From b0fb44b0b97505184140cbc829691e0d7578af3b Mon Sep 17 00:00:00 2001 From: Wenyu Date: Wed, 22 Jun 2022 17:16:48 +0800 Subject: [PATCH] fix convnext init (#6248) --- ppdet/modeling/backbones/convnext.py | 10 +++------- 1 file changed, 3 insertions(+), 7 deletions(-) diff --git a/ppdet/modeling/backbones/convnext.py b/ppdet/modeling/backbones/convnext.py index 5b6f80157..e2a678bab 100644 --- a/ppdet/modeling/backbones/convnext.py +++ b/ppdet/modeling/backbones/convnext.py @@ -29,7 +29,7 @@ import numpy as np from ppdet.core.workspace import register, serializable from ..shape_spec import ShapeSpec -from .transformer_utils import DropPath, trunc_normal_ +from .transformer_utils import DropPath, trunc_normal_, zeros_ __all__ = ['ConvNeXt'] @@ -129,7 +129,6 @@ class ConvNeXt(nn.Layer): dims (int): Feature dimension at each stage. Default: [96, 192, 384, 768] drop_path_rate (float): Stochastic depth rate. Default: 0. layer_scale_init_value (float): Init value for Layer Scale. Default: 1e-6. - head_init_scale (float): Init scaling value for classifier weights and biases. Default: 1. """ def __init__( @@ -139,7 +138,6 @@ class ConvNeXt(nn.Layer): dims=[96, 192, 384, 768], drop_path_rate=0., layer_scale_init_value=1e-6, - head_init_scale=1., return_idx=[1, 2, 3], norm_output=True, pretrained=None, ): @@ -188,8 +186,6 @@ class ConvNeXt(nn.Layer): ]) self.apply(self._init_weights) - # self.head.weight.set_value(self.head.weight.numpy() * head_init_scale) - # self.head.bias.set_value(self.head.weight.numpy() * head_init_scale) if pretrained is not None: if 'http' in pretrained: #URL @@ -201,8 +197,8 @@ class ConvNeXt(nn.Layer): def _init_weights(self, m): if isinstance(m, (nn.Conv2D, nn.Linear)): - trunc_normal_(m.weight, std=.02) - nn.init.constant_(m.bias, 0) + trunc_normal_(m.weight) + zeros_(m.bias) def forward_features(self, x): output = [] -- GitLab