:j
d dl mZ d dlmZmZmZmZ d dlZd dlm Z d dlm
Z
ddl m Z ddl
mZ dd lmZmZmZ dd
lmZ dd lmZmZ g d ZdEd
ededededede j4 f dZdFd
ededede j4 fdZ G d de j: Z G d de j: Z G d de j: Z de!ee e f d e"e d ee d e#d ede f d Z$d!ed"Z% G d# d$e Z& G d% d&e Z' G d' d(e Z( G d) d*e Z) G d+ d,e Z* G d- d.e Z+ G d/ d0e Z, G d1 d2e Z- G d3 d4e Z. G d5 d6e Z/ e ed7e&j` f8 dd9d:d ee& d e#d ede fd; Z1 e ed7e'j` f8 dd9d:d ee' d e#d ede fd< Z2 e ed7e(j` f8 dd9d:d ee( d e#d ede fd= Z3 e ed7e)j` f8 dd9d:d ee) d e#d ede fd> Z4 e ed7e*j` f8 dd9d:d ee* d e#d ede fd? Z5 e ed7e+j` f8 dd9d:d ee+ d e#d ede fd@ Z6 e ed7e,j` f8 dd9d:d ee, d e#d ede fdA Z7 e ed7e-j` f8 dd9d:d ee- d e#d ede fdB Z8 e ed7e.j` f8 dd9d:d ee. d e#d ede fdC Z9 e ed7e/j` f8 dd9d:d ee/ d e#d ede fdD Z:y)G )partial)AnyCallableOptionalUnionN)Tensor )ImageClassification)_log_api_usage_once )register_modelWeights WeightsEnum)_IMAGENET_CATEGORIES)_ovewrite_named_paramhandle_legacy_interface)ResNetResNet18_WeightsResNet34_WeightsResNet50_WeightsResNet101_WeightsResNet152_WeightsResNeXt50_32X4D_WeightsResNeXt101_32X8D_WeightsResNeXt101_64X4D_WeightsWide_ResNet50_2_WeightsWide_ResNet101_2_Weightsresnet18resnet34resnet50 resnet101 resnet152resnext50_32x4dresnext101_32x8dresnext101_64x4dwide_resnet50_2wide_resnet101_2 in_planes
out_planesstridegroupsdilationreturnc
< t j | |d|||d| S )z 3x3 convolution with padding F) kernel_sizer* paddingr+ biasr, nnConv2d)r( r) r* r+ r, s n/home/mcse/projects/srt_converter/srt-converter-venv/lib/python3.12/site-packages/torchvision/models/resnet.pyconv3x3r7 ( s+
99
c 6 t j | |d|d S )z1x1 convolutionr F)r0 r* r2 r3 )r( r) r* s r6 conv1x1r: 6 s
99Y
&u UUr8 c e Zd ZU dZeed< ddedededeej ded ed
ed ee d ej f d
df fd
Z
de d
e fdZ xZ
S )
BasicBlockr expansionNinplanesplanesr*
downsampler+
base_widthr,
norm_layer.r- c \ t | |t j }|dk7 s|dk7 r t d |dkD r t d t
||| | _ || | _ t j d | _
t
|| | _ || | _ || _
|| _ y )Nr @ z3BasicBlock only supports groups=1 and base_width=64z(Dilation > 1 not supported in BasicBlockTinplace)super__init__r4 BatchNorm2d
ValueErrorNotImplementedErrorr7 conv1bn1ReLUreluconv2bn2r@ r* )
selfr> r? r* r@ r+ rA r, rB __class__s
r6 rH zBasicBlock.__init__> s J Q;** RS S a<%&PQ Q Xvv6
f%GGD) VV,
f%$ r8 xc |}| j | }| j | }| j | }| j | }| j | }| j
| j | }||z
}| j | }|S N)rL rM rO rP rQ r@ rR rT identityouts r6 forwardzBasicBlock.forwardY s{ jjmhhsmiinjjohhsm ?? &q)H xiin
r8 r Nr rD r N__name__
__module__ __qualname__r= int__annotations__r r4 Moduler rH r rZ
__classcell__rS s @r6 r< r< ; s Is *. 9=
RYY'
Xc299n 56
6 F r8 r< c e Zd ZU dZeed< ddedededeej ded ed
ed ee d ej f d
df fd
Z
de d
e fdZ xZ
S )
Bottleneck r= Nr> r? r* r@ r+ rA r, rB .r- c t
| |t j }t ||dz z |z } t || | _ || | _ t | | ||| | _ || | _
t | || j z | _ ||| j z | _
t j d | _ || _ || _ y )Ng P@TrE )rG rH r4 rI r` r: rL rM r7 rP rQ r= conv3bn3rN rO r@ r* ) rR r> r? r* r@ r+ rA r, rB widthrS s r6 rH zBottleneck.__init__u s JFj4 /01F: Xu-
e$ UE668D
e$ UFT^^$;<
ft~~ 56GGD) $ r8 rT c |}| j | }| j | }| j | }| j | }| j | }| j | }| j | }| j
| }| j | j | }||z
}| j | }|S rV )rL rM rO rP rQ ri rj r@ rW s r6 rZ zBottleneck.forward s jjmhhsmiinjjohhsmiinjjohhsm ?? &q)H xiin
r8 r[ r\ rd s @r6 rf rf l s Is *. 9=
RYY'
Xc299n 56
4 F r8 rf c e Zd Z ddeeeef dee dede dedede
ee d e
e d
e j f d df fd
Z ddeeeef d
ededede d e j f dZded efdZded efdZ xZS )r Nblocklayers num_classeszero_init_residualr+ width_per_group replace_stride_with_dilationrB .r- c t
| t | |t j }|| _ d| _ d| _ |g d}t | dk7 rt d| || _
|| _ t j d| j dddd | _
|| j | _ t j d
| _ t j" ddd | _ | j' |d|d
| _ | j' |d|d d|d
| _ | j' |d|d d|d | _ | j' |d|d d|d | _ t j0 d | _ t j4 d|j6 z | | _ | j; D ] } t= | t j r-t j> jA | jB dd Jt= | t j t jD f sut j> jG | jB d t j> jG | jH d
|r| j; D ] } t= | tJ rK| jL jB 5t j> jG | jL jB d
^t= | tN so| jP jB t j> jG | jP jB d
y y )NrD r )FFFr/ zFreplace_stride_with_dilation should be None or a 3-element tuple, got r F)r0 r* r1 r2 TrE )r0 r* r1 r )r* dilate i r r fan_outrO )mode nonlinearity))rG rH r r4 rI _norm_layerr> r, lenrJ r+ rA r5 rL rM rN rO MaxPool2dmaxpool _make_layerlayer1layer2layer3layer4AdaptiveAvgPool2davgpoolLinearr= fcmodules
isinstanceinitkaiming_normal_weight GroupNorm constant_r2 rf rj r< rQ ) rR rn ro rp rq r+ rr rs rB mrS s r6 rH zResNet.__init__ s D! J%
' / ,A ( + , 1 --I,JL
)YYq$--QqRSZ_`
dmm,GGD) ||!QG &&ub&)< &&uc6!9QOklmOn&o &&uc6!9QOklmOn&o &&uc6!9QOklmOn&o ++F3 ))C%// 1;? -A!RYY'''yv'VA =>!!!((A.!!!&&!, - \\^
7 a, 1IGG%%aeellA6 :.155<<3KGG%%aeellA6
7 r8 r? blocksr* rw c | j }d }| j }|r| xj |z c_ d}|dk7 s | j ||j z k7 rKt j
t
| j ||j z | |||j z }g } | j || j |||| j | j || ||j z | _ t d| D ]G }
| j || j || j | j | j | I t j
| S )Nr )r+ rA r, rB ) r} r, r> r= r4
Sequentialr: appendr+ rA range) rR rn r? r r* rw rB r@ previous_dilationro _s r6 r zResNet._make_layer s0 %%
MM MMV #MF Q;$--6EOO+CC
v'?H6EOO34J
vvz4;;Yjlv
0
q&!
A MMMM ;; # !]] )
}}f%%r8 rT c | j | }| j | }| j | }| j | }| j | }| j | }| j
| }| j | }| j | }t j |d }| j | }|S )Nr ) rL rM rO r r r r r r torchflattenr rR rT s r6
_forward_implzResNet._forward_impl
s JJqM HHQK IIaL LLO KKN KKN KKN KKN LLO MM!Q GGAJr8 c $ | j | S rV )r r s r6 rZ zResNet.forward s !!!$$r8 )i Fr rD NN)r F)r] r^ r_ typer r< rf listr` boolr r r4 rb rH r r r r rZ rc rd s @r6 r r s$
#( !=A9=87E*j01287 S 87 87
! 87
87 87 '/tDz&:87 Xc299n 5687
87~
'&E*j012'& '& '&
'&
'&
'&Rv & $% %F %r8 r rn ro weightsprogresskwargsc |#t |dt |j d t | |fi |}|"|j |j |d |S )Nrp
categoriesT)r
check_hash)r r~ metar load_state_dictget_state_dict)rn ro r r r models r6 _resnetr s_ fmS l9S5TU 5& +F +E
g 4 4hSW 4 XY Lr8 ry )min_sizer c V e Zd Z ed eed i edddddd id
d d d
ZeZy)r z9https://download.pytorch.org/models/resnet18-f37072fd.pth crop_sizei(^ Lhttps://github.com/pytorch/vision/tree/main/references/classification#resnet ImageNet-1KgnpQ@gEDV@zacc@1zacc@5g/$?gS㥛TF@XThese weights reproduce closely the results of the paper using a simple training recipe.
num_paramsrecipe_metrics_ops
_file_size_docsurl
transformsr N r] r^ r_ r r r
_COMMON_META
IMAGENET1K_V1DEFAULT r8 r6 r r 8 W G.#>
"d # # s
M$ Gr8 r c V e Zd Z ed eed i edddddd id
d d d
ZeZy)r z9https://download.pytorch.org/models/resnet34-b627a593.pthr r i(Lr r gjtTR@g{GV@r g Zd;O
@gT@r r r Nr r r8 r6 r r N r r8 r c e Zd Z ed eed i edddddd id
d d d
Z ed eedd i edddddd id
ddd
ZeZ y)r z9https://download.pytorch.org/models/resnet50-0676ba61.pthr r i(r r g QS@gI +7W@r gB`"[@gDlqX@r r r z9https://download.pytorch.org/models/resnet50-11ad3fa6.pth r resize_sizezEhttps://github.com/pytorch/vision/issues/3995#issuecomment-1013906621gx6T@gW@g(\rX@
These weights improve upon the results of the original paper by using TorchVision's `new training recipe
`_.
N
r] r^ r_ r r r
r r
IMAGENET1K_V2r r r8 r6 r r d s G.#>
"d # # s
M$ G.#3O
"] # #
M* Gr8 r c e Zd Z ed eed i edddddd id
d d d
Z ed eedd i edddddd id
ddd
ZeZ y)r z:https://download.pytorch.org/models/resnet101-63fe2227.pthr r i(ħr r g-WS@gmbW@r gNbX94 @g1 ZPe@r r r z:https://download.pytorch.org/models/resnet101-cd907fc2.pthr r 8https://github.com/pytorch/vision/issues/3995#new-recipegbX9xT@gR W@g)\Pe@r Nr r r8 r6 r r s H.#>
"d # # !s
M$ H.#3O
"P # #
M* Gr8 r c e Zd Z ed eed i edddddd id
d d d
Z ed eedd i edddddd id
ddd
ZeZ y)r z:https://download.pytorch.org/models/resnet152-394f9c45.pthr r i(xr r gS@gmW@r gI +'@gSl@r r r z:https://download.pytorch.org/models/resnet152-f82ba261.pthr r r gV-T@g㥛 X@gI +l@r Nr r r8 r6 r r s H.#>
"d # # !s
M$ H.#3O
"P # # !
M* Gr8 r c e Zd Z ed eed i edddddd id
d d d
Z ed eedd i edddddd id
ddd
ZeZ y)r z@https://download.pytorch.org/models/resnext50_32x4d-7cdf4587.pthr r i(}Mhttps://github.com/pytorch/vision/tree/main/references/classification#resnextr gd;OgS@g&1lW@r gQ @g"~W@r r r z@https://download.pytorch.org/models/resnext50_32x4d-1a0047aa.pthr r r g&1LT@g(\W@gZd;OW@r Nr r r8 r6 r r s N.#>
"e # # s
M$ N.#3O
"P # #
M* Gr8 r c e Zd Z ed eed i edddddd id
d d d
Z ed eedd i edddddd id
ddd
ZeZ y)r zAhttps://download.pytorch.org/models/resnext101_32x8d-8ba56ff5.pthr r i(Jr r gS@glW@r gDli0@gL7A`9u@r r r zAhttps://download.pytorch.org/models/resnext101_32x8d-110c445d.pthr r Dhttps://github.com/pytorch/vision/issues/3995#new-recipe-with-fixresgL7A`T@g;OX@gT㥛:u@r Nr r r8 r6 r r s O.#>
"e # # !s
M$ O.#3O
"\ # # !
M* Gr8 r c X e Zd Z ed eedd i eddddd d
id d d
d ZeZy)r zAhttps://download.pytorch.org/models/resnext101_64x4d-173b62eb.pthr r r i(mz+https://github.com/pytorch/vision/pull/5935r g9vT@gV X@r gQ .@g +s@z
These weights were trained from scratch by using TorchVision's `new training recipe
`_.
r r Nr r r8 r6 r r ; s[ O.#3O
"C # # !
M* Gr8 r c e Zd Z ed eed i edddddd id
d d d
Z ed eedd i edddddd id
ddd
ZeZ y)r z@https://download.pytorch.org/models/wide_resnet50_2-95faca4d.pthr r i(:https://github.com/pytorch/vision/pull/912#issue-445437439r gˡES@g/$W@r g&@g
ףp=z`@r r r z@https://download.pytorch.org/models/wide_resnet50_2-9ba9bcbe.pthr r r gJ +fT@gnW@gDlqp@r Nr r r8 r6 r r T s N.#>
"R # # s
M$ N.#3O
"\ # # !
M* Gr8 r c e Zd Z ed eed i edddddd id
d d d
Z ed eedd i edddddd id
ddd
ZeZ y)r zAhttps://download.pytorch.org/models/wide_resnet101_2-32ee1156.pthr r i(#r r gʡES@gV-W@r gT㥛6@g&1\n@r r r zAhttps://download.pytorch.org/models/wide_resnet101_2-d733dc28.pthr r r gq=
ףT@gzGX@gˡEK~@r Nr r r8 r6 r r s O.#>
#R # # !s
M$ O.#3O
#P # # !
M* Gr8 r
pretrained)r T)r r c V t j | } t t g d| |fi |S )ap ResNet-18 from `Deep Residual Learning for Image Recognition `__.
Args:
weights (:class:`~torchvision.models.ResNet18_Weights`, optional): The
pretrained weights to use. See
:class:`~torchvision.models.ResNet18_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.resnet.ResNet``
base class. Please refer to the `source code
`_
for more details about this class.
.. autoclass:: torchvision.models.ResNet18_Weights
:members:
)r r r r )r verifyr r< r r r s r6 r r * * %%g.G :|Wh I& IIr8 c V t j | } t t g d| |fi |S )ap ResNet-34 from `Deep Residual Learning for Image Recognition `__.
Args:
weights (:class:`~torchvision.models.ResNet34_Weights`, optional): The
pretrained weights to use. See
:class:`~torchvision.models.ResNet34_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.resnet.ResNet``
base class. Please refer to the `source code
`_
for more details about this class.
.. autoclass:: torchvision.models.ResNet34_Weights
:members:
r/ rg r/ )r r r r< r s r6 r r r r8 c V t j | } t t g d| |fi |S )a ResNet-50 from `Deep Residual Learning for Image Recognition `__.
.. note::
The bottleneck of TorchVision places the stride for downsampling to the second 3x3
convolution while the original paper places it to the first 1x1 convolution.
This variant improves the accuracy and is known as `ResNet V1.5
`_.
Args:
weights (:class:`~torchvision.models.ResNet50_Weights`, optional): The
pretrained weights to use. See
:class:`~torchvision.models.ResNet50_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.resnet.ResNet``
base class. Please refer to the `source code
`_
for more details about this class.
.. autoclass:: torchvision.models.ResNet50_Weights
:members:
r )r r r rf r s r6 r r s* 6 %%g.G :|Wh I& IIr8 c V t j | } t t g d| |fi |S )a ResNet-101 from `Deep Residual Learning for Image Recognition `__.
.. note::
The bottleneck of TorchVision places the stride for downsampling to the second 3x3
convolution while the original paper places it to the first 1x1 convolution.
This variant improves the accuracy and is known as `ResNet V1.5
`_.
Args:
weights (:class:`~torchvision.models.ResNet101_Weights`, optional): The
pretrained weights to use. See
:class:`~torchvision.models.ResNet101_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.resnet.ResNet``
base class. Please refer to the `source code
`_
for more details about this class.
.. autoclass:: torchvision.models.ResNet101_Weights
:members:
r/ rg r/ )r r r rf r s r6 r! r! * 6 &&w/G :}gx J6 JJr8 c V t j | } t t g d| |fi |S )a ResNet-152 from `Deep Residual Learning for Image Recognition `__.
.. note::
The bottleneck of TorchVision places the stride for downsampling to the second 3x3
convolution while the original paper places it to the first 1x1 convolution.
This variant improves the accuracy and is known as `ResNet V1.5
`_.
Args:
weights (:class:`~torchvision.models.ResNet152_Weights`, optional): The
pretrained weights to use. See
:class:`~torchvision.models.ResNet152_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.resnet.ResNet``
base class. Please refer to the `source code
`_
for more details about this class.
.. autoclass:: torchvision.models.ResNet152_Weights
:members:
)r/ $ r/ )r r r rf r s r6 r" r" r r8 c t j | } t |dd t |dd t t g d| |fi |S )a ResNeXt-50 32x4d model from
`Aggregated Residual Transformation for Deep Neural Networks `_.
Args:
weights (:class:`~torchvision.models.ResNeXt50_32X4D_Weights`, optional): The
pretrained weights to use. See
:class:`~torchvision.models.ResNext50_32X4D_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.resnet.ResNet``
base class. Please refer to the `source code
`_
for more details about this class.
.. autoclass:: torchvision.models.ResNeXt50_32X4D_Weights
:members:
r+ rr rg r )r r r r rf r s r6 r# r# > sC . &,,W5G&(B/&"3Q7 :|Wh I& IIr8 c t j | } t |dd t |dd t t g d| |fi |S )a ResNeXt-101 32x8d model from
`Aggregated Residual Transformation for Deep Neural Networks `_.
Args:
weights (:class:`~torchvision.models.ResNeXt101_32X8D_Weights`, optional): The
pretrained weights to use. See
:class:`~torchvision.models.ResNeXt101_32X8D_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.resnet.ResNet``
base class. Please refer to the `source code
`_
for more details about this class.
.. autoclass:: torchvision.models.ResNeXt101_32X8D_Weights
:members:
r+ r rr r r )r r r r rf r s r6 r$ r$ \ C . '--g6G&(B/&"3Q7 :}gx J6 JJr8 c t j | } t |dd t |dd t t g d| |fi |S )a ResNeXt-101 64x4d model from
`Aggregated Residual Transformation for Deep Neural Networks `_.
Args:
weights (:class:`~torchvision.models.ResNeXt101_64X4D_Weights`, optional): The
pretrained weights to use. See
:class:`~torchvision.models.ResNeXt101_64X4D_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.resnet.ResNet``
base class. Please refer to the `source code
`_
for more details about this class.
.. autoclass:: torchvision.models.ResNeXt101_64X4D_Weights
:members:
r+ rD rr rg r )r r r r rf r s r6 r% r% z r r8 c p t j | } t |dd t t g d| |fi |S )a Wide ResNet-50-2 model from
`Wide Residual Networks `_.
The model is the same as ResNet except for the bottleneck number of channels
which is twice larger in every block. The number of channels in outer 1x1
convolutions is the same, e.g. last block in ResNet-50 has 2048-512-2048
channels, and in Wide ResNet-50-2 has 2048-1024-2048.
Args:
weights (:class:`~torchvision.models.Wide_ResNet50_2_Weights`, optional): The
pretrained weights to use. See
:class:`~torchvision.models.Wide_ResNet50_2_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.resnet.ResNet``
base class. Please refer to the `source code
`_
for more details about this class.
.. autoclass:: torchvision.models.Wide_ResNet50_2_Weights
:members:
rr rv r )r r r r rf r s r6 r& r& s7 8 &,,W5G&"3V< :|Wh I& IIr8 c p t j | } t |dd t t g d| |fi |S )a Wide ResNet-101-2 model from
`Wide Residual Networks `_.
The model is the same as ResNet except for the bottleneck number of channels
which is twice larger in every block. The number of channels in outer 1x1
convolutions is the same, e.g. last block in ResNet-101 has 2048-512-2048
channels, and in Wide ResNet-101-2 has 2048-1024-2048.
Args:
weights (:class:`~torchvision.models.Wide_ResNet101_2_Weights`, optional): The
pretrained weights to use. See
:class:`~torchvision.models.Wide_ResNet101_2_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.resnet.ResNet``
base class. Please refer to the `source code
`_
for more details about this class.
.. autoclass:: torchvision.models.Wide_ResNet101_2_Weights
:members:
rr rv r )r r r r rf r s r6 r' r' s7 8 '--g6G&"3V< :}gx J6 JJr8 )r r r )r ); functoolsr typingr r r r r torch.nnr4 r transforms._presetsr
utilsr _apir
r r _metar _utilsr r __all__r` r5 r7 r: rb r< rf r r r r r r r r r r r r r r r r r r r r r! r" r# r$ r% r& r' r r8 r6 r st 1 1 5 ' 6 6 ' B 2 s S c Y\ egenen Vs V VS V V
. .b7 7tw%RYY w%t j*,- . I k
"
& & { , { ,( { ( V( ( V( ( V( k ( V( { ( V { 2( k ( V( { ( V ,0@0N0N!OP6:T J"23 Jd J]` Jek J Q J0 ,0@0N0N!OP6:T J"23 Jd J]` Jek J Q J0 ,0@0N0N!OP6:T J"23 Jd J]` Jek J Q J< ,0A0O0O!PQ8