from transformers import RobertaTokenizer, RobertaForSequenceClassification, RobertaModel
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import TensorDataset, DataLoader


device = torch.device("cpu")
class MLP(nn.Module):
    def __init__(self, input_dim):
        super(MLP, self).__init__()
        self.fc1 = nn.Linear(input_dim, 256)
        self.fc2 = nn.Linear(256, 2)
        self.gelu = nn.GELU()

    def forward(self, x):
        x = self.gelu(self.fc1(x))
        x = self.fc2(x)
        return x
def extract_features(text):

    tokenizer = RobertaTokenizer.from_pretrained("roberta-base")
    model = RobertaModel.from_pretrained("roberta-base").to(device)
    tokenized_text = tokenizer.encode(text, truncation=True, max_length=512, return_tensors="pt")
    outputs = model(tokenized_text)
    last_hidden_states = outputs.last_hidden_state
    TClassification = last_hidden_states[:, 0, :].squeeze().detach().numpy()
    return TClassification

def RobertaSentinelOpenGPTInference(input_text):
    features = extract_features(input_text)
    loaded_model = MLP(768).to(device)
    loaded_model.load_state_dict(torch.load("SentinelCheckpoint/RobertaSentinelOpenGPT.pth", map_location=device))

    # Define the tokenizer and model for feature extraction
    with torch.no_grad():
        inputs = torch.tensor(features).to(device)
        outputs = loaded_model(inputs.float())
        _, predicted = torch.max(outputs, 0)

        Probs = (F.softmax(outputs, dim=0).cpu().numpy())

    return Probs

def RobertaSentinelCSAbstractInference(input_text):
    features = extract_features(input_text)
    loaded_model = MLP(768).to(device)
    loaded_model.load_state_dict(torch.load("SentinelCheckpoint/RobertaSentinelCSAbstract.pth", map_location=device))

    # Define the tokenizer and model for feature extraction
    with torch.no_grad():
        inputs = torch.tensor(features).to(device)
        outputs = loaded_model(inputs.float())
        _, predicted = torch.max(outputs, 0)

        Probs = (F.softmax(outputs, dim=0).cpu().numpy())

    return Probs


def RobertaClassifierOpenGPTInference(input_text):
    tokenizer = RobertaTokenizer.from_pretrained("roberta-base")
    model_path = "ClassifierCheckpoint/RobertaClassifierOpenGPT512.pth"
    model = RobertaForSequenceClassification.from_pretrained('roberta-base', num_labels=2)
    model.load_state_dict(torch.load(model_path, map_location=device))
    model = model.to(device)
    model.eval()


    tokenized_input = tokenizer(input_text, truncation=True, padding=True, max_length=512, return_tensors='pt')
    input_ids = tokenized_input['input_ids'].to(device)
    attention_mask = tokenized_input['attention_mask'].to(device)

    # Make a prediction
    with torch.no_grad():
        outputs = model(input_ids, attention_mask=attention_mask)
    logits = outputs.logits
    Probs = F.softmax(logits, dim=1).cpu().numpy()[0]

    return Probs


def RobertaClassifierGPABenchmarkInference(input_text):
    tokenizer = RobertaTokenizer.from_pretrained("roberta-base")
    model_path = "ClassifierCheckpoint/RobertaClassifierGPABenchmark512.pth"
    model = RobertaForSequenceClassification.from_pretrained('roberta-base', num_labels=2)
    model.load_state_dict(torch.load(model_path, map_location=device))
    model = model.to(device)
    model.eval()


    tokenized_input = tokenizer(input_text, truncation=True, padding=True, max_length=512, return_tensors='pt')
    input_ids = tokenized_input['input_ids'].to(device)
    attention_mask = tokenized_input['attention_mask'].to(device)

    # Make a prediction
    with torch.no_grad():
        outputs = model(input_ids, attention_mask=attention_mask)
    logits = outputs.logits
    Probs = F.softmax(logits, dim=1).cpu().numpy()[0]

    return Probs

def RobertaClassifierCHEATInference(input_text):
    tokenizer = RobertaTokenizer.from_pretrained("roberta-base")
    model_path = "ClassifierCheckpoint/RobertaClassifierCHEAT256.pth"
    model = RobertaForSequenceClassification.from_pretrained('roberta-base', num_labels=2)
    model.load_state_dict(torch.load(model_path, map_location=device))
    model = model.to(device)
    model.eval()

    tokenized_input = tokenizer(input_text, truncation=True, padding=True, max_length=256, return_tensors='pt')
    input_ids = tokenized_input['input_ids'].to(device)
    attention_mask = tokenized_input['attention_mask'].to(device)

    # Make a prediction
    with torch.no_grad():
        outputs = model(input_ids, attention_mask=attention_mask)
    logits = outputs.logits
    Probs = F.softmax(logits, dim=1).cpu().numpy()[0]

    return Probs