Fix inference bug and release EfficientNetB0_small in image classification (#3667)

* fix_infer_bug_release_some_model

PaddleCV/image_classification/README.md

@@ -164,6 +164,7 @@ CUDA_VISIBLE_DEVICES=0,1,2,3 python -m paddle.distributed.launch train.py \
 * **label_smoothing_epsilon**: label_smoothing的epsilon， 默认值:0.1
 * **random_seed**: 随机数种子， 默认值: 1000
 * **padding_type**: efficientNet中卷积操作的padding方式, 默认值: "SAME".
+* **use_se**: efficientNet中是否使用Squeeze-and-Excitation模块, 默认值: True.
 * **use_ema**: 是否在更新模型参数时使用ExponentialMovingAverage. 默认值: False.
 * **ema_decay**: ExponentialMovingAverage的decay rate. 默认值: 0.9999.
 
@@ -416,9 +417,12 @@ PaddlePaddle/Models ImageClassification 支持自定义数据
 |[EfficientNetB5](https://paddle-imagenet-models-name.bj.bcebos.com/EfficientNetB5_pretrained.tar)<sup>[1](#trans)</sup> | 83.62% | 96.72% | 88.578 | 32.102 |
 |[EfficientNetB6](https://paddle-imagenet-models-name.bj.bcebos.com/EfficientNetB6_pretrained.tar)<sup>[1](#trans)</sup> | 84.00% | 96.88% | 138.670 | 51.059 |
 |[EfficientNetB7](https://paddle-imagenet-models-name.bj.bcebos.com/EfficientNetB7_pretrained.tar)<sup>[1](#trans)</sup> | 84.30% | 96.89% | 234.364 | 82.107 |
+|[EfficientNetB0_small](https://paddle-imagenet-models-name.bj.bcebos.com/EfficientNetB0_Small_pretrained.tar)<sup>[2](#trans)</sup> | 75.80% | 92.58% | 3.342 | 2.729 |
 
 <a name="trans">[1]</a> 表示该预训练权重是由[官方的代码仓库](https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet)转换来的。
 
+<a name="trans">[2]</a> 表示该预训练权重是在EfficientNetB0的基础上去除se模块，并使用通用的卷积训练的，精度稍稍下降，但是速度大幅提升。
+
 ## FAQ
 
 **Q:** 加载预训练模型报错，Enforce failed. Expected x_dims[1] == labels_dims[1], but received x_dims[1]:1000 != labels_dims[1]:6.

PaddleCV/image_classification/README_en.md

@@ -157,6 +157,7 @@ Switch:
 * **label_smoothing_epsilon**: the label_smoothing_epsilon. Default:0.1.
 * **random_seed**: random seed for debugging, Default: 1000.
 * **padding_type**: padding type of convolution for efficientNet, Default: "SAME".
+* **use_se**: whether to use Squeeze-and-Excitation module in efficientNet, Default: True.
 * **use_ema**: whether to use ExponentialMovingAverage or not. Default: False.
 * **ema_decay**: the value of ExponentialMovingAverage decay rate. Default: 0.9999.
 
@@ -399,9 +400,12 @@ Pretrained models can be downloaded by clicking related model names.
 |[EfficientNetB5](https://paddle-imagenet-models-name.bj.bcebos.com/EfficientNetB5_pretrained.tar)<sup>[1](#trans)</sup> | 83.62% | 96.72% | 88.578 | 32.102 |
 |[EfficientNetB6](https://paddle-imagenet-models-name.bj.bcebos.com/EfficientNetB6_pretrained.tar)<sup>[1](#trans)</sup> | 84.00% | 96.88% | 138.670 | 51.059 |
 |[EfficientNetB7](https://paddle-imagenet-models-name.bj.bcebos.com/EfficientNetB7_pretrained.tar)<sup>[1](#trans)</sup> | 84.30% | 96.89% | 234.364 | 82.107 |
+|[EfficientNetB0_small](https://paddle-imagenet-models-name.bj.bcebos.com/EfficientNetB0_Small_pretrained.tar)<sup>[2](#trans)</sup> | 75.80% | 92.58% | 3.342 | 2.729 |
 
 <a name="trans">[1]</a> means the pretrained weight is converted form [original repository](https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet).
 
+<a name="trans">[2]</a> means the pretrained weight is based on EfficientNetB0, removed Squeeze-and-Excitation module and use general convolution. This model speed is much faster.
+
 ## FAQ
 
 **Q:** How to solve this problem when I try to train a 6-classes dataset with indicating pretrained_model parameter ?

PaddleCV/image_classification/eval.py

@@ -48,6 +48,7 @@ parser.add_argument('--image_std', nargs='+', type=float, default=[0.229, 0.224,
 add_arg('crop_size',        int,  224,                  "The value of crop size")
 add_arg('interpolation',    int,  None,                 "The interpolation mode")
 add_arg('padding_type',     str,  "SAME",               "Padding type of convolution")
+add_arg('use_se',           bool, True,                 "Whether to use Squeeze-and-Excitation module for EfficientNet.")
 # yapf: enable
 
 
@@ -67,7 +68,8 @@ def eval(args):
 
     # model definition
     if args.model.startswith('EfficientNet'):
-        model = models.__dict__[args.model](is_test=True, padding_type=args.padding_type)
+        model = models.__dict__[args.model](is_test=True, padding_type=args.padding_type,
+                                            use_se=args.use_se)
     else:
         model = models.__dict__[args.model]()
 

PaddleCV/image_classification/models/efficientnet.py

@@ -103,7 +103,7 @@ def round_repeats(repeats, global_params):
 
 
 class EfficientNet():
-    def __init__(self, name='b0', padding_type='SAME', override_params=None, is_test=False):
+    def __init__(self, name='b0', padding_type='SAME', override_params=None, is_test=False, use_se=True):
         valid_names = ['b' + str(i) for i in range(8)]
         assert name in valid_names, 'efficient name should be in b0~b7'
         model_name = 'efficientnet-' + name
@@ -112,6 +112,7 @@ class EfficientNet():
         self._bn_eps = self._global_params.batch_norm_epsilon
         self.is_test = is_test
         self.padding_type = padding_type
+        self.use_se = use_se
 
     def net(self, input, class_dim=1000, is_test=False):
 
@@ -243,7 +244,7 @@ class EfficientNet():
     def mb_conv_block(self, inputs, block_args, is_test=False, drop_connect_rate=None, name=None):
         # Expansion and Depthwise Convolution
         oup = block_args.input_filters * block_args.expand_ratio  # number of output channels
-        has_se = (block_args.se_ratio is not None) and (0 < block_args.se_ratio <= 1)
+        has_se = self.use_se and (block_args.se_ratio is not None) and (0 < block_args.se_ratio <= 1)
         id_skip = block_args.id_skip  # skip connection and drop connect
         conv = inputs
         if block_args.expand_ratio != 1:
@@ -413,41 +414,41 @@ class BlockDecoder(object):
         return block_strings
 
 
-def EfficientNetB0(is_test=False, padding_type='SAME', override_params=None):
-    model = EfficientNet(name='b0', is_test=is_test, padding_type=padding_type, override_params=override_params)
+def EfficientNetB0(is_test=False, padding_type='SAME', override_params=None, use_se=True):
+    model = EfficientNet(name='b0', is_test=is_test, padding_type=padding_type, override_params=override_params, use_se=use_se)
     return model
 
 
-def EfficientNetB1(is_test=False, padding_type='SAME', override_params=None):
-    model = EfficientNet(name='b1', is_test=is_test, padding_type=padding_type, override_params=override_params)
+def EfficientNetB1(is_test=False, padding_type='SAME', override_params=None, use_se=True):
+    model = EfficientNet(name='b1', is_test=is_test, padding_type=padding_type, override_params=override_params, use_se=use_se)
     return model
 
 
-def EfficientNetB2(is_test=False, padding_type='SAME', override_params=None):
-    model = EfficientNet(name='b2', is_test=is_test, padding_type=padding_type, override_params=override_params)
+def EfficientNetB2(is_test=False, padding_type='SAME', override_params=None, use_se=True):
+    model = EfficientNet(name='b2', is_test=is_test, padding_type=padding_type, override_params=override_params, use_se=use_se)
     return model
 
 
-def EfficientNetB3(is_test=False, padding_type='SAME', override_params=None):
-    model = EfficientNet(name='b3', is_test=is_test, padding_type=padding_type, override_params=override_params)
+def EfficientNetB3(is_test=False, padding_type='SAME', override_params=None, use_se=True):
+    model = EfficientNet(name='b3', is_test=is_test, padding_type=padding_type, override_params=override_params, use_se=use_se)
     return model
 
 
-def EfficientNetB4(is_test=False, padding_type='SAME', override_params=None):
-    model = EfficientNet(name='b4', is_test=is_test, padding_type=padding_type, override_params=override_params)
+def EfficientNetB4(is_test=False, padding_type='SAME', override_params=None, use_se=True):
+    model = EfficientNet(name='b4', is_test=is_test, padding_type=padding_type, override_params=override_params, use_se=use_se)
     return model
 
 
-def EfficientNetB5(is_test=False, padding_type='SAME', override_params=None):
-    model = EfficientNet(name='b5', is_test=is_test, padding_type=padding_type, override_params=override_params)
+def EfficientNetB5(is_test=False, padding_type='SAME', override_params=None, use_se=True):
+    model = EfficientNet(name='b5', is_test=is_test, padding_type=padding_type, override_params=override_params, use_se=use_se)
     return model
 
 
-def EfficientNetB6(is_test=False, padding_type='SAME', override_params=None):
-    model = EfficientNet(name='b6', is_test=is_test, padding_type=padding_type, override_params=override_params)
+def EfficientNetB6(is_test=False, padding_type='SAME', override_params=None, use_se=True):
+    model = EfficientNet(name='b6', is_test=is_test, padding_type=padding_type, override_params=override_params, use_se=use_se)
     return model
 
 
-def EfficientNetB7(is_test=False, padding_type='SAME', override_params=None):
-    model = EfficientNet(name='b7', is_test=is_test, padding_type=padding_type, override_params=override_params)
+def EfficientNetB7(is_test=False, padding_type='SAME', override_params=None, use_se=True):
+    model = EfficientNet(name='b7', is_test=is_test, padding_type=padding_type, override_params=override_params, use_se=use_se)
     return model
\ No newline at end of file

PaddleCV/image_classification/reader.py

@@ -265,7 +265,10 @@ class ImageNetReader:
                         rotate=False,
                         data_dir=None):
         num_trainers = int(os.environ.get('PADDLE_TRAINERS_NUM', 1))
-        batch_size = settings.batch_size / paddle.fluid.core.get_cuda_device_count()
+        if mode == 'test':
+            batch_size = 1
+        else:
+            batch_size = settings.batch_size / paddle.fluid.core.get_cuda_device_count()
         def reader():
             def read_file_list():
                 with open(file_list) as flist:

PaddleCV/image_classification/train.py

@@ -59,7 +59,9 @@ def build_program(is_train, main_prog, startup_prog, args):
         is_test = False if is_train else True
         override_params = {"drop_connect_rate": args.drop_connect_rate}
         padding_type = args.padding_type
-        model = models.__dict__[args.model](is_test=is_test, override_params=override_params, padding_type=padding_type)
+        use_se = args.use_se
+        model = models.__dict__[args.model](is_test=is_test, override_params=override_params,
+                                            padding_type=padding_type, use_se=use_se)
     else:
         model = models.__dict__[args.model]()
     with fluid.program_guard(main_prog, startup_prog):

PaddleCV/image_classification/utils/utility.py

@@ -139,6 +139,7 @@ def parse_args():
     add_arg('use_ema',                  bool,   False,                  "Whether to use ExponentialMovingAverage.")
     add_arg('ema_decay',                float,  0.9999,                 "The value of ema decay rate")
     add_arg('padding_type',             str,    "SAME",                 "Padding type of convolution")
+    add_arg('use_se',                   bool,   True,                   "Whether to use Squeeze-and-Excitation module for EfficientNet.")
     # yapf: enable
 
     args = parser.parse_args()