watermark

ppcls/arch/backbone/__init__.py

@@ -79,6 +79,7 @@ from .variant_models.resnet_variant import ResNet50_adaptive_max_pool2d
 from .variant_models.vgg_variant import VGG19Sigmoid
 from .variant_models.pp_lcnet_variant import PPLCNet_x2_5_Tanh
 from .variant_models.pp_lcnetv2_variant import PPLCNetV2_base_ShiTu
+from .variant_models.efficientnet_variant import EfficientNetB3_watermark
 from .model_zoo.adaface_ir_net import AdaFace_IR_18, AdaFace_IR_34, AdaFace_IR_50, AdaFace_IR_101, AdaFace_IR_152, AdaFace_IR_SE_50, AdaFace_IR_SE_101, AdaFace_IR_SE_152, AdaFace_IR_SE_200
 from .model_zoo.wideresnet import WideResNet
 from .model_zoo.uniformer import UniFormer_small, UniFormer_small_plus, UniFormer_small_plus_dim64, UniFormer_base, UniFormer_base_ls

ppcls/arch/backbone/model_zoo/efficientnet.py

@@ -26,6 +26,7 @@ import collections
 import re
 import copy
 
+from ..base.theseus_layer import TheseusLayer
 from ....utils.save_load import load_dygraph_pretrain, load_dygraph_pretrain_from_url
 
 MODEL_URLS = {
@@ -289,7 +290,7 @@ def _drop_connect(inputs, prob, is_test):
     return output
 
 
-class Conv2ds(nn.Layer):
+class Conv2ds(TheseusLayer):
     def __init__(self,
                  input_channels,
                  output_channels,
@@ -361,13 +362,14 @@ class Conv2ds(nn.Layer):
         return x
 
 
-class ConvBNLayer(nn.Layer):
+class ConvBNLayer(TheseusLayer):
     def __init__(self,
                  input_channels,
                  filter_size,
                  output_channels,
                  stride=1,
                  num_groups=1,
+                 global_params=None,
                  padding_type="SAME",
                  conv_act=None,
                  bn_act="swish",
@@ -396,12 +398,13 @@ class ConvBNLayer(nn.Layer):
         if use_bn is True:
             bn_name = name + bn_name
             param_attr, bias_attr = init_batch_norm_layer(bn_name)
+            epsilon = global_params.batch_norm_epsilon
 
             self._bn = BatchNorm(
                 num_channels=output_channels,
                 act=bn_act,
                 momentum=0.99,
-                epsilon=0.001,
+                epsilon=epsilon,
                 moving_mean_name=bn_name + "_mean",
                 moving_variance_name=bn_name + "_variance",
                 param_attr=param_attr,
@@ -416,10 +419,11 @@ class ConvBNLayer(nn.Layer):
             return self._conv(inputs)
 
 
-class ExpandConvNorm(nn.Layer):
+class ExpandConvNorm(TheseusLayer):
     def __init__(self,
                  input_channels,
                  block_args,
+                 global_params,
                  padding_type,
                  name=None,
                  model_name=None,
@@ -434,6 +438,7 @@ class ExpandConvNorm(nn.Layer):
                 input_channels,
                 1,
                 self.oup,
+                global_params=global_params,
                 bn_act=None,
                 padding_type=padding_type,
                 name=name,
@@ -449,10 +454,11 @@ class ExpandConvNorm(nn.Layer):
             return inputs
 
 
-class DepthwiseConvNorm(nn.Layer):
+class DepthwiseConvNorm(TheseusLayer):
     def __init__(self,
                  input_channels,
                  block_args,
+                 global_params,
                  padding_type,
                  name=None,
                  model_name=None,
@@ -471,6 +477,7 @@ class DepthwiseConvNorm(nn.Layer):
             oup,
             self.s,
             num_groups=input_channels,
+            global_params=global_params,
             bn_act=None,
             padding_type=padding_type,
             name=name,
@@ -483,10 +490,11 @@ class DepthwiseConvNorm(nn.Layer):
         return self._conv(inputs)
 
 
-class ProjectConvNorm(nn.Layer):
+class ProjectConvNorm(TheseusLayer):
     def __init__(self,
                  input_channels,
                  block_args,
+                 global_params,
                  padding_type,
                  name=None,
                  model_name=None,
@@ -499,6 +507,7 @@ class ProjectConvNorm(nn.Layer):
             input_channels,
             1,
             final_oup,
+            global_params=global_params,
             bn_act=None,
             padding_type=padding_type,
             name=name,
@@ -511,7 +520,7 @@ class ProjectConvNorm(nn.Layer):
         return self._conv(inputs)
 
 
-class SEBlock(nn.Layer):
+class SEBlock(TheseusLayer):
     def __init__(self,
                  input_channels,
                  num_squeezed_channels,
@@ -549,10 +558,11 @@ class SEBlock(nn.Layer):
         return out
 
 
-class MbConvBlock(nn.Layer):
+class MbConvBlock(TheseusLayer):
     def __init__(self,
                  input_channels,
                  block_args,
+                 global_params,
                  padding_type,
                  use_se,
                  name=None,
@@ -573,6 +583,7 @@ class MbConvBlock(nn.Layer):
             self._ecn = ExpandConvNorm(
                 input_channels,
                 block_args,
+                global_params,
                 padding_type=padding_type,
                 name=name,
                 model_name=model_name,
@@ -581,6 +592,7 @@ class MbConvBlock(nn.Layer):
         self._dcn = DepthwiseConvNorm(
             input_channels * block_args.expand_ratio,
             block_args,
+            global_params,
             padding_type=padding_type,
             name=name,
             model_name=model_name,
@@ -601,6 +613,7 @@ class MbConvBlock(nn.Layer):
         self._pcn = ProjectConvNorm(
             input_channels * block_args.expand_ratio,
             block_args,
+            global_params,
             padding_type=padding_type,
             name=name,
             model_name=model_name,
@@ -627,7 +640,7 @@ class MbConvBlock(nn.Layer):
         return x
 
 
-class ConvStemNorm(nn.Layer):
+class ConvStemNorm(TheseusLayer):
     def __init__(self,
                  input_channels,
                  padding_type,
@@ -643,6 +656,7 @@ class ConvStemNorm(nn.Layer):
             filter_size=3,
             output_channels=output_channels,
             stride=2,
+            global_params=_global_params,
             bn_act=None,
             padding_type=padding_type,
             name="",
@@ -655,7 +669,7 @@ class ConvStemNorm(nn.Layer):
         return self._conv(inputs)
 
 
-class ExtractFeatures(nn.Layer):
+class ExtractFeatures(TheseusLayer):
     def __init__(self,
                  input_channels,
                  _block_args,
@@ -708,6 +722,7 @@ class ExtractFeatures(nn.Layer):
                 MbConvBlock(
                     block_args.input_filters,
                     block_args=block_args,
+                    global_params=_global_params,
                     padding_type=padding_type,
                     use_se=use_se,
                     name="_blocks." + str(idx) + ".",
@@ -728,6 +743,7 @@ class ExtractFeatures(nn.Layer):
                     MbConvBlock(
                         block_args.input_filters,
                         block_args,
+                        global_params=_global_params,
                         padding_type=padding_type,
                         use_se=use_se,
                         name="_blocks." + str(idx) + ".",
@@ -746,7 +762,7 @@ class ExtractFeatures(nn.Layer):
         return x
 
 
-class EfficientNet(nn.Layer):
+class EfficientNet(TheseusLayer):
     def __init__(self,
                  name="b0",
                  padding_type="SAME",
@@ -789,6 +805,7 @@ class EfficientNet(nn.Layer):
             oup,
             1,
             output_channels,
+            global_params=self._global_params,
             bn_act="swish",
             padding_type=self.padding_type,
             name="",

ppcls/arch/backbone/variant_models/efficientnet_variant.py

+import paddle
+import paddle.nn as nn
+from paddle.nn import Sigmoid
+from paddle.nn import Tanh
+from ..model_zoo.efficientnet import EfficientNetB3, _load_pretrained
+
+MODEL_URLS = {
+    "EfficientNetB3_watermark":
+    "https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/EfficientNetB3_watermark_pretrained.pdparams"
+}
+
+__all__ = list(MODEL_URLS.keys())
+
+
+def EfficientNetB3_watermark(padding_type='DYNAMIC',
+                             override_params={"batch_norm_epsilon": 0.00001},
+                             use_se=True,
+                             pretrained=False,
+                             use_ssld=False,
+                             **kwargs):
+    def replace_function(_fc, pattern):
+        classifier = nn.Sequential(
+            # 1536 is the orginal in_features
+            nn.Linear(
+                in_features=1536, out_features=625),
+            nn.ReLU(),  # ReLu to be the activation function
+            nn.Dropout(p=0.3),
+            nn.Linear(
+                in_features=625, out_features=256),
+            nn.ReLU(),
+            nn.Linear(
+                in_features=256, out_features=2), )
+        return classifier
+
+    pattern = "_fc"
+    model = EfficientNetB3(
+        padding_type=padding_type,
+        override_params=override_params,
+        use_se=True,
+        pretrained=False,
+        use_ssld=False,
+        **kwargs)
+    model.upgrade_sublayer(pattern, replace_function)
+    _load_pretrained(pretrained, model, MODEL_URLS["EfficientNetB3_watermark"],
+                     use_ssld)
+    return model