from transformers import BertTokenizer, BertForSequenceClassification, TrainingArguments, Trainer
import random
import torch
import numpy as np

with open('train/in.tsv') as f:
    data_train_X = f.readlines()

with open('train/expected.tsv') as f:
    data_train_Y = f.readlines()

with open('dev-0/in.tsv') as f:
    data_dev_X = f.readlines()

with open('test-A/in.tsv') as f:
    data_test_X = f.readlines()

class CustomDataset(torch.utils.data.Dataset):
    def __init__(self, encodings, labels=None):
        self.encodings = encodings
        self.labels = labels

    def __getitem__(self, idx):
        item = {key: torch.tensor(val[idx]) for key, val in self.encodings.items()}
        if self.labels:
            item["labels"] = torch.tensor(self.labels[idx])
        return item

    def __len__(self):
        return len(self.encodings["input_ids"])

tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")

model_path = "model/checkpoint-1500"
model = BertForSequenceClassification.from_pretrained(model_path, num_labels=2)

trainer = Trainer(model)

with open('train/out.tsv', 'w') as writer:
    train_encodings = tokenizer(data_train_X, truncation=True, padding=True)
    train_dataset = CustomDataset(train_encodings)
    raw_pred, _, _ = trainer.predict(train_dataset)
    for result in np.argmax(raw_pred, axis=1):
        writer.write(str(result) + '\n')

with open('dev-0/out.tsv', 'w') as writer:
    dev_encodings = tokenizer(data_dev_X, truncation=True, padding=True)
    dev_dataset = CustomDataset(dev_encodings)
    raw_pred, _, _ = trainer.predict(dev_dataset)
    for result in np.argmax(raw_pred, axis=1):
        writer.write(str(result) + '\n')

with open('test-A/out.tsv', 'w') as writer:
    test_encodings = tokenizer(data_test_X, truncation=True, padding=True)
    test_dataset = CustomDataset(test_encodings)
    raw_pred, _, _ = trainer.predict(test_dataset)
    for result in np.argmax(raw_pred, axis=1):
        writer.write(str(result) + '\n')