# coding:utf8
import torch
from torch import nn, optim   # nn 神經網絡模塊 optim優化函數模塊
from torch.utils.data import DataLoader
from torch.autograd import Variable
from torchvision import transforms, datasets
from visdom import Visdom  # 可視化處理模塊
import time
import numpy as np
# 可視化app
viz = Visdom()
# 超參數
BATCH_SIZE = 40
LR = 1e-3
EPOCH = 2
# 判斷是否使用gpu
USE_GPU = True
if USE_GPU:
    gpu_status = torch.cuda.is_available()
else:
    gpu_status = False
transform=transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.1307,), (0.3081,))])
# 數據引入
train_dataset = datasets.MNIST('../data', True, transform, download=False)
test_dataset = datasets.MNIST('../data', False, transform)
train_loader = DataLoader(train_dataset, BATCH_SIZE, True)
# 為加快測試，把測試數據從10000縮小到2000
test_data = torch.unsqueeze(test_dataset.test_data, 1)[:1500]
test_label = test_dataset.test_labels[:1500]
# visdom可視化部分數據
viz.images(test_data[:100], nrow=10)
#viz.images(test_data[:100], nrow=10)
# 為防止可視化視窗重疊現象，停頓0.5秒
time.sleep(0.5)
if gpu_status:
    test_data = test_data.cuda()
test_data = Variable(test_data, volatile=True).float()
# 創建線圖可視化窗口
line = viz.line(np.arange(10))
# 創建cnn神經網絡
class CNN(nn.Module):
    def __init__(self, in_dim, n_class):
        super(CNN, self).__init__()
        self.conv = nn.Sequential(
            # channel 為信息高度 padding為圖片留白 kernel_size 掃描模塊size（5x5）
            nn.Conv2d(in_channels=in_dim, out_channels=16,kernel_size=5,stride=1, padding=2),
            nn.ReLU(),
            # 平面縮減 28x28 >> 14*14
            nn.MaxPool2d(kernel_size=2),
            nn.Conv2d(16, 32, 3, 1, 1),
            nn.ReLU(),
            # 14x14 >> 7x7
            nn.MaxPool2d(2)
        )
        self.fc = nn.Sequential(
            nn.Linear(32*7*7, 120),
            nn.Linear(120, n_class)
        )
    def forward(self, x):
        out = self.conv(x)
        out = out.view(out.size(0), -1)
        out = self.fc(out)
        return out
net = CNN(1,10)
if gpu_status :
    net = net.cuda()
    #print("#"*26, "使用gpu", "#"*26)
else:
    #print("#" * 26, "使用cpu", "#" * 26)
    pass
# loss、optimizer 函數設置
loss_f = nn.CrossEntropyLoss()
optimizer = optim.Adam(net.parameters(), lr=LR)
# 起始時間設置
start_time = time.time()
# 可視化所需數據點
time_p, tr_acc, ts_acc, loss_p = [], [], [], []
# 創建可視化數據視窗
text = viz.text("h1>convolution Nueral Network/h1>")
for epoch in range(EPOCH):
    # 由于分批次學習，輸出loss為一批平均，需要累積or平均每個batch的loss，acc
    sum_loss, sum_acc, sum_step = 0., 0., 0.
    for i, (tx, ty) in enumerate(train_loader, 1):
        if gpu_status:
            tx, ty = tx.cuda(), ty.cuda()
        tx = Variable(tx)
        ty = Variable(ty)
        out = net(tx)
        loss = loss_f(out, ty)
        #print(tx.size())
        #print(ty.size())
        #print(out.size())
        sum_loss += loss.item()*len(ty)
        #print(sum_loss)
        pred_tr = torch.max(out,1)[1]
        sum_acc += sum(pred_tr==ty).item()
        sum_step += ty.size(0)
        # 學習反饋
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        # 每40個batch可視化一下數據
        if i % 40 == 0:
            if gpu_status:
                test_data = test_data.cuda()
            test_out = net(test_data)
            print(test_out.size())
            # 如果用gpu運行out數據為cuda格式需要.cpu（）轉化為cpu數據 在進行比較
            pred_ts = torch.max(test_out, 1)[1].cpu().data.squeeze()
            print(pred_ts.size())
            rightnum = pred_ts.eq(test_label.view_as(pred_ts)).sum().item()
            #rightnum =sum(pred_tr==ty).item()
            #  sum_acc += sum(pred_tr==ty).item()
            acc =  rightnum/float(test_label.size(0))
            print("epoch: [{}/{}] | Loss: {:.4f} | TR_acc: {:.4f} | TS_acc: {:.4f} | Time: {:.1f}".format(epoch+1, EPOCH,
                                    sum_loss/(sum_step), sum_acc/(sum_step), acc, time.time()-start_time))
            # 可視化部分
            time_p.append(time.time()-start_time)
            tr_acc.append(sum_acc/sum_step)
            ts_acc.append(acc)
            loss_p.append(sum_loss/sum_step)
            viz.line(X=np.column_stack((np.array(time_p), np.array(time_p), np.array(time_p))),
                     Y=np.column_stack((np.array(loss_p), np.array(tr_acc), np.array(ts_acc))),
                     win=line,
                     opts=dict(legend=["Loss", "TRAIN_acc", "TEST_acc"]))
            # visdom text 支持html語句
            viz.text("p style='color:red'>epoch:{}/p>br>p style='color:blue'>Loss:{:.4f}/p>br>"
                     "p style='color:BlueViolet'>TRAIN_acc:{:.4f}/p>br>p style='color:orange'>TEST_acc:{:.4f}/p>br>"
                     "p style='color:green'>Time:{:.2f}/p>".format(epoch, sum_loss/sum_step, sum_acc/sum_step, acc,
                                                                       time.time()-start_time),
                     win=text)
            sum_loss, sum_acc, sum_step = 0., 0., 0.