Source code for towhee.models.uniformer.uniformer
# Code for "UniFormer: Unified Transformer for Efficient Spatiotemporal Representation Learning"
# arXiv:2201.04676
# Kunchang Li, Yali Wang, Peng Gao, Guanglu Song, Yu Liu, Hongsheng Li, Yu Qiao
#
# 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.
#
# modified by Zilliz.
from collections import OrderedDict
import torch
from torch import nn
from torch.utils import checkpoint
from functools import partial
from towhee.models.utils.general_utils import to_2tuple
from towhee.models.utils.weight_init import trunc_normal_
from towhee.models.layers.droppath import DropPath
from towhee.models.layers.mlp import Mlp
from towhee.models.layers.attention import MultiHeadAttention
from towhee.models.uniformer.config import _C
import os
model_path = 'path_to_models'
model_path = {
'uniformer_small_in1k': os.path.join(model_path, 'uniformer_small_in1k.pth'),
'uniformer_small_k400_8x8': os.path.join(model_path, 'uniformer_small_k400_8x8.pth'),
'uniformer_small_k400_16x4': os.path.join(model_path, 'uniformer_small_k400_16x4.pth'),
'uniformer_small_k600_16x4': os.path.join(model_path, 'uniformer_small_k600_16x4.pth'),
'uniformer_base_in1k': os.path.join(model_path, 'uniformer_base_in1k.pth'),
'uniformer_base_k400_8x8': os.path.join(model_path, 'uniformer_base_k400_8x8.pth'),
'uniformer_base_k400_16x4': os.path.join(model_path, 'uniformer_base_k400_16x4.pth'),
'uniformer_base_k600_16x4': os.path.join(model_path, 'uniformer_base_k600_16x4.pth'),
}
[docs]def conv_3xnxn(inp, oup, kernel_size=3, stride=3, groups=1):
return nn.Conv3d(inp, oup, (3, kernel_size, kernel_size), (2, stride, stride), (1, 0, 0), groups=groups)
[docs]def conv_1xnxn(inp, oup, kernel_size=3, stride=3, groups=1):
return nn.Conv3d(inp, oup, (1, kernel_size, kernel_size), (1, stride, stride), (0, 0, 0), groups=groups)
[docs]def conv_3xnxn_std(inp, oup, kernel_size=3, stride=3, groups=1):
return nn.Conv3d(inp, oup, (3, kernel_size, kernel_size), (1, stride, stride), (1, 0, 0), groups=groups)
[docs]def conv_1x1x1(inp, oup, groups=1):
return nn.Conv3d(inp, oup, (1, 1, 1), (1, 1, 1), (0, 0, 0), groups=groups)
[docs]def conv_3x3x3(inp, oup, groups=1):
return nn.Conv3d(inp, oup, (3, 3, 3), (1, 1, 1), (1, 1, 1), groups=groups)
[docs]def conv_5x5x5(inp, oup, groups=1):
return nn.Conv3d(inp, oup, (5, 5, 5), (1, 1, 1), (2, 2, 2), groups=groups)
[docs]class CMlp(nn.Module):
"""
CMlp
"""
[docs] def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
super().__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
self.fc1 = conv_1x1x1(in_features, hidden_features)
self.act = act_layer()
self.fc2 = conv_1x1x1(hidden_features, out_features)
self.drop = nn.Dropout(drop)
[docs] def forward(self, x):
x = self.fc1(x)
x = self.act(x)
x = self.drop(x)
x = self.fc2(x)
x = self.drop(x)
return x
[docs]class CBlock(nn.Module):
"""
CBlock
"""
# pylint: disable=W0613
[docs] def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
super().__init__()
self.pos_embed = conv_3x3x3(dim, dim, groups=dim)
self.norm1 = bn_3d(dim)
self.conv1 = conv_1x1x1(dim, dim, 1)
self.conv2 = conv_1x1x1(dim, dim, 1)
self.attn = conv_5x5x5(dim, dim, groups=dim)
# NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
self.norm2 = bn_3d(dim)
mlp_hidden_dim = int(dim * mlp_ratio)
self.mlp = CMlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
[docs] def forward(self, x):
x = x + self.pos_embed(x)
x = x + self.drop_path(self.conv2(self.attn(self.conv1(self.norm1(x)))))
x = x + self.drop_path(self.mlp(self.norm2(x)))
return x
[docs]class SABlock(nn.Module):
"""
SABlock
"""
[docs] def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
super().__init__()
self.pos_embed = conv_3x3x3(dim, dim, groups=dim)
self.norm1 = norm_layer(dim)
self.attn = MultiHeadAttention(
dim,
num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
attn_drop_ratio=attn_drop, proj_drop_ratio=drop)
# NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
self.norm2 = norm_layer(dim)
mlp_hidden_dim = int(dim * mlp_ratio)
self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
[docs] def forward(self, x):
x = x + self.pos_embed(x)
b, c, t, h, w = x.shape
x = x.flatten(2).transpose(1, 2)
x = x + self.drop_path(self.attn(self.norm1(x)))
x = x + self.drop_path(self.mlp(self.norm2(x)))
x = x.transpose(1, 2).reshape(b, c, t, h, w)
return x
[docs]class SplitSABlock(nn.Module):
"""
SplitSABlock
"""
[docs] def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
super().__init__()
self.pos_embed = conv_3x3x3(dim, dim, groups=dim)
self.t_norm = norm_layer(dim)
self.t_attn = MultiHeadAttention(
dim,
num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
attn_drop_ratio=attn_drop, proj_drop_ratio=drop)
self.norm1 = norm_layer(dim)
self.attn = MultiHeadAttention(
dim,
num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
attn_drop_ratio=attn_drop, proj_drop_ratio=drop)
# NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
self.norm2 = norm_layer(dim)
mlp_hidden_dim = int(dim * mlp_ratio)
self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
[docs] def forward(self, x):
x = x + self.pos_embed(x)
b, c, t, h, w = x.shape
attn = x.view(b, c, t, h * w).permute(0, 3, 2, 1).contiguous()
attn = attn.view(b * h * w, t, c)
attn = attn + self.drop_path(self.t_attn(self.t_norm(attn)))
attn = attn.view(b, h * w, t, c).permute(0, 2, 1, 3).contiguous()
attn = attn.view(b * t, h * w, c)
residual = x.view(b, c, t, h * w).permute(0, 2, 3, 1).contiguous()
residual = residual.view(b * t, h * w, c)
attn = residual + self.drop_path(self.attn(self.norm1(attn)))
attn = attn.view(b, t * h * w, c)
out = attn + self.drop_path(self.mlp(self.norm2(attn)))
out = out.transpose(1, 2).reshape(b, c, t, h, w)
return out
[docs]class SpeicalPatchEmbed(nn.Module):
"""
Image to Patch Embedding
"""
[docs] def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):
super().__init__()
img_size = to_2tuple(img_size)
patch_size = to_2tuple(patch_size)
num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])
self.img_size = img_size
self.patch_size = patch_size
self.num_patches = num_patches
self.norm = nn.LayerNorm(embed_dim)
self.proj = conv_3xnxn(in_chans, embed_dim, kernel_size=patch_size[0], stride=patch_size[0])
[docs] def forward(self, x):
b, _, t, h, w = x.shape
# FIXME look at relaxing size constraints
# assert H == self.img_size[0] and W == self.img_size[1], \
# f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
x = self.proj(x)
b, _, t, h, w = x.shape
x = x.flatten(2).transpose(1, 2)
x = self.norm(x)
x = x.reshape(b, t, h, w, -1).permute(0, 4, 1, 2, 3).contiguous()
return x
[docs]class PatchEmbed(nn.Module):
"""
Image to Patch Embedding
"""
[docs] def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, std=False):
super().__init__()
img_size = to_2tuple(img_size)
patch_size = to_2tuple(patch_size)
num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])
self.img_size = img_size
self.patch_size = patch_size
self.num_patches = num_patches
self.norm = nn.LayerNorm(embed_dim)
if std:
self.proj = conv_3xnxn_std(in_chans, embed_dim, kernel_size=patch_size[0], stride=patch_size[0])
else:
self.proj = conv_1xnxn(in_chans, embed_dim, kernel_size=patch_size[0], stride=patch_size[0])
[docs] def forward(self, x):
b, _, t, h, w = x.shape
# FIXME look at relaxing size constraints
# assert H == self.img_size[0] and W == self.img_size[1], \
# f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
x = self.proj(x)
b, _, t, h, w = x.shape
x = x.flatten(2).transpose(1, 2)
x = self.norm(x)
x = x.reshape(b, t, h, w, -1).permute(0, 4, 1, 2, 3).contiguous()
return x
[docs]class Uniformer(nn.Module):
"""
Code for "UniFormer: Unified Transformer for Efficient Spatiotemporal Representation Learning"
arXiv:2201.04676
Kunchang Li, Yali Wang, Peng Gao, Guanglu Song, Yu Liu, Hongsheng Li, Yu Qiao
Args:
depth (`tuple[int]`)
Depth. Default: (5, 8, 20, 7).
num_classes (`int`):
dimension of head. Default: 400.
img_size (`int`):
The spatial crop size for training. Default: 224.
in_chans (`int`):
Image input channels num. Default: 3.
embed_dim (`tuple[int]`)
Embedding dimension. Default: (64, 128, 320, 512).
head_dim (`int`):
Dimension of head. Default: 64.
mlp_ratio (`int`):
Ratio of mlp hidden dim to embedding dim. Default: 4.
qkv_bias (`bool`):
Enable bias for qkv if True. Default: True.
qk_scale (`float`):
Override default qk scale of head_dim ** -0.5 if set. Default: None.
representation_size (`int`):
Enable and set representation layer (pre-logits) to this value if set. Default: None.
drop_rate (`float`):
Dropout rate. Default: 0.
attn_drop_rate (`float`):
Attention dropout rate. Default: 0.
drop_path_rate (`float`):
Stochastic depth rate. Default: 0.1.
split (`bool`):
Whether use split attention. Default: False.
std (`bool`):
Spatial-temporal downsample. Default: False.
use_checkpoint (`bool`):
Whether use checkpoint. Default: False.
checkpoint_num (`tuple[int]`)
Checkpoint num. Default: (0, 0, 0, 0).
pretrain_name (`string`):
Pretrained name. Default: None.
"""
[docs] def __init__(self,
depth = (5, 8, 20, 7),
num_classes = 400,
img_size = 32,
in_chans = 3,
embed_dim = (64, 128, 320, 512),
head_dim = 64,
mlp_ratio = 4,
qkv_bias = True,
qk_scale = None,
representation_size = None,
drop_rate = 0,
attn_drop_rate = 0,
drop_path_rate = 0.1,
split = False,
std = False,
use_checkpoint = False,
checkpoint_num = (0, 0, 0, 0),
):
super().__init__()
self.use_checkpoint = use_checkpoint
self.checkpoint_num = checkpoint_num
self.num_classes = num_classes
self.num_features = self.embed_dim = embed_dim # num_features for consistency with other models
norm_layer = partial(nn.LayerNorm, eps=1e-6)
self.patch_embed1 = SpeicalPatchEmbed(
img_size=img_size, patch_size=4, in_chans=in_chans, embed_dim=embed_dim[0])
self.patch_embed2 = PatchEmbed(
img_size=img_size // 4, patch_size=2, in_chans=embed_dim[0], embed_dim=embed_dim[1], std=std)
self.patch_embed3 = PatchEmbed(
img_size=img_size // 8, patch_size=2, in_chans=embed_dim[1], embed_dim=embed_dim[2], std=std)
self.patch_embed4 = PatchEmbed(
img_size=img_size // 16, patch_size=2, in_chans=embed_dim[2], embed_dim=embed_dim[3], std=std)
self.pos_drop = nn.Dropout(p=drop_rate)
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depth))] # stochastic depth decay rule
num_heads = [dim // head_dim for dim in embed_dim]
self.blocks1 = nn.ModuleList([
CBlock(
dim=embed_dim[0], num_heads=num_heads[0], mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer)
for i in range(depth[0])])
self.blocks2 = nn.ModuleList([
CBlock(
dim=embed_dim[1], num_heads=num_heads[1], mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i+depth[0]], norm_layer=norm_layer)
for i in range(depth[1])])
if split:
self.blocks3 = nn.ModuleList([
SplitSABlock(
dim=embed_dim[2], num_heads=num_heads[2], mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i+depth[0]+depth[1]], norm_layer=norm_layer)
for i in range(depth[2])])
self.blocks4 = nn.ModuleList([
SplitSABlock(
dim=embed_dim[3], num_heads=num_heads[3], mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i+depth[0]+depth[1]+depth[2]], norm_layer=norm_layer)
for i in range(depth[3])])
else:
self.blocks3 = nn.ModuleList([
SABlock(
dim=embed_dim[2], num_heads=num_heads[2], mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i+depth[0]+depth[1]], norm_layer=norm_layer)
for i in range(depth[2])])
self.blocks4 = nn.ModuleList([
SABlock(
dim=embed_dim[3], num_heads=num_heads[3], mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i+depth[0]+depth[1]+depth[2]], norm_layer=norm_layer)
for i in range(depth[3])])
self.norm = bn_3d(embed_dim[-1])
# Representation layer
if representation_size:
self.num_features = representation_size
self.pre_logits = nn.Sequential(OrderedDict([
('fc', nn.Linear(embed_dim, representation_size)),
('act', nn.Tanh())
]))
else:
self.pre_logits = nn.Identity()
# Classifier head
self.head = nn.Linear(embed_dim[-1], num_classes) if num_classes > 0 else nn.Identity()
self.apply(self._init_weights)
for name, p in self.named_parameters():
# fill proj weight with 1 here to improve training dynamics. Otherwise temporal attention inputs
# are multiplied by 0*0, which is hard for the model to move out of.
if 't_attn.qkv.weight' in name:
nn.init.constant_(p, 0)
if 't_attn.qkv.bias' in name:
nn.init.constant_(p, 0)
if 't_attn.proj.weight' in name:
nn.init.constant_(p, 1)
if 't_attn.proj.bias' in name:
nn.init.constant_(p, 0)
def _init_weights(self, m):
if isinstance(m, nn.Linear):
trunc_normal_(m.weight, std=.02)
if isinstance(m, nn.Linear) and m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.LayerNorm):
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
@torch.jit.ignore
def no_weight_decay(self):
return {'pos_embed', 'cls_token'}
def get_classifier(self):
return self.head
def reset_classifier(self, num_classes):
self.num_classes = num_classes
self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
def inflate_weight(self, weight_2d, time_dim, center=False):
if center:
weight_3d = torch.zeros(*weight_2d.shape)
weight_3d = weight_3d.unsqueeze(2).repeat(1, 1, time_dim, 1, 1)
middle_idx = time_dim // 2
weight_3d[:, :, middle_idx, :, :] = weight_2d
else:
weight_3d = weight_2d.unsqueeze(2).repeat(1, 1, time_dim, 1, 1)
weight_3d = weight_3d / time_dim
return weight_3d
def forward_features(self, x):
x = self.patch_embed1(x)
x = self.pos_drop(x)
for i, blk in enumerate(self.blocks1):
if self.use_checkpoint and i < self.checkpoint_num[0]:
x = checkpoint.checkpoint(blk, x)
else:
x = blk(x)
x = self.patch_embed2(x)
for i, blk in enumerate(self.blocks2):
if self.use_checkpoint and i < self.checkpoint_num[1]:
x = checkpoint.checkpoint(blk, x)
else:
x = blk(x)
x = self.patch_embed3(x)
for i, blk in enumerate(self.blocks3):
if self.use_checkpoint and i < self.checkpoint_num[2]:
x = checkpoint.checkpoint(blk, x)
else:
x = blk(x)
x = self.patch_embed4(x)
for i, blk in enumerate(self.blocks4):
if self.use_checkpoint and i < self.checkpoint_num[3]:
x = checkpoint.checkpoint(blk, x)
else:
x = blk(x)
x = self.norm(x)
x = self.pre_logits(x)
return x
[docs] def forward(self, x):
x = self.forward_features(x)
x = x.flatten(2).mean(-1)
x = self.head(x)
return x
[docs]def create_model(
model_name: str = 'uniformer_k400_s8',
pretrained: bool = False,
weights_path: str = None,
device: str = None,
):
if device is None:
device = 'cuda' if torch.cuda.is_available() else 'cpu'
if model_name == 'uniformer_k400_s8':
config = _C.MODEL.UniFormerS8
elif model_name == 'uniformer_k400_s16':
config = _C.MODEL.UniFormerS16
elif model_name == 'uniformer_k400_b8':
config = _C.MODEL.UniFormerB8
elif model_name == 'uniformer_k400_b16':
config = _C.MODEL.UniFormerB16
else:
raise AttributeError(f'Invalid model_name {model_name}.')
model=Uniformer(
depth = config.depth,
num_classes = config.num_classes,
img_size = config.img_size,
in_chans = config.in_chans,
embed_dim = config.embed_dim,
head_dim = config.head_dim,
mlp_ratio = config.mlp_ratio,
qkv_bias = config.qkv_bias,
qk_scale = config.qk_scale,
representation_size = config.representation_size,
drop_rate = config.drop_rate,
attn_drop_rate = config.attn_drop_rate,
drop_path_rate = config.drop_path_rate,
split = config.split,
std = config.std,
use_checkpoint = config.use_checkpoint,
checkpoint_num = config.checkpoint_num,
)
if pretrained:
model_checkpoint = torch.load(weights_path, map_location=device)
if 'model' in model_checkpoint:
model_checkpoint = model_checkpoint['model']
elif 'model_state' in model_checkpoint:
model_checkpoint = model_checkpoint['model_state']
# pylint: disable=E1136
state_dict_3d = model.state_dict()
for k in model_checkpoint.keys():
# pylint: disable=E1136
if model_checkpoint[k].shape != state_dict_3d[k].shape:
if len(state_dict_3d[k].shape) <= 2:
continue
# pylint: disable=E1136
time_dim = state_dict_3d[k].shape[2]
model_checkpoint[k] = model.inflate_weight(model_checkpoint[k], time_dim)
if model.num_classes != model_checkpoint['head.weight'].shape[0]:
del model_checkpoint['head.weight']
del model_checkpoint['head.bias']
model.load_state_dict(model_checkpoint)
model.to(device)
return model