Add unittest for imperative resnet

Fix the bug of static BatchNorm layer

python/paddle/fluid/imperative/nn.py

@@ -27,6 +27,7 @@ __all__ = [
     'Conv2D',
     'Pool2D',
     'FC',
+    'BatchNorm',
 ]
 
 
@@ -209,14 +210,24 @@ class FC(layers.Layer):
     def __init__(self,
                  size,
                  param_attr=None,
+                 bias_attr=None,
+                 dtype=core.VarDesc.VarType.FP32,
                  num_flatten_dims=1,
-                 dtype=core.VarDesc.VarType.FP32):
+                 act=None,
+                 is_test=False,
+                 name=None):
         super(FC, self).__init__()
+
         self._size = size
         self._num_flatten_dims = num_flatten_dims
         self._dtype = dtype
         from ..layer_helper import LayerHelper
-        self._helper = LayerHelper('FC', param_attr=param_attr)
+        self._helper = LayerHelper(
+            'FC',
+            param_attr=param_attr,
+            bias_attr=bias_attr,
+            act=act,
+            name=name)
 
     def _build_once(self, input):
         input_shape = input.shape
@@ -247,4 +258,132 @@ class FC(layers.Layer):
             inputs={"X": [tmp]},
             outputs={"Out": out},
             attrs={"use_mkldnn": False})
-        return out
+
+        pre_activation = self._helper.append_bias_op(
+            pre_bias, dim_start=num_flatten_dims)
+        return self._helper.append_activation(pre_activation)
+
+
+class BatchNorm(layers.Layer):
+    def __init__(self,
+                 num_channels,
+                 act=None,
+                 is_test=False,
+                 momentum=0.9,
+                 epsilon=1e-05,
+                 param_attr=None,
+                 bias_attr=None,
+                 dtype=core.VarDesc.VarType.FP32,
+                 data_layout='NCHW',
+                 in_place=False,
+                 name=None,
+                 moving_mean_name=None,
+                 moving_variance_name=None,
+                 do_model_average_for_mean_and_var=False,
+                 fuse_with_relu=False,
+                 use_global_stats=False):
+        super(BatchNorm, self).__init__()
+
+        assert bias_attr is not False, "bias_attr should not be False in batch_norm."
+
+        from ..layer_helper import LayerHelper
+        self._helper = LayerHelper(
+            'batch_norm', param_attr=param_attr, bias_attr=bias_attr, name=name)
+
+        if dtype == core.VarDesc.VarType.FP16:
+            self._dtype = core.VarDesc.VarType.FP32
+        else:
+            self._dtype = dtype
+
+        param_shape = [num_channels]
+
+        # create parameter
+        self._scale = self._helper.create_parameter(
+            attr=self._helper.param_attr,
+            shape=param_shape,
+            dtype=self._dtype,
+            default_initializer=Constant(1.0))
+
+        # setting stop_gradient=True to reduce computation
+        if use_global_stats and self._helper.param_attr.learning_rate == 0.:
+            self._scale.stop_gradient = True
+
+        self._bias = self._helper.create_parameter(
+            attr=self._helper.bias_attr,
+            shape=param_shape,
+            dtype=self._dtype,
+            is_bias=True)
+        # setting stop_gradient=True to reduce computation
+        if use_global_stats and self._helper.bias_attr.learning_rate == 0.:
+            self._bias.stop_gradient = True
+
+        self._mean = self._helper.create_parameter(
+            attr=ParamAttr(
+                name=moving_mean_name,
+                initializer=Constant(0.0),
+                trainable=False,
+                do_model_average=do_model_average_for_mean_and_var),
+            shape=param_shape,
+            dtype=self._dtype)
+        self._mean.stop_gradient = True
+
+        self._variance = self._helper.create_parameter(
+            attr=ParamAttr(
+                name=moving_variance_name,
+                initializer=Constant(1.0),
+                trainable=False,
+                do_model_average=do_model_average_for_mean_and_var),
+            shape=param_shape,
+            dtype=self._dtype)
+        self._variance.stop_gradient = True
+
+        self._in_place = in_place
+        self._momentum = momentum
+        self._epsilon = epsilon
+        self._is_test = is_test
+        self._fuse_with_relu = fuse_with_relu
+        self._use_global_stats = use_global_stats
+
+    def _build_once(self, input):
+        pass
+
+    def forward(self, input):
+        # create output
+        # mean and mean_out share the same memory
+        mean_out = self._mean
+        # variance and variance out share the same memory
+        variance_out = self._variance
+
+        saved_mean = self._helper.create_variable_for_type_inference(
+            dtype=dtype, stop_gradient=True)
+        saved_variance = self._helper.create_variable_for_type_inference(
+            dtype=dtype, stop_gradient=True)
+        batch_norm_out = input if self._in_place else self._helper.create_variable_for_type_inference(
+            dtype)
+
+        self._helper.append_op(
+            type="batch_norm",
+            inputs={
+                "X": input,
+                "Scale": self._scale,
+                "Bias": self._bias,
+                "Mean": self._mean,
+                "Variance": self._variance
+            },
+            outputs={
+                "Y": batch_norm_out,
+                "MeanOut": mean_out,
+                "VarianceOut": variance_out,
+                "SavedMean": saved_mean,
+                "SavedVariance": saved_variance
+            },
+            attrs={
+                "momentum": self._momentum,
+                "epsilon": self._epsilon,
+                "is_test": self._is_test,
+                "use_mkldnn": False,
+                "fuse_with_relu": self._fuse_with_relu,
+                "use_global_stats": self._use_global_stats
+            })
+
+        return self._helper.append_activation(batch_norm_out)

python/paddle/fluid/layers/nn.py

@@ -2835,7 +2835,7 @@ def batch_norm(input,
         attr=helper.bias_attr, shape=param_shape, dtype=dtype, is_bias=True)
     # setting stop_gradient=True to reduce computation
     if use_global_stats and helper.bias_attr.learning_rate == 0.:
-        scale.stop_gradient = True
+        bias.stop_gradient = True
 
     mean = helper.create_parameter(
         attr=ParamAttr(

python/paddle/fluid/tests/unittests/test_imperative_resnet.py

+# 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.optimizer import SGDOptimizer
+from paddle.fluid.imperative.nn import Conv2D, Pool2D, BatchNorm, FC
+from paddle.fluid.imperative.base import to_variable
+from test_imperative_base import new_program_scope
+
+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": 256,
+        "epochs": [30, 60, 90],
+        "steps": [0.1, 0.01, 0.001, 0.0001]
+    }
+}
+
+
+def optimizer_setting(params):
+    ls = params["learning_strategy"]
+    if ls["name"] == "piecewise_decay":
+        if "total_images" not in params:
+            total_images = 1281167
+        else:
+            total_images = params["total_images"]
+        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 = []
+        lr = [base_lr * (0.1**i) for i in range(len(bd) + 1)]
+        optimizer = fluid.optimizer.Momentum(
+            learning_rate=fluid.layers.piecewise_decay(
+                boundaries=bd, values=lr),
+            momentum=0.9,
+            regularization=fluid.regularizer.L2Decay(1e-4))
+
+    return optimizer
+
+
+class ConvBNLayer(fluid.imperative.Layer):
+    def __init__(self, num_filters, filter_size, stride=1, groups=1, act=None):
+        super(ConvBNLayer, self).__init__()
+
+        self._conv = Conv2D(
+            3,
+            num_filters,
+            filter_size,
+            stride, (filter_size - 1) // 2,
+            groups=groups,
+            act=None,
+            bias_attr=None)
+
+        self._batch_norm = BatchNorm(num_filters, act=act)
+
+    def forward(self, inputs):
+        y = self._conv(inputs)
+        y = self._batch_norm(y)
+
+        return y
+
+
+class BottleneckBlock(fluid.imperative.Layer):
+    def __init__(self, num_filters, stride, shortcut=False):
+        super(BottleneckBlock, self).__init__()
+
+        self.conv0 = ConvBNLayer(
+            num_filters=num_filters, filter_size=1, act='relu')
+        self.conv1 = ConvBNLayer(
+            num_filters=num_filters, filter_size=3, stride=stride, act='relu')
+        self.conv2 = ConvBNLayer(
+            num_filters=num_filters * 4, filter_size=1, act=None)
+
+        if shortcut:
+            self.short = ConvBNLayer(
+                num_filters=num_filters * 4, filter_size=1, stride=stride)
+
+        self.shortcut = shortcut
+
+    def forward(self, inputs):
+        self.conv0()
+        self.conv1()
+        self.conv2()
+
+        if self.shortcut:
+            self.short()
+
+        return fluid.layers.elementwise_add(
+            x=self.short, y=self.conv2, act='relu')
+
+
+class ResNet(fluid.imperative.Layer):
+    def __init__(self, layers=50, class_dim=1000):
+        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:
+            depth = [3, 4, 6, 3]
+        elif layers == 101:
+            depth = [3, 4, 23, 3]
+        elif layers == 152:
+            depth = [3, 8, 36, 3]
+        num_filters = [64, 128, 256, 512]
+
+        self.conv = ConvBNLayer(
+            num_filters=64, filter_size=7, stride=2, act='relu')
+        self.pool2d_max = Pool2D(
+            pool_size=3, pool_stride=2, pool_padding=1, pool_type='max')
+
+        self.bottleneck_block_list = []
+        for block in range(len(depth)):
+            shortcut = True
+            for i in range(depth[block]):
+                bottleneck_block = BottleneckBlock(
+                    num_filters=num_filters[block],
+                    stride=2 if i == 0 and block != 0 else 1,
+                    shortcut=shortcut)
+                self.bottleneck_block_list.append(bottleneck_block)
+                shortcut = False
+
+        self.pool2d_avg = Pool2D(
+            pool_size=7, pool_type='avg', global_pooling=True)
+
+        import math
+        stdv = 1.0 / math.sqrt(2048 * 1.0)
+
+        self.out = FC(size=class_dim,
+                      act='softmax',
+                      param_attr=fluid.param_attr.ParamAttr(
+                          initializer=fluid.initializer.Uniform(-stdv, stdv)))
+
+    def forward(self, inputs):
+        y = self.conv(inputs)
+        y = self.pool2d_max(y)
+        for bottleneck_block in self.bottleneck_block_list:
+            y = bottleneck_block(y)
+        y = self.pool2d_avg(y)
+        y = self.out()
+        return y
+
+
+class TestImperativeResnet(unittest.TestCase):
+    def test_resnet_cpu_float32(self):
+        seed = 90
+
+        with fluid.imperative.guard():
+            fluid.default_startup_program().random_seed = seed
+            fluid.default_main_program().random_seed = seed
+
+            resnet = ResNet()
+            optimizer = optimizer_setting(train_parameters)
+            train_reader = paddle.batch(
+                paddle.dataset.flowers.train(), batch_size=256)
+
+            dy_param_init_value = {}
+            for batch_id, data in enumerate(train_reader()):
+                if batch_id >= 2:
+                    break
+
+                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(
+                    128, 1)
+
+                img = to_variable(x_data)
+                label = to_variable(y_data)
+                label._stop_gradient = True
+
+                cost = resnet(img)
+                loss = fluid.layers.cross_entropy(input=out, label=label)
+                avg_loss = fluid.layers.mean(x=cost)
+                dy_out = avg_loss._numpy()
+
+                if batch_id == 0:
+                    for param in fluid.default_main_program().global_block(
+                    ).all_parameters():
+                        dy_param_init_value[param.name] = param._numpy()
+
+                avg_loss._backward()
+                optimizer.minimize(avg_loss)
+                dy_param_value = {}
+                for param in fluid.default_main_program().global_block(
+                ).all_parameters():
+                    dy_param_value[param.name] = param._numpy()
+
+        #  with new_program_scope():
+        #  fluid.default_startup_program().random_seed = seed
+        #  fluid.default_main_program().random_seed = seed
+
+        #  exe = fluid.Executor(fluid.CPUPlace())
+
+        #  #  mnist = Conv2D(1, 20, 5)
+        #  mnist = MNIST()
+        #  sgd = SGDOptimizer(learning_rate=1e-3)
+        #  train_reader = paddle.batch(
+        #  paddle.dataset.mnist.train(), batch_size=128)
+
+        #  img = fluid.layers.data(
+        #  name='pixel', shape=[1, 28, 28], dtype='float32')
+        #  label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+        #  cost = mnist(img)
+        #  loss = fluid.layers.reduce_mean(cost)
+        #  sgd.minimize(loss)
+
+        #  # initialize params and fetch them
+        #  static_param_init_value = {}
+        #  static_param_name_list = []
+        #  for param in fluid.default_startup_program().global_block(
+        #  ).all_parameters():
+        #  static_param_name_list.append(param.name)
+
+        #  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]
+
+        #  for batch_id, data in enumerate(train_reader()):
+        #  if batch_id >= 2:
+        #  break
+
+        #  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(
+        #  [128, 1])
+
+        #  fetch_list = [loss.name]
+        #  fetch_list.extend(static_param_name_list)
+        #  out = exe.run(fluid.default_main_program(),
+        #  feed={"pixel": x_data,
+        #  "label": y_data},
+        #  fetch_list=fetch_list)
+
+        #  static_param_value = {}
+        #  static_out = out[0]
+        #  for i in range(1, len(out)):
+        #  static_param_value[static_param_name_list[i - 1]] = out[i]
+
+        #  for key, value in six.iteritems(static_param_init_value):
+        #  self.assertTrue(
+        #  np.allclose(value.all(), dy_param_init_value[key].all()))
+        #  self.assertTrue(np.allclose(static_out.all(), dy_out.all()))
+        #  for key, value in six.iteritems(static_param_value):
+        #  self.assertTrue(np.allclose(value.all(), dy_param_value[key].all()))
+
+
+if __name__ == '__main__':
+    unittest.main()