use LSTM

train.py

@@ -1,6 +1,54 @@
-#!/usr/bin/python3
+#!/usr/bin/env python3
+import torch.nn as nn
+from torch.utils.data import DataLoader, TensorDataset
+import pandas as pd
 import re, sys, pickle, random
 from nltk.corpus import stopwords
+import torch
+import ipdb as ip
+from string import punctuation
+from collections import Counter
+import numpy as np
+
+train_on_gpu = False
+
+class ClassifyLSTM(nn.Module):
+    def __init__(self, vocab_size, output_size, embedding_dim, hidden_dim, n_layers, drop_prob=0.5):
+        super().__init__()
+        self.output_size = output_size
+        self.n_layers = n_layers
+        self.hidden_dim = hidden_dim
+        
+        self.embedding = nn.Embedding(vocab_size, embedding_dim)
+        self.lstm = nn.LSTM(embedding_dim, hidden_dim, n_layers, dropout=drop_prob, batch_first=True)
+        
+        self.dropout = nn.Dropout(0.3)
+        
+        self.fc = nn.Linear(hidden_dim, output_size)
+        self.sig = nn.Sigmoid()
+
+    def forward(self, x, hidden):
+        batch_size = x.size(0)
+        embeds = self.embedding(x)
+        lstm_out, hidden = self.lstm(embeds, hidden)
+        
+        lstm_out = lstm_out.contiguous().view(-1, self.hidden_dim)
+        out = self.dropout(lstm_out)
+        out = self.fc(out)
+        sig_out = self.sig(out)
+        sig_out = sig_out.view(batch_size, -1)
+        sig_out = sig_out[:, -1]
+        return sig_out, hidden
+
+    def init_hidden(self, batch_size):
+        weight = next(self.parameters()).data
+
+        if (train_on_gpu):
+            hidden = (weight.new(self.n_layers, batch_size, self.hidden_dim).zero_().cuda(), weight.new(self.n_layers, batch_size, self.hidden_dim).zero_().cuda())
+        else:
+            hidden = (weight.new(self.n_layers, batch_size, self.hidden_dim).zero_(), weight.new(self.n_layers, batch_size, self.hidden_dim).zero_())
+        return hidden
+
 
 def clear_post(post):
     post = post.replace('\\n', ' ')
@@ -17,94 +65,143 @@ def clear_post(post):
     post_no_stop = [w for w in post if not w in stop_words]
     return post_no_stop
 
-# czy słowa musza byc setem?
-def create_vocabulary_and_documents(in_file, expected_file):
-    vocabulary = set()
-    posts = {}
-    with open(in_file) as in_f, open(expected_file) as exp_f:
-        for line, exp in zip(in_f, exp_f):
-            text, timestap = line.rstrip('\n').split('\t')
-            post = clear_post(text)
-            posts[" ".join(post)] = int(exp)
-            for word in post:
-                vocabulary.add(word)
-    with open('data', 'wb') as f:
-        pickle.dump([vocabulary, posts], f)
-    print("data created")
-    return vocabulary, posts
+def count_all_words(posts):
+    joint_posts = ' '.join(posts)
+    words = joint_posts.split()
+    count_words = Counter(words)
+    total_words = len(words)
+    sorted_words = count_words.most_common(total_words)
+    return sorted_words, total_words, count_words
 
-def create_mappings(vocabulary):
-    word_to_index_mapping = {}
-    index_to_word_mapping = {}
-    xi = 1
-    for word in vocabulary:
-        word_to_index_mapping[word] = xi
-        index_to_word_mapping[xi] = word
-        xi += 1
-    return word_to_index_mapping, index_to_word_mapping
+def pad_features(posts_int, seq_length):
+    features = np.zeros((len(posts_int), seq_length), dtype = int)
+    for i, post in enumerate(posts_int):
+        post_len = len(post)
+        if post_len <= seq_length:
+            zeroes = list(np.zeros(seq_length-post_len))
+            new = zeroes+post
+
+        elif post_len > seq_length:
+            new = post[0:seq_length]
+
+        features[i,:] = np.array(new)
+
+    return features
 
 def main():
-    if len(sys.argv) != 4:
-        print("syntax ./train.py model expected_file in_file")
+    if len(sys.argv) != 3:
         return
-    model = str(sys.argv[1])
-    expected_file = str(sys.argv[2])
-    in_file = str(sys.argv[3])
-    try:
-        with open("data", 'rb') as pos:
-            pickle_list = pickle.load(pos)
-        print("data loaded")
-        vocabulary = pickle_list[0]
-        posts = pickle_list[1]
-    except FileNotFoundError:
-        vocabulary, posts = create_vocabulary_and_documents(in_file, expected_file)
 
-    word_to_index_mapping, index_to_word_mapping = create_mappings(vocabulary)
+    in_file = sys.argv[1]
+    expected_file = sys.argv[2]
 
-    weights = []
-    for xi in range(0, len(vocabulary) + 1):
-        weights.append(random.uniform(-0.01,0.01))
+    posts = []
+    labels = []
+    with open(in_file, 'r') as f:
+        for line in f:
+            post = line.split('\t')[0].rstrip().lower()
+            post = ''.join([c for c in post if c not in punctuation])
+            posts.append(post)
 
-    learning_rate = 0.000000001
-    loss_sum = 0.0
-    loss_sum_counter = 0
-    lowest_loss_sum_weights = []
-    lowest_loss_sum = 10000.0
+    with open(expected_file, 'r') as f:
+        for line in f:
+            labels.append(int(line))
 
-    print(f"len of vocabulary {len(vocabulary)}")
-    # mozna ustawić na bardzo bardzo duzo
-    while loss_sum_counter != 10000:
-        try:
-            d, y = random.choice(list(posts.items()))
-            y_hat = weights[0]
-            tokens = d.split(' ')
-            for word in tokens:
-                # mozna tez cos pomyslec z count aby lepiej dzialalo
-                #print(f"{d.count(word)} : {word}")
-                y_hat += weights[word_to_index_mapping[word]] * tokens.count(word)
-                #print(f"{weights[word_to_index_mapping[word]]} : {word}")
+    sorted_words, total_words, count_words = count_all_words(posts)
+    vocab_to_int = {w:i+1 for i, (w,c) in enumerate(sorted_words)}
 
-            loss = (y_hat - y)**2
-            loss_sum += loss
-            delta = (y_hat - y) * learning_rate
-            if loss_sum_counter % 100 == 0:
-                print(f"{loss_sum_counter} : {loss_sum /1000} : {y_hat} : {delta} : {lowest_loss_sum}")
-                #loss_sum_counter = 0
-                loss_sum = 0
+    posts_int = []
+    for post in posts:
+        p = [vocab_to_int[w] for w in post.split()]
+        posts_int.append(p)
 
-            weights[0] -= delta
-            for word in tokens:
-                weights[word_to_index_mapping[word]] -= tokens.count(word) * delta
+    encoded_labels = np.array(labels)
 
-            if lowest_loss_sum > loss_sum and loss_sum != 0:
-                print(f"it happened, new lowest_sum {loss_sum}")
-                lowest_loss_sum = loss_sum
-                lowest_loss_sum_weights = weights
+    posts_len = [len(p) for p in posts_int]
+    pd.Series(posts_len).hist()
+    print(pd.Series(posts_len).describe())
+    # outliners
+    posts_int = [posts_int[i] for i, l in enumerate(posts_len) if l>2 ]
+    encoded_labels = [ encoded_labels[i] for i, l in enumerate(posts_len) if l> 2 ]
 
-            loss_sum_counter +=1
-        except KeyboardInterrupt:
-            break
-    #print(lowest_loss_sum_weights)
-    with open(model, 'wb') as f:
-        pickle.dump([weights, lowest_loss_sum_weights, word_to_index_mapping], f)
+    seq_length = 63
+    train_x = pad_features(posts_int, seq_length)
+    train_y = np.array(encoded_labels)
+
+    train_data = TensorDataset(torch.from_numpy(train_x), torch.from_numpy(train_y))
+    batch_size = 50
+
+    train_loader = DataLoader(train_data, shuffle=True, batch_size=batch_size, drop_last = True)
+
+    dataiter = iter(train_loader)
+    sample_x, sample_y = dataiter.next()
+
+    vocab_size = len(vocab_to_int)+1
+    output_size = 1
+    embedding_dim = 400
+    hidden_dim = 256
+    n_layers = 2
+
+    model = ClassifyLSTM(vocab_size, output_size, embedding_dim, hidden_dim, n_layers)
+    
+    lr=0.001
+    criterion = nn.BCELoss()
+    optimizer = torch.optim.Adam(model.parameters(), lr=lr)
+    epochs = 4
+    counter = 0
+    print_every = 100
+    clip=5
+    if(train_on_gpu):
+        model.cuda()
+
+    model.train()
+    for e in range(epochs):
+        h = model.init_hidden(batch_size)
+        for inputs, labels in train_loader:
+            counter += 1
+            if(train_on_gpu):
+                inputs, labels = inputs.cuda(), labels.cuda()
+            h = tuple([each.data for each in h])
+            model.zero_grad()
+            inputs = inputs.type(torch.LongTensor)
+            output, h = model(inputs, h)
+            loss = criterion(output.squeeze(), labels.float())
+            loss.backward()
+            nn.utils.clip_grad_norm_(model.parameters(), clip)
+            optimizer.step()
+
+            if counter % print_every == 0:
+                val_h = model.init_hidden(batch_size)
+                val_losses = []
+                model.eval()
+                #for inputs, labels in valid_loader:
+                #    val_h = tuple([each.data for each in val_h])
+                #    if(train_on_gpu):
+                #        inputs, labels = inputs.cuda(), labels.cuda()
+                #    inputs = inputs.type(torch.LongTensor)
+                #    output, val_h = model(inputs, val_h)
+                #    val_loss = criterion(output.squeeze(), labels.float())
+                #    val_losses.append(val_loss.item())
+                #model.train()
+                print("Epoch: {}/{}...".format(e+1, epochs),
+                                        "Step: {}...".format(counter),
+                                        "Loss: {:.6f}...".format(loss.item()),
+                                        "Val Loss: {:.6f}".format(np.mean(val_losses)))
+
+#    test_losses = []
+#    num_correct = 0
+#    h = model.init_hidden(batch_size)
+#    model.eval()
+#    for inputs, labels in test_loader:
+#        h = tuple([each.data for each in h])
+#        if(train_on_gpu):
+#            inputs, labels = inputs.cuda(), labels.cuda()
+#
+#        inputs = inputs.type(torch.LongTensor)
+#        output, h = model(inputs, h)
+#        test_loss = criterion(output.squeeze(), labels.float())
+#        test_losses.append(test_loss.item())
+
+
+    ip.set_trace()
 main()