mirrors
/
PaddleClas
mirror of https://github.com/PaddlePaddle/PaddleClas.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

refactor

pull/1916/head
gaotingquan 2022-05-12 14:12:12 +00:00
parent 0b03606471
commit 4c5ed46b2c
No known key found for this signature in database
GPG Key ID: F3EF7F42536A30B7
1 changed files with 35 additions and 79 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

114
ppcls/arch/backbone/legendary_models/pp_lcnet_v2.py
View File

@ -24,7 +24,6 @@ from paddle.nn.initializer import KaimingNormal
from ppcls.arch.backbone.base.theseus_layer import TheseusLayer from ppcls.arch.backbone.base.theseus_layer import TheseusLayer
from ppcls.utils.save_load import load_dygraph_pretrain, load_dygraph_pretrain_from_url from ppcls.utils.save_load import load_dygraph_pretrain, load_dygraph_pretrain_from_url
#TODO(gaotingquan): upload pretrained to bos
MODEL_URLS = { MODEL_URLS = {
"PPLCNetV2_base": "PPLCNetV2_base":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/legendary_models/PPLCNetV2_base_pretrained.pdparams", "https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/legendary_models/PPLCNetV2_base_pretrained.pdparams",
@ -252,14 +251,16 @@ class RepDepthwiseSeparable(TheseusLayer):
class PPLCNetV2(TheseusLayer): class PPLCNetV2(TheseusLayer):
def __init__(self, def __init__(self,
scale=1.0, scale,
depths=[2, 2, 6, 2], depths,
class_num=1000, class_num=1000,
dropout_prob=0.2, dropout_prob=0,
expand=1280): use_last_conv=True,
class_expand=1280):
super().__init__() super().__init__()
self.scale = scale self.scale = scale
self.expand = expand self.use_last_conv = use_last_conv
self.class_expand = class_expand
self.stem = nn.Sequential(* [ self.stem = nn.Sequential(* [
ConvBNLayer( ConvBNLayer(
@ -273,77 +274,33 @@ class PPLCNetV2(TheseusLayer):
dw_size=3) dw_size=3)
]) ])
# stage1 # stages
in_channels, kernel_size, split_pw, use_rep, use_se, use_shortcut = NET_CONFIG[ self.stages = nn.LayerList()
"stage1"] for depth_idx, k in enumerate(NET_CONFIG):
self.stage1 = nn.Sequential(* [ in_channels, kernel_size, split_pw, use_rep, use_se, use_shortcut = NET_CONFIG[
RepDepthwiseSeparable( k]
in_channels=make_divisible((in_channels if i == 0 else self.stages.append(
in_channels * 2) * scale), nn.Sequential(* [
out_channels=make_divisible(in_channels * 2 * scale), RepDepthwiseSeparable(
stride=2 if i == 0 else 1, in_channels=make_divisible((in_channels if i == 0 else
dw_size=kernel_size, in_channels * 2) * scale),
split_pw=split_pw, out_channels=make_divisible(in_channels * 2 * scale),
use_rep=use_rep, stride=2 if i == 0 else 1,
use_se=use_se, dw_size=kernel_size,
use_shortcut=use_shortcut, ) for i in range(depths[0]) split_pw=split_pw,
]) use_rep=use_rep,
use_se=use_se,
# stage2 use_shortcut=use_shortcut)
in_channels, kernel_size, split_pw, use_rep, use_se, use_shortcut = NET_CONFIG[ for i in range(depths[depth_idx])
"stage2"] ]))
self.stage2 = nn.Sequential(* [
RepDepthwiseSeparable(
in_channels=make_divisible((in_channels if i == 0 else
in_channels * 2) * scale),
out_channels=make_divisible(in_channels * 2 * scale),
stride=2 if i == 0 else 1,
dw_size=kernel_size,
split_pw=split_pw,
use_rep=use_rep,
use_se=use_se,
use_shortcut=use_shortcut, ) for i in range(depths[1])
])
# stage3
in_channels, kernel_size, split_pw, use_rep, use_se, use_shortcut = NET_CONFIG[
"stage3"]
self.stage3 = nn.Sequential(* [
RepDepthwiseSeparable(
in_channels=make_divisible((in_channels if i == 0 else
in_channels * 2) * scale),
out_channels=make_divisible(in_channels * 2 * scale),
stride=2 if i == 0 else 1,
dw_size=kernel_size,
split_pw=split_pw,
use_rep=use_rep,
use_se=use_se,
use_shortcut=use_shortcut, ) for i in range(depths[2])
])
# stage4
in_channels, kernel_size, split_pw, use_rep, use_se, use_shortcut = NET_CONFIG[
"stage4"]
self.stage4 = nn.Sequential(* [
RepDepthwiseSeparable(
in_channels=make_divisible((in_channels if i == 0 else
in_channels * 2) * scale),
out_channels=make_divisible(in_channels * 2 * scale),
stride=2 if i == 0 else 1,
dw_size=kernel_size,
split_pw=split_pw,
use_rep=use_rep,
use_se=use_se,
use_shortcut=use_shortcut, ) for i in range(depths[3])
])
self.avg_pool = AdaptiveAvgPool2D(1) self.avg_pool = AdaptiveAvgPool2D(1)
if self.expand: if self.use_last_conv:
self.last_conv = Conv2D( self.last_conv = Conv2D(
in_channels=make_divisible(NET_CONFIG["stage4"][0] * 2 * in_channels=make_divisible(NET_CONFIG["stage4"][0] * 2 *
scale), scale),
out_channels=self.expand, out_channels=self.class_expand,
kernel_size=1, kernel_size=1,
stride=1, stride=1,
padding=0, padding=0,
@ -352,18 +309,16 @@ class PPLCNetV2(TheseusLayer):
self.dropout = Dropout(p=dropout_prob, mode="downscale_in_infer") self.dropout = Dropout(p=dropout_prob, mode="downscale_in_infer")
self.flatten = nn.Flatten(start_axis=1, stop_axis=-1) self.flatten = nn.Flatten(start_axis=1, stop_axis=-1)
in_features = self.expand if self.expand else NET_CONFIG["stage4"][ in_features = self.class_expand if self.use_last_conv else NET_CONFIG[
0] * 2 * scale "stage4"][0] * 2 * scale
self.fc = Linear(in_features, class_num) self.fc = Linear(in_features, class_num)
def forward(self, x): def forward(self, x):
x = self.stem(x) x = self.stem(x)
x = self.stage1(x) for stage in self.stages:
x = self.stage2(x) x = stage(x)
x = self.stage3(x)
x = self.stage4(x)
x = self.avg_pool(x) x = self.avg_pool(x)
if self.expand: if self.use_last_conv:
x = self.last_conv(x) x = self.last_conv(x)
x = self.act(x) x = self.act(x)
x = self.dropout(x) x = self.dropout(x)
@ -395,6 +350,7 @@ def PPLCNetV2_base(pretrained=False, use_ssld=False, **kwargs):
Returns: Returns:
model: nn.Layer. Specific `PPLCNetV2_base` model depends on args. model: nn.Layer. Specific `PPLCNetV2_base` model depends on args.
""" """
model = PPLCNetV2(scale=1.0, depths=[2, 2, 6, 2], **kwargs) model = PPLCNetV2(
scale=1.0, depths=[2, 2, 6, 2], dropout_prob=0.2, **kwargs)
_load_pretrained(pretrained, model, MODEL_URLS["PPLCNetV2_base"], use_ssld) _load_pretrained(pretrained, model, MODEL_URLS["PPLCNetV2_base"], use_ssld)
return model return model