U
4Aft
@ sH d Z ddlZddlZddlZddlmZmZmZmZ ddl m
Z
ddl m Z m
Z
ddlmZ ddlmZ dd lmZmZmZmZ dd
lmZ eeeeef eeeeeef eeeeef eeeeeef f ZG d d d eZ eeeeef eeeef d
ddZ eeedddZ dd Z dd Z dd Z!d ddZ"dS ) z Processor class for KOSMOS-2. N)ListOptionalTupleUnion ) BatchFeature)
ImageInput
is_batched)ProcessorMixin)
AddedToken)
BatchEncodingPaddingStrategy TextInputTruncationStrategy)
TensorTypec sn e Zd ZdZddgZdgZdZdZd& fdd Zd'e e
e e e f e
ee ee eee
eeef e
eeef ee ee ee eeee
eef edddZdd Zdd Zd(e
e e e f e e
ee e
ee e f dddZdd Zdd Zd)d d Z e d d Z ee
e e e e e e f ed d!d"Z!e
e eef e e e e e f f e eef d#d$d%Z" Z#S )*Kosmos2Processora,
Constructs an KOSMOS-2 processor which wraps a KOSMOS-2 image processor and a KOSMOS-2 tokenizer into a single
processor.
[`Kosmos2Processor`] offers all the functionalities of [`CLIPImageProcessor`] and some functionalities of
[`XLMRobertaTokenizerFast`]. See the docstring of [`~Kosmos2Processor.__call__`] and [`~Kosmos2Processor.decode`]
for more information.
Args:
image_processor (`CLIPImageProcessor`):
An instance of [`CLIPImageProcessor`]. The image processor is a required input.
tokenizer (`XLMRobertaTokenizerFast`):
An instance of ['XLMRobertaTokenizerFast`]. The tokenizer is a required input.
num_patch_index_tokens (`int`, *optional*, defaults to 1024):
The number of tokens that represent patch indices.
image_processor tokenizernum_patch_index_tokensZCLIPImageProcessor)ZXLMRobertaTokenizerZXLMRobertaTokenizerFast c s d|_ d| _d| _d| _d| _d| _d| _d| _d | _d
| _ d | _
d | _ | j| j| j| j| j| j| j| j| j | j
| j g | _ || _
d
d t| j
D }g }| j | D ]}|t|dddd q|| t || d S )NFzzzzzzz z z c S s" g | ]}d t |d dqS )
)strzfill.0x r R/tmp/pip-unpacked-wheel-zw5xktn0/transformers/models/kosmos2/processing_kosmos2.py
a s z-Kosmos2Processor.__init__..T)lstriprstrip
normalized)Zreturn_token_type_idsZ eod_token boi_token eoi_tokenZ eoc_tokenZ eol_tokenZ bop_tokenZ eop_tokenZ boo_tokenZ eoo_tokenZ dom_tokenZ grd_tokenZ
tag_tokensr rangeappendr Z
add_tokenssuper__init__)selfr r r kwargsZpatch_index_tokensZ
tokens_to_addtoken __class__r r r* = s>
zKosmos2Processor.__init__N@ TF)imagestextbboxesnum_image_tokensfirst_image_token_idadd_special_tokens
add_eos_tokenpadding
truncation
max_lengthpad_to_multiple_ofreturn_attention_mask
return_lengthverbosereturn_tensorsreturnc ! sp |dkr|dkrt dt }|dk r>j||d}|| |dk rj||||d}|r|st|tr|jj | }n t|t rfdd|D }jf ||o||o|dk| |
|dkr| n| | ||dkr|ndd |}|| |dk rl|dk rl|dkrjj
d }|}t |d }t t ||| }d gdg| d g }g }g }|d
}t|tr||g}|d g|d < |D ]n}|d| | ||| d }|
| t|} |rd g| } | d gt|t| 7 } |
| qt|t rtd d t|jD d
d d} | d \} } | d \} } jf || g|oJ||| |
| |dd|}t|jd | krjjdkrʇ fdd|D } fdd|D } fdd|d D |d < nNjjdkr fdd|D } fdd|D } fdd|d D |d < t|trN|dkrN|d }|d d |d < |d }|t||d |d|d |S ) a
This method uses [`CLIPImageProcessor.__call__`] method to prepare image(s) for the model, and
[`XLMRobertaTokenizerFast.__call__`] to prepare text for the model.
Please refer to the docstring of the above two methods for more information.
The rest of this documentation shows the arguments specific to `Kosmos2Processor`.
Args:
bboxes (`Union[List[Tuple[int]], List[Tuple[float]], List[List[Tuple[int]]], List[List[Tuple[float]]]]`, *optional*):
The bounding bboxes associated to `texts`.
num_image_tokens (`int`, *optional* defaults to 64):
The number of (consecutive) places that are used to mark the placeholders to store image information.
This should be the same as `latent_query_num` in the instance of `Kosmos2Config` you are using.
first_image_token_id (`int`, *optional*):
The token id that will be used for the first place of the subsequence that is reserved to store image
information. If unset, will default to `self.tokenizer.unk_token_id + 1`.
add_eos_token (`bool`, defaults to `False`):
Whether or not to include `EOS` token id in the encoding when `add_special_tokens=True`.
Nz*You have to specify either images or text.)r? )r4 c s g | ]} j j | qS r )r bos_token)r s)r+ r r r! s z-Kosmos2Processor.__call__..) r2 r6 r8 r9 r: r; r< r> r? r input_idsattention_maskc S s g | ]\}}|t |fqS r len)r idxr r r r r! s c S s | d S Nr )r r r r z+Kosmos2Processor.__call__..)keyrJ )r2 r6 r8 r9 r: r; r> r? rightc s& g | ] }|j jg t| qS r r Z pad_token_idrG r max_len_paddedr+ r r r! s c s" g | ]}|d g t | qS r rF r rQ r r r! s c s" g | ]}|d g t | qS rR rF r rS r r r! s leftc s& g | ] }j jg t| | qS r rO r rP r r r! s c s" g | ]}d g t | | qS rR rF r rS r r r! s c s" g | ]}d g t | | qS rR rF r rS r r r! s )rD rE image_embeds_position_mask)dataZ tensor_type)
ValueErrorr r updatepreprocess_examples
isinstancer r rA listZ unk_token_idintr' r( copyrG sorted enumeraterD Z padding_sider )!r+ r1 r2 r3 r4 r5 r6 r7 r8 r9 r: r; r< r= r> r? r, encodingZimage_encodingZ
text_encodingZwith_bos start_indexZimage_token_idsZ base_image_embeds_position_maskrD rU Z
all_input_idsZtext_idsmaskZ
sorted_length_Zmin_len_not_paddedrH r rP r __call__j s '
zKosmos2Processor.__call__c C s |dkr dS t |ts td|D ]x}|dkr2q"nt |tsB|g}|D ]R}t |trt|dkrrtdd |D sFt|dkrtdd |D sFtdqFq"dS ) a
Check `bboxes` for a single text example. It could be
- `None`: no bounding box associated to a text.
- A list with each element being the bounding boxes associated to one ` ... ` pair found
in a text. This could be:
- `None`: no bounding box associated to a ` ... ` pair.
- A tuple of 2 integers: A single bounding box specified by patch indices.
- A tuple of 4 float point number: A single bounding box specified by (normalized) coordinates.
- A list containing the above 2 tuple types: Multiple bounding boxes for a
` ... ` pair.
Nz@`bboxes` (for a single text example) should be `None` or a list. c s s | ]}t |tV qd S N)rZ r\ r r r r * s zAKosmos2Processor._check_bboxes_for_single_text..r c s s | ]}t |tV qd S rf )rZ floatr r r r rg + s a' Each element in `bboxes` (for a single text example) should be either `None`, a tuple containing 2 integers or 4 float point numbers, or a list containing such tuples. Also make sure the arguments `texts` and `bboxes` passed to `preprocess_text` are both in batches or both for a single example.)rZ r[ rW tuplerG all)r+ r3 bboxelementr r r _check_bboxes_for_single_text s,
z.Kosmos2Processor._check_bboxes_for_single_textc C s. | }|d k r | d| }| ||}|S )N )strip_insert_patch_index_tokens)r+ r2 imager3 img_info_tokensr r r _preprocess_single_example4 s
z+Kosmos2Processor._preprocess_single_example)textsr1 r3 r4 r@ c sD j g| }dj g| jg d}t|tr>d}|g}|dkrVdgt| }nt|sd|g}t|t|krtdt| dt| d|s| |g}n>|dk rt|t std|D ]}| qndgt| }t|t|krtd t| dt| d fd
d t
|||D }|s@|d }|S )
a- Add image and bounding box information to `texts` as image and patch index tokens.
Args:
texts (`Union[TextInput, List[TextInput]]`): The texts to be processed.
images (`ImageInput`, *optional*): The images associated to `texts`.
bboxes (`Union[List[Tuple[int]], List[Tuple[float]], List[List[Tuple[int]]], List[List[Tuple[float]]]]`, *optional*):
The bounding bboxes associated to `texts`.
num_image_tokens (`int`, *optional*, defaults to 64):
The number of image tokens (used as latent queries). This should corresponds to the `latent_query_num`
attribute in `Kosmos2Config`.
Returns:
`Union[TextInput, List[TextInput]]`: The processed texts with image and patch index tokens.
rn TFNzGThe number of examples in `texts` and `images` should be the same. Got v.s. instead.zS`bboxes` should be `None` or a list (as a batch) when `texts` is passed as a batch.zGThe number of examples in `texts` and `bboxes` should be the same. Got c s" g | ]\}}} ||| qS r )rs )r r2 rq rk rr r+ r r r! v s z8Kosmos2Processor.preprocess_examples..r ) r% joinr& rZ r rG r rW rm r[ zip) r+ rt r1 r3 r4 Z
img_tokensZbatchedr resultr rw r rY > sB
z$Kosmos2Processor.preprocess_examplesc O s | j j||S )z
This method forwards all its arguments to PreTrainedTokenizer's [`~PreTrainedTokenizer.batch_decode`]. Please
refer to the docstring of this method for more information.
)r batch_decoder+ argsr, r r r r{ s z Kosmos2Processor.batch_decodec O s | j j||S )z
This method forwards all its arguments to PreTrainedTokenizer's [`~PreTrainedTokenizer.decode`]. Please refer to
the docstring of this method for more information.
)r decoder| r r r r~ s zKosmos2Processor.decodec C s | | jd }|r t|S |S rI )splitr& +clean_text_and_extract_entities_with_bboxes)r+ r2 Zcleanup_and_extractcaptionr r r post_process_generation s z(Kosmos2Processor.post_process_generationc C s" | j j}| jj}tt|| S rf )r model_input_namesr r[ dictfromkeys)r+ Ztokenizer_input_namesZimage_processor_input_namesr r r r s z"Kosmos2Processor.model_input_names)r2 r3 r@ c C sT |d kst |dkr|S ttjd|d}t |t |krXtdt | dt | dd}g }t||D ]\}}| \}} |||| | }|d krqjt|t r|g}g }
t
dd |D std |D ]&} | | \} }
|
| d
|
qt |
dkrqjd |
}|d | d
qj|t |k rF|||d d |}|S )Nr z.+?)stringzuThe number of elements in `bboxes` should be the same as the number of ` ... ` pairs in `text`. Got ru rv c s s | ]}|d k V qd S rf r )r boxr r r rg s z>Kosmos2Processor._insert_patch_index_tokens..zTThe multiple bounding boxes for a single phrase should not contain any `None` value.rn z z )
rG r[ refinditerrW ry spanr( rZ ri rj #_convert_bbox_to_patch_index_tokensrx )r+ r2 r3 Zmatched_phrasescurr_posbuffermatchedrk rc endZpatch_index_stringsr Z
patch_index_1Z
patch_index_2Z position_strr r r rp sB
z+Kosmos2Processor._insert_patch_index_tokens)rk r@ c C sh t |dkr|\}}n tt| j}t||\}}dt|d d}dt|d d}||fS )Nre r r r )rG r\ mathsqrtr coordinate_to_patch_indexr r )r+ rk Zidx_1Zidx_2num_patches_per_sideZtoken_1Ztoken_2r r r r s
z4Kosmos2Processor._convert_bbox_to_patch_index_tokens)r )NNNr0 NTFFNNNNFTN)NNr0 )T)$__name__
__module__ __qualname____doc__
attributesZ valid_kwargsZimage_processor_classZtokenizer_classr* r r r r BboxInputr r\ boolr r
r r r rd rm rs rY r{ r~ r propertyr r rh rp r
__classcell__r r r. r r & sz / (#
C
*.
r )rk r r@ c C s | \}}}}||kr ||ks$t dt|| }t|| }t|| d }t|| d } || | }
| | | } |
| fS )a Convert a bounding box to a pair of patch indices.
Args:
bbox (`Tuple[float, float, float, float]`):
The 4 coordinates of the bounding box, with the format being (x1, y1, x2, y2) specifying the upper-left and
lower-right corners of the box. It should have x2 > x1 and y2 > y1.
num_patches_per_side (`int`): the number of patches along each side.
Returns:
`Tuple[int, int]`: A pair of patch indices representing the upper-left patch and lower-right patch.
zTThe coordinates in `bbox` should be `(x1, y1, x2, y2)` with `x2 > x1` and `y2 > y1`.rC )rW r floorceil) rk r x1y1x2y2ul_xul_ylr_xlr_yul_idxlr_idxr r r r s r )r r r c C s d| }| | }| | }|| }|| }| |krZ|| }|| } || | }
|| | } nz||ksj||kr|| }|| } || | }
|| | } n@|| |d }|| |d } || |d }
|| |d } || |
| fS )a
Given a grid of length `num_patches_per_side` and the indices of the upper-left and lower-right corners of a
bounding box, returns the normalized coordinates of the bounding box, in the form (x1, y1, x2, y2).
Args:
ul_idx (`int`): the index of the grid cell that corresponds to the upper-left corner of the bounding box.
lr_idx (`int`): the index of the grid cell that corresponds to the lower-right corner of the bounding box.
num_patches_per_side (`int`): the number of patches along each side.
Returns:
`Tuple[float]`: the normalized coordinates of the bounding box, in the form (x1, y1, x2, y2).
g ?re r ) r r r Z cell_sizer r r r r r r r r r r patch_index_to_coordinate s( r c C s6 d}t || }g }|D ]}|d}| \}}}|sZd}|dd |dd f}|d} g }
| D ]v} t d| } t d| dd }
| rl|
rl|r|
t| dt|
df ql|
t| dt|
df ql|r||||
f q|
D ]0}d|d d |d d
}||||gf qq|S ) a Extract entities contained in `text`. The bounding bboxes is given in the form of patch indices.
This functioin is only intended to be used within `clean_text_and_extract_entities_with_bboxes` where further
processing happens, including converting to normalized coordinates and whitespace character cleaning up.
Examples:
```python
>>> text = " An image of a snowman warming himself by a fire."
>>> entities = extract_entities_with_patch_indices(text)
>>> entities
[(' a snowman', (31, 41), [(44, 863)]), (' a fire', (130, 137), [(5, 911)])]
```z(?:(([^<]+)))?re Nr r zrC r z>r N)rG r sub)r r2 entity_namestartr Zadjusted_startZ adjusted_endadjusted_entityr r r adjust_entity_positionsb s
r c C s | }t| t| }g }|D ]j\}\}}}t|t| } t|t| }
|| | }|| |
}| }||||f|f q$||fS )z9Remove the spaces around the text and the entities in it.)ro rG r" r# r( ) r2 entitiesZnew_textZleading_spacesZ new_entitiesr r r r3 Zentity_name_leading_spacesZentity_name_trailing_spacesr r r _cleanup_spacesl s r c
sp t dd| }t| }g }|D ]F}|dd |d }}t|| } fdd|D } ||| f q t||S )a Remove the tag tokens from `text`, extract entities in it with some cleaning up of white characters.
Examples:
```python
>>> text = " An image of a snowman warming himself by a fire."
>>> clean_text, entities = clean_text_and_extract_entities_with_bboxes(text)
>>> clean_text
'An image of a snowman warming himself by a fire.'
>>> entities
[('a snowman', (12, 21), [(0.390625, 0.046875, 0.984375, 0.828125)]), ('a fire', (41, 47), [(0.171875, 0.015625, 0.484375, 0.890625)])]
```r r r re c s g | ]}t |d |d qS )r rC )r )r rk r r r r! s z?clean_text_and_extract_entities_with_bboxes..)r r r r r( r )
r2 r Zprocessed_textr r itemr r3 r Zbboxes_in_coordsr r r r s
r )r )#r r] r r typingr r r r Zimage_processing_utilsr Z image_utilsr r Zprocessing_utilsr
Ztokenization_utilsr Ztokenization_utils_baser r
r r utilsr r\ rh r r r r r r r r r r r r s6 9& -: