README.md

@@ -1,9 +0,0 @@
-Challenging America word-gap prediction
-===================================
-
-Guess a word in a gap.
-
-Evaluation metric
------------------
-
-LikelihoodHashed is the metric

config.txt

@@ -1 +0,0 @@
---metric PerplexityHashed --precision 2  --in-header in-header.tsv  --out-header out-header.tsv

create_vocab.py

@@ -0,0 +1,30 @@
+from itertools import islice
+import regex as re
+import sys
+from torchtext.vocab import build_vocab_from_iterator
+import lzma
+import scripts
+
+
+
+def get_word_lines_from_file(file_name):
+  counter=0
+  with lzma.open(file_name, 'r') as fh:
+    for line in fh:
+      counter+=1
+      # if counter == 10000:
+      #   break
+      line = line.decode("utf-8")
+      yield scripts.get_words_from_line(line)
+
+
+vocab_size = scripts.vocab_size
+
+vocab = build_vocab_from_iterator(
+    get_word_lines_from_file('train/in.tsv.xz'),
+    max_tokens = vocab_size,
+    specials = ['<unk>'])
+
+import pickle
+with open("vocab.pickle", 'wb') as handle:
+    pickle.dump(vocab, handle)

gonito.yaml

@@ -0,0 +1,14 @@
+description: Zajęcia 8
+tags:
+  - trigram
+  - neural-network
+  - hidden-layer
+  - hiperparameters
+params:
+  epochs: 1,2,3,4
+  learning-rate: 0.0001,0.00001
+  batch-size: 6400,12800
+  training-set: 100000-lines
+links:
+  - title: "Git WMI"
+    url: "https://git.wmi.amu.edu.pl/s444463/neural_word_gap"

in-header.tsv

@@ -1 +0,0 @@
-FileId	Year	LeftContext	RightContext

inference.py

@@ -0,0 +1,107 @@
+from torch import nn
+import torch
+
+
+from torch.utils.data import IterableDataset
+import itertools
+import lzma
+import regex as re
+import pickle
+import scripts
+import os
+os.environ["CUDA_VISIBLE_DEVICES"] = "1"
+
+class SimpleTrigramNeuralLanguageModel(nn.Module):
+  def __init__(self, vocabulary_size, embedding_size):
+      super(SimpleTrigramNeuralLanguageModel, self).__init__()
+      self.embedings = nn.Embedding(vocabulary_size, embedding_size)
+      self.linear = nn.Linear(embedding_size*2, vocabulary_size)
+
+      self.linear_first_layer = nn.Linear(embedding_size*2, embedding_size*2)
+      self.relu = nn.ReLU()
+      self.softmax = nn.Softmax()
+
+    #   self.model = nn.Sequential(
+    #       nn.Embedding(vocabulary_size, embedding_size),
+    #       nn.Linear(embedding_size, vocabulary_size),
+    #       nn.Softmax()
+    #   )
+
+  def forward(self, x):
+      emb_1 = self.embedings(x[0])
+      emb_2 = self.embedings(x[1])
+
+      first_layer = self.linear_first_layer(torch.cat((emb_1, emb_2), dim=1))
+      after_relu = self.relu(first_layer)
+      concated = self.linear(after_relu)
+
+      y = self.softmax(concated)
+
+      return y
+
+vocab_size = scripts.vocab_size
+embed_size = 100
+device = 'cuda'
+
+model = SimpleTrigramNeuralLanguageModel(vocab_size, embed_size).to(device)
+
+model.load_state_dict(torch.load('batch_model_epoch_0.bin'))
+model.eval()
+
+with open("vocab.pickle", 'rb') as handle:
+    vocab = pickle.load(handle)
+vocab.set_default_index(vocab['<unk>'])
+
+
+step = 0
+
+
+with lzma.open('dev-0/in.tsv.xz', 'rb') as file:
+    for line in file:
+        line = line.decode('utf-8')
+        line = line.rstrip()
+        # line = line.lower()
+        line = line.replace("\\\\n", ' ')
+
+
+        line_splitted = line.split('\t')[-2:] 
+        
+        prev = list(scripts.get_words_from_line(line_splitted[0]))[-1]
+        next =  list(scripts.get_words_from_line(line_splitted[1]))[0]
+
+        # prev = line[0].split(' ')[-1]
+        # next = line[1].split(' ')[0]
+
+
+        x = torch.tensor(vocab.forward([prev]))
+        z = torch.tensor(vocab.forward([next]))
+        x = x.to(device)
+        z = z.to(device)
+        ypredicted = model([x, z])
+
+        try:
+            
+            top = torch.topk(ypredicted[0], 128)
+        except:
+            print(ypredicted[0])
+            raise Exception('aa')
+        top_indices = top.indices.tolist()
+        top_probs = top.values.tolist()
+        top_words = vocab.lookup_tokens(top_indices)
+
+        string_to_print = ''
+        sum_probs = 0
+
+        for w, p in zip(top_words, top_probs):
+            if '<unk>' in w:
+                continue
+            if re.search(r'\p{L}+', w):
+                string_to_print += f"{w}:{p} "
+                sum_probs += p
+        if string_to_print == '':
+            print(f"the:0.2 a:0.3 :0.5")
+            continue
+        unknow_prob = 1 - sum_probs
+        string_to_print += f":{unknow_prob}"
+
+        print(string_to_print)

out-header.tsv

@@ -1 +0,0 @@
-Word

run.py

@@ -1,153 +0,0 @@
-import lzma
-import matplotlib.pyplot as plt
-from math import log
-from collections import OrderedDict
-from collections import Counter
-import regex as re
-from itertools import islice
-
-def freq_list(g, top=None):
-    c = Counter(g)
-
-    if top is None:
-       items = c.items()
-    else:
-       items = c.most_common(top)
-
-    return OrderedDict(sorted(items, key=lambda t: -t[1]))
-
-def get_words(t):
-    for m in re.finditer(r'[\p{L}0-9-\*]+', t):
-        yield m.group(0)
-
-def ngrams(iter, size):
-  ngram = []
-  for item in iter:
-    ngram.append(item)
-    if len(ngram) == size:
-        yield tuple(ngram)
-        ngram = ngram[1:]
-
-PREFIX_TRAIN = 'train' 
-words = []
-
-counter_lines = 0
-with lzma.open(f'{PREFIX_TRAIN}/in.tsv.xz', 'r') as train, open(f'{PREFIX_TRAIN}/expected.tsv', 'r') as expected:
-    for t_line, e_line in zip(train, expected):
-        t_line = t_line.decode("utf-8")
-
-        t_line = t_line.rstrip()
-        e_line = e_line.rstrip()
-
-        t_line_splitted_by_tab = t_line.split('\t')
-        
-        t_line_cleared = t_line_splitted_by_tab[-2] + ' ' + e_line + ' ' + t_line_splitted_by_tab[-1]
-
-        words += t_line_cleared.split()
-
-        counter_lines+=1
-        if counter_lines > 90000:
-            break
-
-# lzmaFile = lzma.open('dev-0/in.tsv.xz', 'rb')
-
-# content = lzmaFile.read().decode("utf-8")
-# words = get_words(trainset)
-
-ngrams_ = ngrams(words, 2)
-
-
-def create_probabilities_bigrams(w_c, b_c):
-    probabilities_bigrams = {}
-    for bigram, bigram_amount in b_c.items():
-        if bigram_amount <=2:
-            continue
-        p_word_before = bigram_amount / w_c[bigram[0]] 
-        p_word_after = bigram_amount / w_c[bigram[1]]
-        probabilities_bigrams[bigram] = (p_word_before, p_word_after)
-
-    return probabilities_bigrams
-
-words_c = Counter(words)
-word_=''
-bigram_c = Counter(ngrams_)
-ngrams_=''
-probabilities = create_probabilities_bigrams(words_c, bigram_c)
-
-
-items = probabilities.items()
-probabilities = OrderedDict(sorted(items, key=lambda t:t[1], reverse=True))
-items=''
-# sorted_by_freq = freq_list(ngrams)
-
-PREFIX_VALID = 'test-A'
-
-def count_probabilities(w_b, w_a, probs, w_c, b_c):
-    results_before = {}
-    results_after = {}
-    for bigram, probses in probs.items():
-        if len(results_before) > 20 or len(results_after) > 20:
-            break
-        if w_b == bigram[0]:
-            results_before[bigram] = probses[0]
-        if w_a == bigram[1]:
-            results_after[bigram] = probses[1]
-    a=1
-    best_ = {}
-
-    for bigram, probses in results_before.items():
-        for bigram_2, probses_2 in results_after.items():
-            best_[bigram[1]] = probses * probses_2
-
-    for bigram, probses in results_after.items():
-            for bigram_2, probses_2 in results_before.items():
-                if bigram[0] in best_:
-                    if probses * probses_2 < probses_2:
-                        continue
-                best_[bigram[0]] = probses * probses_2
-
-    items = best_.items()
-    return OrderedDict(sorted(items, key=lambda t:t[1], reverse=True))
-
-
-with lzma.open(f'{PREFIX_VALID}/in.tsv.xz', 'r') as train:
-    for t_line in train:
-        t_line = t_line.decode("utf-8")
-
-        t_line = t_line.rstrip()
-        t_line = t_line.replace('\\n', ' ')
-
-
-        t_line_splitted_by_tab = t_line.split('\t')
-        
-
-        words_pre = t_line_splitted_by_tab[-2].split()
-
-        words_po = t_line_splitted_by_tab[-1].split()
-
-        w_pre = words_pre[-1]
-        w_po = words_po[0]
-
-        probs_ordered = count_probabilities(w_pre, w_po,probabilities, words_c, bigram_c)
-        if len(probs_ordered) ==0:
-            print(f"the:0.5 a:0.3 :0.2")
-            continue
-        result_string = ''
-        counter_ = 0
-        for word_, p in probs_ordered.items():
-            if counter_>4:
-                break
-            re_ = re.search(r'\p{L}+', word_)
-            if re_:
-                word_cleared = re_.group(0)
-                result_string += f"{word_cleared}:{str(p)} "
-
-            else:
-                if result_string == '':
-                    result_string = f"the:0.5 a:0.3 "
-                continue
-
-            counter_+=1
-        result_string += ':0.1'
-        print(result_string)
-        a=1

train.py

@@ -0,0 +1,124 @@
+
+
+from torch import nn
+import torch
+
+
+from torch.utils.data import IterableDataset
+import itertools
+import lzma
+import regex as re
+import pickle
+import scripts
+
+
+def look_ahead_iterator(gen):
+    prev = None
+    current = None
+    next = None
+    for next in gen:
+        if prev is not None and current is not None:
+            yield (prev, current, next)
+        prev = current
+        current = next
+
+
+def get_word_lines_from_file(file_name):
+  counter=0
+  with lzma.open(file_name, 'r') as fh:
+    for line in fh:
+      counter+=1
+      if counter == 100000:
+        break
+      line = line.decode("utf-8")
+      yield scripts.get_words_from_line(line)
+
+
+
+class Trigrams(IterableDataset):
+  def load_vocab(self):
+    with open("vocab.pickle", 'rb') as handle:
+        vocab = pickle.load( handle)
+    return vocab
+
+  def __init__(self, text_file, vocabulary_size):
+      self.vocab = self.load_vocab()
+      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))))
+
+vocab_size = scripts.vocab_size
+
+train_dataset = Trigrams('train/in.tsv.xz', vocab_size)
+
+
+
+#=== trenowanie
+from torch import nn
+import torch
+from torch.utils.data import DataLoader
+embed_size = 100
+
+class SimpleTrigramNeuralLanguageModel(nn.Module):
+  def __init__(self, vocabulary_size, embedding_size):
+      super(SimpleTrigramNeuralLanguageModel, self).__init__()
+      self.embedings = nn.Embedding(vocabulary_size, embedding_size)
+      self.linear = nn.Linear(embedding_size*2, vocabulary_size)
+
+      self.linear_first_layer = nn.Linear(embedding_size*2, embedding_size*2)
+      self.relu = nn.ReLU()
+      self.softmax = nn.Softmax()
+
+    #   self.model = nn.Sequential(
+    #       nn.Embedding(vocabulary_size, embedding_size),
+    #       nn.Linear(embedding_size, vocabulary_size),
+    #       nn.Softmax()
+    #   )
+
+  def forward(self, x):
+      emb_1 = self.embedings(x[0])
+      emb_2 = self.embedings(x[1])
+
+      first_layer = self.linear_first_layer(torch.cat((emb_1, emb_2), dim=1))
+      after_relu = self.relu(first_layer)
+      concated = self.linear(after_relu)
+
+      y = self.softmax(concated)
+
+      return y
+
+model = SimpleTrigramNeuralLanguageModel(vocab_size, embed_size)
+
+vocab = train_dataset.vocab
+
+
+device = 'cuda'
+model = SimpleTrigramNeuralLanguageModel(vocab_size, embed_size).to(device)
+data = DataLoader(train_dataset, batch_size=12800)
+optimizer = torch.optim.Adam(model.parameters(), lr=scripts.learning_rate)
+criterion = torch.nn.NLLLoss()
+
+model.train()
+step = 0
+epochs = 4
+for i in range(epochs):
+  for x, y, z in data:
+    x = x.to(device)
+    y = y.to(device)
+    z = z.to(device)
+    optimizer.zero_grad()
+    ypredicted = model([x, z])
+    loss = criterion(torch.log(ypredicted), y)
+    if step % 2000 == 0:
+        print(step, loss)
+    #     torch.save(model.state_dict(), f'model1_{step}.bin')
+    step += 1
+    loss.backward()
+    optimizer.step()
+  torch.save(model.state_dict(), f'batch_model_epoch_{i}.bin')
+  print(step, loss, f'model_epoch_{i}.bin')
+torch.save(model.state_dict(), 'model_tri1.bin')