j& d dl mZ d dlmZmZmZmZ d dlZd dlmZm Z ddl
m Z ddl m
Z
ddlmZ d d
lmZmZmZ d d lmZ d d lmZmZmZ g d
Z G d dej Z G d dej Z ddedZ G d de Z e ede j f ddddee d e!d ed e fd Z"dS ) )partial)AnyCallableListOptionalN)nnTensor )Conv2dNormActivation)ImageClassification)_log_api_usage_once )register_modelWeights WeightsEnum)_IMAGENET_CATEGORIES)_make_divisible_ovewrite_named_paramhandle_legacy_interface) MobileNetV2MobileNet_V2_Weights mobilenet_v2c r e Zd Z d dededededeedej f ddf fd
Zd
e de fd Z
xZ S )
InvertedResidualNinpoupstride expand_ratio
norm_layer.returnc p t || _ |dvrt d| | t j }t
t ||z }| j dk o||k | _ g }|dk r1| t ||d|t j | t |||||t j t j
||dddd || g t j | | _ || _ |dk | _ d S ) N)r r
z#stride should be 1 or 2 instead of r kernel_sizer activation_layer)r groupsr r$ r F)bias)super__init__r
ValueErrorr BatchNorm2dintrounduse_res_connectappendr ReLU6extendConv2d
Sequentialconv out_channels_is_cn) selfr r r r r
hidden_dimlayers __class__s OC:\Projects\Historical\venv\Lib\site-packages\torchvision/models/mobilenetv2.pyr( zInvertedResidual.__init__ sW K6 K KLL L Js\ 12233
#{a / >C3J "$ 1 MM$S*!PZmomuvvv
% !% )%'X
*c1a???
3
M6* qj xc j | j r|| | z S | | S N)r- r3 r6 r< s r: forwardzInvertedResidual.forward< s2 tyy||# #99Q<< r; r> ) __name__
__module__ __qualname__r+ r r r Moduler( r r@
__classcell__r9 s @r: r r s sw&! &!&! &!*-&!=@&!NVW_`cegen`nWoNp&!
&! &! &! &! &! &!P F r; r c e Zd Z ddededeeee d ed
eed ej f d eed ej f d
eddf fd
Z
de de fdZ de de fdZ
xZS )r ?N 皙? num_classes
width_multinverted_residual_setting
round_nearestblock.r dropoutr c t t | |t }| t j }d}d} |g dg dg dg dg dg d g d
g}t
| d k st
|d d k rt d
| t ||z | }t | t d| z | | _
t d|d|t j g}
|D ][\ } } }
}t | |z | }t |
D ]1}|d k r|nd}|
||||| | |}2\|
t || j
d|t j t j |
| _ t j t j | t j | j
| | _ | D ]B}t+ |t j rRt j |j d |j $t j |j ot+ |t j t j f rIt j |j t j |j t+ |t j rJt j |j d d t j |j DdS )aw
MobileNet V2 main class
Args:
num_classes (int): Number of classes
width_mult (float): Width multiplier - adjusts number of channels in each layer by this amount
inverted_residual_setting: Network structure
round_nearest (int): Round the number of channels in each layer to be a multiple of this number
Set to 1 to turn off rounding
block: Module specifying inverted residual building block for mobilenet
norm_layer: Module specifying the normalization layer to use
dropout (float): The droupout probability
N i )r r r ) r
r
)rU rS r
)rU @ r
)rU ` rW r )rU rW r
)rU i@ r r r rY zGinverted_residual_setting should be non-empty or a 4-element list, got rI rW r
)r r r$ r )r r r" )pfan_out)modeg{Gz?) r' r( r
r r r* lenr) r max last_channelr r/ ranger. r2 featuresDropoutLinear
classifiermodules
isinstancer1 initkaiming_normal_weightr& zeros_ GroupNormones_normal_)r6 rL rM rN rO rP r rQ
input_channelra rc tcnsoutput_channelir mr9 s r: r( zMobileNetV2.__init__D s> 0 D!!! =$E J
$ ,
) % ( ) )Q . .#6OPQ6R2S2SWX2X2X uZsuu
(
(BMRR
+L3sJ;O;O,OQ^ _ _ M!
egem n n n%
4 / /JAq!Q ,Q^] K KN1XX
/
/ 1ff!m^VZ[hr s s sttt .
/
t 0aJikiq
x0
- J ! ! ! Id' 5 5
'
'A!RY''
'''y'AAA6%GNN16***A =>>
'
ah'''qv&&&&Ary))
'!T222qv&&&
'
'r; r< c | | }t j |d }t j |d }| | }|S )Nr r r )rc r
functionaladaptive_avg_pool2dtorchflattenrf r? s r:
_forward_implzMobileNetV2._forward_impl sS
MM! M - -a 8 8 M!Q OOA r; c , | | S r> )r~ r? s r: r@ zMobileNetV2.forward s !!!$$$r; )rH rI NrJ NNrK )rA rB rC r+ floatr r r r rD r( r r~ r@ rE rF s @r: r r C s( ?C 489= ]' ]']' ]' $,DcO#< ]'
]' bi 01
]' Xc29n 56]' ]'
]' ]' ]' ]' ]' ]'~v & % %F % % % % % % % %r; r iz5 ry )
num_paramsmin_size
categoriesc e Zd Z ed eed i edddddid d
d d
Z ed eedd i edddddid ddd
ZeZ dS )r z=https://download.pytorch.org/models/mobilenet_v2-b0353104.pth ) crop_sizezQhttps://github.com/pytorch/vision/tree/main/references/classification#mobilenetv2z ImageNet-1Kgx&1Q@gMV@)zacc@1zacc@5g$C?g\( +@zXThese weights reproduce closely the results of the paper using a simple training recipe.)recipe_metrics_ops
_file_size_docs)url
transformsmetaz=https://download.pytorch.org/models/mobilenet_v2-7ebf99e0.pth )r resize_sizezHhttps://github.com/pytorch/vision/issues/3995#new-recipe-with-reg-tuningg`" R@gS㥛V@gV-2+@a$
These weights improve upon the results of the original paper by using a modified version of TorchVision's
`new training recipe
`_.
N)
rA rB rC r r r _COMMON_META
IMAGENET1K_V1
IMAGENET1K_V2DEFAULT r; r: r r s G K7.#>>>
i # # s
M" G K7.#3OOO
` # #
M* GGGr; r
pretrained)weightsT)r progressr r kwargsr c t | } | )t |dt | j d t di |}| *| | |d |S )a MobileNetV2 architecture from the `MobileNetV2: Inverted Residuals and Linear
Bottlenecks `_ paper.
Args:
weights (:class:`~torchvision.models.MobileNet_V2_Weights`, optional): The
pretrained weights to use. See
:class:`~torchvision.models.MobileNet_V2_Weights` below for
more details, and possible values. By default, no pre-trained
weights are used.
progress (bool, optional): If True, displays a progress bar of the
download to stderr. Default is True.
**kwargs: parameters passed to the ``torchvision.models.mobilenetv2.MobileNetV2``
base class. Please refer to the `source code
`_
for more details about this class.
.. autoclass:: torchvision.models.MobileNet_V2_Weights
:members:
NrL r T)r
check_hashr )r verifyr r_ r r load_state_dictget_state_dict)r r r models r: r r s 0 #))'22G fmS l9S5T5TUUU ! !& ! !E
g 4 4hSW 4 X XYYY Lr; )# functoolsr typingr r r r r| r r ops.miscr transforms._presetsr utilsr
_apir r r _metar _utilsr r r __all__rD r r r r r boolr r r; r: r s 0 0 0 0 0 0 0 0 0 0 0 0 + + + + + + 5 5 5 5 5 5 ' ' ' ' ' ' 6 6 6 6 6 6 6 6 6 6 ' ' ' ' ' ' S S S S S S S S S S B
A
A- - - - - ry - - - `k% k% k% k% k%") k% k% k%^ & ' ' ' ' ' ; ' ' ' T ,0D0R!STTT15 -. AE X[ UT r;