instrumented.py

from __future__ import print_function
import time
import math
import argparse
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import numpy as np
from torchvision import datasets, transforms
#import matplotlib.pyplot as plt
#from large_margin_softmax_linear import LargeMarginSoftmaxLinear
#from experiment import LargeMarginSoftmaxLinear

torch.set_printoptions(precision=20)

tableau20 = [(31, 119, 180), (174, 199, 232), (255, 127, 14), (255, 187, 120),    
             (44, 160, 44), (152, 223, 138), (214, 39, 40), (255, 152, 150),    
             (148, 103, 189), (197, 176, 213), (140, 86, 75), (196, 156, 148),    
             (227, 119, 194), (247, 182, 210), (127, 127, 127), (199, 199, 199),    
             (188, 189, 34), (219, 219, 141), (23, 190, 207), (158, 218, 229)] 
for i in range(len(tableau20)):    
    r, g, b = tableau20[i]    
    tableau20[i] = (r / 255., g / 255., b / 255.)

# Training settings
parser = argparse.ArgumentParser(description='PyTorch MNIST Example')
parser.add_argument('--batch-size', type=int, default=256, metavar='N',
                    help='input batch size for training (default: 256)')
parser.add_argument('--test-batch-size', type=int, default=1000, metavar='N',
                    help='input batch size for testing (default: 1000)')
parser.add_argument('--epochs', type=int, default=30, metavar='N',
                    help='number of epochs to train (default: 30)')
parser.add_argument('--lr', type=float, default=0.01, metavar='LR',
                    help='learning rate (default: 0.01)')
parser.add_argument('--momentum', type=float, default=0.9, metavar='M',
                    help='SGD momentum (default: 0.9)')
parser.add_argument('--weight_decay', type=float, default=0.0005, metavar='W',
                    help='SGD weight decay (default: 0.0005)')
parser.add_argument('--no-cuda', action='store_true', default=False,
                    help='disables CUDA training')
parser.add_argument('--seed', type=int, default=1, metavar='S',
                    help='random seed (default: 1)')
parser.add_argument('--log-interval', type=int, default=50, metavar='N',
                    help='how many batches to wait before logging training status')
parser.add_argument('--vis_path', type=str, default="visualizations/color6", metavar='S',
                    help='path to save your visualization figures')
args = parser.parse_args()
#use_cuda = False
use_cuda = not args.no_cuda and torch.cuda.is_available()
print(use_cuda)

torch.manual_seed(args.seed)

device = torch.device("cuda" if use_cuda else "cpu")


torch.backends.cudnn.benchmark = True if use_cuda else False

print(device)
kwargs = {'num_workers': 6, 'pin_memory': True} if use_cuda else {}
train_loader = torch.utils.data.DataLoader(
    datasets.MNIST('../data', train=True, download=True,
                   transform=transforms.Compose([
                       transforms.ToTensor(),
                       transforms.Normalize((0.1307,), (0.3081,))
                   ])),
    batch_size=args.batch_size, shuffle=True, **kwargs)
test_loader = torch.utils.data.DataLoader(
    datasets.MNIST('../data', train=False, transform=transforms.Compose([
                       transforms.ToTensor(),
                       transforms.Normalize((0.1307,), (0.3081,))
                   ])),
    batch_size=args.test_batch_size, shuffle=True, **kwargs)


def weights_init(m):
    if isinstance(m, nn.Conv2d):
        n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
        m.weight.data.normal_(0, math.sqrt(2. / n))
        #nn.init.xavier_uniform(m.weight.batch_data)
    #elif isinstance(m, nn.BatchNorm2d):
    #    m.weight.batch_data.fill_(1)
    #    m.bias.batch_data.zero_()    
    #elif isinstance(m, nn.BatchNorm1d):
    #    m.weight.batch_data.fill_(1)
    #    m.bias.batch_data.zero_()    

 
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(1, 32, kernel_size=5, padding=2)
        #self.bn1 = nn.BatchNorm2d(32)
        self.prelu1 = nn.PReLU()
        self.conv2 = nn.Conv2d(32, 32, kernel_size=5, padding=2)
        #self.bn2 = nn.BatchNorm2d(32)
        self.prelu2 = nn.PReLU()
        self.conv3 = nn.Conv2d(32, 64, kernel_size=5, padding=2)
        #self.bn3 = nn.BatchNorm2d(64)
        self.prelu3 = nn.PReLU()
        self.conv4 = nn.Conv2d(64, 64, kernel_size=5, padding=2)
        #self.bn4 = nn.BatchNorm2d(64)
        self.prelu4 = nn.PReLU()
        self.conv5 = nn.Conv2d(64, 128, kernel_size=5, padding=2)
        #self.bn5 = nn.BatchNorm2d(128)
        self.prelu5 = nn.PReLU()
        self.conv6 = nn.Conv2d(128, 128, kernel_size=5, padding=2)
        #self.bn6 = nn.BatchNorm2d(128)
        self.prelu6 = nn.PReLU()
        self.fc1 = nn.Linear(1152, 2)
        #self.bn7 = nn.BatchNorm1d(2)
        self.prelu7 = nn.PReLU()
        self.fc2 = nn.Linear(2, 10)
        self.loss = nn.CrossEntropyLoss()
        #self.fc2 = LargeMarginSoftmaxLinear(2, 10, 2, 0)
        #self.fc2 = LargeMarginSoftmaxLinear(2, 10, 2, 0, use_cuda, device)

    def forward(self, x, y):
        x = self.prelu1(self.conv1(x))
        x = F.max_pool2d(self.prelu2(self.conv2(x)), 2, stride=2)
        x = self.prelu3(self.conv3(x))
        x = F.max_pool2d(self.prelu4(self.conv4(x)), 2, stride=2)
        x = self.prelu5(self.conv5(x))
        x = F.max_pool2d(self.prelu6(self.conv6(x)), 2, stride=2)

        #x = F.relu(self.bn1(self.conv1(x)))
        #x = F.max_pool2d(F.relu(self.bn2(self.conv2(x))), 2, stride=2)
        #x = F.relu(self.bn3(self.conv3(x)))
        #x = F.max_pool2d(F.relu(self.bn4(self.conv4(x))), 2, stride=2)
        #x = F.relu(self.bn5(self.conv5(x)))
        #x = F.max_pool2d(F.relu(self.bn6(self.conv6(x))), 2, stride=2)
        x = x.view(-1, 1152)
        #print(x.shape)
        features = self.fc1(x)
        x = self.prelu7(features)
        #x = F.relu(self.bn7(self.fc1(x)))
        #x = self.fc2(x, y)
        x = self.fc2(x)
        #print(x)
        return self.loss(x, y)

model = Net().to(device)
model.apply(weights_init)

optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay)

def visualize(features, labels, epoch, vis_path):
    plt.clf()
    for i in range(10):
        plt.scatter(features[(labels == i), 0], features[(labels == i), 1], s=3, c=tableau20[i], alpha=0.8) 
    plt.savefig(vis_path + "/epoch_" + str(epoch) + ".jpg")
    return

def train(epoch):
    model.train()
    #features = []
    #predictions = []        
    #labels = []
    for batch_idx, (batch_data, batch_labels) in enumerate(train_loader):

        batch_data, batch_labels = batch_data.to(device), batch_labels.to(device)
        #print(batch_data.dtype)
        #print(batch_labels.dtype)
        #print(batch_data.size())
        #print(batch_labels.size())
        optimizer.zero_grad()
        enabled = batch_idx == 10
        with torch.autograd.profiler.profile(enabled=enabled, use_cuda=True) as prof:
            loss = model(batch_data, batch_labels)
            #loss = F.nll_loss(batch_scores, batch_labels)
            #print(loss)
            
            loss.backward()
            optimizer.step()
        if enabled:
            print(prof.key_averages().table(sort_by='cuda_time_total'))
            import pdb
            pdb.set_trace()
        if batch_idx % args.log_interval == 0:
            print('Train Epoch: {} [{}/{} ({:.2f}%)]\tLoss: {:.6f}'.format(
                epoch, batch_idx * len(batch_data), len(train_loader.dataset),
                100. * batch_idx / len(train_loader), loss.item()))
        #features.append(batch_features)
        #batch_predictions = batch_scores.max(1, keepdim=True)[1]
        #predictions.append(batch_predictions)
        #labels.append(batch_labels)
    
    #features = torch.cat(features, 0).data.to('cpu').numpy()
    #predictions = torch.cat(predictions, 0).data.cpu().numpy()
    #labels = torch.cat(labels, 0).data.to('cpu').numpy()
    #visualize(features, labels, epoch, args.vis_path)

def test():
    model.eval()
    test_loss = 0
    correct = 0
    with torch.no_grad():
        for batch_data, batch_labels in test_loader:
            batch_data, batch_labels = batch_data.to(device), batch_labels.to(device)
            batch_scores, batch_features = model(batch_data)
            test_loss += F.nll_loss(batch_scores, batch_labels, size_average=False).item() # sum up batch loss
            pred = batch_scores.max(1, keepdim=True)[1] # get the index of the max log-probability
            correct += pred.eq(batch_labels.view_as(pred)).sum().item()

    test_loss /= len(test_loader.dataset)
    print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.2f}%)\n'.format(
        test_loss, correct, len(test_loader.dataset),
        100. * correct / len(test_loader.dataset)))

scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=10, gamma=0.1) 
for epoch in range(1, args.epochs + 1):
    scheduler.step()
    #lr = []
    #for param_group in optimizer.param_groups:
    #    lr += [param_group['lr']]
    #print(lr)
    torch.cuda.synchronize()

    start = time.perf_counter()
    train(epoch)
    torch.cuda.synchronize()
    end = time.perf_counter()
    #print(end-start)
    print('Total time taken: {:.2f}s\n'.format(end-start))
    #test()