word2vec.c

//  Copyright 2013 Google Inc. 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.


// ---------------------------------------------------------------------
//
// Word2vec C code commented by Chandler May.  Run
//
//   git diff original
//
// to see the complete set of changes to the original code.
//
// Preliminaries:
// * There are a lot of global variables.  They are important.
// * a, b, and c are used for loop counters and lots of other things;
//   it's not uncommon for one of them to be set to one thing and then
//   re-initialized and used for something completely different.
// * Random number generation is through a custom linear congruential
//   generator.  This in particular facilitates decoupled random number
//   generation across training threads.
//
// Notation:
// * Backticks are used to denote variables (`foo`)
// * The skip-gram model learns co-occurrence information between a
//   word and a fixed-length window of context words on either side of
//   it.  The embedding of the central word is called the "input"
//   representation and the embeddings of the left and right context
//   words are "output" representations.  This distinction (and
//   terminology) are quite important for understanding the code so here
//   is an example, assumping a context window of two words on either
//   side:
//
//     The quick brown fox jumped over the lazy dog.
//                ^     ^    ^     ^    ^
//                |     |    |     |    |
//           output output input output output
//
//   When looking at the input word "jumped" in this sentence, the
//   skip-gram with negative sampling (SGNS) model learns that it
//   co-occurs with "brown" and "fox" and "over" and "the" and does not
//   co-occur with a selection of negative-sample words (perhaps
//   "apple", "of", and "slithered").
//
// Partial call graph:
//
//   main
//   |
//   L- TrainModel
//      |
//      |- ReadVocab
//      |  |- ReadWord
//      |  |- AddWordToVocab
//      |  L- SortVocab
//      |
//      |- LearnVocabFromTrainFile
//      |  |- ReadWord
//      |  |- AddWordToVocab
//      |  |- SearchVocab
//      |  |- ReduceVocab
//      |  L- SortVocab
//      |
//      |- SaveVocab
//      |- InitNet
//      |- InitUnigramTable
//      |
//      L- TrainModelThread
//         L- ReadWordIndex
//            |- ReadWord
//            L- SearchVocab
//
// ---------------------------------------------------------------------


#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <math.h>
#include <pthread.h>


// max length of filenames, vocabulary words (including null terminator)
#define MAX_STRING 100
// size of pre-computed e^x / (e^x + 1) table
#define EXP_TABLE_SIZE 1000
// max exponent x for which to pre-compute e^x / (e^x + 1)
#define MAX_EXP 6
// max length (in words) of a sequence of overlapping word contexts or
// "sentences" (if there is a sequence of words longer than this not
// explicitly broken up into multiple "sentences" in the training data,
// it will be broken up into multiple "sentences" of length no longer
// than `MAX_SENTENCE_LENGTH` each during training)
#define MAX_SENTENCE_LENGTH 1000
// max length of Huffman codes used by hierarchical softmax
#define MAX_CODE_LENGTH 40


// Maximum 30 * 0.7 = 21M words in the vocabulary
// (where 0.7 is the magical load factor beyond which hash table
// performance degrades)
const int vocab_hash_size = 30000000;

// Negative sampling distribution represented by 1e8-element discrete
// sample from smoothed empirical unigram distribution.
const int table_size = 1e8;

// Set precision of real numbers
typedef float real;

// Representation of a word in the vocabulary, including (optional,
// for hierarchical softmax only) Huffman coding
struct vocab_word {
  long long cn;
  int *point;
  char *word, *code, codelen;
};

struct vocab_word *vocab;      // vocabulary
char
  train_file[MAX_STRING],      // training data (text) input file
  output_file[MAX_STRING],     // word vector (or word vector cluster)
                               //   (binary/text) output file
  save_vocab_file[MAX_STRING], // vocabulary (text) output file
  read_vocab_file[MAX_STRING]; // vocabulary (text) input file
int
  binary = 0,                  // 0 for text output, 1 for binary
  cbow = 1,                    // 0 for skip-gram, 1 for CBOW
  debug_mode = 2,              // 1 for extra terminal output, 2 for
                               //   extra extra terminal output
  window = 5,                  // context window radius (use `window`
                               //   output words on left and `window`
                               //   output words on right as context
                               //   of central input word)
  min_count = 5,               // min count of words in vocabulary when
                               //   pruning to mitigate sparsity
  num_threads = 12,            // number of threads to use for training
  min_reduce = 1,              // initial min count of words to keep in
                               //   vocabulary if pruning for space
                               //   (will be incremented as necessary)
                               //   (do not change)
  hs = 0,                      // 1 for hierarchical softmax
  negative = 5;                // number of negative samples to draw
                               //   per word
int *vocab_hash,               // hash table of words to positions in
                               //   vocabulary
  *table;                      // discrete sample of words used as
                               //   negative sampling distribution
long long
  vocab_max_size = 1000,       // capacity of vocabulary
                               //   (will be incremented as necessary)
  vocab_size = 0,              // number of words in vocabulary
                               //   (do not change)
  layer1_size = 100,           // size of embeddings
  train_words = 0,             // number of word tokens in training data
                               //   (do not change)
  word_count_actual = 0,       // number of word tokens seen so far
                               //   during training, over all
                               //   iterations; updated infrequently
                               //   (used for terminal output and
                               //   learning rate)
                               //   (do not change)
  iter = 5,                    // number of passes to take through
                               //   training data
  file_size = 0,               // size (in bytes) of training data file
  classes = 0;                 // number of k-means clusters to learn
                               //   of word vectors and write to output
                               //   file (0 to write word vectors to
                               //   output file, no clustering)
real
  alpha = 0.025,               // linear-decay learning rate
  starting_alpha,              // initial learning rate
                               //   (do not change; initialized at
                               //   beginning of training)
  sample = 1e-3;               // word subsampling threshold
real
  *syn0,                       // input word embeddings
  *syn1,                       // (used by hierarchical softmax)
  *syn1neg,                    // output word embeddings
  *expTable;                   // precomputed table of
                               //   e^x / (e^x + 1) for x in
                               //   [-MAX_EXP, MAX_EXP)
clock_t start;                 // start time of training algorithm

// Allocate and populate negative-sampling data structure `table`, an
// array of `table_size` words distributed approximately according to
// the empirical unigram distribution (smoothed by raising all
// probabilities to the power of 0.75 and re-normalizing), from `vocab`,
// an array of `vocab_size` words represented as `vocab_word` structs.
void InitUnigramTable() {
  int a, i;
  double train_words_pow = 0;
  double d1, power = 0.75;
  // allocate memory
  table = (int *)malloc(table_size * sizeof(int));
  // compute normalizer, `train_words_pow`
  for (a = 0; a < vocab_size; a++) train_words_pow += pow(vocab[a].cn, power);
  // initialize vocab position `i` and cumulative probability mass `d1`
  i = 0;
  d1 = pow(vocab[i].cn, power) / train_words_pow;
  for (a = 0; a < table_size; a++) {
    table[a] = i;
    if (a / (double)table_size > d1) {
      i++;
      d1 += pow(vocab[i].cn, power) / train_words_pow;
    }
    if (i >= vocab_size) i = vocab_size - 1;
  }
}

// Read a single word from file `fin` into length `MAX_STRING` array
// `word`, terminating with a null byte, treating space, tab, and
// newline as word boundaries.  Ignore carriage returns.  If the first
// character read (excluding carriage returns) is a newline, set `word`
// to "</s>" and return.  If a newline is encountered after reading one
// or more non-boundary characters, put that newline back into the
// stream and leave word as the non-boundary characters up to that point
// (so that the next time this function is called the newline will be
// read and `word` will be set to "</s>").  If the number of
// non-boundary characters is equal to or greater than `MAX_STRING`,
// read and ignore the trailing characters.  If we reach end of file,
// set `eof` to 1.
void ReadWord(char *word, FILE *fin, char *eof) {
  int a = 0, ch;
  while (1) {
    ch = fgetc_unlocked(fin);
    if (ch == EOF) {
      *eof = 1;
      break;
    }
    if (ch == '\r') continue;  // skip carriage returns
    if ((ch == ' ') || (ch == '\t') || (ch == '\n')) {
      if (a > 0) {
        if (ch == '\n') ungetc(ch, fin);
        break;
      }
      if (ch == '\n') {
        strcpy(word, (char *)"</s>");
        return;
      } else continue;
    }
    word[a] = ch;
    a++;
    if (a >= MAX_STRING - 1) a--;   // Truncate too-long words
  }
  word[a] = 0;
}

// Return hash (integer between 0, inclusive, and `vocab_hash_size`,
// exclusive) of `word`
int GetWordHash(char *word) {
  unsigned long long a, hash = 0;
  for (a = 0; a < strlen(word); a++) hash = hash * 257 + word[a];
  hash = hash % vocab_hash_size;
  return hash;
}

// Return position of `word` in vocabulary `vocab` using `vocab_hash`,
// a linear-probing hash table; if the word is not found, return -1.
int SearchVocab(char *word) {
  // compute initial hash
  unsigned int hash = GetWordHash(word);
  while (1) {
    // return -1 if cell is empty
    if (vocab_hash[hash] == -1) return -1;
    // return position at current hash if word is a match
    if (!strcmp(word, vocab[vocab_hash[hash]].word)) return vocab_hash[hash];
    // no match, increment hash
    hash = (hash + 1) % vocab_hash_size;
  }
  return -1;
}

// Read a word from file `fin` and return its position in vocabulary
// `vocab`.  If the next thing in the file is a newline, return 0 (the
// index of "</s>").  If the word is not in the vocabulary, return -1.
// If the end of file is reached, set `eof` to 1 and return -1.  TODO:
// if the file is not newline-terminated, the last word will be
// swallowed (-1 will be returned because we have reached EOF, even if
// a word was read).
int ReadWordIndex(FILE *fin, char *eof) {
  char word[MAX_STRING], eof_l = 0;
  ReadWord(word, fin, &eof_l);
  if (eof_l) {
    *eof = 1;
    return -1;
  }
  return SearchVocab(word);
}

// Add word `word` to first empty slot in vocabulary `vocab`, increment
// vocabulary length `vocab_size`, increase `vocab_max_size` by
// increments of 1000 (increasing `vocab` allocation) as needed, and set
// position of that slot in vocabulary hash table `vocab_hash`.
// Return index of `word` in `vocab`.
int AddWordToVocab(char *word) {
  unsigned int hash, length = strlen(word) + 1;
  if (length > MAX_STRING) length = MAX_STRING;
  // add word to `vocab` and increment `vocab_size`
  vocab[vocab_size].word = (char *)calloc(length, sizeof(char));
  // TODO: this may blow up if strlen(word) + 1 > MAX_STRING
  strcpy(vocab[vocab_size].word, word);
  vocab[vocab_size].cn = 0;
  vocab_size++;
  // Reallocate memory if needed
  if (vocab_size + 2 >= vocab_max_size) {
    vocab_max_size += 1000;
    vocab = (struct vocab_word *)realloc(vocab, vocab_max_size * sizeof(struct vocab_word));
  }
  // add word vocabulary position to `vocab_hash`
  hash = GetWordHash(word);
  while (vocab_hash[hash] != -1) hash = (hash + 1) % vocab_hash_size;
  vocab_hash[hash] = vocab_size - 1;
  return vocab_size - 1;
}

// Given pointers `a` and `b` to `vocab_word` structs, return 1 if the
// stored count `cn` of b is greater than that of a, -1 if less than,
// and 0 if equal.  (Comparator for a reverse sort by word count.)
int VocabCompare(const void *a, const void *b) {
  long long l = ((struct vocab_word *)b)->cn - ((struct vocab_word *)a)->cn;
  if (l > 0) return 1;
  if (l < 0) return -1;
  return 0;
}

// Sort vocabulary `vocab` by word count, decreasing, while removing
// words that have count less than `min_count`; re-compute `vocab_hash`
// accordingly; shrink vocab memory allocation to minimal size.
void SortVocab() {
  int a, size;
  unsigned int hash;
  // Sort the vocabulary but keep "</s>" at the first position
  qsort(&vocab[1], vocab_size - 1, sizeof(struct vocab_word), VocabCompare);
  // clear `vocab_hash` cells
  for (a = 0; a < vocab_hash_size; a++) vocab_hash[a] = -1;
  // save initial `vocab_size` (we will decrease `vocab_size` as we
  // prune infrequent words)
  size = vocab_size;
  // initialize total word count
  train_words = 0;
  for (a = 0; a < size; a++) {
    if ((vocab[a].cn < min_count) && (a != 0)) {
      // word is infrequent and not "</s>", discard it
      vocab_size--;
      free(vocab[a].word);
    } else {
      // word is frequent or "</s>", add to hash table
      hash=GetWordHash(vocab[a].word);
      while (vocab_hash[hash] != -1) hash = (hash + 1) % vocab_hash_size;
      vocab_hash[hash] = a;
      train_words += vocab[a].cn;
    }
  }
  // shrink vocab memory allocation to minimal size
  // TODO: to be safe we should probably update vocab_max_size which
  // seems to be interpreted as the allocation size
  vocab = (struct vocab_word *)realloc(vocab, (vocab_size + 1) * sizeof(struct vocab_word));
  // Allocate memory for the binary tree construction
  for (a = 0; a < vocab_size; a++) {
    vocab[a].code = (char *)calloc(MAX_CODE_LENGTH, sizeof(char));
    vocab[a].point = (int *)calloc(MAX_CODE_LENGTH, sizeof(int));
  }
}

// Reduce vocabulary `vocab` size by removing words with count equal to
// `min_reduce` or less, in order to make room in hash table (not for
// mitigating data sparsity).  Increment `min_reduce` by one, so that
// this function can be called in a loop until there is enough space.
void ReduceVocab() {
  int a, b = 0;
  unsigned int hash;
  for (a = 0; a < vocab_size; a++) if (vocab[a].cn > min_reduce) {
    vocab[b].cn = vocab[a].cn;
    vocab[b].word = vocab[a].word;
    b++;
  } else free(vocab[a].word);
  vocab_size = b;
  for (a = 0; a < vocab_hash_size; a++) vocab_hash[a] = -1;
  // recompute `vocab_hash` as we have removed some items and it may
  // now be broken
  for (a = 0; a < vocab_size; a++) {
    hash = GetWordHash(vocab[a].word);
    while (vocab_hash[hash] != -1) hash = (hash + 1) % vocab_hash_size;
    vocab_hash[hash] = a;
  }
  fflush(stdout);
  min_reduce++;
}

// Create binary Huffman tree from word counts in `vocab`, storing
// codes in `vocab`; frequent words will have short uniqe binary codes.
// Used by hierarchical softmax.
void CreateBinaryTree() {
  long long a, b, i, min1i, min2i, pos1, pos2, point[MAX_CODE_LENGTH];
  char code[MAX_CODE_LENGTH];
  long long *count = (long long *)calloc(vocab_size * 2 + 1, sizeof(long long));
  long long *binary = (long long *)calloc(vocab_size * 2 + 1, sizeof(long long));
  long long *parent_node = (long long *)calloc(vocab_size * 2 + 1, sizeof(long long));
  for (a = 0; a < vocab_size; a++) count[a] = vocab[a].cn;
  for (a = vocab_size; a < vocab_size * 2; a++) count[a] = 1e15;
  pos1 = vocab_size - 1;
  pos2 = vocab_size;
  // Following algorithm constructs the Huffman tree by adding one node at a time
  for (a = 0; a < vocab_size - 1; a++) {
    // First, find two smallest nodes 'min1, min2'
    if (pos1 >= 0) {
      if (count[pos1] < count[pos2]) {
        min1i = pos1;
        pos1--;
      } else {
        min1i = pos2;
        pos2++;
      }
    } else {
      min1i = pos2;
      pos2++;
    }
    if (pos1 >= 0) {
      if (count[pos1] < count[pos2]) {
        min2i = pos1;
        pos1--;
      } else {
        min2i = pos2;
        pos2++;
      }
    } else {
      min2i = pos2;
      pos2++;
    }
    count[vocab_size + a] = count[min1i] + count[min2i];
    parent_node[min1i] = vocab_size + a;
    parent_node[min2i] = vocab_size + a;
    binary[min2i] = 1;
  }
  // Now assign binary code to each vocabulary word
  for (a = 0; a < vocab_size; a++) {
    b = a;
    i = 0;
    while (1) {
      code[i] = binary[b];
      point[i] = b;
      i++;
      b = parent_node[b];
      if (b == vocab_size * 2 - 2) break;
    }
    vocab[a].codelen = i;
    vocab[a].point[0] = vocab_size - 2;
    for (b = 0; b < i; b++) {
      vocab[a].code[i - b - 1] = code[b];
      vocab[a].point[i - b] = point[b] - vocab_size;
    }
  }
  free(count);
  free(binary);
  free(parent_node);
}

// Compute vocabulary `vocab` and corresponding hash table `vocab_hash`
// from text in `train_file`.  Insert </s> as vocab item 0.  Prune vocab
// incrementally (as needed to keep number of items below effective hash
// table capacity).  After reading file, sort vocabulary by word count,
// decreasing.
void LearnVocabFromTrainFile() {
  char word[MAX_STRING], eof = 0;
  FILE *fin;
  long long a, i, wc = 0;
  for (a = 0; a < vocab_hash_size; a++) vocab_hash[a] = -1;
  fin = fopen(train_file, "rb");
  if (fin == NULL) {
    printf("ERROR: training data file not found!\n");
    exit(1);
  }
  vocab_size = 0;
  AddWordToVocab((char *)"</s>");
  while (1) {
    // TODO: if file is not newline-terminated, last word may be
    // swallowed
    ReadWord(word, fin, &eof);
    if (eof) break;
    train_words++;
    wc++;
    if ((debug_mode > 1) && (wc >= 1000000)) {
      printf("%lldM%c", train_words / 1000000, 13);
      fflush(stdout);
      wc = 0;
    }
    i = SearchVocab(word);
    if (i == -1) {
      a = AddWordToVocab(word);
      vocab[a].cn = 1;
    } else vocab[i].cn++;
    if (vocab_size > vocab_hash_size * 0.7) ReduceVocab();
  }
  SortVocab();
  if (debug_mode > 0) {
    printf("Vocab size: %lld\n", vocab_size);
    printf("Words in train file: %lld\n", train_words);
  }
  file_size = ftell(fin);
  fclose(fin);
}

// Write vocabulary `vocab` to file `save_vocab_file`, one word per
// line, with each line containing a word, a space, the word count, and
// a newline.
void SaveVocab() {
  long long i;
  FILE *fo = fopen(save_vocab_file, "wb");
  for (i = 0; i < vocab_size; i++) fprintf(fo, "%s %lld\n", vocab[i].word, vocab[i].cn);
  fclose(fo);
}

// Read vocabulary from file `read_vocab_file` that has one word per
// line, where each line contains a word, a space, the word count, and a
// newline.  Store vocabulary in `vocab` and update `vocab_hash`
// accordingly.  After reading, sort vocabulary by word count,
// decreasing.
void ReadVocab() {
  long long a, i = 0;
  char c, eof = 0;
  char word[MAX_STRING];
  FILE *fin = fopen(read_vocab_file, "rb");
  if (fin == NULL) {
    printf("Vocabulary file not found\n");
    exit(1);
  }
  for (a = 0; a < vocab_hash_size; a++) vocab_hash[a] = -1;
  vocab_size = 0;
  while (1) {
    // TODO: if file is not newline-terminated, last word may be
    // swallowed
    ReadWord(word, fin, &eof);
    if (eof) break;
    a = AddWordToVocab(word);
    fscanf(fin, "%lld%c", &vocab[a].cn, &c);
    i++;
  }
  SortVocab();
  if (debug_mode > 0) {
    printf("Vocab size: %lld\n", vocab_size);
    printf("Words in train file: %lld\n", train_words);
  }
  fin = fopen(train_file, "rb");
  if (fin == NULL) {
    printf("ERROR: training data file not found!\n");
    exit(1);
  }
  fseek(fin, 0, SEEK_END);
  file_size = ftell(fin);
  fclose(fin);
}

// Allocate memory for and initialize neural network parameters.  Each
// array has size `vocab_size` x `layer1_size`.
//
//   syn0: input word embeddings, initialized uniformly on
//         [-0.5/`layer1_size`, 0.5/`layer1_size`)
//   syn1: only used by hierarchical softmax; initialized to 0
//   syn1neg: output word embeddings; initialized to zero
void InitNet() {
  long long a, b;
  unsigned long long next_random = 1;
  a = posix_memalign((void **)&syn0, 128, (long long)vocab_size * layer1_size * sizeof(real));
  if (syn0 == NULL) {printf("Memory allocation failed\n"); exit(1);}
  if (hs) {
    a = posix_memalign((void **)&syn1, 128, (long long)vocab_size * layer1_size * sizeof(real));
    if (syn1 == NULL) {printf("Memory allocation failed\n"); exit(1);}
    for (a = 0; a < vocab_size; a++) for (b = 0; b < layer1_size; b++)
     syn1[a * layer1_size + b] = 0;
  }
  if (negative>0) {
    a = posix_memalign((void **)&syn1neg, 128, (long long)vocab_size * layer1_size * sizeof(real));
    if (syn1neg == NULL) {printf("Memory allocation failed\n"); exit(1);}
    for (a = 0; a < vocab_size; a++) for (b = 0; b < layer1_size; b++)
     syn1neg[a * layer1_size + b] = 0;
  }
  for (a = 0; a < vocab_size; a++) for (b = 0; b < layer1_size; b++) {
    next_random = next_random * (unsigned long long)25214903917 + 11;
    syn0[a * layer1_size + b] = (((next_random & 0xFFFF) / (real)65536) - 0.5) / layer1_size;
  }
  CreateBinaryTree();
}

// Given allocated and initialized vocabulary `vocab`, corresponding
// hash `vocab_hash`, and neural network parameters `syn0`, `syn1`, and
// `syn1neg`, train word2vec model on 1 / `num_threads` fraction of text
// in `train_file` (storing learned parameters in `syn0`, `syn1`, and
// `syn1neg`).  When cast to long long, `id` should be an integer
// between 0 (inclusive) and `num_threads` (exclusive) representing
// which training thread this is.
//
// Note main loop is broken down into procedures for hierarchical
// softmax versus negative sampling and those for continuous BOW versus
// skip-gram; ensure you are looking in the right code block (for your
// purposes)!  The added comments focus on the SGNS case.
void *TrainModelThread(void *id) {
  long long
    a,                     // loop counter among other things
    b,                     // offset of dynamic window in max window
                           //   (if 0, use all `window` output
                           //   words on each side of input word;
                           //   otherwise use `window - b` words
                           //   on each side)
    d,                     // loop counter among other things
    cw,                    // (used by CBOW)
    word,                  // index of output word in vocabulary
    last_word,             // index of input word in vocabulary
    sentence_length = 0,   // number of words read into the current
                           //   sentence
    sentence_position = 0, // position of current output word in
                           //   sentence
    word_count = 0,        // number of words seen so far in this
                           //   iteration
    last_word_count = 0,   // number of words seen so far in this
                           //   iteration, as of the most recent
                           //   update of terminal output and learning
                           //   rate
    l1,                    // input word row offset in syn0
    l2,                    // output word row offset in syn1, syn1neg
    c,                     // loop counter among other things
    target,                // index of output word or negatively-sampled
                           //   word in vocabulary
    label,                 // switch between output word (1) and
                           //   negatively-sampled word (0)
    local_iter = iter;     // iterations over this thread's chunk of the
                           //   data set left
  long long
    sen[MAX_SENTENCE_LENGTH + 1]; // index of word in vocabulary for
                                  //   each word in current sentence
  unsigned long long
    next_random = (long long)id;  // thread-specific RNG state
  char eof = 0;            // 1 if end of file has been reached
  real f, g;               // work space (values of sub-expressions in
                           //   gradient computation)
  clock_t now;             // current time during training
  // allocate memory for gradients
  real
    *neu1 = (real *)calloc(layer1_size, sizeof(real)),
    *neu1e = (real *)calloc(layer1_size, sizeof(real));
  FILE *fi = fopen(train_file, "rb");

  fseek(fi, file_size / (long long)num_threads * (long long)id, SEEK_SET);

  // iteratively read a sentence and train (update gradients) over it;
  // read over all sentences in this thread's chunk of the training
  // data `iter` times, then break
  while (1) {
    // every 10k words, update progress in terminal and update learning
    // rate
    if (word_count - last_word_count > 10000) {
      word_count_actual += word_count - last_word_count;
      last_word_count = word_count;
      if ((debug_mode > 1)) {
        now=clock();
        printf("%cAlpha: %f  Progress: %.2f%%  Words/thread/sec: %.2fk  ", 13, alpha,
         word_count_actual / (real)(iter * train_words + 1) * 100,
         word_count_actual / ((real)(now - start + 1) / (real)CLOCKS_PER_SEC * 1000));
        fflush(stdout);
      }
      // linear-decay learning rate (decreases from one toward zero
      // linearly in number of words seen, but thresholded below
      // at one ten-thousandth of initial learning rate)
      // (each word in the data is to be seen `iter` times)
      alpha = starting_alpha * (1 - word_count_actual / (real)(iter * train_words + 1));
      if (alpha < starting_alpha * 0.0001) alpha = starting_alpha * 0.0001;
    }

    // if we have finished training on the most recently-read sentence
    // (or we are just starting training), read a new sentence;
    // truncate each sentence at `MAX_SENTENCE_LENGTH` words (sentences
    // longer than that will be broken up into smaller sentences)
    if (sentence_length == 0) {
      // iteratively read word and add to sentence
      while (1) {
        word = ReadWordIndex(fi, &eof);
        if (eof) break;
        // skip OOV
        if (word == -1) continue;
        word_count++;
        // if EOS, we're done reading this sentence
        if (word == 0) break;
        // The subsampling randomly discards frequent words while keeping the ranking same
        if (sample > 0) {
          // discard w.p. 1 - [ sqrt(t / p_{word}) + t / p_{word} ]
          // (t is the subsampling threshold `sample`, p_{word} is the
          // ML estimate of the probability of `word` in a unigram LM
          // (normalized frequency)
          // TODO: why is this not merely 1 - sqrt(t / p_{word}) as in
          // the paper?
          real ran = (sqrt(vocab[word].cn / (sample * train_words)) + 1) * (sample * train_words) / vocab[word].cn;
          next_random = next_random * (unsigned long long)25214903917 + 11;
          if (ran < (next_random & 0xFFFF) / (real)65536) continue;
        }

        sen[sentence_length] = word;
        sentence_length++;
        // truncate long sentences
        if (sentence_length >= MAX_SENTENCE_LENGTH) break;
      }

      // set output word position to first word in sentence
      sentence_position = 0;
    }

    // if we are at the end of this iteration (sweep over the data),
    // restart and decrement `local_iter`
    if (eof || (word_count > train_words / num_threads)) {
      word_count_actual += word_count - last_word_count;
      local_iter--;
      if (local_iter == 0) break;
      word_count = 0;
      last_word_count = 0;
      // signal to read new sentence
      sentence_length = 0;
      fseek(fi, file_size / (long long)num_threads * (long long)id, SEEK_SET);
      continue;
    }

    // get index of output word
    word = sen[sentence_position];
    // skip OOV (TODO, checked OOV already when reading sentence?)
    if (word == -1) continue;
    // reset gradients to zero
    for (c = 0; c < layer1_size; c++) neu1[c] = 0;
    for (c = 0; c < layer1_size; c++) neu1e[c] = 0;
    // pick dynamic window offset (uniformly at random, between 0
    // (inclusive) and max window size `window` (exclusive))
    next_random = next_random * (unsigned long long)25214903917 + 11;
    b = next_random % window;

    // CBOW
    if (cbow) {
      // in -> hidden
      cw = 0;
      for (a = b; a < window * 2 + 1 - b; a++) if (a != window) {
        c = sentence_position - window + a;
        if (c < 0) continue;
        if (c >= sentence_length) continue;
        last_word = sen[c];
        if (last_word == -1) continue;
        for (c = 0; c < layer1_size; c++) neu1[c] += syn0[c + last_word * layer1_size];
        cw++;
      }
      if (cw) {
        for (c = 0; c < layer1_size; c++) neu1[c] /= cw;

        // CBOW HIERARCHICAL SOFTMAX
        if (hs) for (d = 0; d < vocab[word].codelen; d++) {
          f = 0;
          l2 = vocab[word].point[d] * layer1_size;
          // Propagate hidden -> output
          for (c = 0; c < layer1_size; c++) f += neu1[c] * syn1[c + l2];
          if (f <= -MAX_EXP) continue;
          else if (f >= MAX_EXP) continue;
          else f = expTable[(int)((f + MAX_EXP) * (EXP_TABLE_SIZE / MAX_EXP / 2))];
          // 'g' is the gradient multiplied by the learning rate
          g = (1 - vocab[word].code[d] - f) * alpha;
          // Propagate errors output -> hidden
          for (c = 0; c < layer1_size; c++) neu1e[c] += g * syn1[c + l2];
          // Learn weights hidden -> output
          for (c = 0; c < layer1_size; c++) syn1[c + l2] += g * neu1[c];
        }

        // CBOW NEGATIVE SAMPLING
        if (negative > 0) for (d = 0; d < negative + 1; d++) {
          if (d == 0) {
            target = word;
            label = 1;
          } else {
            next_random = next_random * (unsigned long long)25214903917 + 11;
            target = table[(next_random >> 16) % table_size];
            if (target == 0) target = next_random % (vocab_size - 1) + 1;
            if (target == word) continue;
            label = 0;
          }
          l2 = target * layer1_size;
          f = 0;
          for (c = 0; c < layer1_size; c++) f += neu1[c] * syn1neg[c + l2];
          if (f > MAX_EXP) g = (label - 1) * alpha;
          else if (f < -MAX_EXP) g = (label - 0) * alpha;
          else g = (label - expTable[(int)((f + MAX_EXP) * (EXP_TABLE_SIZE / MAX_EXP / 2))]) * alpha;
          for (c = 0; c < layer1_size; c++) neu1e[c] += g * syn1neg[c + l2];
          for (c = 0; c < layer1_size; c++) syn1neg[c + l2] += g * neu1[c];
        }

        // hidden -> in
        for (a = b; a < window * 2 + 1 - b; a++) if (a != window) {
          c = sentence_position - window + a;
          if (c < 0) continue;
          if (c >= sentence_length) continue;
          last_word = sen[c];
          if (last_word == -1) continue;
          for (c = 0; c < layer1_size; c++) syn0[c + last_word * layer1_size] += neu1e[c];
        }
      }

    // SKIP-GRAM
    } else {
      // loop over offsets within dynamic window
      // (relative to max window size)
      for (a = b; a < window * 2 + 1 - b; a++) if (a != window) {
        // compute position in sentence of input word
        // (output word pos - max window size + rel offset)
        c = sentence_position - window + a;
        // skip if input word position OOB
        // (note this is our main constraint on word position w.r.t. sentence
        // bounds)
        if (c < 0) continue;
        if (c >= sentence_length) continue;
        // compute input word index
        last_word = sen[c];
        // skip OOV (TODO checked already, should never fire)
        if (last_word == -1) continue;
        // compute input word row offset
        l1 = last_word * layer1_size;
        // initialize gradient for input word (work space)
        for (c = 0; c < layer1_size; c++) neu1e[c] = 0;

        // SKIP-GRAM HIERARCHICAL SOFTMAX
        if (hs) for (d = 0; d < vocab[word].codelen; d++) {
          f = 0;
          l2 = vocab[word].point[d] * layer1_size;
          // Propagate hidden -> output
          for (c = 0; c < layer1_size; c++) f += syn0[c + l1] * syn1[c + l2];
          if (f <= -MAX_EXP) continue;
          else if (f >= MAX_EXP) continue;
          else f = expTable[(int)((f + MAX_EXP) * (EXP_TABLE_SIZE / MAX_EXP / 2))];
          // 'g' is the gradient multiplied by the learning rate
          g = (1 - vocab[word].code[d] - f) * alpha;
          // Propagate errors output -> hidden
          for (c = 0; c < layer1_size; c++) neu1e[c] += g * syn1[c + l2];
          // Learn weights hidden -> output
          for (c = 0; c < layer1_size; c++) syn1[c + l2] += g * syn0[c + l1];
        }

        // SKIP-GRAM NEGATIVE SAMPLING
        if (negative > 0) for (d = 0; d < negative + 1; d++) {
          if (d == 0) {
            // fetch output word
            target = word;
            label = 1;
          } else {
            // fetch negative-sampled word
            next_random = next_random * (unsigned long long)25214903917 + 11;
            target = table[(next_random >> 16) % table_size];
            if (target == 0) target = next_random % (vocab_size - 1) + 1;
            if (target == word) continue;
            label = 0;
          }
          l2 = target * layer1_size; // output/neg-sample word row offset
          // compute f = < v_{w_I}', v_{w_O} >
          // (inner product for neg sample)
          f = 0;
          for (c = 0; c < layer1_size; c++) f += syn0[c + l1] * syn1neg[c + l2];
          // compute gradient coeff g = alpha * (label - 1 / (e^-f + 1))
          // (alpha is learning rate, label is 1 for output and 0 for neg)
          if (f > MAX_EXP) g = (label - 1) * alpha;
          else if (f < -MAX_EXP) g = (label - 0) * alpha;
          else g = (label - expTable[(int)((f + MAX_EXP) * (EXP_TABLE_SIZE / MAX_EXP / 2))]) * alpha;
          // contribute to gradient for input word
          for (c = 0; c < layer1_size; c++) neu1e[c] += g * syn1neg[c + l2];
          // perform gradient step for output/neg-sample word
          for (c = 0; c < layer1_size; c++) syn1neg[c + l2] += g * syn0[c + l1];
        }

        // now that we've taken gradient step for output and all neg sample
        // words, take gradient step for input word
        for (c = 0; c < layer1_size; c++) syn0[c + l1] += neu1e[c];
      }
    }

    // update to next output word; if we are at end of sentence, signal
    // to read new sentence
    sentence_position++;
    if (sentence_position >= sentence_length) {
      sentence_length = 0;
      continue;
    }
  }

  // clean up
  fclose(fi);
  free(neu1);
  free(neu1e);
  pthread_exit(NULL);
}

// Train word embeddings on text in `train_file` using one or more
// threads, either learning vocabulary from that training data (in a
// separate pass over the data) or loading the vocabulary from a file
// `read_vocab_file`.  Optionally save vocabulary to a file
// `save_vocab_file`; either save word embeddings to a file
// `output_file` or, if `classes` is greater than zero, run k-means
// clustering and save those clusters to `output_file`.
//
// If `output_file` is empty (first byte is null), do not train; this
// can be used to learn the vocabulary only from a training text file.
void TrainModel() {
  long a, b, c, d; // loop counters among other things
  FILE *fo;        // output file
  pthread_t *pt = (pthread_t *)malloc(num_threads * sizeof(pthread_t));

  printf("Starting training using file %s\n", train_file);

  // initialize learning rate
  starting_alpha = alpha;

  // read vocab from file or learn from training data
  if (read_vocab_file[0] != 0) ReadVocab(); else LearnVocabFromTrainFile();
  // save vocab to file
  if (save_vocab_file[0] != 0) SaveVocab();
  // if no `output_file` is specified, exit (do not train)
  if (output_file[0] == 0) return;

  // initialize network parameters
  InitNet();
  // initialize negative sampling distribution
  if (negative > 0) InitUnigramTable();

  start = clock();
  for (a = 0; a < num_threads; a++) pthread_create(&pt[a], NULL, TrainModelThread, (void *)a);
  for (a = 0; a < num_threads; a++) pthread_join(pt[a], NULL);
  fo = fopen(output_file, "wb");
  if (classes == 0) {
    // Save the word vectors
    fprintf(fo, "%lld %lld\n", vocab_size, layer1_size);
    for (a = 0; a < vocab_size; a++) {
      fprintf(fo, "%s ", vocab[a].word);
      if (binary) for (b = 0; b < layer1_size; b++) fwrite(&syn0[a * layer1_size + b], sizeof(real), 1, fo);
      else for (b = 0; b < layer1_size; b++) fprintf(fo, "%lf ", syn0[a * layer1_size + b]);
      fprintf(fo, "\n");
    }
  } else {
    // Run K-means on the word vectors
    int clcn = classes, iter = 10, closeid;
    int *centcn = (int *)malloc(classes * sizeof(int));
    int *cl = (int *)calloc(vocab_size, sizeof(int));
    real closev, x;
    real *cent = (real *)calloc(classes * layer1_size, sizeof(real));
    for (a = 0; a < vocab_size; a++) cl[a] = a % clcn;
    for (a = 0; a < iter; a++) {
      for (b = 0; b < clcn * layer1_size; b++) cent[b] = 0;
      for (b = 0; b < clcn; b++) centcn[b] = 1;
      for (c = 0; c < vocab_size; c++) {
        for (d = 0; d < layer1_size; d++) cent[layer1_size * cl[c] + d] += syn0[c * layer1_size + d];
        centcn[cl[c]]++;
      }
      for (b = 0; b < clcn; b++) {
        closev = 0;
        for (c = 0; c < layer1_size; c++) {
          cent[layer1_size * b + c] /= centcn[b];
          closev += cent[layer1_size * b + c] * cent[layer1_size * b + c];
        }
        closev = sqrt(closev);
        for (c = 0; c < layer1_size; c++) cent[layer1_size * b + c] /= closev;
      }
      for (c = 0; c < vocab_size; c++) {
        closev = -10;
        closeid = 0;
        for (d = 0; d < clcn; d++) {
          x = 0;
          for (b = 0; b < layer1_size; b++) x += cent[layer1_size * d + b] * syn0[c * layer1_size + b];
          if (x > closev) {
            closev = x;
            closeid = d;
          }
        }
        cl[c] = closeid;
      }
    }
    // Save the K-means classes
    for (a = 0; a < vocab_size; a++) fprintf(fo, "%s %d\n", vocab[a].word, cl[a]);
    free(centcn);
    free(cent);
    free(cl);
  }
  fclose(fo);
}

int ArgPos(char *str, int argc, char **argv) {
  int a;
  for (a = 1; a < argc; a++) if (!strcmp(str, argv[a])) {
    if (a == argc - 1) {
      printf("Argument missing for %s\n", str);
      exit(1);
    }
    return a;
  }
  return -1;
}

int main(int argc, char **argv) {
  int i;
  if (argc == 1) {
    printf("WORD VECTOR estimation toolkit v 0.1c\n\n");
    printf("Options:\n");
    printf("Parameters for training:\n");
    printf("\t-train <file>\n");
    printf("\t\tUse text data from <file> to train the model\n");
    printf("\t-output <file>\n");
    printf("\t\tUse <file> to save the resulting word vectors / word clusters\n");
    printf("\t-size <int>\n");
    printf("\t\tSet size of word vectors; default is 100\n");
    printf("\t-window <int>\n");
    printf("\t\tSet max skip length between words; default is 5\n");
    printf("\t-sample <float>\n");
    printf("\t\tSet threshold for occurrence of words. Those that appear with higher frequency in the training data\n");
    printf("\t\twill be randomly down-sampled; default is 1e-3, useful range is (0, 1e-5)\n");
    printf("\t-hs <int>\n");
    printf("\t\tUse Hierarchical Softmax; default is 0 (not used)\n");
    printf("\t-negative <int>\n");
    printf("\t\tNumber of negative examples; default is 5, common values are 3 - 10 (0 = not used)\n");
    printf("\t-threads <int>\n");
    printf("\t\tUse <int> threads (default 12)\n");
    printf("\t-iter <int>\n");
    printf("\t\tRun more training iterations (default 5)\n");
    printf("\t-min-count <int>\n");
    printf("\t\tThis will discard words that appear less than <int> times; default is 5\n");
    printf("\t-alpha <float>\n");
    printf("\t\tSet the starting learning rate; default is 0.025 for skip-gram and 0.05 for CBOW\n");
    printf("\t-classes <int>\n");
    printf("\t\tOutput word classes rather than word vectors; default number of classes is 0 (vectors are written)\n");
    printf("\t-debug <int>\n");
    printf("\t\tSet the debug mode (default = 2 = more info during training)\n");
    printf("\t-binary <int>\n");
    printf("\t\tSave the resulting vectors in binary moded; default is 0 (off)\n");
    printf("\t-save-vocab <file>\n");
    printf("\t\tThe vocabulary will be saved to <file>\n");
    printf("\t-read-vocab <file>\n");
    printf("\t\tThe vocabulary will be read from <file>, not constructed from the training data\n");
    printf("\t-cbow <int>\n");
    printf("\t\tUse the continuous bag of words model; default is 1 (use 0 for skip-gram model)\n");
    printf("\nExamples:\n");
    printf("./word2vec -train data.txt -output vec.txt -size 200 -window 5 -sample 1e-4 -negative 5 -hs 0 -binary 0 -cbow 1 -iter 3\n\n");
    return 0;
  }
  output_file[0] = 0;
  save_vocab_file[0] = 0;
  read_vocab_file[0] = 0;
  if ((i = ArgPos((char *)"-size", argc, argv)) > 0) layer1_size = atoi(argv[i + 1]);
  if ((i = ArgPos((char *)"-train", argc, argv)) > 0) strcpy(train_file, argv[i + 1]);
  if ((i = ArgPos((char *)"-save-vocab", argc, argv)) > 0) strcpy(save_vocab_file, argv[i + 1]);
  if ((i = ArgPos((char *)"-read-vocab", argc, argv)) > 0) strcpy(read_vocab_file, argv[i + 1]);
  if ((i = ArgPos((char *)"-debug", argc, argv)) > 0) debug_mode = atoi(argv[i + 1]);
  if ((i = ArgPos((char *)"-binary", argc, argv)) > 0) binary = atoi(argv[i + 1]);
  if ((i = ArgPos((char *)"-cbow", argc, argv)) > 0) cbow = atoi(argv[i + 1]);
  if (cbow) alpha = 0.05;
  if ((i = ArgPos((char *)"-alpha", argc, argv)) > 0) alpha = atof(argv[i + 1]);
  if ((i = ArgPos((char *)"-output", argc, argv)) > 0) strcpy(output_file, argv[i + 1]);
  if ((i = ArgPos((char *)"-window", argc, argv)) > 0) window = atoi(argv[i + 1]);
  if ((i = ArgPos((char *)"-sample", argc, argv)) > 0) sample = atof(argv[i + 1]);
  if ((i = ArgPos((char *)"-hs", argc, argv)) > 0) hs = atoi(argv[i + 1]);
  if ((i = ArgPos((char *)"-negative", argc, argv)) > 0) negative = atoi(argv[i + 1]);
  if ((i = ArgPos((char *)"-threads", argc, argv)) > 0) num_threads = atoi(argv[i + 1]);
  if ((i = ArgPos((char *)"-iter", argc, argv)) > 0) iter = atoi(argv[i + 1]);
  if ((i = ArgPos((char *)"-min-count", argc, argv)) > 0) min_count = atoi(argv[i + 1]);
  if ((i = ArgPos((char *)"-classes", argc, argv)) > 0) classes = atoi(argv[i + 1]);
  vocab = (struct vocab_word *)calloc(vocab_max_size, sizeof(struct vocab_word));
  vocab_hash = (int *)calloc(vocab_hash_size, sizeof(int));
  // precompute e^x / (e^x + 1) for x in [-MAX_EXP, MAX_EXP)
  // TODO extra element (+ 1) seems unused?
  expTable = (real *)malloc((EXP_TABLE_SIZE + 1) * sizeof(real));
  for (i = 0; i < EXP_TABLE_SIZE; i++) {
    // Precompute the exp() table
    expTable[i] = exp((i / (real)EXP_TABLE_SIZE * 2 - 1) * MAX_EXP);
    // Precompute f(x) = x / (x + 1)
    expTable[i] = expTable[i] / (expTable[i] + 1);
  }
  TrainModel();
  return 0;
}