test_imperative_qat.py 16.3 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 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 59 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256
#   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.

from __future__ import print_function

import os
import numpy as np
import random
import unittest
import logging
import paddle
import paddle.fluid as fluid
from paddle.fluid import core
from paddle.fluid.optimizer import AdamOptimizer
from paddle.fluid.framework import IrGraph
from paddle.fluid.contrib.slim.quantization import ImperativeQuantAware
from paddle.fluid.contrib.slim.quantization import QuantizationTransformPass
from paddle.fluid.dygraph.container import Sequential
from paddle.fluid.dygraph.nn import Conv2D
from paddle.fluid.dygraph.nn import Pool2D
from paddle.fluid.dygraph.nn import Linear
from paddle.fluid.log_helper import get_logger

os.environ["CPU_NUM"] = "1"
if core.is_compiled_with_cuda():
    fluid.set_flags({"FLAGS_cudnn_deterministic": True})

_logger = get_logger(
    __name__, logging.INFO, fmt='%(asctime)s-%(levelname)s: %(message)s')


def StaticLenet(data, num_classes=10, classifier_activation='softmax'):
    conv2d_w1_attr = fluid.ParamAttr(name="conv2d_w_1")
    conv2d_w2_attr = fluid.ParamAttr(name="conv2d_w_2")
    fc_w1_attr = fluid.ParamAttr(name="fc_w_1")
    fc_w2_attr = fluid.ParamAttr(name="fc_w_2")
    fc_w3_attr = fluid.ParamAttr(name="fc_w_3")
    conv2d_b1_attr = fluid.ParamAttr(name="conv2d_b_1")
    conv2d_b2_attr = fluid.ParamAttr(name="conv2d_b_2")
    fc_b1_attr = fluid.ParamAttr(name="fc_b_1")
    fc_b2_attr = fluid.ParamAttr(name="fc_b_2")
    fc_b3_attr = fluid.ParamAttr(name="fc_b_3")
    conv1 = fluid.layers.conv2d(
        data,
        num_filters=6,
        filter_size=3,
        stride=1,
        padding=1,
        param_attr=conv2d_w1_attr,
        bias_attr=conv2d_b1_attr)
    pool1 = fluid.layers.pool2d(
        conv1, pool_size=2, pool_type='max', pool_stride=2)
    conv2 = fluid.layers.conv2d(
        pool1,
        num_filters=16,
        filter_size=5,
        stride=1,
        padding=0,
        param_attr=conv2d_w2_attr,
        bias_attr=conv2d_b2_attr)
    pool2 = fluid.layers.pool2d(
        conv2, pool_size=2, pool_type='max', pool_stride=2)

    fc1 = fluid.layers.fc(input=pool2,
                          size=120,
                          param_attr=fc_w1_attr,
                          bias_attr=fc_b1_attr)
    fc2 = fluid.layers.fc(input=fc1,
                          size=84,
                          param_attr=fc_w2_attr,
                          bias_attr=fc_b2_attr)
    fc3 = fluid.layers.fc(input=fc2,
                          size=num_classes,
                          act=classifier_activation,
                          param_attr=fc_w3_attr,
                          bias_attr=fc_b3_attr)

    return fc3


class ImperativeLenet(fluid.dygraph.Layer):
    def __init__(self, num_classes=10, classifier_activation='softmax'):
        super(ImperativeLenet, self).__init__()
        conv2d_w1_attr = fluid.ParamAttr(name="conv2d_w_1")
        conv2d_w2_attr = fluid.ParamAttr(name="conv2d_w_2")
        fc_w1_attr = fluid.ParamAttr(name="fc_w_1")
        fc_w2_attr = fluid.ParamAttr(name="fc_w_2")
        fc_w3_attr = fluid.ParamAttr(name="fc_w_3")
        conv2d_b1_attr = fluid.ParamAttr(name="conv2d_b_1")
        conv2d_b2_attr = fluid.ParamAttr(name="conv2d_b_2")
        fc_b1_attr = fluid.ParamAttr(name="fc_b_1")
        fc_b2_attr = fluid.ParamAttr(name="fc_b_2")
        fc_b3_attr = fluid.ParamAttr(name="fc_b_3")
        self.features = Sequential(
            Conv2D(
                num_channels=1,
                num_filters=6,
                filter_size=3,
                stride=1,
                padding=1,
                param_attr=conv2d_w1_attr,
                bias_attr=conv2d_b1_attr),
            Pool2D(
                pool_size=2, pool_type='max', pool_stride=2),
            Conv2D(
                num_channels=6,
                num_filters=16,
                filter_size=5,
                stride=1,
                padding=0,
                param_attr=conv2d_w2_attr,
                bias_attr=conv2d_b2_attr),
            Pool2D(
                pool_size=2, pool_type='max', pool_stride=2))

        self.fc = Sequential(
            Linear(
                input_dim=400,
                output_dim=120,
                param_attr=fc_w1_attr,
                bias_attr=fc_b1_attr),
            Linear(
                input_dim=120,
                output_dim=84,
                param_attr=fc_w2_attr,
                bias_attr=fc_b2_attr),
            Linear(
                input_dim=84,
                output_dim=num_classes,
                act=classifier_activation,
                param_attr=fc_w3_attr,
                bias_attr=fc_b3_attr))

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

        x = fluid.layers.flatten(x, 1)
        x = self.fc(x)
        return x


class TestImperativeQat(unittest.TestCase):
    """
    QAT = quantization-aware training
    """

    def test_qat_save(self):
        imperative_qat = ImperativeQuantAware(
            weight_quantize_type='abs_max',
            activation_quantize_type='moving_average_abs_max')

        with fluid.dygraph.guard():
            lenet = ImperativeLenet()
            imperative_qat.quantize(lenet)
            adam = AdamOptimizer(
                learning_rate=0.001, parameter_list=lenet.parameters())
            train_reader = paddle.batch(
                paddle.dataset.mnist.train(), batch_size=32, drop_last=True)
            test_reader = paddle.batch(
                paddle.dataset.mnist.test(), batch_size=32)

            epoch_num = 1
            for epoch in range(epoch_num):
                lenet.train()
                for batch_id, data in enumerate(train_reader()):
                    x_data = np.array([x[0].reshape(1, 28, 28)
                                       for x in data]).astype('float32')
                    y_data = np.array(
                        [x[1] for x in data]).astype('int64').reshape(-1, 1)

                    img = fluid.dygraph.to_variable(x_data)
                    label = fluid.dygraph.to_variable(y_data)

                    out = lenet(img)
                    acc = fluid.layers.accuracy(out, label)
                    loss = fluid.layers.cross_entropy(out, label)
                    avg_loss = fluid.layers.mean(loss)
                    avg_loss.backward()
                    adam.minimize(avg_loss)
                    lenet.clear_gradients()
                    if batch_id % 100 == 0:
                        _logger.info(
                            "Train | At epoch {} step {}: loss = {:}, acc= {:}".
                            format(epoch, batch_id,
                                   avg_loss.numpy(), acc.numpy()))

                lenet.eval()
                for batch_id, data in enumerate(test_reader()):
                    x_data = np.array([x[0].reshape(1, 28, 28)
                                       for x in data]).astype('float32')
                    y_data = np.array(
                        [x[1] for x in data]).astype('int64').reshape(-1, 1)

                    img = fluid.dygraph.to_variable(x_data)
                    label = fluid.dygraph.to_variable(y_data)

                    out = lenet(img)
                    acc_top1 = fluid.layers.accuracy(
                        input=out, label=label, k=1)
                    acc_top5 = fluid.layers.accuracy(
                        input=out, label=label, k=5)

                    if batch_id % 100 == 0:
                        _logger.info(
                            "Test | At epoch {} step {}: acc1 = {:}, acc5 = {:}".
                            format(epoch, batch_id,
                                   acc_top1.numpy(), acc_top5.numpy()))

            # save weights
            model_dict = lenet.state_dict()
            fluid.save_dygraph(model_dict, "save_temp")

            # test the correctness of `save_quantized_model`
            data = next(test_reader())
            test_data = np.array([x[0].reshape(1, 28, 28)
                                  for x in data]).astype('float32')
            test_img = fluid.dygraph.to_variable(test_data)
            lenet.eval()
            before_save = lenet(test_img)

        # save inference quantized model
        path = "./mnist_infer_model"
        imperative_qat.save_quantized_model(
            dirname=path,
            model=lenet,
            input_shape=[(1, 28, 28)],
            input_dtype=['float32'],
            feed=[0],
            fetch=[0])
        if core.is_compiled_with_cuda():
            place = core.CUDAPlace(0)
        else:
            place = core.CPUPlace()
        exe = fluid.Executor(place)
        [inference_program, feed_target_names, fetch_targets] = (
            fluid.io.load_inference_model(
                dirname=path, executor=exe))
        after_save, = exe.run(inference_program,
                              feed={feed_target_names[0]: test_data},
                              fetch_list=fetch_targets)

        self.assertTrue(
            np.allclose(after_save, before_save.numpy()),
            msg='Failed to save the inference quantized model.')

257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422
    def test_qat_acc(self):
        def _build_static_lenet(main, startup, is_test=False, seed=1000):
            with fluid.unique_name.guard():
                with fluid.program_guard(main, startup):
                    main.random_seed = seed
                    startup.random_seed = seed
                    img = fluid.layers.data(
                        name='image', shape=[1, 28, 28], dtype='float32')
                    label = fluid.layers.data(
                        name='label', shape=[1], dtype='int64')
                    prediction = StaticLenet(img)
                    if not is_test:
                        loss = fluid.layers.cross_entropy(
                            input=prediction, label=label)
                        avg_loss = fluid.layers.mean(loss)
                    else:
                        avg_loss = prediction
            return img, label, avg_loss

        reader = paddle.batch(
            paddle.dataset.mnist.test(), batch_size=32, drop_last=True)
        weight_quantize_type = 'abs_max'
        activation_quant_type = 'moving_average_abs_max'
        param_init_map = {}
        seed = 1000
        lr = 0.1

        # imperative train
        _logger.info(
            "--------------------------dynamic graph qat--------------------------"
        )
        imperative_qat = ImperativeQuantAware(
            weight_quantize_type=weight_quantize_type,
            activation_quantize_type=activation_quant_type)

        with fluid.dygraph.guard():
            np.random.seed(seed)
            fluid.default_main_program().random_seed = seed
            fluid.default_startup_program().random_seed = seed
            lenet = ImperativeLenet()
            fixed_state = {}
            for name, param in lenet.named_parameters():
                p_shape = param.numpy().shape
                p_value = param.numpy()
                if name.endswith("bias"):
                    value = np.zeros_like(p_value).astype('float32')
                else:
                    value = np.random.normal(
                        loc=0.0, scale=0.01, size=np.product(p_shape)).reshape(
                            p_shape).astype('float32')
                fixed_state[name] = value
                param_init_map[param.name] = value
            lenet.set_dict(fixed_state)

            imperative_qat.quantize(lenet)
            adam = AdamOptimizer(
                learning_rate=lr, parameter_list=lenet.parameters())
            dynamic_loss_rec = []
            lenet.train()
            for batch_id, data in enumerate(reader()):
                x_data = np.array([x[0].reshape(1, 28, 28)
                                   for x in data]).astype('float32')
                y_data = np.array(
                    [x[1] for x in data]).astype('int64').reshape(-1, 1)

                img = fluid.dygraph.to_variable(x_data)
                label = fluid.dygraph.to_variable(y_data)

                out = lenet(img)
                loss = fluid.layers.cross_entropy(out, label)
                avg_loss = fluid.layers.mean(loss)
                avg_loss.backward()
                adam.minimize(avg_loss)
                lenet.clear_gradients()
                dynamic_loss_rec.append(avg_loss.numpy()[0])
                if batch_id % 100 == 0:
                    _logger.info('{}: {}'.format('loss', avg_loss.numpy()))

        imperative_qat.save_quantized_model(
            dirname="./dynamic_mnist",
            model=lenet,
            input_shape=[(1, 28, 28)],
            input_dtype=['float32'],
            feed=[0],
            fetch=[0])

        # static graph train
        _logger.info(
            "--------------------------static graph qat--------------------------"
        )
        static_loss_rec = []
        if core.is_compiled_with_cuda():
            place = core.CUDAPlace(0)
        else:
            place = core.CPUPlace()
        exe = fluid.Executor(place)

        main = fluid.Program()
        infer = fluid.Program()
        startup = fluid.Program()
        static_img, static_label, static_loss = _build_static_lenet(
            main, startup, False, seed)
        infer_img, _, infer_pre = _build_static_lenet(infer, startup, True,
                                                      seed)
        with fluid.unique_name.guard():
            with fluid.program_guard(main, startup):
                opt = AdamOptimizer(learning_rate=lr)
                opt.minimize(static_loss)

        scope = core.Scope()
        with fluid.scope_guard(scope):
            exe.run(startup)
        for param in main.all_parameters():
            param_tensor = scope.var(param.name).get_tensor()
            param_tensor.set(param_init_map[param.name], place)

        main_graph = IrGraph(core.Graph(main.desc), for_test=False)
        infer_graph = IrGraph(core.Graph(infer.desc), for_test=True)
        transform_pass = QuantizationTransformPass(
            scope=scope,
            place=place,
            activation_quantize_type=activation_quant_type,
            weight_quantize_type=weight_quantize_type,
            quantizable_op_type=['conv2d', 'depthwise_conv2d', 'mul'])
        transform_pass.apply(main_graph)
        transform_pass.apply(infer_graph)
        build_strategy = fluid.BuildStrategy()
        build_strategy.fuse_all_reduce_ops = False
        binary = fluid.CompiledProgram(main_graph.graph).with_data_parallel(
            loss_name=static_loss.name, build_strategy=build_strategy)

        feeder = fluid.DataFeeder(
            feed_list=[static_img, static_label], place=place)
        with fluid.scope_guard(scope):
            for batch_id, data in enumerate(reader()):
                loss_v, = exe.run(binary,
                                  feed=feeder.feed(data),
                                  fetch_list=[static_loss])
                static_loss_rec.append(loss_v[0])
                if batch_id % 100 == 0:
                    _logger.info('{}: {}'.format('loss', loss_v))

        save_program = infer_graph.to_program()
        with fluid.scope_guard(scope):
            fluid.io.save_inference_model("./static_mnist", [infer_img.name],
                                          [infer_pre], exe, save_program)
        rtol = 1e-05
        atol = 1e-08
        for i, (loss_d,
                loss_s) in enumerate(zip(dynamic_loss_rec, static_loss_rec)):
            diff = np.abs(loss_d - loss_s)
            if diff > (atol + rtol * np.abs(loss_s)):
                _logger.info(
                    "diff({}) at {}, dynamic loss = {}, static loss = {}".
                    format(diff, i, loss_d, loss_s))
                break

        self.assertTrue(
            np.allclose(
                np.array(dynamic_loss_rec),
                np.array(static_loss_rec),
                rtol=rtol,
                atol=atol,
                equal_nan=True),
            msg='Failed to do the imperative qat.')

423 424 425

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