PaddleClas/ppcls/arch/backbone/model_zoo/nextvit.py

644 lines
21 KiB
Python

# copyright (c) 2022 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Code was based on https://github.com/bytedance/Next-ViT/blob/main/classification/nextvit.py
# reference: https://arxiv.org/abs/2207.05501
from functools import partial
import paddle
from paddle import nn
from paddle.nn.initializer import TruncatedNormal, Constant, Normal
from .vision_transformer import trunc_normal_, zeros_, ones_, to_2tuple, DropPath, Identity
from ....utils.save_load import load_dygraph_pretrain
MODEL_URLS = {
"NextViT_small_224":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/NextViT_small_224_pretrained.pdparams",
"NextViT_base_224":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/NextViT_base_224_pretrained.pdparams",
"NextViT_large_224":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/NextViT_large_224_pretrained.pdparams",
"NextViT_small_384":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/NextViT_small_384_pretrained.pdparams",
"NextViT_base_384":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/NextViT_base_384_pretrained.pdparams",
"NextViT_large_384":
"https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/NextViT_large_384_pretrained.pdparams",
}
__all__ = list(MODEL_URLS.keys())
NORM_EPS = 1e-5
def rearrange(x, pattern, **axes_lengths):
if 'b (h w) c -> b c h w' == pattern:
b, n, c = x.shape
h = axes_lengths.pop('h', -1)
w = axes_lengths.pop('w', -1)
h = h if w == -1 else n // w
w = w if h == -1 else n // h
return x.transpose([0, 2, 1]).reshape([b, c, h, w])
if 'b c h w -> b (h w) c' == pattern:
b, c, h, w = x.shape
return x.reshape([b, c, h * w]).transpose([0, 2, 1])
if 'b t (h d) -> b h t d' == pattern:
b, t, h_d = x.shape
h = axes_lengths['h']
return x.reshape([b, t, h, h_d // h]).transpose([0, 2, 1, 3])
if 'b h t d -> b t (h d)' == pattern:
b, h, t, d = x.shape
return x.transpose([0, 2, 1, 3]).reshape([b, t, h * d])
raise NotImplementedError(
"Rearrangement '{}' has not been implemented.".format(pattern))
def merge_pre_bn(layer, pre_bn_1, pre_bn_2=None):
""" Merge pre BN to reduce inference runtime.
"""
weight = layer.weight
if isinstance(layer, nn.Linear):
weight = weight.transpose([1, 0])
bias = layer.bias
if pre_bn_2 is None:
scale_invstd = (pre_bn_1._variance + pre_bn_1._epsilon).pow(-0.5)
extra_weight = scale_invstd * pre_bn_1.weight
extra_bias = pre_bn_1.bias - pre_bn_1.weight * pre_bn_1._mean * scale_invstd
else:
scale_invstd_1 = (pre_bn_1._variance + pre_bn_1._epsilon).pow(-0.5)
scale_invstd_2 = (pre_bn_2._variance + pre_bn_2._epsilon).pow(-0.5)
extra_weight = scale_invstd_1 * pre_bn_1.weight * scale_invstd_2 * pre_bn_2.weight
extra_bias = scale_invstd_2 * pre_bn_2.weight * (
pre_bn_1.bias - pre_bn_1.weight * pre_bn_1._mean * scale_invstd_1 -
pre_bn_2._mean) + pre_bn_2.bias
if isinstance(layer, nn.Linear):
extra_bias = weight @extra_bias
weight = weight.multiply(
extra_weight.reshape([1, weight.shape[1]]).expand_as(weight))
weight = weight.transpose([1, 0])
elif isinstance(layer, nn.Conv2D):
assert weight.shape[2] == 1 and weight.shape[3] == 1
weight = weight.reshape([weight.shape[0], weight.shape[1]])
extra_bias = weight @extra_bias
weight = weight.multiply(
extra_weight.reshape([1, weight.shape[1]]).expand_as(weight))
weight = weight.reshape([weight.shape[0], weight.shape[1], 1, 1])
bias = bias.add(extra_bias)
layer.weight.set_value(weight)
layer.bias.set_value(bias)
def _make_divisible(v, divisor, min_value=None):
if min_value is None:
min_value = divisor
new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
# Make sure that round down does not go down by more than 10%.
if new_v < 0.9 * v:
new_v += divisor
return new_v
class ConvBNReLU(nn.Layer):
def __init__(self,
in_channels,
out_channels,
kernel_size,
stride,
groups=1):
super(ConvBNReLU, self).__init__()
self.conv = nn.Conv2D(
in_channels,
out_channels,
kernel_size=kernel_size,
stride=stride,
padding=1,
groups=groups,
bias_attr=False)
self.norm = nn.BatchNorm2D(out_channels, epsilon=NORM_EPS)
self.act = nn.ReLU()
def forward(self, x):
x = self.conv(x)
x = self.norm(x)
x = self.act(x)
return x
class PatchEmbed(nn.Layer):
def __init__(self, in_channels, out_channels, stride=1):
super(PatchEmbed, self).__init__()
norm_layer = partial(nn.BatchNorm2D, epsilon=NORM_EPS)
if stride == 2:
self.avgpool = nn.AvgPool2D((2, 2), stride=2, ceil_mode=True)
self.conv = nn.Conv2D(
in_channels,
out_channels,
kernel_size=1,
stride=1,
bias_attr=False)
self.norm = norm_layer(out_channels)
elif in_channels != out_channels:
self.avgpool = nn.Identity()
self.conv = nn.Conv2D(
in_channels,
out_channels,
kernel_size=1,
stride=1,
bias_attr=False)
self.norm = norm_layer(out_channels)
else:
self.avgpool = nn.Identity()
self.conv = nn.Identity()
self.norm = nn.Identity()
def forward(self, x):
return self.norm(self.conv(self.avgpool(x)))
class MHCA(nn.Layer):
"""
Multi-Head Convolutional Attention
"""
def __init__(self, out_channels, head_dim):
super(MHCA, self).__init__()
norm_layer = partial(nn.BatchNorm2D, epsilon=NORM_EPS)
self.group_conv3x3 = nn.Conv2D(
out_channels,
out_channels,
kernel_size=3,
stride=1,
padding=1,
groups=out_channels // head_dim,
bias_attr=False)
self.norm = norm_layer(out_channels)
self.act = nn.ReLU()
self.projection = nn.Conv2D(
out_channels, out_channels, kernel_size=1, bias_attr=False)
def forward(self, x):
out = self.group_conv3x3(x)
out = self.norm(out)
out = self.act(out)
out = self.projection(out)
return out
class Mlp(nn.Layer):
def __init__(self,
in_features,
out_features=None,
mlp_ratio=None,
drop=0.,
bias=True):
super().__init__()
out_features = out_features or in_features
hidden_dim = _make_divisible(in_features * mlp_ratio, 32)
self.conv1 = nn.Conv2D(
in_features,
hidden_dim,
kernel_size=1,
bias_attr=None if bias == True else False)
self.act = nn.ReLU()
self.conv2 = nn.Conv2D(
hidden_dim,
out_features,
kernel_size=1,
bias_attr=None if bias == True else False)
self.drop = nn.Dropout(drop)
def merge_bn(self, pre_norm):
merge_pre_bn(self.conv1, pre_norm)
self.is_bn_merged = True
def forward(self, x):
x = self.conv1(x)
x = self.act(x)
x = self.drop(x)
x = self.conv2(x)
x = self.drop(x)
return x
class NCB(nn.Layer):
"""
Next Convolution Block
"""
def __init__(self,
in_channels,
out_channels,
stride=1,
path_dropout=0.0,
drop=0.0,
head_dim=32,
mlp_ratio=3):
super(NCB, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
norm_layer = partial(nn.BatchNorm2D, epsilon=NORM_EPS)
assert out_channels % head_dim == 0
self.patch_embed = PatchEmbed(in_channels, out_channels, stride)
self.mhca = MHCA(out_channels, head_dim)
self.attention_path_dropout = DropPath(path_dropout)
self.norm = norm_layer(out_channels)
self.mlp = Mlp(out_channels, mlp_ratio=mlp_ratio, drop=drop, bias=True)
self.mlp_path_dropout = DropPath(path_dropout)
self.is_bn_merged = False
def merge_bn(self):
if not self.is_bn_merged:
self.mlp.merge_bn(self.norm)
self.is_bn_merged = True
def forward(self, x):
x = self.patch_embed(x)
x = x + self.attention_path_dropout(self.mhca(x))
if not self.is_bn_merged:
out = self.norm(x)
else:
out = x
x = x + self.mlp_path_dropout(self.mlp(out))
return x
class E_MHSA(nn.Layer):
"""
Efficient Multi-Head Self Attention
"""
def __init__(self,
dim,
out_dim=None,
head_dim=32,
qkv_bias=True,
qk_scale=None,
attn_drop=0,
proj_drop=0.,
sr_ratio=1):
super().__init__()
self.dim = dim
self.out_dim = out_dim if out_dim is not None else dim
self.num_heads = self.dim // head_dim
self.scale = qk_scale or head_dim**-0.5
self.q = nn.Linear(dim, self.dim, bias_attr=qkv_bias)
self.k = nn.Linear(dim, self.dim, bias_attr=qkv_bias)
self.v = nn.Linear(dim, self.dim, bias_attr=qkv_bias)
self.proj = nn.Linear(self.dim, self.out_dim)
self.attn_drop = nn.Dropout(attn_drop)
self.proj_drop = nn.Dropout(proj_drop)
self.sr_ratio = sr_ratio
self.N_ratio = sr_ratio**2
if sr_ratio > 1:
self.sr = nn.AvgPool1D(
kernel_size=self.N_ratio, stride=self.N_ratio)
self.norm = nn.BatchNorm1D(dim, epsilon=NORM_EPS)
self.is_bn_merged = False
def merge_bn(self, pre_bn):
merge_pre_bn(self.q, pre_bn)
if self.sr_ratio > 1:
merge_pre_bn(self.k, pre_bn, self.norm)
merge_pre_bn(self.v, pre_bn, self.norm)
else:
merge_pre_bn(self.k, pre_bn)
merge_pre_bn(self.v, pre_bn)
self.is_bn_merged = True
def forward(self, x):
B, N, C = x.shape
q = self.q(x)
q = q.reshape(
[B, N, self.num_heads, int(C // self.num_heads)]).transpose(
[0, 2, 1, 3])
if self.sr_ratio > 1:
x_ = x.transpose([0, 2, 1])
x_ = self.sr(x_)
if not self.is_bn_merged:
x_ = self.norm(x_)
x_ = x_.transpose([0, 2, 1])
k = self.k(x_)
k = k.reshape(
[B, k.shape[1], self.num_heads, int(C // self.num_heads)
]).transpose([0, 2, 3, 1])
v = self.v(x_)
v = v.reshape(
[B, v.shape[1], self.num_heads, int(C // self.num_heads)
]).transpose([0, 2, 1, 3])
else:
k = self.k(x)
k = k.reshape(
[B, k.shape[1], self.num_heads, int(C // self.num_heads)
]).transpose([0, 2, 3, 1])
v = self.v(x)
v = v.reshape(
[B, v.shape[1], self.num_heads, int(C // self.num_heads)
]).transpose([0, 2, 1, 3])
attn = (q @k) * self.scale
attn = nn.functional.softmax(attn, axis=-1)
attn = self.attn_drop(attn)
x = (attn @v).transpose([0, 2, 1, 3]).reshape([B, N, C])
x = self.proj(x)
x = self.proj_drop(x)
return x
class NTB(nn.Layer):
"""
Next Transformer Block
"""
def __init__(
self,
in_channels,
out_channels,
path_dropout,
stride=1,
sr_ratio=1,
mlp_ratio=2,
head_dim=32,
mix_block_ratio=0.75,
attn_drop=0.0,
drop=0.0, ):
super(NTB, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.mix_block_ratio = mix_block_ratio
norm_func = partial(nn.BatchNorm2D, epsilon=NORM_EPS)
self.mhsa_out_channels = _make_divisible(
int(out_channels * mix_block_ratio), 32)
self.mhca_out_channels = out_channels - self.mhsa_out_channels
self.patch_embed = PatchEmbed(in_channels, self.mhsa_out_channels,
stride)
self.norm1 = norm_func(self.mhsa_out_channels)
self.e_mhsa = E_MHSA(
self.mhsa_out_channels,
head_dim=head_dim,
sr_ratio=sr_ratio,
attn_drop=attn_drop,
proj_drop=drop)
self.mhsa_path_dropout = DropPath(path_dropout * mix_block_ratio)
self.projection = PatchEmbed(
self.mhsa_out_channels, self.mhca_out_channels, stride=1)
self.mhca = MHCA(self.mhca_out_channels, head_dim=head_dim)
self.mhca_path_dropout = DropPath(path_dropout * (1 - mix_block_ratio))
self.norm2 = norm_func(out_channels)
self.mlp = Mlp(out_channels, mlp_ratio=mlp_ratio, drop=drop)
self.mlp_path_dropout = DropPath(path_dropout)
self.is_bn_merged = False
def merge_bn(self):
if not self.is_bn_merged:
self.e_mhsa.merge_bn(self.norm1)
self.mlp.merge_bn(self.norm2)
self.is_bn_merged = True
def forward(self, x):
x = self.patch_embed(x)
B, C, H, W = x.shape
if not self.is_bn_merged:
out = self.norm1(x)
else:
out = x
out = rearrange(out, "b c h w -> b (h w) c") # b n c
out = self.e_mhsa(out)
out = self.mhsa_path_dropout(out)
x = x + rearrange(out, "b (h w) c -> b c h w", h=H)
out = self.projection(x)
out = out + self.mhca_path_dropout(self.mhca(out))
x = paddle.concat([x, out], axis=1)
if not self.is_bn_merged:
out = self.norm2(x)
else:
out = x
x = x + self.mlp_path_dropout(self.mlp(out))
return x
class NextViT(nn.Layer):
def __init__(self,
stem_chs,
depths,
path_dropout,
attn_drop=0,
drop=0,
class_num=1000,
strides=[1, 2, 2, 2],
sr_ratios=[8, 4, 2, 1],
head_dim=32,
mix_block_ratio=0.75):
super(NextViT, self).__init__()
self.stage_out_channels = [
[96] * (depths[0]), [192] * (depths[1] - 1) + [256],
[384, 384, 384, 384, 512] * (depths[2] // 5),
[768] * (depths[3] - 1) + [1024]
]
# Next Hybrid Strategy
self.stage_block_types = [[NCB] * depths[0],
[NCB] * (depths[1] - 1) + [NTB],
[NCB, NCB, NCB, NCB, NTB] * (depths[2] // 5),
[NCB] * (depths[3] - 1) + [NTB]]
self.stem = nn.Sequential(
ConvBNReLU(
3, stem_chs[0], kernel_size=3, stride=2),
ConvBNReLU(
stem_chs[0], stem_chs[1], kernel_size=3, stride=1),
ConvBNReLU(
stem_chs[1], stem_chs[2], kernel_size=3, stride=1),
ConvBNReLU(
stem_chs[2], stem_chs[2], kernel_size=3, stride=2), )
input_channel = stem_chs[-1]
features = []
idx = 0
dpr = [
x.item() for x in paddle.linspace(0, path_dropout, sum(depths))
] # stochastic depth decay rule
for stage_id in range(len(depths)):
numrepeat = depths[stage_id]
output_channels = self.stage_out_channels[stage_id]
block_types = self.stage_block_types[stage_id]
for block_id in range(numrepeat):
if strides[stage_id] == 2 and block_id == 0:
stride = 2
else:
stride = 1
output_channel = output_channels[block_id]
block_type = block_types[block_id]
if block_type is NCB:
layer = NCB(input_channel,
output_channel,
stride=stride,
path_dropout=dpr[idx + block_id],
drop=drop,
head_dim=head_dim)
features.append(layer)
elif block_type is NTB:
layer = NTB(input_channel,
output_channel,
path_dropout=dpr[idx + block_id],
stride=stride,
sr_ratio=sr_ratios[stage_id],
head_dim=head_dim,
mix_block_ratio=mix_block_ratio,
attn_drop=attn_drop,
drop=drop)
features.append(layer)
input_channel = output_channel
idx += numrepeat
self.features = nn.Sequential(*features)
self.norm = nn.BatchNorm2D(output_channel, epsilon=NORM_EPS)
self.avgpool = nn.AdaptiveAvgPool2D((1, 1))
self.proj_head = nn.Sequential(nn.Linear(output_channel, class_num), )
self.stage_out_idx = [
sum(depths[:idx + 1]) - 1 for idx in range(len(depths))
]
self._initialize_weights()
def merge_bn(self):
self.eval()
for idx, layer in self.named_sublayers():
if isinstance(layer, NCB) or isinstance(layer, NTB):
layer.merge_bn()
def _initialize_weights(self):
for n, m in self.named_sublayers():
if isinstance(m, (nn.BatchNorm2D, nn.GroupNorm, nn.LayerNorm,
nn.BatchNorm1D)):
ones_(m.weight)
zeros_(m.bias)
elif isinstance(m, nn.Linear):
trunc_normal_(m.weight)
if hasattr(m, 'bias') and m.bias is not None:
zeros_(m.bias)
elif isinstance(m, nn.Conv2D):
trunc_normal_(m.weight)
if hasattr(m, 'bias') and m.bias is not None:
zeros_(m.bias)
def forward(self, x):
x = self.stem(x)
for layer in self.features:
x = layer(x)
x = self.norm(x)
x = self.avgpool(x)
x = paddle.flatten(x, 1)
x = self.proj_head(x)
return x
def _load_pretrained(pretrained, model, model_url, use_ssld=False):
if pretrained is False:
pass
elif pretrained is True:
load_dygraph_pretrain(model, model_url, use_ssld=use_ssld)
elif isinstance(pretrained, str):
load_dygraph_pretrain(model, pretrained)
else:
raise RuntimeError(
"pretrained type is not available. Please use `string` or `boolean` type."
)
def NextViT_small_224(pretrained=False, use_ssld=False, **kwargs):
model = NextViT(
stem_chs=[64, 32, 64],
depths=[3, 4, 10, 3],
path_dropout=0.1,
**kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["NextViT_small_224"], use_ssld=use_ssld)
return model
def NextViT_base_224(pretrained=False, use_ssld=False, **kwargs):
model = NextViT(
stem_chs=[64, 32, 64],
depths=[3, 4, 20, 3],
path_dropout=0.2,
**kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["NextViT_base_224"], use_ssld=use_ssld)
return model
def NextViT_large_224(pretrained=False, use_ssld=False, **kwargs):
model = NextViT(
stem_chs=[64, 32, 64],
depths=[3, 4, 30, 3],
path_dropout=0.2,
**kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["NextViT_large_224"], use_ssld=use_ssld)
return model
def NextViT_small_384(pretrained=False, use_ssld=False, **kwargs):
model = NextViT(
stem_chs=[64, 32, 64],
depths=[3, 4, 10, 3],
path_dropout=0.1,
**kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["NextViT_small_384"], use_ssld=use_ssld)
return model
def NextViT_base_384(pretrained=False, use_ssld=False, **kwargs):
model = NextViT(
stem_chs=[64, 32, 64],
depths=[3, 4, 20, 3],
path_dropout=0.2,
**kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["NextViT_base_384"], use_ssld=use_ssld)
return model
def NextViT_large_384(pretrained=False, use_ssld=False, **kwargs):
model = NextViT(
stem_chs=[64, 32, 64],
depths=[3, 4, 30, 3],
path_dropout=0.2,
**kwargs)
_load_pretrained(
pretrained, model, MODEL_URLS["NextViT_large_384"], use_ssld=use_ssld)
return model