Neural bigrams

run.py

@@ -1,76 +1,212 @@
+# -*- coding: utf-8 -*-
+"""Untitled12.ipynb
+
+Automatically generated by Colaboratory.
+
+Original file is located at
+    https://colab.research.google.com/drive/1ia944iiX5i5KOxESwbcHksNJrG4L12U6
+"""
+
+!pip install torch regexp pandas
+
+!git clone --single-branch git://gonito.net/challenging-america-word-gap-prediction -b master
+
+!xzcat ./challenging-america-word-gap-prediction/train/in.tsv.xz > ./challenging-america-word-gap-prediction/train/in.tsv
+
+!xzcat ./challenging-america-word-gap-prediction/dev-0/in.tsv.xz > ./challenging-america-word-gap-prediction/dev-0/in.tsv
+
+!xzcat ./challenging-america-word-gap-prediction/test-A/in.tsv.xz > ./challenging-america-word-gap-prediction/test-A/in.tsv
+
 import pandas as pd
-import csv
+
+def read_train_data(file):
+    data = pd.read_csv(file, sep="\t", error_bad_lines=False, index_col=0, header=None)
+    with open('input_train.txt', 'w') as f:
+        for index, row in data[:500000].iterrows():
+            first_part = str(row[6]).replace('\\n', '')
+            sec_part = str(row[7]).replace('\\n', '')
+            if first_part != 'nan':
+                f.write(first_part + '\n')
+            if sec_part != 'nan':
+                f.write(sec_part + '\n')
+
+read_train_data('./challenging-america-word-gap-prediction/train/in.tsv')
+
+!head -10 input_train.txt
+
+from itertools import islice
+  import regex as re
+  import sys
+  from torchtext.vocab import build_vocab_from_iterator
+
+
+  def get_words_from_line(line):
+    line = line.rstrip()
+    yield '<s>'
+    for m in re.finditer(r'[\p{L}0-9\*]+|\p{P}+', line):
+       yield m.group(0).lower()
+    yield '</s>'
+
+
+  def get_word_lines_from_file(file_name):
+    with open(file_name, 'r') as fh:
+      for line in fh:
+         yield get_words_from_line(line)
+
+  vocab_size = 30000
+
+  vocab = build_vocab_from_iterator(
+      get_word_lines_from_file('input_train.txt'),
+      max_tokens = vocab_size,
+      specials = ['<unk>'])
+
+vocab['is']
+
+print(vocab['is'])
+
+from torch import nn
+  import torch
+
+  embed_size = 100
+
+  class SimpleBigramNeuralLanguageModel(nn.Module):
+    def __init__(self, vocabulary_size, embedding_size):
+        super(SimpleBigramNeuralLanguageModel, self).__init__()
+        self.model = nn.Sequential(
+            nn.Embedding(vocabulary_size, embedding_size),
+            nn.Linear(embedding_size, vocabulary_size),
+            nn.Softmax()
+        )
+
+    def forward(self, x):
+        return self.model(x)
+
+  model = SimpleBigramNeuralLanguageModel(vocab_size, embed_size)
+
+  vocab.set_default_index(vocab['<unk>'])
+
+!shuf < input_train.txt > input_train.shuf.txt
+
+from torch.utils.data import IterableDataset
+import itertools
+
+def look_ahead_iterator(gen):
+    prev = None
+    for item in gen:
+      if prev is not None:
+          yield (prev, item)
+      prev = item
+
+class Bigrams(IterableDataset):
+  def __init__(self, text_file, vocabulary_size):
+      self.vocab = build_vocab_from_iterator(
+          get_word_lines_from_file(text_file),
+          max_tokens = vocabulary_size,
+          specials = ['<unk>'])
+      self.vocab.set_default_index(self.vocab['<unk>'])
+      self.vocabulary_size = vocabulary_size
+      self.text_file = text_file
+
+  def __iter__(self):
+      return look_ahead_iterator(
+          (self.vocab[t] for t in itertools.chain.from_iterable(get_word_lines_from_file(self.text_file))))
+
+train_dataset = Bigrams('input_train.shuf.txt', vocab_size)
+
+from torch.utils.data import DataLoader
+
+
+device = 'cuda'
+model = SimpleBigramNeuralLanguageModel(vocab_size, embed_size).to(device)
+data = DataLoader(train_dataset, batch_size=5000)
+optimizer = torch.optim.Adam(model.parameters())
+criterion = torch.nn.NLLLoss()
+
+model.train()
+step = 0
+for x, y in data:
+    x = x.to(device)
+    y = y.to(device)
+    optimizer.zero_grad()
+    ypredicted = model(x)
+    loss = criterion(torch.log(ypredicted), y)
+    if step % 100 == 0:
+      print(step, loss)
+    step += 1
+    loss.backward()
+    optimizer.step()
+
+torch.save(model.state_dict(), 'model1.bin')
+
+device = 'cuda'
+model = SimpleBigramNeuralLanguageModel(vocab_size, embed_size).to(device)
+model.load_state_dict(torch.load('model1.bin'))
+model.eval()
+
+ixs = torch.tensor(vocab.forward(['he'])).to(device)
+
+out = model(ixs)
+top = torch.topk(out[0], 10)
+top_indices = top.indices.tolist()
+top_probs = top.values.tolist()
+top_words = vocab.lookup_tokens(top_indices)
+list(zip(top_words, top_indices, top_probs))
+
+import regex as re
+
+def predict_word(word):
+  device = 'cuda'
+  model = SimpleBigramNeuralLanguageModel(vocab_size, embed_size).to(device)
+  model.load_state_dict(torch.load('model1.bin'))
+  model.eval()
+
+  ixs = torch.tensor(vocab.forward([word])).to(device)
+
+  out = model(ixs)
+  top = torch.topk(out[0], 8)
+  top_indices = top.indices.tolist()
+  top_probs = top.values.tolist()
+  top_words = vocab.lookup_tokens(top_indices)
+  to_return = ''
+  total = 1.0
+  for el in list(zip(top_words, top_indices, top_probs)):
+    pattern = re.compile("^([A-Za-z0-9])+$")
+    if re.match(pattern, el[0]):
+      if total - top_probs[0] >= 0:
+        to_return += f'{el[0]}:{top_probs[0]} '
+        total -= top_probs[0]
+  if total != 1.0:
+    to_return += f':{total}'
+  return to_return
+
+!pip install nltk
+
 from nltk.tokenize import RegexpTokenizer
-from english_words import english_words_set
-from nltk import trigrams
-import os
-import kenlm
-from math import log10
+tokenizer = RegexpTokenizer(r"\w+")
 
+import csv
 
-class WordGapPrediction:
-
-    def __init__(self):
-        self.tokenizer = RegexpTokenizer(r"\w+")
-        self.model = None
-        self.vocab = set()
-        self.alpha = 0.6
-
-    def read_train_data(self, file):
-        data = pd.read_csv(file, sep="\t", error_bad_lines=False, index_col=0, header=None)
-        with open('input_train.txt', 'w') as f:
-            for index, row in data[:500000].iterrows():
-                first_part = str(row[6])
-                sec_part = str(row[7])
-                if first_part != 'nan':
-                    f.write(first_part + '\n')
-                if sec_part != 'nan':
-                    f.write(sec_part + '\n')
-        os.system('sh ./kenlm.sh')
-        self.model = kenlm.Model("model.binary")
-        
-    def generate_outputs(self, input_file, output_file):
-        data = pd.read_csv(input_file, sep='\t', error_bad_lines=False, index_col=0, header=None, quoting=csv.QUOTE_NONE)
-        with open(output_file, 'w') as f:
-            for index, row in data.iterrows():
-                first_context = row[6]
-                sec_context = row[7]
-                first_context_tokens = self.tokenizer.tokenize(first_context)
-                sec_context_tokens = self.tokenizer.tokenize(sec_context)
-                if len(first_context_tokens) + len(sec_context_tokens) < 4:
-                    prediction = 'the:0.2 be:0.2 to:0.2 of:0.1 and:0.1 a:0.1 :0.1'
-                else:
-                    prediction = word_gap_prediction.predict_probs(first_context_tokens[-1], sec_context_tokens[0])
-                f.write(prediction + '\n')
-
-    def predict_probs(self, word1, word2):
-
-        predictions = []
-        for word in english_words_set:
-            sentence = word1 + ' ' + word + ' ' + word2
-            text_score = self.model.score(sentence, bos=False, eos=False)
-
-            if len(predictions) < 12:
-                predictions.append((word, text_score))
+def generate_outputs(input_file, output_file):
+    data = pd.read_csv(input_file, sep='\t', error_bad_lines=False, index_col=0, header=None, quoting=csv.QUOTE_NONE)
+    with open(output_file, 'w') as f:
+        for index, row in data.iterrows():
+            first_context = row[6]
+            sec_context = row[7]
+            first_context_tokens = tokenizer.tokenize(first_context)
+            sec_context_tokens = tokenizer.tokenize(sec_context)
+            if len(first_context_tokens) + len(sec_context_tokens) < 4:
+                prediction = 'the:0.2 be:0.2 to:0.2 of:0.1 and:0.1 a:0.1 :0.1'
             else:
-                worst_score = None
-                for score in predictions:
-                    if not worst_score:
-                        worst_score = score
-                    else:
-                        if worst_score[1] > score[1]:
-                            worst_score = score
-                if worst_score[1] < text_score:
-                    predictions.remove(worst_score)
-                    predictions.append((word, text_score))
-        probs = sorted(predictions, key=lambda tup: tup[1], reverse=True)
-        pred_str = ''
-        for word, prob in probs:
-            pred_str += f'{word}:{prob} '
-        pred_str += f':{log10(0.99)}'
-        return pred_str
+                prediction = predict_word(first_context_tokens[-1])
+                if not prediction:
+                  prediction = 'the:0.2 be:0.2 to:0.2 of:0.1 and:0.1 a:0.1 :0.1'
+            f.write(prediction + '\n')
 
-word_gap_prediction = WordGapPrediction()
-word_gap_prediction.read_train_data('./train/in.tsv')
-word_gap_prediction.generate_outputs('dev-0/in.tsv', 'dev-0/out.tsv')
-word_gap_prediction.generate_outputs('test-A/in.tsv', 'test-A/out.tsv')
+generate_outputs('./challenging-america-word-gap-prediction/dev-0/in.tsv', './challenging-america-word-gap-prediction/dev-0/out.tsv')
+generate_outputs('./challenging-america-word-gap-prediction/test-A/in.tsv', './challenging-america-word-gap-prediction/test-A/out.tsv')
+
+!wget https://gonito.net/get/bin/geval
+
+!chmod u+x geval
+
+!./geval -t ./challenging-america-word-gap-prediction/dev-0/ --metric PerplexityHashed