add hyperparameters search

run.py

@@ -15,19 +15,12 @@ from torch import nn
 import torch
 
 from tqdm.notebook import tqdm
-
-embed_size = 300
-vocab_size = 30_000
-num_epochs = 1
 device = 'cuda'
-batch_size = 8192
-train_file_path = 'train/train.txt'
 
 
 # In[2]:
 
 
-# Function to extract words from a line of text
 def get_words_from_line(line):
     line = line.rstrip()
     yield '<s>'
@@ -35,13 +28,11 @@ def get_words_from_line(line):
         yield m.group(0).lower()
     yield '</s>'
 
-# Generator to read lines from a file
 def get_word_lines_from_file(file_name):
     with open(file_name, 'r', encoding='utf8') as fh:
         for line in fh:
             yield get_words_from_line(line)
 
-# Function to create 5-grams from a sequence
 def look_ahead_iterator(gen):
     prev2, prev1, next1, next2 = None, None, None, None
     for item in gen:
@@ -49,7 +40,10 @@ def look_ahead_iterator(gen):
             yield (prev2, prev1, next2, item, next1)
         prev2, prev1, next1, next2 = prev1, next1, next2, item
 
-# Dataset class for 5-grams
+
+# In[3]:
+
+
 class FiveGrams(IterableDataset):
     def __init__(self, text_file, vocabulary_size):
         self.vocab = build_vocab_from_iterator(
@@ -66,20 +60,6 @@ class FiveGrams(IterableDataset):
             (self.vocab[t] for t in itertools.chain.from_iterable(get_word_lines_from_file(self.text_file)))
         )
 
-# Instantiate the dataset
-train_dataset = FiveGrams(train_file_path, vocab_size)
-
-
-# In[3]:
-
-
-i = 0
-for x in train_dataset:
-    print(train_dataset.vocab.lookup_tokens(x))
-    if i >= 1:
-        break
-    i += 1
-
 
 # In[4]:
 
@@ -99,12 +79,14 @@ class SimpleFiveGramNeuralLanguageModel(nn.Module):
         out = self.linear2(out)
         return self.softmax(out)
 
-model = SimpleFiveGramNeuralLanguageModel(vocab_size, embed_size).to(device)
-
 
 # In[5]:
 
 
+def train(embed_size,vocab_size,num_epochs,batch_size,train_file_path):
+    train_dataset = FiveGrams(train_file_path, vocab_size)
+    model = SimpleFiveGramNeuralLanguageModel(vocab_size, embed_size).to(device)
+    
     data = DataLoader(train_dataset, batch_size=batch_size)
     optimizer = torch.optim.Adam(model.parameters())
     criterion = torch.nn.CrossEntropyLoss()
@@ -125,13 +107,16 @@ for _ in range(num_epochs):
             loss.backward()
             optimizer.step()
         step = 0
+        break
     model.eval()
 
-
+    return model, train_dataset.vocab
-# In[8]:
 
 
-def get_gap_candidates(words, n=20, vocab=train_dataset.vocab):
+# In[6]:
+
+
+def get_gap_candidates(words, model, vocab, n=20):
     ixs = vocab(words)
     ixs = torch.tensor(ixs).unsqueeze(0).to(device)
 
@@ -151,11 +136,11 @@ def clean(text):
     text = text.strip()
     return text
     
-def predictor(prefix, suffix):
+def predictor(prefix, suffix, model, vocab):
     prefix = clean(prefix)
     suffix = clean(suffix)
     words = prefix.split(' ')[-2:] + suffix.split(' ')[:2]
-    candidates = get_gap_candidates(words)
+    candidates = get_gap_candidates(words, model, vocab)
 
     probs_sum = 0
     output = ''
@@ -169,10 +154,10 @@ def predictor(prefix, suffix):
     return output
 
 
-# In[9]:
+# In[7]:
 
 
-def generate_result(input_path, output_path='out.tsv'):
+def generate_result(input_path,model, vocab, output_path='out.tsv'):
     lines = []
     with open(input_path, encoding='utf-8') as f:
         for line in f:
@@ -183,8 +168,113 @@ def generate_result(input_path, output_path='out.tsv'):
 
     with open(output_path, 'w', encoding='utf-8') as output_file:
         for prefix, suffix in tqdm(lines):
-            result = predictor(prefix, suffix)
+            result = predictor(prefix, suffix, model, vocab)
             output_file.write(result + '\n')
 
-generate_result('dev-0/in.tsv', output_path='dev-0/out.tsv')
+
+# In[8]:
+
+
+import subprocess
+
+def evaluate():
+    cmd = 'wsl bash -c "cd /mnt/d/UAM/MODELOWANIE/5GRAM && ./geval -t dev-0"'
+    result = subprocess.run(cmd, shell=True, capture_output=True, text=True)
+    return float(result.stdout)
+
+
+# In[9]:
+
+
+embed_sizes = [100,200,300]
+vocab_sizes = [10_000, 20_000, 30_000]
+num_epochss = [1]
+batch_sizes = [8192]
+train_file_paths = ['train/nano.txt', 'train/train.txt']
+
+results = []
+
+for embed_size in embed_sizes:
+    for vocab_size in vocab_sizes:
+        for num_epochs in num_epochss:
+            for batch_size in batch_sizes:
+                for train_file_path in train_file_paths:
+                    model, vocab = train(embed_size,vocab_size,num_epochs,batch_size,train_file_path)
+                    generate_result('dev-0/in.tsv', model, vocab, output_path='dev-0/out.tsv')
+                    result = evaluate()
+
+                    config = {"embed_size": embed_size, "vocab_size": vocab_size, "num_epochs": num_epochs, "batch_size": batch_size, "train_file_path": train_file_path, "perplexity": result }
+                    print(config)
+                    results.append( config  )
+
+
+# In[10]:
+
+
+results
+
+
+# In[23]:
+
+
+import matplotlib.pyplot as plt
+import numpy as np
+from scipy.interpolate import griddata
+
+# Sample data
+data = results
+
+# Extracting data
+vocab_size = [item['vocab_size'] for item in data if 'nano' not in item['train_file_path'] ]
+embed_size = [item['embed_size'] for item in data if 'nano' not in item['train_file_path'] ]
+perplexity = [item['perplexity'] for item in data if 'nano' not in item['train_file_path'] ]
+
+# Plotting
+grid_x, grid_y = np.meshgrid(np.linspace(min(vocab_size), max(vocab_size), 100),
+                             np.linspace(min(embed_size), max(embed_size), 100))
+grid_z = griddata((vocab_size, embed_size), perplexity, (grid_x, grid_y), method='cubic')
+
+# Plotting
+plt.figure(figsize=(10, 6))
+contour = plt.contourf(grid_x, grid_y, grid_z, cmap='viridis')
+plt.colorbar(contour, label='Perplexity')
+plt.scatter(vocab_size, embed_size, c='red')  # Optional: plot actual data points
+plt.xlabel('Vocab Size')
+plt.ylabel('Embed Size')
+plt.title('Embed Size vs Vocab Size with Perplexity for whole training set')
+plt.show()
+
+
+# In[22]:
+
+
+# Extracting data
+vocab_size = [item['vocab_size'] for item in data if 'nano' in item['train_file_path'] ]
+embed_size = [item['embed_size'] for item in data if 'nano' in item['train_file_path'] ]
+perplexity = [item['perplexity'] for item in data if 'nano' in item['train_file_path'] ]
+
+# Plotting
+grid_x, grid_y = np.meshgrid(np.linspace(min(vocab_size), max(vocab_size), 100),
+                             np.linspace(min(embed_size), max(embed_size), 100))
+grid_z = griddata((vocab_size, embed_size), perplexity, (grid_x, grid_y), method='cubic')
+
+# Plotting
+plt.figure(figsize=(10, 6))
+contour = plt.contourf(grid_x, grid_y, grid_z, cmap='viridis')
+plt.colorbar(contour, label='Perplexity')
+plt.scatter(vocab_size, embed_size, c='red')  # Optional: plot actual data points
+plt.xlabel('Vocab Size')
+plt.ylabel('Embed Size')
+plt.title('Embed Size vs Vocab Size with Perplexity for nano training set')
+plt.show()
+
+
+# In[26]:
+
+
+from math import log
+
+best_model_parameters = min(results, key=lambda x: x['perplexity'])
+best_model_parameters['logPerplexity'] = log(best_model_parameters['perplexity'])
+best_model_parameters