Name	Size	Mode	Actions
config.cpython-36.pyc	20837	0644	editdlrm
convert.cpython-36.pyc	12845	0644	editdlrm
features.cpython-36.pyc	10214	0644	editdlrm
utils.cpython-36.pyc	1524	0644	editdlrm
__init__.cpython-36.pyc	876	0644	editdlrm
__main__.cpython-36.pyc	2702	0644	editdlrm

3 �Eg\f�@s:
dd
lZdd
l
Z
dd
lZddlmZmZ
ddlmZ
ddlm Z m Z mZmZm Z mZmZmZmZmZ
dd
lZddlmZ
ddlmZmZmZmZ
dd lmZmZmZ
e r�dd lm Z
ddl!m"Z"
ddl#m$Z$
e�r�dd l%m&Z&
ej'e(�
Z)dZ*dZ+e
j,Gdd�d��
Z-Gdd�de�Z.Gdd�de.e�Z/Gdd�de/�Z0d
S)�N)�ABC�abstractmethod)
�OrderedDict) � TYPE_CHECKING�Any�Callable�Dict�Iterable�List�Mapping�Optional�Tuple�Union)
�version�)� TensorType�is_torch_available�is_vision_available�logging�
)�ParameterFormat� compute_effective_axis_dimension�"compute_serialized_parameters_size)
�PretrainedConfig)
�FeatureExtractionMixin)
�PreTrainedTokenizerBase)
�Image�ic@s6eZ
dZUd
Zeee dZ ee dZeedS)�PatchingSpeca�
Data class that holds patching specifications. Args: o: Module / object where the op to patch is located name: Name of the op to monkey patch custom_op: Custom op that patches the original op orig_op: Original op that is being patched op_wrapper: Wrapper (optional) that wraps both the original and custom ops. It is useful for ops that are class or static methods for instance. N) �__name__� __module__�__qualname__�__doc__r�
o�str�namer� custom_op�orig_opr� op_wrapper�r)r)�B/usr/local/lib/python3.6/site-packages/transformers/onnx/config.pyr.s 



rc @s:eZ
dZd
ZdZdZejd�
Ze dddd�i
�
e d ddd�i
�
e d ddd�i
�
e d dd d�i
�
e d ddi
i
�
e d ddd�i
�
e d ddi
i
�
e ddd�ddd�d��
e d ddd�i
�
d � Z d>deee d�dd�Zed?dedd�dd��
Zeeeeeeeffd�
dd��
�
Zeeeeeeffd�
dd��
Zeeeeefd�
dd��
Zeed�
dd��
Zeed�
d d!��
Zeed�
d"d#��
Zeed�
d$d%��
Zeed�
d&d'��
Z e!eed(�d)d*��
Z"d@eeeed-�d.d/�Z#dCe$dDeeeee%eeed2eeefd4� d5d6�Z&d7d8�Z'd9d:�Z(eee)ee*eefd;�d 0 )r+�default_fixed_batch)
r=r)r)r*�default_batch_size�s zOnnxConfig.default_batch_sizec


Cstj
S)
zu The default sequence length to use if no other indication Returns: Integer > 0 )r+�default_fixed_sequence)
r=r)r)r*�default_sequence_length�sz"OnnxConfig.default_sequence_lengthc


CstS)
z{ Which onnx opset to use when exporting the model Returns: Integer ONNX Opset version )
�DEFAULT_ONNX_OPSET)
r=r)r)r*�default_onnx_opset�szOnnxConfig.default_onnx_opsetc


Csd
S)z� What absolute tolerance value to use during model conversion validation. Returns: Float absolute tolerance value. g�h㈵��>r))
r=r)r)r*�atol_for_validation�szOnnxConfig.atol_for_validationc
Cs$t�rd
dl
m}

|
|jkSdSdS)z� The minimum PyTorch version required to export the model. Returns: `bool`: Whether the installed version of PyTorch is compatible with the model. r
)
� torch_versionFN)rZtransformers.utilsrS�torch_onnx_minimum_version)r=rSr)r)r*�is_torch_support_available�s
 z%OnnxConfig.is_torch_support_available)�num_parametersr@c

Cst|t
j�tkS)
a
Flag indicating if the model requires using external data format Args: num_parameters: Number of parameter on the model Returns: True if model.num_parameters() * size_of(float32) >= 2Gb False otherwise )rrZFloat�EXTERNAL_DATA_FORMAT_SIZE_LIMIT)
rVr)r)r*�use_external_data_format�s
z#OnnxConfig.use_external_data_format��()� batch_size�num_channels�image_height�image_widthcCsJg}x@t|
�
D]4}t
jj|||�d
}|jtj|jd�
�
jd�
�

qW|S)N��Zuint8ZRGB) �range�np�randomZrandr<rZ fromarrayZastype�convert)r=r[r\r]r^�images�
_�datar)r)r*�_generate_dummy_images�s 



z!OnnxConfig._generate_dummy_imagesrFrr) �preprocessorr[� seq_length�is_pair� frameworkr\r^r]� tokenizerr@c Cs
d
dlm
}
d
dlm}
t|
|�r2| dk r2td�
�
| dk rTtjdt�
t j d�

| }
t|
|�r�t|tj dd �}|
j|�
}t|tj|d �}d j|
jg
�
|g
|} t|
| |d��
St|
| �r�|
jddkr�t|tj d �}|j||||�} t|
| |d��
Std�
�
dS)a, Generate inputs to provide to the ONNX exporter for the specific framework Args: preprocessor: ([`PreTrainedTokenizerBase`] or [`FeatureExtractionMixin`]): The preprocessor associated with this model configuration. batch_size (`int`, *optional*, defaults to -1): The batch size to export the model for (-1 means dynamic axis). seq_length (`int`, *optional*, defaults to -1): The sequence length to export the model for (-1 means dynamic axis). is_pair (`bool`, *optional*, defaults to `False`): Indicate if the input is a pair (sentence 1, sentence 2) framework (`TensorType`, *optional*, defaults to `None`): The framework (PyTorch or TensorFlow) that the tokenizer will generate tensors for. num_channels (`int`, *optional*, defaults to 3): The number of channels of the generated images. image_width (`int`, *optional*, defaults to 40): The width of the generated images. image_height (`int`, *optional*, defaults to 40): The height of the generated images. Returns: Mapping[str, Tensor] holding the kwargs to provide to the model's forward function r)
r)
rNzPYou cannot provide both a tokenizer and a preprocessor to generate dummy inputs.ztThe `tokenizer` argument is deprecated and will be removed in version 5 of Transformers. Use `preprocessor` instead.zSOverwriting the `preprocessor` argument with `tokenizer` to generate dummmy inputs.r
)�fixed_dimensionZnum_token_to_add�
)
�return_tensorsZpixel_values)
rm)rdroz\Unable to generate dummy inputs for the model. Please provide a tokenizer or a preprocessor.)�feature_extraction_utilsr�tokenization_utils_baser� isinstancer6�warnings�warn� FutureWarning�logger�warningrr+rLZnum_special_tokens_to_addrN�joinZ unk_token�dictZmodel_input_namesrg)r=rhr[rirjrkr\r^r]rlrrZtoken_to_addZdummy_inputr)r)r*�generate_dummy_inputs�s0$















z OnnxConfig.generate_dummy_inputsc
Cs@x:|jD]0}
|
j
dkr|
jn |
j
|
j�
}t|
j|
j|�
qWdS)
N)r8r(r&�setattrr#r%)r=r>r&r)r)r*� patch_ops5
s


zOnnxConfig.patch_opsc
Cs@x:|jD]0}
|
j
dkr|
jn |
j
|
j�
}t|
j|
j|�
qWdS)
N)r8r(r'r{r#r%)r=r>r'r)r)r*�restore_ops:
s


zOnnxConfig.restore_ops)r%�fieldr@cs(d
dlm
}
�f
dd�t|j|�
�
D�
S)a�
Flatten any potential nested structure expanding the name of the field with the index of the element within the structure. Args: name: The name of the nested structure field: The structure to, potentially, be flattened Returns: (Dict[str, Any]): Outputs with flattened structure and key mapping this new structure. r
)
�chainc
s i|]\}
}|��d|
���qS)
�
.r))�.0�idx�item)
r%r)r*� O
szAOnnxConfig.flatten_output_collection_property..)� itertoolsr� enumerate� from_iterable)rAr%r~rr))
r%r*�"flatten_output_collection_property?
sz-OnnxConfig.flatten_output_collection_property)r0N)
r0)rrYrZrZ�����r�)r�r�FNrYrZrZN)rr),rr r!r"rLrNr�parserTrr5r$r rr?�classmethodrB�propertyrr�intrDrHrrrKrMrOrQ�floatrR�boolrU�staticmethodrXrgrrrzr|r}r rr�r)r)r)r*r+Csf












" 


 
 
 
 









<
r+c s"
eZ
dZd!deeeed��f
dd� Zed"dedd�d d ��
Z e eeeeeffd�
�f
dd ��
Z e eeeefd�
dd��
Ze ed�
dd��
Ze ed�
dd��
Zd%deeeeeeeefd��f
dd� Zeeeeeffed�dd�Zdd�Zeeeeeefd��f
dd �Z�ZS)&�OnnxConfigWithPastr0NFr)r1r2r3�use_pastcst�j
|
||d
�
||_dS)N)r2r3)�superr?r�)r=r1r2r3r�)
� __class__r)r*r?S
s
zOnnxConfigWithPast.__init__)r1r2r@cCs||
|d
d�S)z� Instantiate a OnnxConfig with `use_past` attribute set to True Args: config: The underlying model's config to use when exporting to ONNX Returns: OnnxConfig with `.use_past = True` T)r2r�r))rAr1r2r)r)r*� with_past]
szOnnxConfigWithPast.with_past)
r@c
s t�j
}
|jr|j|
d
d�
|
S)NrH)
� direction)r�rHr��fill_with_past_key_values_)r=rG)
r�r)r*rHj
s

zOnnxConfigWithPast.outputsc

Cst|j
d
�rd
|ji
SdS)NrI)rJr4r�)
r=r)r)r*rKr
s
z"OnnxConfigWithPast.values_overridec

Cst|j
d
�std�
�
|j
jS)z� The number of layers attribute retrieved from the model config. Override this for model configs where the number of layers attribute is not called `num_layers`. � num_layersz�could not find the number of layers attribute in the model configuration, override the num_layers property of the model OnnxConfig to solve this)rJr4�AttributeErrorr�)
r=r)r)r*r�y
s

zOnnxConfigWithPast.num_layersc

Cst|j
d
�std�
�
|j
jS)z� The number of attention heads attribute retrieved from the model config. Override this for model configs where the number of attention heads attribute is not called `num_attention_heads`. �num_attention_headsz�could not find the number of attention heads attribute in the model configuration, override the num_attention_heads property of the model OnnxConfig to solve this)rJr4r�r�)
r=r)r)r*r��
s

z&OnnxConfigWithPast.num_attention_headsrr)rlr[rirjrkr@c s�t�j
|
||||�}|jr�t�s*td
�
�
nddl}|dj\}} | d} ||j| |jj |jf}d|kr�|j |d|j|| �gdd�|d<g|d<x0t|j �
D]"}|dj|j|�
|j|�
f�

q�W|S) NzACannot generate dummy past_keys inputs without PyTorch installed.r
� input_idsrZattention_maskr)
Zdim�past_key_values)r�rzr�rr6�torch�shaper�r4�hidden_size�catZonesr`r�r<�zeros) r=rlr[rirjrk� common_inputsr�r-ZseqlenZpast_key_values_lengthr�re)
r�r)r*rz�
s$ 








"z(OnnxConfigWithPast.generate_dummy_inputs)�inputs_or_outputsr�cCst|dkrtd|�d��
�
|d
kr$dnd}xFt
|j�
D]8}ddd �|
|�d |�d�<ddd �|
|�d |�d�<q4Wd S)a_
Fill the input_or_ouputs mapping with past_key_values dynamic axes considering. Args: inputs_or_outputs: The mapping to fill. direction: either "inputs" or "outputs", it specifies whether input_or_outputs is the input mapping or the output mapping, this is important for axes naming. rDrHz4direction must either be "inputs" or "outputs", but z was givenr��presentr-zpast_sequence + sequence)r
rr�z.keyz.valueN)rDrH)r6r`r�)r=r�r�r%�
ir)r)r*r��
s 



z-OnnxConfigWithPast.fill_with_past_key_values_cCs4|d
|
|�d|�d�<|d|
|�d|�d�<dS)Nr
r�z.keyrz.valuer))r=�flattened_outputr%r��
tr)r)r*�_flatten_past_key_values_�
s

z,OnnxConfigWithPast._flatten_past_key_values_)r%r~r@csFi}|
dkr4x4t|�
D]\}}|j
||
||�
qWnt�j|
|�}|S)Nr�r�)r�r�)r�r�r�r�)r=r%r~r�r�r�)
r�r)r*r��
s



z5OnnxConfigWithPast.flatten_output_collection_property)r0NF)
r0r�r�)r�r�FN)rr r!r$r rr�r?r�r�r�rr�rHrrrKr�r�rrzr�r�r rr�� __classcell__r)r))
r�r*r�R
s*



$






 r�c s�eZ
dZeeeeeeffd
�
�f
dd��
Zeeed
�
�f
dd��
Z eeed
�
�f
dd��
Z ddeeeee eeefd��f
d d� Zeeeeeffed�dd�Zdd�Z�ZS)�OnnxSeq2SeqConfigWithPast)
r@c
sztt
|�j}
xT|
j�D]H\}}d
|kr*dnd}x.|j�D]"\}}d|krR|||<q8|||<q8WqW|jrv|j|
dd�
|
S)N�encoder�encoder_sequencer/r.rH)
r�)r�r�rH�itemsr�r�)r=rGr%Z axes_namesZ sequence_nameZaxis_idx)
r�r)r*rH�
s





z!OnnxSeq2SeqConfigWithPast.outputsc
s`yt�j
}
|
|
f}
WnFtk rZ


t|jd
�rNt|jd�rN|jj|jjf}
ntd�
�
YnX|
S)N�encoder_layers�decoder_layersz�could not find the number of encoder and decoder layers attributes in the model configuration, override the num_layers property of the model OnnxConfig to solve this)r�r�r�rJr4r�r�)r=r�)
r�r)r*r��
s





z$OnnxSeq2SeqConfigWithPast.num_layersc
s`yt�j
}
|
|
f}
WnFtk rZ


t|jd
�rNt|jd�rN|jj|jjf}
ntd�
�
YnX|
S)N�encoder_attention_heads�decoder_attention_headsz�could not find the number of attention heads for the encoder and the decoder attributes in the model configuration, override the num_attention_heads property of the model OnnxConfig to solve this)r�r�r�rJr4r�r�)r=r�)
r�r)r*r��
s





z-OnnxSeq2SeqConfigWithPast.num_attention_headsrFNr)rlr[rirjrkr@cs�
tt
|�j|
||||�}|js"|nd
}tt
|�j|
||||�}dd�|j�D�
}tf||��
} |j�
r�t�svtd�
�
nddl} | dj d}| dj d
}| dj d
}|j \} }|| ||jj| f}|||d|jj|f}g| d <|j \}}t||�}t||�|}||k�
rd nd}x@t|�
D]4}| d j| j|�
| j|�
| j|�
| j|�
f�

�
q$W|d k�
rj|n|}x2t||�D]$}| d j| j|�
| j|�
f�

�
qzW| S)Nrc
Ssi|]\}
}|d|
���qS)
Zdecoder_r))r�r%Ztensorr)r)r*r�szCOnnxSeq2SeqConfigWithPast.generate_dummy_inputs..zACannot generate dummy past_keys inputs without PyTorch installed.r
r�Zdecoder_input_idsrYr�r��decoder)r�r�rzr�r�ryrr6r�r�r�r4r�r��min�maxr`r<r�)r=rlr[rirjrkZencoder_inputsZdecoder_seq_lengthZdecoder_inputsr�r�r-Zencoder_seq_lengthZnum_encoder_attention_headsZnum_decoder_attention_headsZ encoder_shapeZ decoder_shape�num_encoder_layers�num_decoder_layers�min_num_layers�max_num_layers�remaining_side_namerer�)
r�r)r*rzsL






















$z/OnnxSeq2SeqConfigWithPast.generate_dummy_inputs)r�r�c Cs:
|dkrtd|�d��
�
|d
kr$dnd}|j
\}}t||�}t||�|}||krVdnd}d } |d
krjd nd} xxt|�
D]l}d| d �|
|�d|�d�<d| d �|
|�d|�d�<d| d �|
|�d|�d�<d| d �|
|�d|�d�<qxWxLt||�D]>}|dk�
rd| d �}n d| d �}||
|�d|�d|�d�<q�WdS)NrDrHz4direction must either be "inputs" or "outputs", but z was givenr�r�r�r�Zpast_encoder_sequenceZpast_decoder_sequencez past_decoder_sequence + sequencer-)r
rr�z.decoder.keyz.decoder.valuez.encoder.keyz.encoder.valuez.key)rDrH)r6r�r�r�r`) r=r�r�r%r�r�r�r�r�r�r/r�Z axes_infor)r)r*r�Ns&












z4OnnxSeq2SeqConfigWithPast.fill_with_past_key_values_cCsd|d
|
|�d|�d�<|d|
|�d|�d�<|d|
|�d|�d�<|d|
|�d|�d �<dS) Nr
r�z.decoder.keyrz.decoder.valuerz.encoder.keyrYz.encoder.valuer))r=r�r%r�r�r)r)r*r�js



z3OnnxSeq2SeqConfigWithPast._flatten_past_key_values_r�r�)r�r�FN)rr r!r�rr$r�rHr r�r�r�rrrrzr�r�r�r)r))
r�r*r��
s

$





? r�ii l)1rEr9rs�abcrr�collectionsr�typingrrrrr r rrr rZnumpyra� packagingr�utilsrrrrrrrZconfiguration_utilsrrprrqrZPILr� get_loggerrrvrPrWZ dataclassrr+r�r�r)r)r)r*�s2




0