Track-Anything/model/__pycache__/network.cpython-39.pyc

a
_{3dp<00>@sJdZd
dl
Z
d
dlmZ
d
dlmZ
d
dlTd
dlTGdd<06>dej<08>Z	dS)z<>
This file defines XMem, the highest level nn.Module interface
During training, it is used by trainer.py
During evaluation, it is used by inference_core.py

It further depends on modules.py which gives more detailed implementations of sub-modules
<EFBFBD>N)
<01>	aggregate)
<01>
*cs`eZ
dZd<14>f
dd<03>	Zddd<06>
Zddd<08>
Zd	d
<EFBFBD>Zddd<0C>
Zd
d<0E>Zddd<10>
Z	ddd<13>
Z
<EFBFBD>ZS)<1A>XMemNcs<>t<00><00>
<01>
|<00>|
||<03>}|
<01>d
d<02>|_td|j<04><00><02>

t<06>|_t|j	|j
|j<04>|_td|j
<0A>|_t|j	|j
<EFBFBD>|_|du
r<>|j|dd<07>
dS)z<>
        model_path/map_location are used in evaluation only
        map_location is for converting models saved in cuda to cpu
        <20>
single_objectFzSingle object mode: iNT)
<01>init_as_zero_if_needed)<12>super<65>__init__<5F>init_hyperparameters<72>getr<00>printZ
KeyEncoder<EFBFBD>key_encoderZValueEncoder<65>	value_dim<69>
hidden_dim<EFBFBD>
value_encoderZ
KeyProjection<6F>key_dim<69>key_proj<6F>Decoder<65>decoder<65>load_weights<74><05>self<6C>config<69>
model_path<EFBFBD>map_locationZ
model_weights<74>
<01>	__class__<5F><00>-/home/gaomingqi/SAM/xmem-sam/model/network.pyrs




z
XMem.__init__Tc
Csp
t|
j
<01>
d
kr4d}|
j
dd<03>\}}|
jddd<06>}
nt|
j
<01>
dkrHd}nt<03>
|<00>|
<01>
\}}}	|<00>|||<03>\}
}}|<04>
r`|
j||g|
j
d	d<00><00>
R<00><00>dd<03><02><08>}
|du
r<>|j||g|j
d	d<00><00>
R<00><00>dd<03><02><08>}|du
<01>
r|j||g|j
d	d<00><00>
R<00><00>dd<03><02><08>}|j||g|j
d	d<00><00>
R<00>}|j||g|j
d	d<00><00>
R<00>}|	j||g|	j
d	d<00><00>
R<00>}	|
|||||	fS)
N<EFBFBD>T<>r
<00>
<00><02>	start_dim<69>end_dim<69>F<><46><EFBFBD><EFBFBD><EFBFBD>)	<09>len<65>shape<70>flatten<65>NotImplementedErrorrr<00>view<65>	transpose<73>
contiguous)
r<00>frameZneed_skZneed_ekZneed_reshape<70>
b<>
t<>f16<31>f8<66>f4<66>key<65>	shrinkage<67>	selectionrrr<00>
encode_key(s&


*

*


*

zXMem.encode_keycs^<00>jd
<00>
<01>
d
kr4t
<01><02><00>
fdd<03>t<03>
<01>
D<00>
d
<01>}n
t
<01><04><00>
}|<00>|
||<03>||<05>\}}||fS)Nr c

s:g|]2<>tj
<01>
dd<00><02>f
d
d<02>t<02><02>
D<00>
fddd<05><03>qS)Nc
sg|]}
<01>|
kr|
<01>qSrr)<02>.0<EFBFBD>
j<>
<01>
irr<00>
<listcomp>M<00>z0XMem.encode_value.<locals>.<listcomp>.<listcomp>r T)<02>dim<69>keepdim)<03>torch<63>sum<75>range)
r7<00><02>masks<6B>num_objectsr9rr;Ks
<02>
 
<04>z%XMem.encode_value.<locals>.<listcomp>)r'r?<00>catrA<00>
zeros_liker)rr-Zimage_feat_f16Zh16rC<00>is_deep_update<74>othersZg16rrBr<00>encode_valueHs



<06><02>
zXMem.encode_valuec
Cs^|jd
d<02>\}}|j
ddd<04>}t|||
|<02>}t||<05>}	|	j|||jg|	jdd
<01><00>
R<00>}	|	S)z<>
        query_key       : B * CK * H * W
        query_selection : B * CK * H * W
        memory_key      : B * CK * T * H * W
        memory_shrinkage: B * 1  * T * H * W
        memory_value    : B * num_objects * CV * T * H * W
        Nrr r!<00><><EFBFBD><EFBFBD><EFBFBD>)r'r(Zget_affinityZreadoutr*r
)
rZ	query_keyZquery_selectionZ
memory_keyZmemory_shrinkageZmemory_value<75>
batch_sizerDZaffinity<74>memoryrrr<00>read_memoryYs	



"zXMem.read_memoryc	Csp|jg|
<01>
|<03>
|<02>
Rd
|i
<01>
\}}t
<01>|<07>
}|du
r<||}t|ddd<04>\}}|rf|dd<00>dd<00>f}|||fS)N<>h_outr T)r=Z
return_logits)rr?<00>sigmoidr)	rZmulti_scale_features<65>memory_readoutZhidden_state<74>selectorrNZstrip_bg<62>logits<74>probrrr<00>segmentks"




zXMem.segmentcOsh|
d
kr|j|i|<03>
<01>
S|
dkr0|j
|i|<03>
<01>
S|
dkrH|j|i|<03>
<01>
S|
dkr`|j|i|<03>
<01>
St<04>
dS)Nr6rIrMrT)r6rIrMrTr))r<00>mode<64>args<67>kwargsrrr<00>forwardzs







zXMem.forwardcCs4
|d
u
r<>tj
||d<02>}|djd|_|djd|_d|v
|_|jrNd|_n|djdd|_td|j<03>d	|j<04>d
|j<06><00><06>

n<>d
}d|
v
r<>d|_td
|j<03><00><02>

n
|
d|_d|
v
r<>d|_td|j<04><00><02>

n
|
d|_d|
v
r<>d|_td|j<06><00><02>

n
|
d|_|jdk
|_|j|
d<|j|
d<|j|
d<|S)a

        Init three hyperparameters: key_dim, value_dim, and hidden_dim
        If model_path is provided, we load these from the model weights
        The actual parameters are then updated to the config in-place

        Otherwise we load it either from the config or default
        N)
rzkey_proj.key_proj.weightr
z'value_encoder.fuser.block2.conv2.weightz&decoder.hidden_update.transform.weight<68>z1Hyperparameters read from the model weights: C^k=z, C^v=z, C^h=r<00>@z,key_dim not found in config. Set to default r
iz.value_dim not found in config. Set to default rz/hidden_dim not found in config. Set to default )r?<00>loadr'rr
Zdisable_hiddenrrrrrrr	<00>sB







<04>
<04>
<04>














zXMem.init_hyperparametersFcCs<>t|
<01>
<01><00>
D]t}|d
kr|
|jddkrtd<04>

tjd|
|jd<06>}|s`td<07>

tj<08>	|<04>

ntd<08>

t<04>
|
||gd<02>|
|<q|<00>|
<01>

dS)	Nzvalue_encoder.conv1.weightr r$z:Converting weights from single object to multiple objects.)rZr <00>r\)
<01>devicezRandomly initialized padding.zZero-initialized padding.)<0C>list<73>keysr'rr?<00>zerosr]<00>nn<6E>init<69>orthogonal_rE<00>load_state_dict)r<00>src_dictr<00>
kZpadsrrrr<00>s







zXMem.load_weights)NN)TT)
T)NTT)NN)
F)<0C>__name__<5F>
__module__<EFBFBD>__qualname__rr6rIrMrTrXr	r<00>
__classcell__rrrrrs

 
<00>

2r)
<EFBFBD>__doc__r?<00>torch.nnraZmodel.aggregaterZ
model.modulesZmodel.memory_util<69>Modulerrrrr<00><module>
s