up to paddle 1.7 (#4216)

dygraph/mobilenet/mobilenet_v1.py

@@ -23,7 +23,7 @@ import paddle.fluid as fluid
 from paddle.fluid.initializer import MSRA
 from paddle.fluid.param_attr import ParamAttr
 from paddle.fluid.layer_helper import LayerHelper
-from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, FC
+from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear
 from paddle.fluid.dygraph.base import to_variable
 from paddle.fluid import framework
 import math
@@ -32,7 +32,7 @@ import sys
 
 class ConvBNLayer(fluid.dygraph.Layer):
     def __init__(self,
-                 name_scope,
+                 num_channels,
                  filter_size,
                  num_filters,
                  stride,
@@ -42,10 +42,10 @@ class ConvBNLayer(fluid.dygraph.Layer):
                  act='relu',
                  use_cudnn=True,
                  name=None):
-        super(ConvBNLayer, self).__init__(name_scope)
+        super(ConvBNLayer, self).__init__()
 
         self._conv = Conv2D(
-            self.full_name(),
+            num_channels=num_channels,
             num_filters=num_filters,
             filter_size=filter_size,
             stride=stride,
@@ -58,13 +58,12 @@ class ConvBNLayer(fluid.dygraph.Layer):
             bias_attr=False)
 
         self._batch_norm = BatchNorm(
-            self.full_name(),
             num_filters,
             act=act,
-            param_attr=ParamAttr(name="_bn" + "_scale"),
-            bias_attr=ParamAttr(name="_bn" + "_offset"),
-            moving_mean_name="_bn" + '_mean',
-            moving_variance_name="_bn" + '_variance')
+            param_attr=ParamAttr(name=self.full_name() + "_bn" + "_scale"),
+            bias_attr=ParamAttr(name=self.full_name() + "_bn" + "_offset"),
+            moving_mean_name=self.full_name() + "_bn" + '_mean',
+            moving_variance_name=self.full_name() + "_bn" + '_variance')
 
     def forward(self, inputs):
         y = self._conv(inputs)
@@ -74,17 +73,17 @@ class ConvBNLayer(fluid.dygraph.Layer):
 
 class DepthwiseSeparable(fluid.dygraph.Layer):
     def __init__(self,
-                 name_scope,
+                 num_channels,
                  num_filters1,
                  num_filters2,
                  num_groups,
                  stride,
                  scale,
                  name=None):
-        super(DepthwiseSeparable, self).__init__(name_scope)
+        super(DepthwiseSeparable, self).__init__()
 
         self._depthwise_conv = ConvBNLayer(
-            name_scope="dw",
+            num_channels=num_channels,
             num_filters=int(num_filters1 * scale),
             filter_size=3,
             stride=stride,
@@ -93,7 +92,7 @@ class DepthwiseSeparable(fluid.dygraph.Layer):
             use_cudnn=False)
 
         self._pointwise_conv = ConvBNLayer(
-            name_scope="sep",
+            num_channels=int(num_filters1 * scale),
             filter_size=1,
             num_filters=int(num_filters2 * scale),
             stride=1,
@@ -106,13 +105,13 @@ class DepthwiseSeparable(fluid.dygraph.Layer):
 
 
 class MobileNetV1(fluid.dygraph.Layer):
-    def __init__(self, name_scope, scale=1.0, class_dim=102):
-        super(MobileNetV1, self).__init__(name_scope)
+    def __init__(self, scale=1.0, class_dim=1000):
+        super(MobileNetV1, self).__init__()
         self.scale = scale
         self.dwsl = []
 
         self.conv1 = ConvBNLayer(
-            name_scope="conv1",
+            num_channels=3,
             filter_size=3,
             channels=3,
             num_filters=int(32 * scale),
@@ -121,7 +120,7 @@ class MobileNetV1(fluid.dygraph.Layer):
 
         dws21 = self.add_sublayer(
             sublayer=DepthwiseSeparable(
-                name_scope="conv2_1",
+                num_channels=int(32 * scale),
                 num_filters1=32,
                 num_filters2=64,
                 num_groups=32,
@@ -132,7 +131,7 @@ class MobileNetV1(fluid.dygraph.Layer):
 
         dws22 = self.add_sublayer(
             sublayer=DepthwiseSeparable(
-                name_scope="conv2_2",
+                num_channels=int(64 * scale),
                 num_filters1=64,
                 num_filters2=128,
                 num_groups=64,
@@ -143,7 +142,7 @@ class MobileNetV1(fluid.dygraph.Layer):
 
         dws31 = self.add_sublayer(
             sublayer=DepthwiseSeparable(
-                name_scope="conv3_1",
+                num_channels=int(128 * scale),
                 num_filters1=128,
                 num_filters2=128,
                 num_groups=128,
@@ -154,7 +153,7 @@ class MobileNetV1(fluid.dygraph.Layer):
 
         dws32 = self.add_sublayer(
             sublayer=DepthwiseSeparable(
-                name_scope="conv3_2",
+                num_channels=int(128 * scale),
                 num_filters1=128,
                 num_filters2=256,
                 num_groups=128,
@@ -165,7 +164,7 @@ class MobileNetV1(fluid.dygraph.Layer):
 
         dws41 = self.add_sublayer(
             sublayer=DepthwiseSeparable(
-                name_scope="conv4_1",
+                num_channels=int(256 * scale),
                 num_filters1=256,
                 num_filters2=256,
                 num_groups=256,
@@ -176,7 +175,7 @@ class MobileNetV1(fluid.dygraph.Layer):
 
         dws42 = self.add_sublayer(
             sublayer=DepthwiseSeparable(
-                name_scope="conv4_2",
+                num_channels=int(256 * scale),
                 num_filters1=256,
                 num_filters2=512,
                 num_groups=256,
@@ -188,7 +187,7 @@ class MobileNetV1(fluid.dygraph.Layer):
         for i in range(5):
             tmp = self.add_sublayer(
                 sublayer=DepthwiseSeparable(
-                    name_scope="conv5_" + str(i + 1),
+                    num_channels=int(512 * scale),
                     num_filters1=512,
                     num_filters2=512,
                     num_groups=512,
@@ -199,7 +198,7 @@ class MobileNetV1(fluid.dygraph.Layer):
 
         dws56 = self.add_sublayer(
             sublayer=DepthwiseSeparable(
-                name_scope="conv5_6",
+                num_channels=int(512 * scale),
                 num_filters1=512,
                 num_filters2=1024,
                 num_groups=512,
@@ -210,7 +209,7 @@ class MobileNetV1(fluid.dygraph.Layer):
 
         dws6 = self.add_sublayer(
             sublayer=DepthwiseSeparable(
-                name_scope="conv6",
+                num_channels=int(1024 * scale),
                 num_filters1=1024,
                 num_filters2=1024,
                 num_groups=1024,
@@ -219,21 +218,20 @@ class MobileNetV1(fluid.dygraph.Layer):
             name="conv6")
         self.dwsl.append(dws6)
 
-        self.pool2d_avg = Pool2D(
-            name_scope="pool", pool_type='avg', global_pooling=True)
+        self.pool2d_avg = Pool2D(pool_type='avg', global_pooling=True)
 
-        self.out = FC(name_scope="fc",
-                      size=class_dim,
-                      param_attr=ParamAttr(
-                          initializer=MSRA(),
-                          name=self.full_name() + "fc7_weights"),
-                      bias_attr=ParamAttr(name="fc7_offset"))
+        self.out = Linear(
+            int(1024 * scale),
+            class_dim,
+            param_attr=ParamAttr(
+                initializer=MSRA(), name=self.full_name() + "fc7_weights"),
+            bias_attr=ParamAttr(name="fc7_offset"))
 
     def forward(self, inputs):
         y = self.conv1(inputs)
-        idx = 0
         for dws in self.dwsl:
             y = dws(y)
         y = self.pool2d_avg(y)
+        y = fluid.layers.reshape(y, shape=[-1, 1024])
         y = self.out(y)
         return y

dygraph/mobilenet/mobilenet_v2.py

@@ -25,7 +25,7 @@ import paddle.fluid as fluid
 from paddle.fluid.initializer import MSRA
 from paddle.fluid.param_attr import ParamAttr
 from paddle.fluid.layer_helper import LayerHelper
-from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, FC
+from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, Linear
 from paddle.fluid.dygraph.base import to_variable
 
 from paddle.fluid import framework
@@ -36,19 +36,19 @@ import sys
 
 class ConvBNLayer(fluid.dygraph.Layer):
     def __init__(self,
+                 num_channels,
                  filter_size,
                  num_filters,
                  stride,
                  padding,
                  channels=None,
                  num_groups=1,
-                 name=None,
                  use_cudnn=True):
-        super(ConvBNLayer, self).__init__(name)
+        super(ConvBNLayer, self).__init__()
 
-        tmp_param = ParamAttr(name=name + "_weights")
+        tmp_param = ParamAttr(name=self.full_name() + "_weights")
         self._conv = Conv2D(
-            self.full_name(),
+            num_channels=num_channels,
             num_filters=num_filters,
             filter_size=filter_size,
             stride=stride,
@@ -60,12 +60,11 @@ class ConvBNLayer(fluid.dygraph.Layer):
             bias_attr=False)
 
         self._batch_norm = BatchNorm(
-            self.full_name(),
             num_filters,
-            param_attr=ParamAttr(name=name + "_bn" + "_scale"),
-            bias_attr=ParamAttr(name=name + "_bn" + "_offset"),
-            moving_mean_name=name + "_bn" + '_mean',
-            moving_variance_name=name + "_bn" + '_variance')
+            param_attr=ParamAttr(name=self.full_name() + "_bn" + "_scale"),
+            bias_attr=ParamAttr(name=self.full_name() + "_bn" + "_offset"),
+            moving_mean_name=self.full_name() + "_bn" + '_mean',
+            moving_variance_name=self.full_name() + "_bn" + '_variance')
 
     def forward(self, inputs, if_act=True):
         y = self._conv(inputs)
@@ -76,18 +75,19 @@ class ConvBNLayer(fluid.dygraph.Layer):
 
 
 class InvertedResidualUnit(fluid.dygraph.Layer):
-    def __init__(self,
-                 num_in_filter,
-                 num_filters,
-                 stride,
-                 filter_size,
-                 padding,
-                 expansion_factor,
-                 name=None):
-        super(InvertedResidualUnit, self).__init__(name)
+    def __init__(
+            self,
+            num_channels,
+            num_in_filter,
+            num_filters,
+            stride,
+            filter_size,
+            padding,
+            expansion_factor, ):
+        super(InvertedResidualUnit, self).__init__()
         num_expfilter = int(round(num_in_filter * expansion_factor))
         self._expand_conv = ConvBNLayer(
-            name=name + "_expand",
+            num_channels=num_channels,
             num_filters=num_expfilter,
             filter_size=1,
             stride=1,
@@ -95,7 +95,7 @@ class InvertedResidualUnit(fluid.dygraph.Layer):
             num_groups=1)
 
         self._bottleneck_conv = ConvBNLayer(
-            name=name + "_dwise",
+            num_channels=num_expfilter,
             num_filters=num_expfilter,
             filter_size=filter_size,
             stride=stride,
@@ -104,7 +104,7 @@ class InvertedResidualUnit(fluid.dygraph.Layer):
             use_cudnn=False)
 
         self._linear_conv = ConvBNLayer(
-            name=name + "_linear",
+            num_channels=num_expfilter,
             num_filters=num_filters,
             filter_size=1,
             stride=1,
@@ -121,11 +121,11 @@ class InvertedResidualUnit(fluid.dygraph.Layer):
 
 
 class InvresiBlocks(fluid.dygraph.Layer):
-    def __init__(self, in_c, t, c, n, s, name=None):
-        super(InvresiBlocks, self).__init__(name)
+    def __init__(self, in_c, t, c, n, s):
+        super(InvresiBlocks, self).__init__()
 
         self._first_block = InvertedResidualUnit(
-            name=name + "_1",
+            num_channels=in_c,
             num_in_filter=in_c,
             num_filters=c,
             stride=s,
@@ -137,14 +137,14 @@ class InvresiBlocks(fluid.dygraph.Layer):
         for i in range(1, n):
             tmp = self.add_sublayer(
                 sublayer=InvertedResidualUnit(
-                    name=name + "_" + str(i + 1),
+                    num_channels=c,
                     num_in_filter=c,
                     num_filters=c,
                     stride=1,
                     filter_size=3,
                     padding=1,
                     expansion_factor=t),
-                name=name + "_" + str(i + 1))
+                name=self.full_name() + "_" + str(i + 1))
             self._inv_blocks.append(tmp)
 
     def forward(self, inputs):
@@ -155,8 +155,8 @@ class InvresiBlocks(fluid.dygraph.Layer):
 
 
 class MobileNetV2(fluid.dygraph.Layer):
-    def __init__(self, name, class_dim=1000, scale=1.0):
-        super(MobileNetV2, self).__init__(name)
+    def __init__(self, class_dim=1000, scale=1.0):
+        super(MobileNetV2, self).__init__()
         self.scale = scale
         self.class_dim = class_dim
 
@@ -172,7 +172,7 @@ class MobileNetV2(fluid.dygraph.Layer):
 
         #1. conv1 
         self._conv1 = ConvBNLayer(
-            name="conv1_1",
+            num_channels=3,
             num_filters=int(32 * scale),
             filter_size=3,
             stride=2,
@@ -187,34 +187,30 @@ class MobileNetV2(fluid.dygraph.Layer):
             i += 1
             tmp = self.add_sublayer(
                 sublayer=InvresiBlocks(
-                    name='conv' + str(i),
-                    in_c=in_c,
-                    t=t,
-                    c=int(c * scale),
-                    n=n,
-                    s=s),
+                    in_c=in_c, t=t, c=int(c * scale), n=n, s=s),
                 name='conv' + str(i))
             self._invl.append(tmp)
             in_c = int(c * scale)
 
         #3. last_conv
+        self._out_c = int(1280 * scale) if scale > 1.0 else 1280
         self._conv9 = ConvBNLayer(
-            name="conv9",
-            num_filters=int(1280 * scale) if scale > 1.0 else 1280,
+            num_channels=in_c,
+            num_filters=self._out_c,
             filter_size=1,
             stride=1,
             padding=0)
 
         #4. pool
-        self._pool2d_avg = Pool2D(
-            name_scope="pool", pool_type='avg', global_pooling=True)
+        self._pool2d_avg = Pool2D(pool_type='avg', global_pooling=True)
 
         #5. fc
-        tmp_param = ParamAttr(name="fc10_weights")
-        self._fc = FC(name_scope="fc",
-                      size=class_dim,
-                      param_attr=tmp_param,
-                      bias_attr=ParamAttr(name="fc10_offset"))
+        tmp_param = ParamAttr(name=self.full_name() + "fc10_weights")
+        self._fc = Linear(
+            self._out_c,
+            class_dim,
+            param_attr=tmp_param,
+            bias_attr=ParamAttr(name="fc10_offset"))
 
     def forward(self, inputs):
         y = self._conv1(inputs, if_act=True)
@@ -222,5 +218,6 @@ class MobileNetV2(fluid.dygraph.Layer):
             y = inv(y)
         y = self._conv9(y, if_act=True)
         y = self._pool2d_avg(y)
+        y = fluid.layers.reshape(y, shape=[-1, self._out_c])
         y = self._fc(y)
         return y

dygraph/mobilenet/run_mul_v1.sh

 export CUDA_VISIBLE_DEVICES=0,1,2,3
-python -m paddle.distributed.launch --log_dir ./mylog.time train.py --use_data_parallel 1 --batch_size=256     --reader_thread=8    --total_images=1281167    --class_dim=1000 --image_shape=3,224,224 --model_save_dir=output/ --lr_strategy=piecewise_decay --lr=0.1   --data_dir=../../PaddleCV/image_classification/data/ILSVRC2012 --l2_decay=3e-5  --model=MobileNetV1
+python3 -m paddle.distributed.launch --log_dir ./mylog.time train.py --use_data_parallel 1 --batch_size=256     --reader_thread=8    --total_images=1281167    --class_dim=1000 --image_shape=3,224,224 --model_save_dir=output/ --lr_strategy=piecewise_decay --lr=0.1   --data_dir=../../PaddleCV/image_classification/data/ILSVRC2012 --l2_decay=3e-5  --model=MobileNetV1

dygraph/mobilenet/run_mul_v2.sh

 export CUDA_VISIBLE_DEVICES=0,1,2,3
-python -m paddle.distributed.launch --log_dir ./mylog.time train.py --use_data_parallel 1 --batch_size=256     --reader_thread=8    --total_images=1281167    --class_dim=1000 --image_shape=3,224,224 --model_save_dir=output/ --lr_strategy=piecewise_decay --lr=0.1   --data_dir=../../PaddleCV/image_classification/data/ILSVRC2012 --l2_decay=3e-5  --model=MobileNetV2
+python3 -m paddle.distributed.launch --log_dir ./mylog.time train.py --use_data_parallel 1 --batch_size=256     --reader_thread=8    --total_images=1281167    --class_dim=1000 --image_shape=3,224,224 --model_save_dir=output/ --lr_strategy=piecewise_decay --lr=0.1   --data_dir=../../PaddleCV/image_classification/data/ILSVRC2012 --l2_decay=3e-5  --model=MobileNetV2

dygraph/mobilenet/run_sing_v1.sh

 export CUDA_VISIBLE_DEVICES=0
-python train.py      --batch_size=256        --total_images=1281167    --class_dim=1000 --image_shape=3,224,224 --model_save_dir=output/ --lr_strategy=piecewise_decay --lr=0.1   --data_dir=../../PaddleCV/image_classification/data/ILSVRC2012  --l2_decay=3e-5  --model=MobileNetV1 
+python3 train.py      --batch_size=256        --total_images=1281167    --class_dim=1000 --image_shape=3,224,224 --model_save_dir=output/ --lr_strategy=piecewise_decay --lr=0.1   --data_dir=../../PaddleCV/image_classification/data/ILSVRC2012  --l2_decay=3e-5  --model=MobileNetV1 

dygraph/mobilenet/run_sing_v2.sh

 export CUDA_VISIBLE_DEVICES=0
-python train.py      --batch_size=128        --total_images=1281167    --class_dim=1000 --image_shape=3,224,224 --model_save_dir=output/ --lr_strategy=piecewise_decay --lr=0.1   --data_dir=../../PaddleCV/image_classification/data/ILSVRC2012  --model=MobileNetV2
+python3 train.py      --batch_size=128        --total_images=1281167    --class_dim=1000 --image_shape=3,224,224 --model_save_dir=output/ --lr_strategy=piecewise_decay --lr=0.1   --data_dir=../../PaddleCV/image_classification/data/ILSVRC2012  --model=MobileNetV2

dygraph/mobilenet/train.py

@@ -27,7 +27,7 @@ import paddle.fluid as fluid
 from paddle.fluid.initializer import MSRA
 from paddle.fluid.param_attr import ParamAttr
 from paddle.fluid.layer_helper import LayerHelper
-from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, FC
+#from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, FC
 from paddle.fluid.dygraph.base import to_variable
 
 from paddle.fluid import framework
@@ -95,11 +95,10 @@ def train_mobilenet():
 
         net = None
         if args.model == "MobileNetV1":
-            net = MobileNetV1("mobilenet_v1", class_dim=args.class_dim)
+            net = MobileNetV1(class_dim=args.class_dim)
             para_name = 'mobilenet_v1_params'
         elif args.model == "MobileNetV2":
-            net = MobileNetV2(
-                name="mobilenet_v2", class_dim=args.class_dim, scale=1.0)
+            net = MobileNetV2(class_dim=args.class_dim, scale=1.0)
             para_name = 'mobilenet_v2_params'
         else:
             print(
@@ -107,7 +106,7 @@ def train_mobilenet():
             )
             exit()
 
-        optimizer = create_optimizer(args)
+        optimizer = create_optimizer(args=args, parameter_list=net.parameters())
         if args.use_data_parallel:
             net = fluid.dygraph.parallel.DataParallel(net, strategy)
         train_data_loader, train_data = utility.create_data_loader(

dygraph/mobilenet/utils/optimizer.py

@@ -148,7 +148,8 @@ class Optimizer(object):
         
     """
 
-    def __init__(self, args):
+    def __init__(self, args, parameter_list):
+        self.parameter_list = parameter_list
         self.batch_size = args.batch_size
         self.lr = args.lr
         self.lr_strategy = args.lr_strategy
@@ -175,7 +176,8 @@ class Optimizer(object):
         optimizer = fluid.optimizer.Momentum(
             learning_rate=learning_rate,
             momentum=self.momentum_rate,
-            regularization=fluid.regularizer.L2Decay(self.l2_decay))
+            regularization=fluid.regularizer.L2Decay(self.l2_decay),
+            parameter_list=self.parameter_list)
         return optimizer
 
     def cosine_decay(self):
@@ -192,7 +194,8 @@ class Optimizer(object):
         optimizer = fluid.optimizer.Momentum(
             learning_rate=learning_rate,
             momentum=self.momentum_rate,
-            regularization=fluid.regularizer.L2Decay(self.l2_decay))
+            regularization=fluid.regularizer.L2Decay(self.l2_decay),
+            parameter_list=self.parameter_list)
         return optimizer
 
     def cosine_decay_warmup(self):
@@ -209,7 +212,8 @@ class Optimizer(object):
         optimizer = fluid.optimizer.Momentum(
             learning_rate=learning_rate,
             momentum=self.momentum_rate,
-            regularization=fluid.regularizer.L2Decay(self.l2_decay))
+            regularization=fluid.regularizer.L2Decay(self.l2_decay),
+            parameter_list=self.parameter_list)
         return optimizer
 
     def exponential_decay_warmup(self):
@@ -230,7 +234,8 @@ class Optimizer(object):
             regularization=fluid.regularizer.L2Decay(self.l2_decay),
             momentum=self.momentum_rate,
             rho=0.9,
-            epsilon=0.001)
+            epsilon=0.001,
+            parameter_list=self.parameter_list)
         return optimizer
 
     def linear_decay(self):
@@ -246,7 +251,8 @@ class Optimizer(object):
         optimizer = fluid.optimizer.Momentum(
             learning_rate=learning_rate,
             momentum=self.momentum_rate,
-            regularization=fluid.regularizer.L2Decay(self.l2_decay))
+            regularization=fluid.regularizer.L2Decay(self.l2_decay),
+            parameter_list=self.parameter_list)
 
         return optimizer
 
@@ -257,7 +263,8 @@ class Optimizer(object):
             an adam_decay optimizer
         """
 
-        return fluid.optimizer.Adam(learning_rate=self.lr)
+        return fluid.optimizer.Adam(
+            learning_rate=self.lr, parameter_list=self.parameter_list)
 
     def cosine_decay_RMSProp(self):
         """cosine decay with RMSProp optimizer
@@ -275,7 +282,8 @@ class Optimizer(object):
             momentum=self.momentum_rate,
             regularization=fluid.regularizer.L2Decay(self.l2_decay),
             # Apply epsilon=1 on ImageNet dataset.
-            epsilon=1)
+            epsilon=1,
+            parameter_list=self.parameter_list)
         return optimizer
 
     def default_decay(self):
@@ -288,12 +296,13 @@ class Optimizer(object):
         optimizer = fluid.optimizer.Momentum(
             learning_rate=self.lr,
             momentum=self.momentum_rate,
-            regularization=fluid.regularizer.L2Decay(self.l2_decay))
+            regularization=fluid.regularizer.L2Decay(self.l2_decay),
+            parameter_list=self.parameter_list)
         return optimizer
 
 
-def create_optimizer(args):
-    Opt = Optimizer(args)
+def create_optimizer(args, parameter_list):
+    Opt = Optimizer(args, parameter_list)
     optimizer = getattr(Opt, args.lr_strategy)()
 
     return optimizer