jooni22/dst-question

Label info:

0: "fragment",
1: "statement",
2: "question",
3: "command",
4: "rhetorical question",
5: "rhetorical command",
6: "intonation-dependent utterance"

Training process:

{'loss': 1.8008, 'grad_norm': 7.2770233154296875, 'learning_rate': 1e-05, 'epoch': 0.03}
{'loss': 0.894, 'grad_norm': 27.84651756286621, 'learning_rate': 2e-05, 'epoch': 0.06}
{'loss': 0.6504, 'grad_norm': 30.617990493774414, 'learning_rate': 3e-05, 'epoch': 0.09}
{'loss': 0.5939, 'grad_norm': 34.73934555053711, 'learning_rate': 4e-05, 'epoch': 0.12}
{'loss': 0.5786, 'grad_norm': 6.585583209991455, 'learning_rate': 5e-05, 'epoch': 0.15}
{'eval_loss': 0.5915874242782593, 'eval_accuracy': 0.8297766749379653, 'eval_f1': 0.8315132136625163, 'eval_precision': 0.8410462605264737, 'eval_recall': 0.8297766749379653, 'eval_runtime': 265.1144, 'eval_samples_per_second': 22.801, 'eval_steps_per_second': 1.426, 'epoch': 0.15}
{'loss': 0.5928, 'grad_norm': 10.66515064239502, 'learning_rate': 4.8276456394346784e-05, 'epoch': 0.18}
{'loss': 0.5611, 'grad_norm': 3.804234266281128, 'learning_rate': 4.655291278869355e-05, 'epoch': 0.21}
{'loss': 0.5151, 'grad_norm': 8.275078773498535, 'learning_rate': 4.4829369183040333e-05, 'epoch': 0.24}
{'loss': 0.4696, 'grad_norm': 2.44854474067688, 'learning_rate': 4.310582557738711e-05, 'epoch': 0.26}
{'loss': 0.5183, 'grad_norm': 8.534456253051758, 'learning_rate': 4.138228197173389e-05, 'epoch': 0.29}
{'eval_loss': 0.5429911017417908, 'eval_accuracy': 0.8415219189412738, 'eval_f1': 0.8231674368620022, 'eval_precision': 0.8383674385161947, 'eval_recall': 0.8415219189412738, 'eval_runtime': 268.1016, 'eval_samples_per_second': 22.547, 'eval_steps_per_second': 1.41, 'epoch': 0.29}
{'loss': 0.4802, 'grad_norm': 10.636425018310547, 'learning_rate': 3.965873836608066e-05, 'epoch': 0.32}
{'loss': 0.4877, 'grad_norm': 6.05213737487793, 'learning_rate': 3.793519476042744e-05, 'epoch': 0.35}
{'loss': 0.5093, 'grad_norm': 5.5984015464782715, 'learning_rate': 3.621165115477422e-05, 'epoch': 0.38}
{'loss': 0.496, 'grad_norm': 7.945780277252197, 'learning_rate': 3.4488107549120996e-05, 'epoch': 0.41}
{'loss': 0.5005, 'grad_norm': 5.778200626373291, 'learning_rate': 3.276456394346777e-05, 'epoch': 0.44}
{'eval_loss': 0.41184064745903015, 'eval_accuracy': 0.8684863523573201, 'eval_f1': 0.8635611747282996, 'eval_precision': 0.8629771033516368, 'eval_recall': 0.8684863523573201, 'eval_runtime': 270.0108, 'eval_samples_per_second': 22.388, 'eval_steps_per_second': 1.4, 'epoch': 0.44}
{'loss': 0.4436, 'grad_norm': 4.413114070892334, 'learning_rate': 3.1041020337814545e-05, 'epoch': 0.47}
{'loss': 0.4899, 'grad_norm': 18.563016891479492, 'learning_rate': 2.9317476732161327e-05, 'epoch': 0.5}
{'loss': 0.4637, 'grad_norm': 26.92985725402832, 'learning_rate': 2.7593933126508105e-05, 'epoch': 0.53}
{'loss': 0.4387, 'grad_norm': 7.494612693786621, 'learning_rate': 2.5870389520854876e-05, 'epoch': 0.56}
{'loss': 0.4401, 'grad_norm': 20.5152530670166, 'learning_rate': 2.4146845915201654e-05, 'epoch': 0.59}
{'eval_loss': 0.42229706048965454, 'eval_accuracy': 0.8663358147229115, 'eval_f1': 0.859666580414163, 'eval_precision': 0.8638930298685418, 'eval_recall': 0.8663358147229115, 'eval_runtime': 272.7465, 'eval_samples_per_second': 22.163, 'eval_steps_per_second': 1.386, 'epoch': 0.59}
{'loss': 0.4289, 'grad_norm': 10.1361665725708, 'learning_rate': 2.2423302309548433e-05, 'epoch': 0.62}
{'loss': 0.4193, 'grad_norm': 8.068666458129883, 'learning_rate': 2.0699758703895207e-05, 'epoch': 0.65}
{'loss': 0.4038, 'grad_norm': 8.713869094848633, 'learning_rate': 1.8976215098241985e-05, 'epoch': 0.68}
{'loss': 0.4073, 'grad_norm': 12.182595252990723, 'learning_rate': 1.7252671492588764e-05, 'epoch': 0.71}
{'loss': 0.4095, 'grad_norm': 13.43953800201416, 'learning_rate': 1.5529127886935542e-05, 'epoch': 0.74}
{'eval_loss': 0.3974127173423767, 'eval_accuracy': 0.8726220016542597, 'eval_f1': 0.8677290061110087, 'eval_precision': 0.8672987137526573, 'eval_recall': 0.8726220016542597, 'eval_runtime': 270.2975, 'eval_samples_per_second': 22.364, 'eval_steps_per_second': 1.398, 'epoch': 0.74}
{'loss': 0.3473, 'grad_norm': 16.423139572143555, 'learning_rate': 1.3805584281282317e-05, 'epoch': 0.76}
{'loss': 0.3982, 'grad_norm': 6.357703685760498, 'learning_rate': 1.2082040675629095e-05, 'epoch': 0.79}
{'loss': 0.3286, 'grad_norm': 4.977189064025879, 'learning_rate': 1.0358497069975871e-05, 'epoch': 0.82}
{'loss': 0.3712, 'grad_norm': 4.068944454193115, 'learning_rate': 8.634953464322648e-06, 'epoch': 0.85}
{'loss': 0.345, 'grad_norm': 6.266202926635742, 'learning_rate': 6.911409858669425e-06, 'epoch': 0.88}
{'eval_loss': 0.3740645945072174, 'eval_accuracy': 0.8822167080231597, 'eval_f1': 0.8780706451391699, 'eval_precision': 0.877925468669178, 'eval_recall': 0.8822167080231597, 'eval_runtime': 270.0795, 'eval_samples_per_second': 22.382, 'eval_steps_per_second': 1.4, 'epoch': 0.88}
{'loss': 0.4049, 'grad_norm': 10.76927375793457, 'learning_rate': 5.187866253016201e-06, 'epoch': 0.91}
{'loss': 0.3919, 'grad_norm': 12.331282615661621, 'learning_rate': 3.4643226473629783e-06, 'epoch': 0.94}
{'loss': 0.3576, 'grad_norm': 8.6154203414917, 'learning_rate': 1.7407790417097554e-06, 'epoch': 0.97}
{'loss': 0.3544, 'grad_norm': 10.01504135131836, 'learning_rate': 1.723543605653223e-08, 'epoch': 1.0}
{'train_runtime': 7076.4012, 'train_samples_per_second': 7.688, 'train_steps_per_second': 0.481, 'train_loss': 0.5087223172678522, 'epoch': 1.0}
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3401/3401 [1:57:56<00:00, 2.08s/it]
Training completed. Model saved.

Classification Report:

                                precision    recall  f1-score   support

                      fragment       0.95      0.92      0.94       597
                     statement       0.84      0.91      0.87      1811
                      question       0.95      0.94      0.94      1786
                       command       0.88      0.91      0.90      1296
           rhetorical question       0.73      0.62      0.67       174
            rhetorical command       0.86      0.56      0.68       108
intonation-dependent utterance       0.57      0.38      0.46       273

                      accuracy                           0.88      6045
                     macro avg       0.83      0.75      0.78      6045
                  weighted avg       0.88      0.88      0.88      6045

Predictions saved

Train code:

import pandas as pd
from sklearn.model_selection import train_test_split
from transformers import (
    RobertaTokenizerFast, 
    RobertaForSequenceClassification, 
    Trainer, 
    TrainingArguments,
    EarlyStoppingCallback
)
from datasets import Dataset
import torch
import numpy as np
from sklearn.metrics import accuracy_score, precision_recall_fscore_support, classification_report
from tensorflow.python.keras.optimizer_v2.adam import Adam

# Load and prepare data
train_df = pd.read_csv("./train_fix_v1.csv")
test_df = pd.read_csv("./test_fix_v1.csv")

# Convert to Dataset objects
train_dataset = Dataset.from_pandas(train_df)
test_dataset = Dataset.from_pandas(test_df)

# Initialize tokenizer and model
model_name = "FacebookAI/roberta-base"
tokenizer = RobertaTokenizerFast.from_pretrained(model_name)
model = RobertaForSequenceClassification.from_pretrained(
    model_name,
    num_labels=7,
    id2label={
        0: "fragment",
        1: "statement",
        2: "question",
        3: "command",
        4: "rhetorical question",
        5: "rhetorical command",
        6: "intonation-dependent utterance"
    },
    label2id={
        "fragment": 0,
        "statement": 1,
        "question": 2,
        "command": 3,
        "rhetorical question": 4,
        "rhetorical command": 5,
        "intonation-dependent utterance": 6
    }
)

# Tokenize function
def tokenize_function(examples):
    return tokenizer(examples["text"], padding="max_length", truncation=True, max_length=512)

# Tokenize datasets
tokenized_train = train_dataset.map(tokenize_function, batched=True)
tokenized_test = test_dataset.map(tokenize_function, batched=True)

# Compute metrics function for evaluation
def compute_metrics(pred):
    labels = pred.label_ids
    preds = pred.predictions.argmax(-1)
    precision, recall, f1, _ = precision_recall_fscore_support(labels, preds, average='weighted')
    acc = accuracy_score(labels, preds)
    return {
        'accuracy': acc,
        'f1': f1,
        'precision': precision,
        'recall': recall
    }

# Training arguments
training_args = TrainingArguments(
    output_dir="./roberta_base_stock",
    num_train_epochs=1,  # Ustawione na 10, ale z early stopping
    per_device_train_batch_size=16,
    per_device_eval_batch_size=16,
    warmup_steps=500,
    weight_decay=0.01,
    logging_dir='./logs',
    logging_steps=100,
    evaluation_strategy="steps",
    eval_steps=500,
    save_strategy="steps",
    save_steps=500,
    load_best_model_at_end=True,
    metric_for_best_model="f1",
    learning_rate=5e-05,
)

# Initialize Trainer
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_train,
    eval_dataset=tokenized_test,
    compute_metrics=compute_metrics,
    callbacks=[EarlyStoppingCallback(early_stopping_patience=3)]
)

# Train the model
trainer.train()

# Save the fine-tuned model
model.save_pretrained("./roberta_base_stock")
tokenizer.save_pretrained("./roberta_base_stock")

print("Training completed. Model saved.")

# Evaluate the model on the test set
print("Evaluating model on test set...")
test_results = trainer.evaluate(tokenized_test)

print("Test set evaluation results:")
for key, value in test_results.items():
    print(f"{key}: {value}")

# Perform predictions on the test set
test_predictions = trainer.predict(tokenized_test)

# Get predicted labels
predicted_labels = np.argmax(test_predictions.predictions, axis=1)
true_labels = test_predictions.label_ids

# Print classification report
print("\nClassification Report:")
print(classification_report(true_labels, predicted_labels, 
                            target_names=model.config.id2label.values()))

# Optional: Save predictions to CSV
test_df['predicted_label'] = predicted_labels
test_df.to_csv("./roberta_base_stock/test_predictions.csv", index=False)
print("Predictions saved")