"""Visualizer for TAPAS
Implementation heavily based on
`EncodingVisualizer` from `tokenizers.tools`.
"""
import os
from typing import Any, List, Dict
from collections import defaultdict
import pandas as pd
from transformers import TapasTokenizer
dirname = os.path.dirname(__file__)
css_filename = os.path.join(dirname, "tapas-styles.css")
with open(css_filename) as f:
css = f.read()
def HTMLBody(table_html: str, css_styles: str = css) -> str:
"""
Generates the full html with css from a list of html spans
Args:
table_html (str):
The html string of the table
css_styles (str):
CSS styling to be embedded inline
Returns:
:obj:`str`: An HTML string with style markup
"""
return f"""
{table_html}
"""
class TapasVisualizer:
def __init__(self, tokenizer: TapasTokenizer) -> None:
self.tokenizer = tokenizer
def normalize_token_str(self, token_str: str) -> str:
# Normalize subword tokens to org subword str
return token_str.replace("##", "")
def style_span(self, span_text: str, css_classes: List[str]) -> str:
css = f'''class="{' '.join(css_classes)}"'''
return f"{span_text}"
def text_to_html(self, org_text: str, tokens: List[str]) -> str:
"""Create html based on the original text and its tokens.
Note: The tokens need to be in same order as in the original text
Args:
org_text (str): Original string before tokenization
tokens (List[str]): The tokens of org_text
Returns:
str: html with styling for the tokens
"""
if len(tokens) == 0:
print(f"Empty tokens for: {org_text}")
return ""
cur_token_id = 0
cur_token = self.normalize_token_str(tokens[cur_token_id])
# Loop through each character
next_start = 0
last_end = 0
spans = []
while next_start < len(org_text):
candidate = org_text[next_start : next_start + len(cur_token)]
# The tokenizer performs lowercasing; so check against lowercase
if candidate.lower() == cur_token:
if last_end != next_start:
# There was token-less text (probably whitespace)
# in the middle
spans.append(
self.style_span(org_text[last_end:next_start], ["non-token"])
)
odd_or_even = "even-token" if cur_token_id % 2 == 0 else "odd-token"
spans.append(self.style_span(candidate, ["token", odd_or_even]))
next_start += len(cur_token)
last_end = next_start
cur_token_id += 1
if cur_token_id >= len(tokens):
break
cur_token = self.normalize_token_str(tokens[cur_token_id])
else:
next_start += 1
if last_end != len(org_text):
spans.append(self.style_span(org_text[last_end:next_start], ["non-token"]))
return spans
def cells_to_html(
self,
cell_vals: List[List[str]],
cell_tokens: Dict,
row_id_start: int = 0,
cell_element: str = "td",
cumulative_cnt: int = 0,
table_html: str = "",
) -> str:
for row_id, row in enumerate(cell_vals, start=row_id_start):
row_html = ""
row_token_cnt = 0
for col_id, cell in enumerate(row, start=1):
cur_cell_tokens = cell_tokens[(row_id, col_id)]
span_htmls = self.text_to_html(cell, cur_cell_tokens)
cell_html = "".join(span_htmls)
row_html += f"<{cell_element}>{cell_html}{cell_element}>"
row_token_cnt += len(cur_cell_tokens)
cumulative_cnt += row_token_cnt
cnt_html = (
f''
f'{self.style_span(str(cumulative_cnt), ["non-token", "count"])}'
" | "
f''
f'{self.style_span(f"<+{row_token_cnt}", ["non-token", "count"])}'
" | "
)
row_html = cnt_html + row_html
table_html += f"{row_html}
"
return table_html, cumulative_cnt
def __call__(self, table: pd.DataFrame) -> Any:
tokenized = self.tokenizer(table)
cell_tokens = defaultdict(list)
for id_ind, input_id in enumerate(tokenized["input_ids"]):
input_id = int(input_id)
# 'prev_label', 'column_rank', 'inv_column_rank', 'numeric_relation'
# not required
segment_id, col_id, row_id, *_ = tokenized["token_type_ids"][id_ind]
token_text = self.tokenizer._convert_id_to_token(input_id)
if int(segment_id) == 1:
cell_tokens[(row_id, col_id)].append(token_text)
table_html, cumulative_cnt = self.cells_to_html(
cell_vals=[table.columns],
cell_tokens=cell_tokens,
row_id_start=0,
cell_element="th",
cumulative_cnt=0,
table_html="",
)
table_html, cumulative_cnt = self.cells_to_html(
cell_vals=table.values,
cell_tokens=cell_tokens,
row_id_start=1,
cell_element="td",
cumulative_cnt=cumulative_cnt,
table_html=table_html,
)
top_label = self.style_span("#Tokens", ["count"])
top_label_cnt = self.style_span(f"(Total: {cumulative_cnt})", ["count"])
table_html = (
''
f'{top_label} | '
f'{top_label_cnt} | '
"
"
f"{table_html}"
)
table_html = f""
return HTMLBody(table_html)