# 13_feedforward with tensorboard
import torch
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms
import matplotlib.pyplot as plt

import sys
from torch.utils.tensorboard import SummaryWriter
writer = SummaryWriter('runs/mnist')

# device config
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(f'Device is {device}')

# hyper parameters
input_size = 784 # 28x28 pixel images
hidden_size = 100 # PLAY WITH THIS
num_classes = 10 # digits 0..9
num_epochs = 2 # PLAY WITH THIS
batch_size = 100
learning_rate = .001

# MNIST
train_dataset = torchvision.datasets.MNIST(root='./data', train=True,
    download=True, transform=transforms.ToTensor())

test_dataset = torchvision.datasets.MNIST(root='./data', train=False,
    transform=transforms.ToTensor())

train_loader = torch.utils.data.DataLoader(dataset=train_dataset, batch_size=batch_size, 
    shuffle=True)

test_loader = torch.utils.data.DataLoader(dataset=test_dataset, batch_size=batch_size)

examples = iter(train_loader)
samples, labels = examples.next()
print(samples.shape, labels.shape)

for i in range(6):
    plt.subplot(2, 3, i+1)
    plt.imshow(samples[i][0], cmap='gray')
# plt.show()
img_grid = torchvision.utils.make_grid(samples)
writer.add_image('mnist_images', img_grid)
writer.close()
# sys.exit()

model = nn.Sequential(
            nn.Linear(input_size, hidden_size),
            nn.ReLU(),
            nn.Linear(hidden_size, num_classes)
            # no softmax because its included in the CE loss function
        ).to(device)
# print(model)

# loss and optimizer
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)

writer.add_graph(model.to('cpu'), samples.to('cpu').reshape(-1, 28*28))
writer.close()
# sys.exit()

# training loop
num_total_steps = len(train_loader)

running_loss = 0.0
running_correct = 0
for epoch in range(num_epochs):
    for batch, (images, labels) in enumerate(train_loader):
        # reshape 100, 1, 28, 28 -> 100, 784
        images = images.reshape(-1, 28*28).to(device) # reshape and send to gpu if available
        labels = labels.to(device)
        model = model.to(device)

        # forward
        outputs = model(images)
        loss = criterion(outputs, labels)

        # backward + update
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

        running_loss += loss.item()
        _, predictions = torch.max(outputs, 1)
        running_correct += (predictions == labels).sum().item()
        writer.add_scalar('training loss each', loss.item(), epoch * num_total_steps + batch)
        writer.add_scalar('accuracy each', (predictions == labels).sum().item(), epoch * num_total_steps + batch)


        if (batch+1) % 100 == 0:
            writer.add_scalar('training loss', running_loss/100, epoch * num_total_steps + batch)
            writer.add_scalar('accuracy', running_correct/100, epoch * num_total_steps + batch)
            print(f'Epoch {epoch+1}/{num_epochs}, step {batch+1}/{num_total_steps}, loss = {loss.item():.4f}')
            running_loss = 0.0
            running_correct = 0

# test

b_labels = []
b_preds = []
with torch.no_grad():
    n_correct = 0
    n_samples = 0
    for images, labels in test_loader:
        images = images.reshape(-1, 28*28).to(device)
        labels = labels.to(device)
        outputs = model(images)

        # value, index (index is class label)
        _, predictions = torch.max(outputs.data, 1)
        n_samples += labels.shape[0]
        n_correct += (predictions==labels).sum().item()

        sm = nn.Softmax(dim=0)
        class_predictions = [sm(output) for output in outputs]
        b_preds.append(class_predictions)
        b_labels.append(predictions)

    b_preds = torch.cat([torch.stack(batch) for batch in b_preds])
    b_labels = torch.cat(b_labels)

    acc = 100.*n_correct/n_samples
    print(f'Accuracy = {acc}%')

    classes = range(10)
    for i in classes:
        labels_i = b_labels == i
        preds_i = b_preds[:,i]
        writer.add_pr_curve(str(i), labels_i, preds_i, global_step=0)

writer.close()