well...

.gitignore

@@ -6,3 +6,8 @@
 *.o
 .DS_Store
 .token
+venv/*
+*.pickle
+.idea/*
+.vscode/*
+in.tsv

main.py

@@ -0,0 +1,54 @@
+"""
+Zadanie domowe
+wybrać jedno z poniższych repozytoriów i je sforkować:
+https://git.wmi.amu.edu.pl/kubapok/paranormal-or-skeptic-ISI-public
+https://git.wmi.amu.edu.pl/kubapok/sport-text-classification-ball-ISI-public
+stworzyć klasyfikator bazujący na prostej sieci neuronowej feed forward w pytorchu (można bazować na tym jupyterze). 
+Zamiast tfidf proszę skorzystać z jakieś reprezentacji gęstej (np. word2vec).
+stworzyć predykcje w plikach dev-0/out.tsv oraz test-A/out.tsv
+wynik accuracy sprawdzony za pomocą narzędzia geval (patrz poprzednie zadanie) powinien wynosić conajmniej 0.67
+proszę umieścić predykcję oraz skrypty generujące (w postaci tekstowej a nie jupyter) w repo, a w MS TEAMS umieścić link do
+ swojego repo termin 25.05, 70 punktów
+
+"""
+
+import pandas as pd
+import spacy
+from net import FFN
+import numpy as np
+import torch
+from utils import create_embeddings_file, load_embeddings_file
+from nltk.tokenize import word_tokenize
+# sp = spacy.load('en_core_web_sm')
+
+# def word2vec(word):
+#     return sp(word).vector
+    # return np.random.uniform(low=0.0, high=1.0, size=(384,))
+
+train_data = pd.read_csv("train/in.tsv", sep='\t')
+train_data.columns = ['PostText', 'Timestamp']
+train_expected = pd.read_csv("train/expected.tsv", sep='\t')
+train_expected.columns = ['Label']
+
+# test_data = pd.read_csv("test-A/in.tsv", sep='\t')
+# test_data.columns = ['PostText', 'Timestamp']
+
+# dev_data = pd.read_csv('dev-0/in.tsv', sep='\t')
+# dev_data.columns = ['PostText', 'Timestamp']
+# dev_expected = pd.read_csv('dev-0/expected.tsv', sep='\t')
+# dev_expected.columns = ['Label']
+
+# create_embeddings_file(dev_data['PostText'], 'dev-0/embeddings.csv', word2vec)
+# create_embeddings_file(test_data['PostText'], 'test-A/embeddings.csv', word2vec)
+# create_embeddings_file(train_data['PostText'], 'train/embeddings.csv', word2vec)
+
+# train_data = load_embeddings_file('train/embeddings.csv').to_numpy()
+# dev_data = load_embeddings_file('dev-0/embeddings.csv').to_numpy()
+# test_data = load_embeddings_file('test-A/embeddings.csv').to_numpy()
+
+model = FFN(300, 1, 300, 300, 0.01, 4, 100)
+# model.double()
+# model.train([np.asarray(word_tokenize(x)) for x in train_data['PostText']], train_expected['Label'])
+model.load()
+model.double()
+model.test([np.asarray(word_tokenize(x)) for x in train_data['PostText']], train_expected['Label'], "train/out.tsv")

net.py

@@ -0,0 +1,103 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.optim as optim
+from torchvision import transforms
+import pickle
+import numpy as np
+import pandas as pd
+from word2vec import Word2Vec
+
+class FFN(nn.Module):
+
+    def __init__(self, input_dim, output_dim, hidden1_size, hidden2_size, lr, epochs, batch_size):
+        super(FFN, self).__init__()
+        self.path = 'model1.pickle'
+        self.lr = lr
+        self.epochs = epochs
+        self.output_dim = output_dim
+        self.word2vec = Word2Vec()
+        self.word2vec.load()
+        self.batch_size = batch_size
+        self.input_dim = input_dim
+        self.fc1 = nn.Linear(batch_size, hidden1_size)
+        self.fc2 = nn.Linear(hidden1_size, hidden2_size)
+        self.fc3 = nn.Linear(hidden2_size, hidden2_size)
+        self.fc4 = nn.Linear(hidden2_size, hidden2_size)
+        self.fc5 = nn.Linear(hidden2_size, batch_size)
+
+    def forward(self, data):
+        data = F.relu(self.fc1(data))
+        data = F.relu(self.fc2(data))
+        data = F.relu(self.fc3(data))
+        data = F.relu(self.fc4(data))
+        data = F.sigmoid(self.fc5(data))
+        return data
+
+    def serialize(self):
+        with open(self.path, 'wb') as file:
+            pickle.dump(self, file)
+
+    def load(self):
+        with open(self.path, 'rb') as file:
+            self = pickle.load(file)
+
+    def batch(self, iterable, n=1):
+        l = len(iterable)
+        for ndx in range(0, l, n):
+            yield iterable[ndx:min(ndx + n, l)]
+
+    """
+    data is a tuple of embedding vector and a label of 0/1
+    """
+    def train(self, data, expected):
+        self.zero_grad()
+        criterion = torch.nn.BCELoss()
+        optimizer = optim.Adam(self.parameters(), lr=self.lr)
+        batch_size = self.batch_size
+        num_of_classes = self.output_dim
+        for epoch in range(self.epochs):
+            epoch_loss = 0.0
+            idx = 0
+            for i in range(0, int(len(data)/batch_size)*batch_size, batch_size):
+                inputs = data[i:i + batch_size]
+                labels = expected[i:i+ batch_size]
+                optimizer.zero_grad()
+                outputs = self.forward(torch.tensor(self.word2vec.list_of_sentences2vec(inputs)))
+                target = torch.tensor(labels.values).double()
+                loss = criterion(outputs.view(batch_size), target.view(-1,))
+                loss.backward()
+                optimizer.step()
+
+                epoch_loss += loss.item()
+                if(idx % 1000 == 0):
+                    print('epoch: {}, idx: {}, loss: {}'.format(epoch, idx, epoch_loss/1000))
+                    epoch_loss = 0
+                idx += 1
+        self.serialize()
+
+    def test(self, data, expected, path):
+        correct = 0
+        incorrect = 0
+        total = 0
+        predictions = []
+        batch_size = self.batch_size
+        for i in range(0, int(len(data)/batch_size)*batch_size, batch_size):
+            inputs = data[i:i + batch_size]
+            labels = expected[i:i+ batch_size]
+            predicted = self.forward(torch.tensor(self.word2vec.list_of_sentences2vec(inputs)))
+            score = [1 if x > 0.5 else 0 for x in predicted]
+
+            for x, y in zip(score, labels):
+                if(x == y):
+                    correct += 1
+                else:
+                    incorrect += 1
+            predictions.append(score)
+
+        print(correct)
+        print(incorrect)
+        print(correct/(incorrect + correct))
+        df = pd.DataFrame(np.asarray(predictions).reshape(int(len(data)/batch_size)*batch_size))
+        df.reset_index(drop=True, inplace=True)
+        df.to_csv(path, sep="\t", index=False)

utils.py

@@ -0,0 +1,11 @@
+import pandas as pd
+
+def create_embeddings_file(data, path, func):
+    out = []
+    for line in data:
+        out.append(func(line))
+    df = pd.DataFrame(out)
+    df.to_csv(path)
+
+def load_embeddings_file(path):
+    return pd.read_csv(path)

word2vec.py

@@ -0,0 +1,15 @@
+import gensim.downloader
+import numpy as np
+
+class Word2Vec():
+    def __init__(self) -> None:
+        pass
+
+    def load(self):
+        self.model = gensim.downloader.load('word2vec-google-news-300')
+
+    def sentence2vec(self, sentence):
+        return np.mean([self.model[word] if word in self.model else np.zeros(300) for word in sentence])
+
+    def list_of_sentences2vec(self, sentences):
+        return [self.sentence2vec(x) for x in sentences]