SDL_LSTM / LSTM+fasttext.py
kbeslic's picture
Upload LSTM+fasttext.py
4a2ff97 verified
Raw
History Blame Contribute Delete
8.72 kB
import collections
import pandas as pd
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
import tqdm
from sklearn.model_selection import GroupShuffleSplit
from sklearn.metrics import accuracy_score, precision_recall_fscore_support, confusion_matrix
seed = 1234
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cudnn.deterministic = True
# 1. Load and Clean Dataset
df = pd.read_csv("Projekt1.csv", sep=";")
df = df[["review_id", "text", "label"]].dropna()
df["label"] = df["label"].str.lower().str.strip()
# Map the 5 Croatian classes
label_mapping = {"negative": 0, "neutral": 1, "positive": 2, "mixed": 3, "sarcasm": 4}
df["label"] = df["label"].map(label_mapping)
df = df.dropna().reset_index(drop=True)
# 2. GroupShuffleSplit (80% Train -> 20% Val, 20% Test)
gss_test = GroupShuffleSplit(n_splits=1, test_size=0.20, random_state=seed)
train_idx, test_idx = next(gss_test.split(df, groups=df["review_id"]))
df_train_tmp, df_test = df.iloc[train_idx], df.iloc[test_idx]
gss_val = GroupShuffleSplit(n_splits=1, test_size=0.20, random_state=seed)
final_train_idx, val_idx = next(gss_val.split(df_train_tmp, groups=df_train_tmp["review_id"]))
df_train = df_train_tmp.iloc[final_train_idx].reset_index(drop=True)
df_val = df_train_tmp.iloc[val_idx].reset_index(drop=True)
df_test = df_test.reset_index(drop=True)
# 3. Simple Tokenization & Vocab Build (Closest to your original style)
def tokenizer(text):
return str(text).lower().split()
max_length = 256
min_freq = 5
special_tokens = ["<unk>", "<pad>"]
# Build vocab from training tokens
token_counts = collections.Counter([tok for text in df_train["text"] for tok in tokenizer(text)])
vocab_words = [word for word, freq in token_counts.items() if freq >= min_freq]
vocab = special_tokens + vocab_words
word_to_id = {word: idx for idx, word in enumerate(vocab)}
unk_index, pad_index = word_to_id["<unk>"], word_to_id["<pad>"]
# 4. Numericalize Dataset
def process_dataset(dataframe):
data = []
for _, row in dataframe.iterrows():
tokens = tokenizer(row["text"])[:max_length]
length = max(len(tokens), 1) # Prevent 0 length crashes
ids = [word_to_id.get(tok, unk_index) for tok in tokens]
data.append({"ids": torch.tensor(ids, dtype=torch.long),
"length": torch.tensor(length),
"label": torch.tensor(row["label"], dtype=torch.long)})
return data
train_data = process_dataset(df_train)
valid_data = process_dataset(df_val)
test_data = process_dataset(df_test)
# 5. Collate and Data Loaders (Exactly like your original style)
def get_collate_fn(pad_index):
def collate_fn(batch):
batch_ids = nn.utils.rnn.pad_sequence([i["ids"] for i in batch], padding_value=pad_index, batch_first=True)
batch_length = torch.stack([i["length"] for i in batch])
batch_label = torch.stack([i["label"] for i in batch])
return {"ids": batch_ids, "length": batch_length, "label": batch_label}
return collate_fn
batch_size = 256
collate_fn = get_collate_fn(pad_index)
train_data_loader = torch.utils.data.DataLoader(train_data, batch_size=batch_size, shuffle=True, collate_fn=collate_fn)
valid_data_loader = torch.utils.data.DataLoader(valid_data, batch_size=batch_size, collate_fn=collate_fn)
test_data_loader = torch.utils.data.DataLoader(test_data, batch_size=batch_size, collate_fn=collate_fn)
# 6. LSTM Model Architecture
class LSTM(nn.Module):
def __init__(self, vocab_size, embedding_dim, hidden_dim, output_dim, n_layers, bidirectional, dropout_rate, pad_index):
super().__init__()
self.embedding = nn.Embedding(vocab_size, embedding_dim, padding_idx=pad_index)
self.lstm = nn.LSTM(embedding_dim, hidden_dim, n_layers, bidirectional=bidirectional, dropout=dropout_rate if n_layers > 1 else 0, batch_first=True)
self.fc = nn.Linear(hidden_dim * 2 if bidirectional else hidden_dim, output_dim)
self.dropout = nn.Dropout(dropout_rate)
def forward(self, ids, length):
embedded = self.dropout(self.embedding(ids))
packed_embedded = nn.utils.rnn.pack_padded_sequence(embedded, length.to("cpu"), batch_first=True, enforce_sorted=False)
packed_output, (hidden, cell) = self.lstm(packed_embedded)
if self.lstm.bidirectional:
hidden = self.dropout(torch.cat([hidden[-1], hidden[-2]], dim=-1))
else:
hidden = self.dropout(hidden[-1])
return self.fc(hidden)
# Initialize Model (5 output classes)
model = LSTM(len(vocab), 300, 256, 5, 2, True, 0.5, pad_index)
# 7. Simplified FastText Loading
# Place your 'cc.hr.300.vec' file in the same folder
try:
with open("cc.hr.300.vec", "r", encoding="utf-8") as f:
next(f) # Skip header
ft_embeddings = {}
for line in f:
parts = line.strip().split(" ")
ft_embeddings[parts[0]] = np.array(parts[1:], dtype=np.float32)
weights = np.random.normal(scale=0.6, size=(len(vocab), 300))
for idx, word in enumerate(vocab):
if word in ft_embeddings:
weights[idx] = ft_embeddings[word]
weights[pad_index] = np.zeros(300)
model.embedding.weight.data.copy_(torch.from_numpy(weights))
print("FastText Embeddings Loaded!")
except FileNotFoundError:
print("FastText file not found. Training with random weights.")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
optimizer = optim.Adam(model.parameters(), lr=5e-4)
criterion = nn.CrossEntropyLoss().to(device)
model = model.to(device)
# 8. Updated Metrics Engine (Accuracy, Precision, Recall, F1)
def compute_metrics(all_preds, all_labels):
preds = np.concatenate(all_preds)
labels = np.concatenate(all_labels)
acc = accuracy_score(labels, preds)
prec, rec, f1, _ = precision_recall_fscore_support(labels, preds, average="weighted", zero_division=0)
cm = confusion_matrix(labels, preds, labels=[0, 1, 2, 3, 4])
return acc, prec, rec, f1, cm
# 9. Train and Evaluate Functions
def run_epoch(dataloader, model, criterion, optimizer=None, is_train=True):
if is_train:
model.train()
else:
model.eval()
epoch_losses = []
all_preds, all_labels = [], []
context = torch.enable_grad() if is_train else torch.no_grad()
with context:
for batch in tqdm.tqdm(dataloader, desc="Processing..."):
ids = batch["ids"].to(device)
length = batch["length"]
label = batch["label"].to(device)
prediction = model(ids, length)
loss = criterion(prediction, label)
if is_train:
optimizer.zero_grad()
loss.backward()
optimizer.step()
epoch_losses.append(loss.item())
all_preds.append(prediction.argmax(dim=-1).cpu().numpy())
all_labels.append(label.cpu().numpy())
acc, prec, rec, f1,cm = compute_metrics(all_preds, all_labels)
return np.mean(epoch_losses), acc, prec, rec, f1,cm
# 10. Training Loop
n_epochs = 10
best_valid_loss = float("inf")
for epoch in range(n_epochs):
train_loss, train_acc, _, _, train_f1, _ = run_epoch(train_data_loader, model, criterion, optimizer, is_train=True)
valid_loss, valid_acc, _, _, valid_f1, _ = run_epoch(valid_data_loader, model, criterion, is_train=False)
if valid_loss < best_valid_loss:
best_valid_loss = valid_loss
torch.save(model.state_dict(), "lstm_croatian.pt")
print(f"Epoch: {epoch+1} | Train Loss: {train_loss:.3f} | Train Acc: {train_acc*100:.1f}% | Train F1: {train_f1:.2f}")
print(f"Val Loss: {valid_loss:.3f} | Val Acc: {valid_acc*100:.1f}% | Val F1: {valid_f1:.2f}")
# 11. Final Test Evaluation
model.load_state_dict(torch.load("lstm_croatian.pt"))
test_loss, test_acc, test_prec, test_rec, test_f1, test_cm = run_epoch(test_data_loader, model, criterion, is_train=False)
print("\n=== FINAL TEST METRICS ===")
print(f"Accuracy: {test_acc*100:.2f}%")
print(f"Precision: {test_prec:.3f}")
print(f"Recall: {test_rec:.3f}")
print(f"F1-Score: {test_f1:.3f}")
label_list = ["negative", "neutral", "positive", "mixed", "sarcasm"]
print("\nConfusion matrix:")
print(" " * 12 + " ".join(f"{lbl:10}" for lbl in label_list))
for i, lbl in enumerate(label_list):
print(f"{lbl:12} " + " ".join(f"{test_cm[i, j]:10d}" for j in range(len(label_list))))