io.DataLoader(
            self.val_dataset,
            batch_sampler=sampler,
            places=self.device,
            return_list=True)

        model.evaluate(val_loader)

        fluid.disable_dygraph() if dynamic else None

    def predict(self, dynamic):
        fluid.enable_dygraph(self.device) if dynamic else None
294 295
        model = Model(LeNet(), self.inputs)
        model.prepare()
296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317
        model.load(self.weight_path)
        output = model.predict(
            self.test_dataset, batch_size=64, stack_outputs=True)
        np.testing.assert_equal(output[0].shape[0], len(self.test_dataset))

        acc = compute_acc(output[0], self.val_dataset.labels)
        np.testing.assert_allclose(acc, self.acc1)

        sampler = DistributedBatchSampler(
            self.test_dataset, batch_size=64, shuffle=False)

        test_loader = fluid.io.DataLoader(
            self.test_dataset,
            batch_sampler=sampler,
            places=self.device,
            return_list=True)

        model.evaluate(test_loader)

        fluid.disable_dygraph() if dynamic else None


318
class MyModel(paddle.nn.Layer):
319
    def __init__(self, classifier_activation='softmax'):
320
        super(MyModel, self).__init__()
321 322
        self._fc = Linear(20, 10)
        self._act = Softmax()  #Todo: accept any activation
323 324 325

    def forward(self, x):
        y = self._fc(x)
326
        y = self._act(y)
327 328 329 330 331
        return y


class TestModelFunction(unittest.TestCase):
    def set_seed(self, seed=1024):
L
Leo Chen 已提交
332 333
        paddle.manual_seed(seed)
        paddle.framework.random._manual_program_seed(seed)
334 335 336 337 338 339 340 341 342

    def test_train_batch(self, dynamic=True):
        dim = 20
        data = np.random.random(size=(4, dim)).astype(np.float32)
        label = np.random.randint(0, 10, size=(4, 1)).astype(np.int64)

        def get_expect():
            fluid.enable_dygraph(fluid.CPUPlace())
            self.set_seed()
343
            m = MyModel(classifier_activation=None)
344 345 346
            optim = fluid.optimizer.SGD(learning_rate=0.001,
                                        parameter_list=m.parameters())
            m.train()
347 348
            output = m(to_tensor(data))
            loss = CrossEntropyLoss(reduction='sum')(output, to_tensor(label))
349 350 351 352 353 354 355 356 357
            avg_loss = fluid.layers.reduce_sum(loss)
            avg_loss.backward()
            optim.minimize(avg_loss)
            m.clear_gradients()
            fluid.disable_dygraph()
            return avg_loss.numpy()

        ref = get_expect()
        for dynamic in [True, False]:
358
            device = paddle.set_device('cpu')
359 360 361
            fluid.enable_dygraph(device) if dynamic else None
            self.set_seed()

362
            net = MyModel(classifier_activation=None)
363
            optim2 = fluid.optimizer.SGD(learning_rate=0.001,
364
                                         parameter_list=net.parameters())
365

366 367
            inputs = [InputSpec([None, dim], 'float32', 'x')]
            labels = [InputSpec([None, 1], 'int64', 'label')]
368
            model = Model(net, inputs, labels)
369
            model.prepare(optim2, loss=CrossEntropyLoss(reduction="sum"))
370 371 372 373
            loss, = model.train_batch([data], [label])
            np.testing.assert_allclose(loss.flatten(), ref.flatten())
            fluid.disable_dygraph() if dynamic else None

374
    def test_test_batch(self):
375 376 377 378 379 380 381 382
        dim = 20
        data = np.random.random(size=(4, dim)).astype(np.float32)

        def get_expect():
            fluid.enable_dygraph(fluid.CPUPlace())
            self.set_seed()
            m = MyModel()
            m.eval()
383
            output = m(to_tensor(data))
384 385 386 387 388
            fluid.disable_dygraph()
            return output.numpy()

        ref = get_expect()
        for dynamic in [True, False]:
389
            device = paddle.set_device('cpu')
390 391
            fluid.enable_dygraph(device) if dynamic else None
            self.set_seed()
392
            net = MyModel()
393
            inputs = [InputSpec([None, dim], 'float32', 'x')]
394 395
            model = Model(net, inputs)
            model.prepare()
396 397
            out, = model.test_batch([data])

398
            np.testing.assert_allclose(out, ref, rtol=1e-6)
399 400 401 402 403
            fluid.disable_dygraph() if dynamic else None

    def test_save_load(self):
        path = tempfile.mkdtemp()
        for dynamic in [True, False]:
404
            device = paddle.set_device('cpu')
405
            fluid.enable_dygraph(device) if dynamic else None
406
            net = MyModel(classifier_activation=None)
407 408
            inputs = [InputSpec([None, 20], 'float32', 'x')]
            labels = [InputSpec([None, 1], 'int64', 'label')]
409
            optim = fluid.optimizer.SGD(learning_rate=0.001,
410 411
                                        parameter_list=net.parameters())
            model = Model(net, inputs, labels)
412
            model.prepare(
413
                optimizer=optim, loss=CrossEntropyLoss(reduction="sum"))
414 415 416 417 418
            model.save(path + '/test')
            model.load(path + '/test')
            shutil.rmtree(path)
            fluid.disable_dygraph() if dynamic else None

419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441
    def test_dynamic_load(self):
        mnist_data = MnistDataset(mode='train')
        for new_optimizer in [True, False]:
            path = tempfile.mkdtemp()
            paddle.disable_static()
            net = LeNet()
            inputs = [InputSpec([None, 1, 28, 28], 'float32', 'x')]
            labels = [InputSpec([None, 1], 'int64', 'label')]
            if new_optimizer:
                optim = paddle.optimizer.Adam(
                    learning_rate=0.001, parameters=net.parameters())
            else:
                optim = fluid.optimizer.Adam(
                    learning_rate=0.001, parameter_list=net.parameters())
            model = Model(net, inputs, labels)
            model.prepare(
                optimizer=optim, loss=CrossEntropyLoss(reduction="sum"))
            model.fit(mnist_data, batch_size=64, verbose=0)
            model.save(path + '/test')
            model.load(path + '/test')
            shutil.rmtree(path)
            paddle.enable_static()

442 443
    def test_dynamic_save_static_load(self):
        path = tempfile.mkdtemp()
444
        # dynamic saving
445
        device = paddle.set_device('cpu')
446
        fluid.enable_dygraph(device)
447
        model = Model(MyModel(classifier_activation=None))
448 449
        optim = fluid.optimizer.SGD(learning_rate=0.001,
                                    parameter_list=model.parameters())
450
        model.prepare(optimizer=optim, loss=CrossEntropyLoss(reduction="sum"))
451 452
        model.save(path + '/test')
        fluid.disable_dygraph()
453

454 455
        inputs = [InputSpec([None, 20], 'float32', 'x')]
        labels = [InputSpec([None, 1], 'int64', 'label')]
456
        model = Model(MyModel(classifier_activation=None), inputs, labels)
457 458
        optim = fluid.optimizer.SGD(learning_rate=0.001,
                                    parameter_list=model.parameters())
459
        model.prepare(optimizer=optim, loss=CrossEntropyLoss(reduction="sum"))
460 461 462 463 464 465
        model.load(path + '/test')
        shutil.rmtree(path)

    def test_static_save_dynamic_load(self):
        path = tempfile.mkdtemp()

466
        net = MyModel(classifier_activation=None)
467 468
        inputs = [InputSpec([None, 20], 'float32', 'x')]
        labels = [InputSpec([None, 1], 'int64', 'label')]
469
        optim = fluid.optimizer.SGD(learning_rate=0.001,
470 471
                                    parameter_list=net.parameters())
        model = Model(net, inputs, labels)
472
        model.prepare(optimizer=optim, loss=CrossEntropyLoss(reduction="sum"))
473 474
        model.save(path + '/test')

475
        device = paddle.set_device('cpu')
476 477
        fluid.enable_dygraph(device)  #if dynamic else None

478
        net = MyModel(classifier_activation=None)
479 480
        inputs = [InputSpec([None, 20], 'float32', 'x')]
        labels = [InputSpec([None, 1], 'int64', 'label')]
481
        optim = fluid.optimizer.SGD(learning_rate=0.001,
482 483
                                    parameter_list=net.parameters())
        model = Model(net, inputs, labels)
484
        model.prepare(optimizer=optim, loss=CrossEntropyLoss(reduction="sum"))
485 486 487 488 489 490
        model.load(path + '/test')
        shutil.rmtree(path)
        fluid.disable_dygraph()

    def test_parameters(self):
        for dynamic in [True, False]:
491
            device = paddle.set_device('cpu')
492
            fluid.enable_dygraph(device) if dynamic else None
493
            net = MyModel()
494
            inputs = [InputSpec([None, 20], 'float32', 'x')]
495 496
            model = Model(net, inputs)
            model.prepare()
497 498 499 500 501
            params = model.parameters()
            self.assertTrue(params[0].shape[0] == 20)
            self.assertTrue(params[0].shape[1] == 10)
            fluid.disable_dygraph() if dynamic else None

L
LielinJiang 已提交
502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521
    def test_summary(self):
        def _get_param_from_state_dict(state_dict):
            params = 0
            for k, v in state_dict.items():
                params += np.prod(v.numpy().shape)
            return params

        for dynamic in [True, False]:
            device = paddle.set_device('cpu')
            fluid.enable_dygraph(device) if dynamic else None
            net = MyModel()
            inputs = [InputSpec([None, 20], 'float32', 'x')]
            model = Model(net, inputs)
            model.prepare()
            params_info = model.summary()
            gt_params = _get_param_from_state_dict(net.state_dict())

            np.testing.assert_allclose(params_info['total_params'], gt_params)
            print(params_info)

522 523 524 525
            model.summary(input_size=(20))
            model.summary(input_size=[(20)])
            model.summary(input_size=(20), batch_size=2)

526
    def test_export_deploy_model(self):
527 528 529 530 531 532
        for dynamic in [True, False]:
            fluid.enable_dygraph() if dynamic else None
            # paddle.disable_static() if dynamic else None
            prog_translator = ProgramTranslator()
            prog_translator.enable(False) if not dynamic else None
            net = LeNetDeclarative()
533
            inputs = [InputSpec([None, 1, 28, 28], 'float32', 'x')]
534 535 536 537 538 539 540 541 542 543
            model = Model(net, inputs)
            model.prepare()
            save_dir = tempfile.mkdtemp()
            if not os.path.exists(save_dir):
                os.makedirs(save_dir)
            tensor_img = np.array(
                np.random.random((1, 1, 28, 28)), dtype=np.float32)
            ori_results = model.test_batch(tensor_img)
            model.save(save_dir, training=False)
            fluid.disable_dygraph() if dynamic else None
544

545 546 547 548 549 550 551 552 553 554 555 556 557 558
            place = fluid.CPUPlace() if not fluid.is_compiled_with_cuda(
            ) else fluid.CUDAPlace(0)
            new_scope = fluid.Scope()
            with fluid.scope_guard(new_scope):
                exe = fluid.Executor(place)
                [inference_program, feed_target_names, fetch_targets] = (
                    fluid.io.load_inference_model(
                        dirname=save_dir, executor=exe))
                results = exe.run(inference_program,
                                  feed={feed_target_names[0]: tensor_img},
                                  fetch_list=fetch_targets)
                np.testing.assert_allclose(
                    results, ori_results, rtol=1e-5, atol=1e-7)
                shutil.rmtree(save_dir)
559 560


561 562 563 564
class TestRaiseError(unittest.TestCase):
    def test_input_without_name(self):
        net = MyModel(classifier_activation=None)

565 566
        inputs = [InputSpec([None, 10], 'float32')]
        labels = [InputSpec([None, 1], 'int64', 'label')]
567 568 569 570
        with self.assertRaises(ValueError):
            model = Model(net, inputs, labels)


571 572
if __name__ == '__main__':
    unittest.main()