128 lines
4.5 KiB
Python
128 lines
4.5 KiB
Python
# copyright (c) 2021 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.
|
|
|
|
import paddle
|
|
import paddle.nn as nn
|
|
from .vision_transformer import VisionTransformer, Identity, trunc_normal_, zeros_
|
|
|
|
|
|
__all__ = [
|
|
'DeiT_tiny_patch16_224', 'DeiT_small_patch16_224', 'DeiT_base_patch16_224',
|
|
'DeiT_tiny_distilled_patch16_224', 'DeiT_small_distilled_patch16_224',
|
|
'DeiT_base_distilled_patch16_224', 'DeiT_base_patch16_384',
|
|
'DeiT_base_distilled_patch16_384'
|
|
]
|
|
|
|
|
|
class DistilledVisionTransformer(VisionTransformer):
|
|
def __init__(self, img_size=224, patch_size=16, class_dim=1000, embed_dim=768, depth=12,
|
|
num_heads=12, mlp_ratio=4, qkv_bias=False, norm_layer='nn.LayerNorm', epsilon=1e-5,
|
|
**kwargs):
|
|
super().__init__(img_size=img_size, patch_size=patch_size, class_dim=class_dim, embed_dim=embed_dim, depth=depth,
|
|
num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, norm_layer=norm_layer, epsilon=epsilon,
|
|
**kwargs)
|
|
self.pos_embed = self.create_parameter(
|
|
shape=(1, self.patch_embed.num_patches + 2, self.embed_dim), default_initializer=zeros_)
|
|
self.add_parameter("pos_embed", self.pos_embed)
|
|
|
|
self.dist_token = self.create_parameter(
|
|
shape=(1, 1, self.embed_dim), default_initializer=zeros_)
|
|
self.add_parameter("cls_token", self.cls_token)
|
|
|
|
self.head_dist = nn.Linear(
|
|
self.embed_dim, self.class_dim) if self.class_dim > 0 else Identity()
|
|
|
|
trunc_normal_(self.dist_token)
|
|
trunc_normal_(self.pos_embed)
|
|
self.head_dist.apply(self._init_weights)
|
|
|
|
def forward_features(self, x):
|
|
B = x.shape[0]
|
|
x = self.patch_embed(x)
|
|
|
|
cls_tokens = self.cls_token.expand((B, -1, -1))
|
|
dist_token = self.dist_token.expand((B, -1, -1))
|
|
x = paddle.concat((cls_tokens, dist_token, x), axis=1)
|
|
|
|
x = x + self.pos_embed
|
|
x = self.pos_drop(x)
|
|
|
|
for blk in self.blocks:
|
|
x = blk(x)
|
|
|
|
x = self.norm(x)
|
|
return x[:, 0], x[:, 1]
|
|
|
|
def forward(self, x):
|
|
x, x_dist = self.forward_features(x)
|
|
x = self.head(x)
|
|
x_dist = self.head_dist(x_dist)
|
|
return (x + x_dist) / 2
|
|
|
|
|
|
def DeiT_tiny_patch16_224(**kwargs):
|
|
model = VisionTransformer(
|
|
patch_size=16, embed_dim=192, depth=12, num_heads=3, mlp_ratio=4, qkv_bias=True,
|
|
epsilon=1e-6, **kwargs)
|
|
return model
|
|
|
|
|
|
def DeiT_small_patch16_224(**kwargs):
|
|
model = VisionTransformer(
|
|
patch_size=16, embed_dim=384, depth=12, num_heads=6, mlp_ratio=4, qkv_bias=True,
|
|
epsilon=1e-6, **kwargs)
|
|
return model
|
|
|
|
|
|
def DeiT_base_patch16_224(**kwargs):
|
|
model = VisionTransformer(
|
|
patch_size=16, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4, qkv_bias=True,
|
|
epsilon=1e-6, **kwargs)
|
|
return model
|
|
|
|
|
|
def DeiT_tiny_distilled_patch16_224(**kwargs):
|
|
model = DistilledVisionTransformer(
|
|
patch_size=16, embed_dim=192, depth=12, num_heads=3, mlp_ratio=4, qkv_bias=True,
|
|
epsilon=1e-6, **kwargs)
|
|
return model
|
|
|
|
|
|
def DeiT_small_distilled_patch16_224(**kwargs):
|
|
model = DistilledVisionTransformer(
|
|
patch_size=16, embed_dim=384, depth=12, num_heads=6, mlp_ratio=4, qkv_bias=True,
|
|
epsilon=1e-6, **kwargs)
|
|
return model
|
|
|
|
|
|
def DeiT_base_distilled_patch16_224(**kwargs):
|
|
model = DistilledVisionTransformer(
|
|
patch_size=16, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4, qkv_bias=True,
|
|
epsilon=1e-6, **kwargs)
|
|
return model
|
|
|
|
|
|
def DeiT_base_patch16_384(**kwargs):
|
|
model = VisionTransformer(
|
|
img_size=384, patch_size=16, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4, qkv_bias=True,
|
|
epsilon=1e-6, **kwargs)
|
|
return model
|
|
|
|
|
|
def DeiT_base_distilled_patch16_384(**kwargs):
|
|
model = DistilledVisionTransformer(
|
|
img_size=384, patch_size=16, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4, qkv_bias=True,
|
|
epsilon=1e-6, **kwargs)
|
|
return model
|