remove inplace=True and pool_size when using global pooling (#3298)

* remove use_inplace=True and add save_step params

* remove useless params when using global pooling

PaddleCV/image_classification/README.md

@@ -96,9 +96,10 @@ bash run.sh train 模型名
 
 * **data_dir**: 数据存储路径，默认值: "./data/ILSVRC2012/"
 * **model_save_dir**: 模型存储路径，默认值: "output/"
-* **save_param**: params存储路径，默认值: None
 * **pretrained_model**: 加载预训练模型路径，默认值: None
 * **checkpoint**: 加载用于继续训练的检查点（指定具体模型存储路径，如"output/ResNet50/100/"），默认值: None
+* **print_step**: 打印训练信息的batch步数，默认值：10
+* **save_step**: 保存模型的epoch步数，默认值：1
 
 模型类型和超参配置：
 
@@ -134,7 +135,6 @@ bash run.sh train 模型名
 一些开关：
 
 * **use_gpu**: 是否在GPU上运行，默认值: True
-* **use_inplace**: 是否开启inplace显存优化，默认值: True
 * **use_label_smoothing**: 是否对数据进行label smoothing处理，默认值: False
 * **label_smoothing_epsilon**: label_smoothing的epsilon， 默认值:0.2
 * **random_seed**: 随机数种子， 默认值: 1000
@@ -207,7 +207,11 @@ FP16相关内容已经迁移至PaddlePaddle/Fleet 中
 PaddlePaddle/Models ImageClassification 支持自定义数据
 
 1. 组织自定义数据，调整数据读取器以正确的传入数据
-2. 注意更改训练脚本中 --data_dim --total_image 等参数
+2. 注意更改训练脚本中
+--data_dim 类别数为自定义数据类别数
+--total_image 图片数量
+3. 当进行finetune时，
+指定--pretrained_model 加载预训练模型，注意：本模型库提供的是基于ImageNet 1000类数据的预训练模型，当使用不同类别数的数据时，请删除预训练模型中fc_weight 和fc_offset参数
 
 
 ## 已发布模型及其性能

PaddleCV/image_classification/README_en.md

@@ -88,9 +88,10 @@ Environment settings:
 
 * **data_dir**: the data root directory Default: "./data/ILSVRC2012".
 * **model_save_dir**: the directory to save trained model. Default: "output".
-* **save_param**: the path to save params. Default: None.
 * **pretrained_model**: load model path for pretraining. Default: None.
 * **checkpoint**: load the checkpoint path to resume. Default: None.
+* **print_step**: the batch steps interval to print log. Default: 10.
+* **save_step**: the epoch steps interval to save checkpoints. Default:1.
 
 Solver and hyperparameters:
 
@@ -105,7 +106,7 @@ Solver and hyperparameters:
 * **lr**: initialized learning rate. Default: 0.1.
 * **l2_decay**: L2_decay parameter. Default: 1e-4.
 * **momentum_rate**: momentum_rate. Default: 0.9.
-* **step_epochs**: piecewise dacay的decay step, Default: [30,60,90]
+* **step_epochs**: decay step of piecewise step, Default: [30,60,90]
 
 Reader and preprocess:
 
@@ -126,7 +127,6 @@ Reader and preprocess:
 Switch:
 
 * **use_gpu**: whether to use GPU or not. Default: True.
-* **use_inplace**: whether to use inplace memory optimization or not. Default: True.
 * **use_label_smoothing**: whether to use label_smoothing or not. Default:False.
 * **label_smoothing_epsilon**: the label_smoothing_epsilon. Default:0.2.
 * **random_seed**: random seed for debugging, Default: 1000

PaddleCV/image_classification/models/resnet_dist.py → PaddleCV/image_classification/legacy/models/resnet_dist.py

@@ -79,15 +79,14 @@ class DistResNet():
         pool = fluid.layers.pool2d(
             input=conv, pool_size=7, pool_type='avg', global_pooling=True)
         stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0)
-        out = fluid.layers.fc(input=pool,
+        out = fluid.layers.fc(
+            input=pool,
             size=class_dim,
             param_attr=fluid.param_attr.ParamAttr(
-                                  initializer=fluid.initializer.Uniform(-stdv,
-                                                                        stdv),
+                initializer=fluid.initializer.Uniform(-stdv, stdv),
                 regularizer=fluid.regularizer.L2Decay(self.weight_decay)),
             bias_attr=fluid.ParamAttr(
-                                  regularizer=fluid.regularizer.L2Decay(self.weight_decay))
-                              )
+                regularizer=fluid.regularizer.L2Decay(self.weight_decay)))
         return out
 
     def conv_bn_layer(self,
@@ -107,9 +106,12 @@ class DistResNet():
             groups=groups,
             act=None,
             bias_attr=False,
-            param_attr=fluid.ParamAttr(regularizer=fluid.regularizer.L2Decay(self.weight_decay)))
+            param_attr=fluid.ParamAttr(
+                regularizer=fluid.regularizer.L2Decay(self.weight_decay)))
         return fluid.layers.batch_norm(
-                input=conv, act=act, is_test=not self.is_train,
+            input=conv,
+            act=act,
+            is_test=not self.is_train,
             param_attr=fluid.ParamAttr(
                 initializer=fluid.initializer.Constant(bn_init_value),
                 regularizer=None))
@@ -132,9 +134,12 @@ class DistResNet():
             act='relu')
         # NOTE: default bias is 0.0 already
         conv2 = self.conv_bn_layer(
-            input=conv1, num_filters=num_filters * 4, filter_size=1, act=None, bn_init_value=0.0)
+            input=conv1,
+            num_filters=num_filters * 4,
+            filter_size=1,
+            act=None,
+            bn_init_value=0.0)
 
         short = self.shortcut(input, num_filters * 4, stride)
 
         return fluid.layers.elementwise_add(x=short, y=conv2, act='relu')
-

PaddleCV/image_classification/models/__init__.py

@@ -22,7 +22,6 @@ from .resnet_vc import ResNet50_vc, ResNet101_vc, ResNet152_vc
 from .resnet_vd import ResNet50_vd, ResNet101_vd, ResNet152_vd, ResNet200_vd
 from .resnext import ResNeXt50_64x4d, ResNeXt101_64x4d, ResNeXt152_64x4d, ResNeXt50_32x4d, ResNeXt101_32x4d, ResNeXt152_32x4d
 from .resnext_vd import ResNeXt50_vd_64x4d, ResNeXt101_vd_64x4d, ResNeXt152_vd_64x4d, ResNeXt50_vd_32x4d, ResNeXt101_vd_32x4d, ResNeXt152_vd_32x4d
-from .resnet_dist import DistResNet
 from .inception_v4 import InceptionV4
 from .se_resnext import SE_ResNeXt50_32x4d, SE_ResNeXt101_32x4d, SE_ResNeXt152_32x4d
 from .se_resnext_vd import SE_ResNeXt50_32x4d_vd, SE_ResNeXt101_32x4d_vd, SE_154_vd

PaddleCV/image_classification/models/fast_imagenet.py

@@ -29,6 +29,7 @@ import utils
 
 __all__ = ["FastImageNet"]
 
+
 class FastImageNet():
     def __init__(self, layers=50, is_train=True):
         self.layers = layers
@@ -65,8 +66,9 @@ class FastImageNet():
                     stride=2 if i == 0 and block != 0 else 1)
         pool_size = int(img_size / 32)
         pool = fluid.layers.pool2d(
-            input=conv, pool_size=pool_size, pool_type='avg', global_pooling=True)
-        out = fluid.layers.fc(input=pool,
+            input=conv, pool_type='avg', global_pooling=True)
+        out = fluid.layers.fc(
+            input=pool,
             size=class_dim,
             act=None,
             param_attr=fluid.param_attr.ParamAttr(
@@ -93,8 +95,12 @@ class FastImageNet():
             groups=groups,
             act=None,
             bias_attr=False,
-            param_attr=fluid.ParamAttr(regularizer=fluid.regularizer.L2Decay(1e-4)))
-        return fluid.layers.batch_norm(input=conv, act=act, is_test=not self.is_train,
+            param_attr=fluid.ParamAttr(
+                regularizer=fluid.regularizer.L2Decay(1e-4)))
+        return fluid.layers.batch_norm(
+            input=conv,
+            act=act,
+            is_test=not self.is_train,
             param_attr=fluid.param_attr.ParamAttr(
                 initializer=fluid.initializer.Constant(bn_init_value),
                 regularizer=None))
@@ -117,12 +123,17 @@ class FastImageNet():
             act='relu')
         # init bn-weight0
         conv2 = self.conv_bn_layer(
-            input=conv1, num_filters=num_filters * 4, filter_size=1, act=None, bn_init_value=0.0)
+            input=conv1,
+            num_filters=num_filters * 4,
+            filter_size=1,
+            act=None,
+            bn_init_value=0.0)
 
         short = self.shortcut(input, num_filters * 4, stride)
 
         return fluid.layers.elementwise_add(x=short, y=conv2, act='relu')
 
+
 def lr_decay(lrs, epochs, bs, total_image):
     boundaries = []
     values = []
@@ -130,7 +141,7 @@ def lr_decay(lrs, epochs, bs, total_image):
         step = total_image // bs[idx]
         if step * bs[idx] < total_image:
             step += 1
-        ratio = (lrs[idx][1] - lrs[idx][0])*1.0 / (epoch[1] - epoch[0])
+        ratio = (lrs[idx][1] - lrs[idx][0]) * 1.0 / (epoch[1] - epoch[0])
         lr_base = lrs[idx][0]
         for s in range(epoch[0], epoch[1]):
             if boundaries:
@@ -139,7 +150,9 @@ def lr_decay(lrs, epochs, bs, total_image):
                 boundaries = [step]
             lr = lr_base + ratio * (s - epoch[0])
             values.append(lr)
-            print("epoch: [%d], steps: [%d], lr: [%f]" % (s, boundaries[-1], values[-1])) 
+            print("epoch: [%d], steps: [%d], lr: [%f]" %
+                  (s, boundaries[-1], values[-1]))
     values.append(lrs[-1])
-    print("epoch: [%d:], steps: [%d:], lr:[%f]" % (epochs[-1][-1], boundaries[-1], values[-1]))
+    print("epoch: [%d:], steps: [%d:], lr:[%f]" %
+          (epochs[-1][-1], boundaries[-1], values[-1]))
     return boundaries, values

PaddleCV/image_classification/models/inception_v4.py

@@ -45,7 +45,7 @@ class InceptionV4():
             x = self.inceptionC(x, name=str(i + 1))
 
         pool = fluid.layers.pool2d(
-            input=x, pool_size=8, pool_type='avg', global_pooling=True)
+            input=x, pool_type='avg', global_pooling=True)
 
         drop = fluid.layers.dropout(x=pool, dropout_prob=0.2)
 

PaddleCV/image_classification/models/mobilenet.py

@@ -125,11 +125,7 @@ class MobileNet():
             name="conv6")
 
         input = fluid.layers.pool2d(
-            input=input,
-            pool_size=0,
-            pool_stride=1,
-            pool_type='avg',
-            global_pooling=True)
+            input=input, pool_type='avg', global_pooling=True)
 
         output = fluid.layers.fc(input=input,
                                  size=class_dim,

PaddleCV/image_classification/models/mobilenet_v2.py

@@ -89,11 +89,7 @@ class MobileNetV2():
             name='conv9')
 
         input = fluid.layers.pool2d(
-            input=input,
-            pool_size=7,
-            pool_stride=1,
-            pool_type='avg',
-            global_pooling=True)
+            input=input, pool_type='avg', global_pooling=True)
 
         output = fluid.layers.fc(input=input,
                                  size=class_dim,

PaddleCV/image_classification/models/resnet.py

@@ -77,7 +77,7 @@ class ResNet():
                         name=conv_name)
 
             pool = fluid.layers.pool2d(
-                input=conv, pool_size=7, pool_type='avg', global_pooling=True)
+                input=conv, pool_type='avg', global_pooling=True)
             stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0)
             out = fluid.layers.fc(
                 input=pool,
@@ -96,7 +96,7 @@ class ResNet():
                         name=conv_name)
 
             pool = fluid.layers.pool2d(
-                input=conv, pool_size=7, pool_type='avg', global_pooling=True)
+                input=conv, pool_type='avg', global_pooling=True)
             stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0)
             out = fluid.layers.fc(
                 input=pool,

PaddleCV/image_classification/models/resnet_vc.py

@@ -57,11 +57,26 @@ class ResNet():
         num_filters = [64, 128, 256, 512]
 
         conv = self.conv_bn_layer(
-                input=input, num_filters=32, filter_size=3, stride=2, act='relu', name='conv1_1')
+            input=input,
+            num_filters=32,
+            filter_size=3,
+            stride=2,
+            act='relu',
+            name='conv1_1')
         conv = self.conv_bn_layer(
-                input=conv, num_filters=32, filter_size=3, stride=1, act='relu', name='conv1_2')
+            input=conv,
+            num_filters=32,
+            filter_size=3,
+            stride=1,
+            act='relu',
+            name='conv1_2')
         conv = self.conv_bn_layer(
-                input=conv, num_filters=64, filter_size=3, stride=1, act='relu', name='conv1_3')
+            input=conv,
+            num_filters=64,
+            filter_size=3,
+            stride=1,
+            act='relu',
+            name='conv1_3')
 
         conv = fluid.layers.pool2d(
             input=conv,
@@ -74,18 +89,19 @@ class ResNet():
             for i in range(depth[block]):
                 if layers in [101, 152] and block == 2:
                     if i == 0:
-                        conv_name="res"+str(block+2)+"a"
+                        conv_name = "res" + str(block + 2) + "a"
                     else:
-                        conv_name="res"+str(block+2)+"b"+str(i)
+                        conv_name = "res" + str(block + 2) + "b" + str(i)
                 else:
-                    conv_name="res"+str(block+2)+chr(97+i)
+                    conv_name = "res" + str(block + 2) + chr(97 + i)
                 conv = self.bottleneck_block(
                     input=conv,
                     num_filters=num_filters[block],
-                    stride=2 if i == 0 and block != 0 else 1,name=conv_name)
+                    stride=2 if i == 0 and block != 0 else 1,
+                    name=conv_name)
 
         pool = fluid.layers.pool2d(
-            input=conv, pool_size=7, pool_type='avg', global_pooling=True)
+            input=conv, pool_type='avg', global_pooling=True)
         stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0)
         out = fluid.layers.fc(input=pool,
                               size=class_dim,
@@ -117,13 +133,14 @@ class ResNet():
             bn_name = "bn_" + name
         else:
             bn_name = "bn" + name[3:]
-        return fluid.layers.batch_norm(input=conv, 
+        return fluid.layers.batch_norm(
+            input=conv,
             act=act,
-                                       name=bn_name+'.output.1',
+            name=bn_name + '.output.1',
             param_attr=ParamAttr(name=bn_name + '_scale'),
             bias_attr=ParamAttr(bn_name + '_offset'),
             moving_mean_name=bn_name + '_mean',
-                                       moving_variance_name=bn_name + '_variance',)
+            moving_variance_name=bn_name + '_variance', )
 
     def shortcut(self, input, ch_out, stride, name):
         ch_in = input.shape[1]
@@ -134,20 +151,30 @@ class ResNet():
 
     def bottleneck_block(self, input, num_filters, stride, name):
         conv0 = self.conv_bn_layer(
-            input=input, num_filters=num_filters, filter_size=1, act='relu',name=name+"_branch2a")
+            input=input,
+            num_filters=num_filters,
+            filter_size=1,
+            act='relu',
+            name=name + "_branch2a")
         conv1 = self.conv_bn_layer(
             input=conv0,
             num_filters=num_filters,
             filter_size=3,
             stride=stride,
             act='relu',
-        name=name+"_branch2b")
+            name=name + "_branch2b")
         conv2 = self.conv_bn_layer(
-            input=conv1, num_filters=num_filters * 4, filter_size=1, act=None, name=name+"_branch2c")
+            input=conv1,
+            num_filters=num_filters * 4,
+            filter_size=1,
+            act=None,
+            name=name + "_branch2c")
 
-        short = self.shortcut(input, num_filters * 4, stride, name=name + "_branch1")
+        short = self.shortcut(
+            input, num_filters * 4, stride, name=name + "_branch1")
 
-        return fluid.layers.elementwise_add(x=short, y=conv2, act='relu',name=name+".add.output.5")
+        return fluid.layers.elementwise_add(
+            x=short, y=conv2, act='relu', name=name + ".add.output.5")
 
 
 def ResNet50_vc():

PaddleCV/image_classification/models/resnet_vd.py

@@ -102,7 +102,7 @@ class ResNet():
                     name=conv_name)
 
         pool = fluid.layers.pool2d(
-            input=conv, pool_size=7, pool_type='avg', global_pooling=True)
+            input=conv, pool_type='avg', global_pooling=True)
         stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0)
 
         out = fluid.layers.fc(

PaddleCV/image_classification/models/resnext.py

@@ -82,7 +82,7 @@ class ResNeXt():
                     name=conv_name)
 
         pool = fluid.layers.pool2d(
-            input=conv, pool_size=7, pool_type='avg', global_pooling=True)
+            input=conv, pool_type='avg', global_pooling=True)
         stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0)
         out = fluid.layers.fc(
             input=pool,

PaddleCV/image_classification/models/resnext101_wsl.py

@@ -65,7 +65,7 @@ class ResNeXt101_wsl():
                     name=conv_name)
 
         pool = fluid.layers.pool2d(
-            input=conv, pool_size=7, pool_type='avg', global_pooling=True)
+            input=conv, pool_type='avg', global_pooling=True)
         stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0)
         out = fluid.layers.fc(
             input=pool,

PaddleCV/image_classification/models/resnext_vd.py

@@ -102,7 +102,7 @@ class ResNeXt():
                     name=conv_name)
 
         pool = fluid.layers.pool2d(
-            input=conv, pool_size=7, pool_type='avg', global_pooling=True)
+            input=conv, pool_type='avg', global_pooling=True)
         stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0)
         out = fluid.layers.fc(
             input=pool,

PaddleCV/image_classification/models/se_resnext.py

@@ -120,11 +120,7 @@ class SE_ResNeXt():
                     name=str(n) + '_' + str(i + 1))
 
         pool = fluid.layers.pool2d(
-            input=conv,
-            pool_size=7,
-            pool_type='avg',
-            global_pooling=True,
-            use_cudnn=False)
+            input=conv, pool_type='avg', global_pooling=True, use_cudnn=False)
         drop = fluid.layers.dropout(x=pool, dropout_prob=0.5)
         stdv = 1.0 / math.sqrt(drop.shape[1] * 1.0)
         out = fluid.layers.fc(
@@ -215,11 +211,7 @@ class SE_ResNeXt():
                            reduction_ratio,
                            name=None):
         pool = fluid.layers.pool2d(
-            input=input,
-            pool_size=0,
-            pool_type='avg',
-            global_pooling=True,
-            use_cudnn=False)
+            input=input, pool_type='avg', global_pooling=True, use_cudnn=False)
         stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0)
         squeeze = fluid.layers.fc(
             input=pool,

PaddleCV/image_classification/models/se_resnext_vd.py

@@ -146,7 +146,7 @@ class SE_ResNeXt():
                     name=str(n) + '_' + str(i + 1))
 
         pool = fluid.layers.pool2d(
-            input=conv, pool_size=7, pool_type='avg', global_pooling=True)
+            input=conv, pool_type='avg', global_pooling=True)
         if layers == 152:
             pool = fluid.layers.dropout(x=pool, dropout_prob=0.2)
         stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0)
@@ -289,7 +289,7 @@ class SE_ResNeXt():
                            reduction_ratio,
                            name=None):
         pool = fluid.layers.pool2d(
-            input=input, pool_size=0, pool_type='avg', global_pooling=True)
+            input=input, pool_type='avg', global_pooling=True)
         stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0)
         squeeze = fluid.layers.fc(
             input=pool,

PaddleCV/image_classification/train.py

@@ -100,10 +100,8 @@ def train(args):
         args=args)
     train_py_reader = train_out[-1]
     train_fetch_vars = train_out[:-1]
-    train_fetch_list = []
-    for var in train_fetch_vars:
-        var.persistable = True
-        train_fetch_list.append(var.name)
+
+    train_fetch_list = [var.name for var in train_fetch_vars]
 
     test_out = build_program(
         is_train=False,
@@ -112,10 +110,8 @@ def train(args):
         args=args)
     test_py_reader = test_out[-1]
     test_fetch_vars = test_out[:-1]
-    test_fetch_list = []
-    for var in test_fetch_vars:
-        var.persistable = True
-        test_fetch_list.append(var.name)
+
+    test_fetch_list = [var.name for var in test_fetch_vars]
 
     #Create test_prog and set layers' is_test params to True
     test_prog = test_prog.clone(for_test=True)
@@ -208,6 +204,7 @@ def train(args):
                    list(train_epoch_metrics_avg) + list(test_epoch_metrics_avg),
                    0, "epoch")
         #For now, save model per epoch.
+        if pass_id % args.save_step == 0:
             save_model(args, exe, train_prog, pass_id)
 
 

PaddleCV/image_classification/utils/utility.py

@@ -88,10 +88,8 @@ def parse_args():
     add_arg('data_dir',                 str,    "./data/ILSVRC2012/",   "The ImageNet dataset root directory.")
     add_arg('pretrained_model',         str,    None,                   "Whether to load pretrained model.")
     add_arg('checkpoint',               str,    None,                   "Whether to resume checkpoint.")
-    add_arg('save_params',              str,    "./output",                   "Whether to save params.")
-
     add_arg('print_step',               int,    10,                     "The steps interval to print logs")
-    add_arg('save_step',                int,    100,                    "The steps interval to save checkpoints")
+    add_arg('save_step',                int,    1,                      "The steps interval to save checkpoints")
 
     # SOLVER AND HYPERPARAMETERS
     add_arg('model',                    str,    "ResNet50",   "The name of network.")
@@ -121,7 +119,6 @@ def parse_args():
     parser.add_argument('--image_std', nargs='+', type=float, default=[0.229, 0.224, 0.225], help="The std of input image data")
 
     # SWITCH
-    add_arg('use_inplace',              bool,   True,                   "Whether to use inplace memory optimization.")
     #NOTE: (2019/08/08) FP16 is moving to PaddlePaddle/Fleet now
     #add_arg('use_fp16',                 bool,   False,                  "Whether to enable half precision training with fp16." )
     #add_arg('scale_loss',               float,  1.0,                    "The value of scale_loss for fp16." )
@@ -371,7 +368,7 @@ def best_strategy_compiled(args, program, loss):
         return program
     else:
         build_strategy = fluid.compiler.BuildStrategy()
-        build_strategy.enable_inplace = args.use_inplace
+        build_strategy.enable_inplace = True
 
         exec_strategy = fluid.ExecutionStrategy()
         exec_strategy.num_threads = fluid.core.get_cuda_device_count()