# Copyright 2021 The HuggingFace Team. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import json import matplotlib.image as mpimg import matplotlib.pyplot as plt import nltk import numpy as np import pandas as pd import plotly.graph_objects as go import seaborn as sns import statistics import utils import utils.dataset_utils as ds_utils from data_measurements.tokenize import Tokenize from data_measurements.labels import labels from data_measurements.perplexity import perplexity from data_measurements.lengths import lengths from data_measurements.text_duplicates import text_duplicates as td from data_measurements.npmi import npmi from data_measurements.zipf import zipf from datasets import load_from_disk from nltk.corpus import stopwords from os import mkdir, getenv from os.path import exists, isdir from os.path import join as pjoin from pathlib import Path from sklearn.feature_extraction.text import CountVectorizer from utils.dataset_utils import (CNT, LENGTH_FIELD, TEXT_FIELD, PERPLEXITY_FIELD, PROP, TEXT_NAN_CNT, TOKENIZED_FIELD, TOT_OPEN_WORDS, TOT_WORDS, VOCAB, WORD) logs = utils.prepare_logging(__file__) # TODO: Read this in depending on chosen language / expand beyond english nltk.download("stopwords", quiet=True) _CLOSED_CLASS = ( stopwords.words("english") + ["t", "n", "ll", "d", "s"] + ["wasn", "weren", "won", "aren", "wouldn", "shouldn", "didn", "don", "hasn", "ain", "couldn", "doesn", "hadn", "haven", "isn", "mightn", "mustn", "needn", "shan", "would", "could", "dont"] + [str(i) for i in range(0, 99)] ) IDENTITY_TERMS = [ "man", "woman", "non-binary", "gay", "lesbian", "queer", "trans", "straight", "cis", "she", "her", "hers", "he", "him", "his", "they", "them", "their", "theirs", "himself", "herself", ] # treating inf values as NaN as well pd.set_option("use_inf_as_na", True) MIN_VOCAB_COUNT = 10 _NUM_VOCAB_BATCHES = 2000 _TOP_N = 100 class DatasetStatisticsCacheClass: def __init__( self, dset_name, dset_config, split_name, text_field, label_field, label_names, cache_dir="cache_dir", dataset_cache_dir=None, use_cache=False, save=True, ): ### What are we analyzing? # name of the Hugging Face dataset self.dset_name = dset_name # name of the dataset config self.dset_config = dset_config # name of the split to analyze self.split_name = split_name # which text/feature fields are we analysing? self.text_field = text_field ## Label variables # which label fields are we analysing? self.label_field = label_field # what are the names of the classes? self.label_names = label_names # save label pie chart in the class so it doesn't ge re-computed self.fig_labels = None ## Hugging Face dataset objects self.dset = None # original dataset # HF dataset with all of the self.text_field instances in self.dset self.text_dset = None self.dset_peek = None # HF dataset with text embeddings in the same order as self.text_dset self.embeddings_dset = None # HF dataset with all of the self.label_field instances in self.dset # TODO: Not being used anymore; make sure & remove. self.label_dset = None self.length_obj = None ## Data frames # Tokenized text self.tokenized_df = None # Data Frame version of self.label_dset # TODO: Not being used anymore. Make sure and remove self.label_df = None # where are they being cached? self.label_files = {} # label pie chart used in the UI self.fig_labels = None # results self.label_results = None ## Caching if not dataset_cache_dir: _, self.dataset_cache_dir = ds_utils.get_cache_dir_naming(cache_dir, dset_name, dset_config, split_name, text_field) else: self.dataset_cache_dir = dataset_cache_dir # Use stored data if there; otherwise calculate afresh self.use_cache = use_cache # Save newly calculated results. self.save = save self.dset_peek = None # Tokenized text self.tokenized_df = None ## Zipf # Save zipf fig so it doesn't need to be recreated. self.zipf_fig = None # Zipf object self.z = None ## Vocabulary # Vocabulary with word counts in the dataset self.vocab_counts_df = None # Vocabulary filtered to remove stopwords self.vocab_counts_filtered_df = None self.sorted_top_vocab_df = None # Text Duplicates self.duplicates_results = None self.duplicates_files = {} self.dups_frac = 0 self.dups_dict = {} ## Perplexity self.perplexities_df = None ## Lengths self.avg_length = None self.std_length = None self.length_stats_dict = None self.length_df = None self.fig_tok_length = None self.num_uniq_lengths = 0 ## "General" stats self.general_stats_dict = {} self.total_words = 0 self.total_open_words = 0 # Number of NaN values (NOT empty strings) self.text_nan_count = 0 # nPMI self.npmi_obj = None # The minimum amount of times a word should occur to be included in # word-count-based calculations (currently just relevant to nPMI) self.min_vocab_count = MIN_VOCAB_COUNT self.hf_dset_cache_dir = pjoin(self.dataset_cache_dir, "base_dset") self.tokenized_df_fid = pjoin(self.dataset_cache_dir, "tokenized_df.json") self.text_dset_fid = pjoin(self.dataset_cache_dir, "text_dset") self.dset_peek_json_fid = pjoin(self.dataset_cache_dir, "dset_peek.json") ## Length cache files self.length_df_fid = pjoin(self.dataset_cache_dir, "length_df.json") self.length_stats_json_fid = pjoin(self.dataset_cache_dir, "length_stats.json") self.vocab_counts_df_fid = pjoin(self.dataset_cache_dir, "vocab_counts.json") self.dup_counts_df_fid = pjoin(self.dataset_cache_dir, "dup_counts_df.json") self.fig_tok_length_fid = pjoin(self.dataset_cache_dir, "fig_tok_length.png") ## General text stats self.general_stats_json_fid = pjoin(self.dataset_cache_dir, "general_stats_dict.json") # Needed for UI self.sorted_top_vocab_df_fid = pjoin( self.dataset_cache_dir, "sorted_top_vocab.json" ) # Set the HuggingFace dataset object with the given arguments. self.dset = self._get_dataset() self.text_dset = None # Defines self.text_dset, a HF Dataset with just the TEXT_FIELD instances in self.dset extracted self.load_or_prepare_text_dataset() def _get_dataset(self): """ Gets the HuggingFace Dataset object. First tries to use the given cache directory if specified; otherwise saves to the given cache directory if specified. """ dset = ds_utils.load_truncated_dataset(self.dset_name, self.dset_config, self.split_name, cache_dir=self.hf_dset_cache_dir, save=self.save) return dset def load_or_prepare_text_dataset(self, load_only=False): """ Prepares the HF dataset text/feature based on given config, split, etc. Args: load_only: Whether only a cached dataset can be used. """ logs.info("Doing text dset.") if self.use_cache and exists(self.text_dset_fid): # load extracted text self.text_dset = load_from_disk(self.text_dset_fid) logs.info("Loaded dataset from disk") logs.info(self.text_dset) # ...Or load it from the server and store it anew elif not load_only: # Defines self.text_dset self.prepare_text_dset() if self.save: # save extracted text instances logs.info("Saving dataset to disk") self.text_dset.save_to_disk(self.text_dset_fid) def prepare_text_dset(self): logs.info("Working with dataset:") logs.info(self.dset) # Extract all text instances from the user-specified self.text_field, # which is a dataset-specific text/feature field; # create a new feature called TEXT_FIELD, which is a constant shared # across DMT logic. self.text_dset = self.dset.map( lambda examples: ds_utils.extract_field( examples, self.text_field, TEXT_FIELD ), batched=True, remove_columns=list(self.dset.features), ) self.text_dset = self.text_dset.filter(lambda ex: ex["text"] is not None) def load_or_prepare_general_stats(self, load_only=False): """ Content for expander_general_stats widget. Provides statistics for total words, total open words, the sorted top vocab, the NaN count, and the duplicate count. Args: Returns: """ # General statistics # For the general statistics, text duplicates are not saved in their # own files, but rather just the text duplicate fraction is saved in the # "general" file. We therefore set save=False for # the text duplicate files in this case. # Similarly, we don't get the full list of duplicates # in general stats, so set list_duplicates to False self.load_or_prepare_text_duplicates(load_only=load_only, save=False, list_duplicates=False) logs.info("Duplicates results:") logs.info(self.duplicates_results) self.general_stats_dict.update(self.duplicates_results) # TODO: Tighten the rest of this similar to text_duplicates. if ( self.use_cache and exists(self.general_stats_json_fid) and exists(self.sorted_top_vocab_df_fid) ): logs.info("Loading cached general stats") self.load_general_stats() elif not load_only: logs.info("Preparing general stats") self.prepare_general_stats() if self.save: ds_utils.write_df(self.sorted_top_vocab_df, self.sorted_top_vocab_df_fid) ds_utils.write_json(self.general_stats_dict, self.general_stats_json_fid) def load_or_prepare_text_lengths(self, load_only=False): """ The text length widget relies on this function, which provides a figure of the text lengths, some text length statistics, and a text length dataframe to peruse. Args: load_only (Bool): Whether we can compute anew, or just need to try to grab cache. Returns: """ # We work with the already tokenized dataset self.load_or_prepare_tokenized_df() self.length_obj = lengths.DMTHelper(self, load_only=load_only, save=self.save) self.length_obj.run_DMT_processing() ## Labels functions def load_or_prepare_labels(self, load_only=False): """Uses a generic Labels class, with attributes specific to this project as input. Computes results for each label column, or else uses what's available in the cache. Currently supports Datasets with just one label column. """ label_obj = labels.DMTHelper(self, load_only=load_only, save=self.save) self.label_files = label_obj.get_label_filenames() if self.use_cache and exists(self.label_files["figure json"]) and exists(self.label_files["statistics"]): self.fig_labels = ds_utils.read_plotly(self.label_files["figure json"]) self.label_results = ds_utils.read_json(self.label_files["statistics"]) elif not load_only: label_obj.run_DMT_processing() self.fig_labels = label_obj.fig_labels self.label_results = label_obj.label_results # Get vocab with word counts def load_or_prepare_vocab(self, load_only=False): """ Calculates the vocabulary count from the tokenized text. The resulting dataframes may be used in nPMI calculations, zipf, etc. :param :return: """ if self.use_cache and exists(self.vocab_counts_df_fid): logs.info("Reading vocab from cache") self.load_vocab() self.vocab_counts_filtered_df = filter_vocab(self.vocab_counts_df) elif not load_only: if self.tokenized_df is None: # Building the vocabulary starts with tokenizing. self.load_or_prepare_tokenized_df(load_only=False) logs.info("Calculating vocab afresh") word_count_df = count_vocab_frequencies(self.tokenized_df) logs.info("Making dfs with proportion.") self.vocab_counts_df = calc_p_word(word_count_df) self.vocab_counts_filtered_df = filter_vocab(self.vocab_counts_df) if self.save: logs.info("Writing out.") ds_utils.write_df(self.vocab_counts_df, self.vocab_counts_df_fid) logs.info("unfiltered vocab") logs.info(self.vocab_counts_df) logs.info("filtered vocab") logs.info(self.vocab_counts_filtered_df) def load_vocab(self): self.vocab_counts_df = ds_utils.read_df(self.vocab_counts_df_fid) def load_or_prepare_text_duplicates(self, load_only=False, save=True, list_duplicates=True): """Uses a text duplicates library, which returns strings with their counts, fraction of data that is duplicated, or else uses what's available in the cache. """ dups_obj = td.DMTHelper(self, load_only=load_only, save=save) dups_obj.run_DMT_processing(list_duplicates=list_duplicates) self.duplicates_results = dups_obj.duplicates_results self.dups_frac = self.duplicates_results[td.DUPS_FRAC] if list_duplicates and td.DUPS_DICT in self.duplicates_results: self.dups_dict = self.duplicates_results[td.DUPS_DICT] self.duplicates_files = dups_obj.get_duplicates_filenames() def load_or_prepare_text_perplexities(self, load_only=False): perplex_obj = perplexity.DMTHelper(self, load_only=load_only) perplex_obj.run_DMT_processing() self.perplexities_df = perplex_obj.df def load_general_stats(self): self.general_stats_dict = json.load( open(self.general_stats_json_fid, encoding="utf-8") ) self.sorted_top_vocab_df = ds_utils.read_df(self.sorted_top_vocab_df_fid) self.text_nan_count = self.general_stats_dict[TEXT_NAN_CNT] self.dups_frac = self.general_stats_dict[td.DUPS_FRAC] self.total_words = self.general_stats_dict[TOT_WORDS] self.total_open_words = self.general_stats_dict[TOT_OPEN_WORDS] def prepare_general_stats(self): if self.tokenized_df is None: logs.warning("Tokenized dataset not yet loaded; doing so.") self.load_or_prepare_tokenized_df() if self.vocab_counts_df is None: logs.warning("Vocab not yet loaded; doing so.") self.load_or_prepare_vocab() self.sorted_top_vocab_df = self.vocab_counts_filtered_df.sort_values( "count", ascending=False ).head(_TOP_N) self.total_words = len(self.vocab_counts_df) self.total_open_words = len(self.vocab_counts_filtered_df) self.text_nan_count = int(self.tokenized_df.isnull().sum().sum()) self.load_or_prepare_text_duplicates() self.general_stats_dict = { TOT_WORDS: self.total_words, TOT_OPEN_WORDS: self.total_open_words, TEXT_NAN_CNT: self.text_nan_count, td.DUPS_FRAC: self.dups_frac } def load_or_prepare_dataset(self, load_only=False): """ Prepares the HF dataset text/feature based on given config, split, etc. Args: load_only: Whether only a cached dataset can be used. """ logs.info("Doing text dset.") if self.use_cache and exists(self.text_dset_fid): # load extracted text self.text_dset = load_from_disk(self.text_dset_fid) logs.warning("Loaded dataset from disk") logs.warning(self.text_dset) # ...Or load it from the server and store it anew elif not load_only: self.prepare_text_dset() if self.save: # save extracted text instances logs.warning("Saving dataset to disk") self.text_dset.save_to_disk(self.text_dset_fid) # TODO: Are we not using this anymore? def load_or_prepare_dset_peek(self, load_only=False): if self.use_cache and exists(self.dset_peek_json_fid): with open(self.dset_peek_json_fid, "r") as f: self.dset_peek = json.load(f)["dset peek"] elif not load_only: self.dset_peek = self.dset[:100] if self.save: ds_utils.write_json({"dset peek": self.dset_peek}, self.dset_peek_json_fid) def load_or_prepare_tokenized_df(self, load_only=False): if self.use_cache and exists(self.tokenized_df_fid): self.tokenized_df = ds_utils.read_df(self.tokenized_df_fid) elif not load_only: # tokenize all text instances self.tokenized_df = Tokenize(self.text_dset, feature=TEXT_FIELD, tok_feature=TOKENIZED_FIELD).get_df() logs.info("tokenized df is") logs.info(self.tokenized_df) if self.save: logs.warning("Saving tokenized dataset to disk") # save tokenized text ds_utils.write_df(self.tokenized_df, self.tokenized_df_fid) def load_or_prepare_npmi(self, load_only=False): npmi_obj = npmi.DMTHelper(self, IDENTITY_TERMS, load_only=load_only, use_cache=self.use_cache, save=self.save) npmi_obj.run_DMT_processing() self.npmi_obj = npmi_obj self.npmi_results = npmi_obj.results_dict self.npmi_files = npmi_obj.get_filenames() def load_or_prepare_zipf(self, load_only=False): zipf_json_fid, zipf_fig_json_fid, zipf_fig_html_fid = zipf.get_zipf_fids( self.dataset_cache_dir) if self.use_cache and exists(zipf_json_fid): # Zipf statistics # Read Zipf statistics: Alpha, p-value, etc. with open(zipf_json_fid, "r") as f: zipf_dict = json.load(f) self.z = zipf.Zipf(self.vocab_counts_df) self.z.load(zipf_dict) # Zipf figure if exists(zipf_fig_json_fid): self.zipf_fig = ds_utils.read_plotly(zipf_fig_json_fid) elif not load_only: self.zipf_fig = zipf.make_zipf_fig(self.z) if self.save: ds_utils.write_plotly(self.zipf_fig) elif not load_only: self.prepare_zipf() if self.save: zipf_dict = self.z.get_zipf_dict() ds_utils.write_json(zipf_dict, zipf_json_fid) ds_utils.write_plotly(self.zipf_fig, zipf_fig_json_fid) self.zipf_fig.write_html(zipf_fig_html_fid) def prepare_zipf(self): # Calculate zipf from scratch # TODO: Does z even need to be self? self.z = zipf.Zipf(self.vocab_counts_df) self.z.calc_fit() self.zipf_fig = zipf.make_zipf_fig(self.z) def dummy(doc): return doc def count_vocab_frequencies(tokenized_df): """ Based on an input pandas DataFrame with a 'text' column, this function will count the occurrences of all words. :return: [num_words x num_sentences] DataFrame with the rows corresponding to the different vocabulary words and the column to the presence (0 or 1) of that word. """ cvec = CountVectorizer( tokenizer=dummy, preprocessor=dummy, ) # We do this to calculate per-word statistics # Fast calculation of single word counts logs.info( "Fitting dummy tokenization to make matrix using the previous tokenization" ) cvec.fit(tokenized_df[TOKENIZED_FIELD]) document_matrix = cvec.transform(tokenized_df[TOKENIZED_FIELD]) batches = np.linspace(0, tokenized_df.shape[0], _NUM_VOCAB_BATCHES).astype( int) i = 0 tf = [] while i < len(batches) - 1: if i % 100 == 0: logs.info("%s of %s vocab batches" % (str(i), str(len(batches)))) batch_result = np.sum( document_matrix[batches[i]: batches[i + 1]].toarray(), axis=0 ) tf.append(batch_result) i += 1 word_count_df = pd.DataFrame( [np.sum(tf, axis=0)], columns=cvec.get_feature_names_out() ).transpose() # Now organize everything into the dataframes word_count_df.columns = [CNT] word_count_df.index.name = WORD return word_count_df def calc_p_word(word_count_df): # p(word) word_count_df[PROP] = word_count_df[CNT] / float(sum(word_count_df[CNT])) vocab_counts_df = pd.DataFrame( word_count_df.sort_values(by=CNT, ascending=False)) vocab_counts_df[VOCAB] = vocab_counts_df.index return vocab_counts_df def filter_vocab(vocab_counts_df): # TODO: Add warnings (which words are missing) to log file? filtered_vocab_counts_df = vocab_counts_df.drop(_CLOSED_CLASS, errors="ignore") filtered_count = filtered_vocab_counts_df[CNT] filtered_count_denom = float(sum(filtered_vocab_counts_df[CNT])) filtered_vocab_counts_df[PROP] = filtered_count / filtered_count_denom return filtered_vocab_counts_df