import os
import json
import torch
from torch import nn, optim
from torch.nn import functional as F
from tqdm import tqdm


class ModelWithTemperature(nn.Module):
    """
    A thin decorator, which wraps a model with temperature scaling
    model (nn.Module):
        A classification neural network
        NB: Output of the neural network should be the classification logits,
            NOT the softmax (or log softmax)!
    """
    def __init__(self, model, n_class, opt_bias=False):
        super(ModelWithTemperature, self).__init__()
        self.model = model
        self.temperature = nn.Parameter(torch.ones(1))
        self.bias = nn.Parameter(torch.zeros(1, n_class))
        self.opt_bias = opt_bias
        
    def forward(self, input):
        logits = self.model(input)
        return self.temperature_scale(logits)

    def temperature_scale(self, logits):
        """
        Perform temperature scaling on logits
        """
        return logits / self.temperature + self.bias
    
    def set_temperature(self, temp, bias):
        self.temperature = nn.Parameter(torch.ones(1) * temp)
        self.bias = nn.Parameter(torch.as_tensor(bias))
        self.cuda()

    # This function probably should live outside of this class, but whatever
    def find_temperature(self, valid_loader):
        """
        Tune the tempearature of the model (using the validation set).
        We're going to set it to optimize NLL.
        valid_loader (DataLoader): validation set loader
        """
        self.cuda()
        nll_criterion = nn.CrossEntropyLoss().cuda()
        ece_criterion = _ECELoss().cuda()

        # First: collect all the logits and labels for the validation set
        logits_list = []
        labels_list = []
        with torch.no_grad():
            for items in tqdm(valid_loader):
                input, label = items[0], items[1]
                input = input.cuda()
                logits = self.model(input)
                logits_list.append(logits)
                labels_list.append(label)
            logits = torch.cat(logits_list).cuda()
            labels = torch.cat(labels_list).cuda()

        # Calculate NLL and ECE before temperature scaling
        before_temperature_nll = nll_criterion(logits, labels).item()
        before_temperature_ece = ece_criterion(logits, labels).item()
        print('Before temperature - NLL: %.3f, ECE: %.3f' % (before_temperature_nll, before_temperature_ece))

        # Next: optimize the temperature w.r.t. NLL
        if self.opt_bias:
            params = [self.temperature, self.bias]
        else:
            params = [self.temperature]
        
        optimizer = optim.LBFGS(params, lr=0.01, max_iter=500000)

        def eval():
            optimizer.zero_grad()
            loss = nll_criterion(self.temperature_scale(logits), labels)
            loss.backward()
            return loss
        optimizer.step(eval)

        # Calculate NLL and ECE after temperature scaling
        after_temperature_nll = nll_criterion(self.temperature_scale(logits), labels).item()
        after_temperature_ece = ece_criterion(self.temperature_scale(logits), labels).item()
        print('Optimal temperature: %.3f' % self.temperature.item())
        print('After temperature - NLL: %.3f, ECE: %.3f' % (after_temperature_nll, after_temperature_ece))
        
        conf = round( torch.nn.functional.softmax(self.temperature_scale(logits), dim=1).amax(1).mean().item(), 3 )
        acc = round( (torch.argmax(logits, dim=1) == labels).float().mean().item(), 3 )

        print('TS Conf:', conf)
        print('TS Acc:', acc)

        return self


class _ECELoss(nn.Module):
    """
    Calculates the Expected Calibration Error of a model.
    (This isn't necessary for temperature scaling, just a cool metric).
    The input to this loss is the logits of a model, NOT the softmax scores.
    This divides the confidence outputs into equally-sized interval bins.
    In each bin, we compute the confidence gap:
    bin_gap = | avg_confidence_in_bin - accuracy_in_bin |
    We then return a weighted average of the gaps, based on the number
    of samples in each bin
    See: Naeini, Mahdi Pakdaman, Gregory F. Cooper, and Milos Hauskrecht.
    "Obtaining Well Calibrated Probabilities Using Bayesian Binning." AAAI.
    2015.
    """
    def __init__(self, n_bins=15):
        """
        n_bins (int): number of confidence interval bins
        """
        super(_ECELoss, self).__init__()
        bin_boundaries = torch.linspace(0, 1, n_bins + 1)
        self.bin_lowers = bin_boundaries[:-1]
        self.bin_uppers = bin_boundaries[1:]

    def forward(self, logits, labels):
        softmaxes = F.softmax(logits, dim=1)
        confidences, predictions = torch.max(softmaxes, 1)
        accuracies = predictions.eq(labels)

        ece = torch.zeros(1, device=logits.device)
        for bin_lower, bin_upper in zip(self.bin_lowers, self.bin_uppers):
            # Calculated |confidence - accuracy| in each bin
            in_bin = confidences.gt(bin_lower.item()) * confidences.le(bin_upper.item())
            prop_in_bin = in_bin.float().mean()
            if prop_in_bin.item() > 0:
                accuracy_in_bin = accuracies[in_bin].float().mean()
                avg_confidence_in_bin = confidences[in_bin].mean()
                ece += torch.abs(avg_confidence_in_bin - accuracy_in_bin) * prop_in_bin

        return ece
    

def calibrate(model, n_class, opt_bias, val_loader, temp_dir):
    if os.path.exists(temp_dir):
        with open(temp_dir, 'r') as f:
            temp = json.load(f)

        model = ModelWithTemperature(model, n_class, opt_bias)
        model.set_temperature(temp['t'], temp['bias'])
    else:
        model = ModelWithTemperature(model, n_class, opt_bias)
        model.find_temperature(val_loader)
        temp = model.temperature.item()
        bias = model.bias.tolist()

        with open(temp_dir, 'w') as f:
            json.dump({'t': temp, 'bias': bias}, f)

    return model