import h5py
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from h5py import Dataset
from sklearn.preprocessing import MinMaxScaler
from torch.utils.data import DataLoader
from torch.optim import SGD, Adam
from torch.nn import Linear, Sigmoid, Module, MSELoss, BCELoss
from tqdm import tqdm

# f hat 100k keys, jeder key hat eine hdf5 group als value
# jede group hat 2 keys. in key 'X' ist ein 2d numpy array
# welches die 200x200 pixel des bildes speichert (dset[70:71,:]
# zugriff auf zeile 71) und in key'Y' sind die drei
# korrespondierenden achsenwerte der kamera X, Y und Z


def prepare_data(path):
    # load h5 file
    sims = 3000  # only half the data
    x = []
    y = []
    with h5py.File(path, 'r') as f:
        for key in range(sims):
            X = np.array(f['/' + str(key) + '/X'])
            Y = np.array(f['/' + str(key) + '/Y'])
            X = MinMaxScaler().fit_transform(X)  # normalise the data between 0 and 1
            # Y = MinMaxScaler().fit_transform(Y)
            # print(Y[:])
            if np.amax(X) > 0:  # ignore the blank images (no data)
                x.append(X)
                y.append(Y)
    device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
    print(device)
    x = torch.tensor(x, dtype=torch.float32, device=device)  # input
    y = torch.tensor(y, dtype=torch.float32, device=device)  # output
    print('x.shape', np.shape(x))
    print('y.shape', np.shape(y))
    print(x.element_size() * x.nelement())

    # split test and train data
    data_length = np.shape(x)[0]
    train_length = data_length - 1000
    train_x = x[0:train_length]
    test_x = x[train_length:data_length]
    train_y = y[0:train_length]
    test_y = y[train_length:data_length]
    train = Data(train_x, train_y)
    test = Data(test_x, test_y)
    return train, test

    # train, test = dataset.get_splits()
    # trainloader = torch.utils.data.DataLoader(x, batch_size=8, shuffle=True, num_workers=0)


def train_model(train_dl, model, lr, numEpochs, optimizerFlag, modelSaveFileName, criterionFlag, startEpoch):
    if optimizerFlag == 'A':
        optimizer = Adam(model.parameters(), lr=lr, betas=(0.9, 0.999), eps=1e-08, weight_decay=0, amsgrad=False)
    else:
        optimizer = SGD(model.parameters(), lr=lr, momentum=0.9)
    if criterionFlag == 'MSE':
        criterion = MSELoss()
    else:
        criterion = BCELoss()
    for epoch in range(startEpoch + 1, numEpochs + 1):
        # enumerate mini batches
        loop = tqdm(enumerate(train_dl), total=len(train_dl), leave=False)
        for i, (inputs, targets) in loop:
            # clear the gradients
            optimizer.zero_grad()
            # compute the model output
            yhat = model(inputs)
            # calculate loss
            loss = criterion(yhat, targets)
            # credit assignment
            loss.backward()
            # update model weights
            optimizer.step()
            # upgrade progress bar
            loop.set_description(f"Epoch [{epoch}/{numEpochs}]")
            # loop.set_postfix(loss=loss.item(),acc=torch.rand(1).item())
        if epoch % 10 == 0:
            torch.save({
                'epoch': epoch,
                'model_state_dict': model.state_dict(),
                'optimizer_state_dict': optimizer.state_dict(),
                'lr': lr,
            }, modelSaveFileName + str(epoch) + ".pt")


# TODO anpassen an unsere Situation
class Net(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv1 = nn.Conv2d(1, 8, kernel_size=5)  # 1d pictures, 8 pictures at a time, 5x5 kernel size
        self.pool = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(1, 8, kernel_size=5)
        self.fc1 = nn.Linear(200, 120)
        self.fc2 = nn.Linear(120, 84)
        self.fc3 = nn.Linear(84, 10)

    def forward(self, x):
        x = self.pool(F.relu(self.conv1(x)))
        x = self.pool(F.relu(self.conv2(x)))
        x = torch.flatten(x, 1)  # flatten all dimensions except batch
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x


class Data(Dataset):
    def __init__(self, x, y):
        self.x = x
        self.y = y
        self.len = len(x)

    def __getitem__(self, index):
        self.xc = self.x[index]
        self.yc = self.y[index]
        return self.xc, self.yc

    def __len__(self):
        return self.len


path = "training_data_alignment.h5"
batch_size = 8
num_Epochs = 100
modelSaveFileName = 'model.pt'
optimizerFlag = 'A'
criterionFlag = 'MSE' # MSE oder BCE
startEpoch = 1
learning_rate = 0.01

train, test = prepare_data(path)
train_dl = DataLoader(train, batch_size=batch_size, shuffle=True, num_workers=0, pin_memory=False)
test_dl = DataLoader(test, batch_size=batch_size, shuffle=True, num_workers=0, pin_memory=False)
net = Net()

train_model(train_dl, net, learning_rate, num_Epochs, optimizerFlag, modelSaveFileName, criterionFlag, startEpoch)

# print(net)