update resnet&pp-lcnet

ppcls/arch/backbone/legendary_models/pp_lcnet.py

@@ -17,7 +17,7 @@ from __future__ import absolute_import, division, print_function
 import paddle
 import paddle.nn as nn
 from paddle import ParamAttr
-from paddle.nn import AdaptiveAvgPool2D, BatchNorm, Conv2D, Dropout, Linear
+from paddle.nn import AdaptiveAvgPool2D, BatchNorm2D, Conv2D, Dropout, Linear
 from paddle.regularizer import L2Decay
 from paddle.nn.initializer import KaimingNormal
 from ppcls.arch.backbone.base.theseus_layer import TheseusLayer
@@ -83,7 +83,8 @@ class ConvBNLayer(TheseusLayer):
                  filter_size,
                  num_filters,
                  stride,
-                 num_groups=1):
+                 num_groups=1,
+                 lr_mult=1.0):
         super().__init__()
 
         self.conv = Conv2D(
@@ -93,13 +94,13 @@ class ConvBNLayer(TheseusLayer):
             stride=stride,
             padding=(filter_size - 1) // 2,
             groups=num_groups,
-            weight_attr=ParamAttr(initializer=KaimingNormal()),
+            weight_attr=ParamAttr(initializer=KaimingNormal(), learning_rate=lr_mult),
             bias_attr=False)
 
-        self.bn = BatchNorm(
+        self.bn = BatchNorm2D(
             num_filters,
-            param_attr=ParamAttr(regularizer=L2Decay(0.0)),
-            bias_attr=ParamAttr(regularizer=L2Decay(0.0)))
+            weight_attr=ParamAttr(regularizer=L2Decay(0.0), learning_rate=lr_mult),
+            bias_attr=ParamAttr(regularizer=L2Decay(0.0), learning_rate=lr_mult))
         self.hardswish = nn.Hardswish()
 
     def forward(self, x):
@@ -115,7 +116,8 @@ class DepthwiseSeparable(TheseusLayer):
                  num_filters,
                  stride,
                  dw_size=3,
-                 use_se=False):
+                 use_se=False,
+                 lr_mult=1.0):
         super().__init__()
         self.use_se = use_se
         self.dw_conv = ConvBNLayer(
@@ -123,14 +125,17 @@ class DepthwiseSeparable(TheseusLayer):
             num_filters=num_channels,
             filter_size=dw_size,
             stride=stride,
-            num_groups=num_channels)
+            num_groups=num_channels,
+            lr_mult=lr_mult)
         if use_se:
-            self.se = SEModule(num_channels)
+            self.se = SEModule(num_channels,
+                               lr_mult=lr_mult)
         self.pw_conv = ConvBNLayer(
             num_channels=num_channels,
             filter_size=1,
             num_filters=num_filters,
-            stride=1)
+            stride=1,
+            lr_mult=lr_mult)
 
     def forward(self, x):
         x = self.dw_conv(x)
@@ -141,7 +146,7 @@ class DepthwiseSeparable(TheseusLayer):
 
 
 class SEModule(TheseusLayer):
-    def __init__(self, channel, reduction=4):
+    def __init__(self, channel, reduction=4, lr_mult=1.0):
         super().__init__()
         self.avg_pool = AdaptiveAvgPool2D(1)
         self.conv1 = Conv2D(
@@ -149,14 +154,18 @@ class SEModule(TheseusLayer):
             out_channels=channel // reduction,
             kernel_size=1,
             stride=1,
-            padding=0)
+            padding=0,
+            weight_attr=ParamAttr(learning_rate=lr_mult),
+            bias_attr=ParamAttr(learning_rate=lr_mult))
         self.relu = nn.ReLU()
         self.conv2 = Conv2D(
             in_channels=channel // reduction,
             out_channels=channel,
             kernel_size=1,
             stride=1,
-            padding=0)
+            padding=0,
+            weight_attr=ParamAttr(learning_rate=lr_mult),
+            bias_attr=ParamAttr(learning_rate=lr_mult))
         self.hardsigmoid = nn.Hardsigmoid()
 
     def forward(self, x):
@@ -177,17 +186,44 @@ class PPLCNet(TheseusLayer):
                  class_num=1000,
                  dropout_prob=0.2,
                  class_expand=1280,
+                 lr_mult_list=[1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
+                 stride_list=[2, 2, 2, 2, 2],
+                 use_last_conv=True,
                  return_patterns=None,
                  return_stages=None):
         super().__init__()
         self.scale = scale
         self.class_expand = class_expand
-
+        self.lr_mult_list = lr_mult_list
+        self.use_last_conv = use_last_conv
+        self.stride_list = stride_list
+        self.net_config = NET_CONFIG
+        if isinstance(self.lr_mult_list, str):
+            self.lr_mult_list = eval(self.lr_mult_list)
+
+        assert isinstance(self.lr_mult_list, (
+            list, tuple
+        )), "lr_mult_list should be in (list, tuple) but got {}".format(
+            type(self.lr_mult_list))
+        assert len(self.lr_mult_list
+                   ) == 6, "lr_mult_list length should be 6 but got {}".format(
+                       len(self.lr_mult_list))
+
+        assert isinstance(self.stride_list, (
+            list, tuple
+        )), "stride_list should be in (list, tuple) but got {}".format(
+            type(self.stride_list))
+        assert len(self.stride_list
+                   ) == 5, "stride_list length should be 5 but got {}".format(
+                       len(self.stride_list))
+        for i, stride in enumerate(stride_list[1:]):
+            self.net_config["blocks{}".format(i+3)][0][3] = stride
         self.conv1 = ConvBNLayer(
             num_channels=3,
             filter_size=3,
             num_filters=make_divisible(16 * scale),
-            stride=2)
+            stride=stride_list[0],
+            lr_mult=self.lr_mult_list[0])
 
         self.blocks2 = nn.Sequential(* [
             DepthwiseSeparable(
@@ -195,8 +231,9 @@ class PPLCNet(TheseusLayer):
                 num_filters=make_divisible(out_c * scale),
                 dw_size=k,
                 stride=s,
-                use_se=se)
-            for i, (k, in_c, out_c, s, se) in enumerate(NET_CONFIG["blocks2"])
+                use_se=se,
+                lr_mult=self.lr_mult_list[1])
+            for i, (k, in_c, out_c, s, se) in enumerate(self.net_config["blocks2"])
         ])
 
         self.blocks3 = nn.Sequential(* [
@@ -205,8 +242,9 @@ class PPLCNet(TheseusLayer):
                 num_filters=make_divisible(out_c * scale),
                 dw_size=k,
                 stride=s,
-                use_se=se)
-            for i, (k, in_c, out_c, s, se) in enumerate(NET_CONFIG["blocks3"])
+                use_se=se,
+                lr_mult=self.lr_mult_list[2])
+            for i, (k, in_c, out_c, s, se) in enumerate(self.net_config["blocks3"])
         ])
 
         self.blocks4 = nn.Sequential(* [
@@ -215,8 +253,9 @@ class PPLCNet(TheseusLayer):
                 num_filters=make_divisible(out_c * scale),
                 dw_size=k,
                 stride=s,
-                use_se=se)
-            for i, (k, in_c, out_c, s, se) in enumerate(NET_CONFIG["blocks4"])
+                use_se=se,
+                lr_mult=self.lr_mult_list[3])
+            for i, (k, in_c, out_c, s, se) in enumerate(self.net_config["blocks4"])
         ])
 
         self.blocks5 = nn.Sequential(* [
@@ -225,8 +264,9 @@ class PPLCNet(TheseusLayer):
                 num_filters=make_divisible(out_c * scale),
                 dw_size=k,
                 stride=s,
-                use_se=se)
-            for i, (k, in_c, out_c, s, se) in enumerate(NET_CONFIG["blocks5"])
+                use_se=se,
+                lr_mult=self.lr_mult_list[4])
+            for i, (k, in_c, out_c, s, se) in enumerate(self.net_config["blocks5"])
         ])
 
         self.blocks6 = nn.Sequential(* [
@@ -235,25 +275,26 @@ class PPLCNet(TheseusLayer):
                 num_filters=make_divisible(out_c * scale),
                 dw_size=k,
                 stride=s,
-                use_se=se)
-            for i, (k, in_c, out_c, s, se) in enumerate(NET_CONFIG["blocks6"])
+                use_se=se,
+                lr_mult=self.lr_mult_list[5])
+            for i, (k, in_c, out_c, s, se) in enumerate(self.net_config["blocks6"])
         ])
 
         self.avg_pool = AdaptiveAvgPool2D(1)
-
-        self.last_conv = Conv2D(
-            in_channels=make_divisible(NET_CONFIG["blocks6"][-1][2] * scale),
-            out_channels=self.class_expand,
-            kernel_size=1,
-            stride=1,
-            padding=0,
-            bias_attr=False)
-
-        self.hardswish = nn.Hardswish()
-        self.dropout = Dropout(p=dropout_prob, mode="downscale_in_infer")
+        if self.use_last_conv:
+            self.last_conv = Conv2D(
+                in_channels=make_divisible(self.net_config["blocks6"][-1][2] * scale),
+                out_channels=self.class_expand,
+                kernel_size=1,
+                stride=1,
+                padding=0,
+                bias_attr=False)
+            self.hardswish = nn.Hardswish()
+            self.dropout = Dropout(p=dropout_prob, mode="downscale_in_infer")
+        else:
+            self.last_conv = None
         self.flatten = nn.Flatten(start_axis=1, stop_axis=-1)
-
-        self.fc = Linear(self.class_expand, class_num)
+        self.fc = Linear(self.class_expand if self.use_last_conv else make_divisible(self.net_config["blocks6"][-1][2]), class_num)
 
         super().init_res(
             stages_pattern,
@@ -270,9 +311,10 @@ class PPLCNet(TheseusLayer):
         x = self.blocks6(x)
 
         x = self.avg_pool(x)
-        x = self.last_conv(x)
-        x = self.hardswish(x)
-        x = self.dropout(x)
+        if self.last_conv is not None:
+            x = self.last_conv(x)
+            x = self.hardswish(x)
+            x = self.dropout(x)
         x = self.flatten(x)
         x = self.fc(x)
         return x

ppcls/arch/backbone/legendary_models/resnet.py

@@ -20,9 +20,10 @@ import numpy as np
 import paddle
 from paddle import ParamAttr
 import paddle.nn as nn
-from paddle.nn import Conv2D, BatchNorm, Linear
+from paddle.nn import Conv2D, BatchNorm, Linear, BatchNorm2D
 from paddle.nn import AdaptiveAvgPool2D, MaxPool2D, AvgPool2D
 from paddle.nn.initializer import Uniform
+from paddle.regularizer import L2Decay
 import math
 
 from ppcls.arch.backbone.base.theseus_layer import TheseusLayer
@@ -121,17 +122,21 @@ class ConvBNLayer(TheseusLayer):
         self.is_vd_mode = is_vd_mode
         self.act = act
         self.avg_pool = AvgPool2D(
-            kernel_size=2, stride=2, padding=0, ceil_mode=True)
+            kernel_size=2, stride=stride, padding="SAME", ceil_mode=True)
         self.conv = Conv2D(
             in_channels=num_channels,
             out_channels=num_filters,
             kernel_size=filter_size,
-            stride=stride,
+            stride=1 if is_vd_mode else stride,
             padding=(filter_size - 1) // 2,
             groups=groups,
             weight_attr=ParamAttr(learning_rate=lr_mult),
             bias_attr=False,
             data_format=data_format)
+
+        weight_attr = ParamAttr(learning_rate=lr_mult, trainable=True)
+        bias_attr = ParamAttr(learning_rate=lr_mult, trainable=True)
+
         self.bn = BatchNorm(
             num_filters,
             param_attr=ParamAttr(learning_rate=lr_mult),
@@ -159,7 +164,6 @@ class BottleneckBlock(TheseusLayer):
                  lr_mult=1.0,
                  data_format="NCHW"):
         super().__init__()
-
         self.conv0 = ConvBNLayer(
             num_channels=num_channels,
             num_filters=num_filters,
@@ -188,10 +192,11 @@ class BottleneckBlock(TheseusLayer):
                 num_channels=num_channels,
                 num_filters=num_filters * 4,
                 filter_size=1,
-                stride=stride if if_first else 1,
+                stride=stride,
                 is_vd_mode=False if if_first else True,
                 lr_mult=lr_mult,
                 data_format=data_format)
+
         self.relu = nn.ReLU()
         self.shortcut = shortcut
 
@@ -242,7 +247,7 @@ class BasicBlock(TheseusLayer):
                 num_channels=num_channels,
                 num_filters=num_filters,
                 filter_size=1,
-                stride=stride if if_first else 1,
+                stride=stride,
                 is_vd_mode=False if if_first else True,
                 lr_mult=lr_mult,
                 data_format=data_format)
@@ -281,14 +286,17 @@ class ResNet(TheseusLayer):
                  stem_act="relu",
                  class_num=1000,
                  lr_mult_list=[1.0, 1.0, 1.0, 1.0, 1.0],
+                 stride_list=[2, 2, 2, 2, 2],
                  data_format="NCHW",
                  input_image_channel=3,
                  return_patterns=None,
-                 return_stages=None):
+                 return_stages=None,
+                 **kargs):
         super().__init__()
 
         self.cfg = config
         self.lr_mult_list = lr_mult_list
+        self.stride_list = stride_list
         self.is_vd_mode = version == "vd"
         self.class_num = class_num
         self.num_filters = [64, 128, 256, 512]
@@ -305,14 +313,22 @@ class ResNet(TheseusLayer):
                    ) == 5, "lr_mult_list length should be 5 but got {}".format(
                        len(self.lr_mult_list))
 
+        assert isinstance(self.stride_list, (
+            list, tuple
+        )), "stride_list should be in (list, tuple) but got {}".format(
+            type(self.stride_list))
+        assert len(self.stride_list
+                   ) == 5, "stride_list length should be 5 but got {}".format(
+                       len(self.stride_list))
+
         self.stem_cfg = {
             #num_channels, num_filters, filter_size, stride
-            "vb": [[input_image_channel, 64, 7, 2]],
+            "vb": [[input_image_channel, 64, 7, self.stride_list[0]]],
             "vd":
-            [[input_image_channel, 32, 3, 2], [32, 32, 3, 1], [32, 64, 3, 1]]
+            [[input_image_channel, 32, 3, self.stride_list[0]], [32, 32, 3, 1], [32, 64, 3, 1]]
         }
 
-        self.stem = nn.Sequential(*[
+        self.stem = nn.Sequential(* [
             ConvBNLayer(
                 num_channels=in_c,
                 num_filters=out_c,
@@ -325,7 +341,7 @@ class ResNet(TheseusLayer):
         ])
 
         self.max_pool = MaxPool2D(
-            kernel_size=3, stride=2, padding=1, data_format=data_format)
+            kernel_size=3, stride=stride_list[1], padding=1, data_format=data_format)
         block_list = []
         for block_idx in range(len(self.block_depth)):
             shortcut = False
@@ -334,7 +350,7 @@ class ResNet(TheseusLayer):
                     num_channels=self.num_channels[block_idx] if i == 0 else
                     self.num_filters[block_idx] * self.channels_mult,
                     num_filters=self.num_filters[block_idx],
-                    stride=2 if i == 0 and block_idx != 0 else 1,
+                    stride=self.stride_list[block_idx+1] if i == 0 and block_idx != 0 else 1,
                     shortcut=shortcut,
                     if_first=block_idx == i == 0 if version == "vd" else True,
                     lr_mult=self.lr_mult_list[block_idx + 1],