main_cifar100.py

'''Train CIFAR10 with PyTorch.'''
from __future__ import print_function
import os
os.environ['CUDA_LAUNCH_BLOCKING'] = "1"
import time
from time import sleep
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
import torch.backends.cudnn as cudnn
import copy
import torchvision
import torchvision.transforms as transforms
import math
import os
import argparse
# from torchviz import make_dot
# from IPython import embed
import sys
cwd = os.getcwd()
model_dir = os.path.join(cwd, 'models')
sys.path.append(model_dir)
from models import *
from utils import progress_bar
from tensorboardX import SummaryWriter
from random import randint
from PCANorm import *
from ZCANorm import *
import numpy as np


def isnan(x):
    return x != x

# torch.backends.cudnn.deterministic = True
# torch.manual_seed(999)
# torch.cuda.manual_seed_all(999)

parser = argparse.ArgumentParser(description='PyTorch CIFAR100 Training')
parser.add_argument('--norm', default='batchnorm', type=str, help='norm layer type')
parser.add_argument('--batch_size', default=128, type=int, help='batch size')
parser.add_argument('--lr', default=0.1, type=float, help='learning rate')
parser.add_argument('--resume', '-r', action='store_true', help='resume from checkpoint')
args = parser.parse_args()

device = 'cuda' if torch.cuda.is_available() else 'cpu'
best_acc = 0  # best test accuracy
start_epoch = 0  # start from epoch 0 or last checkpoint epoch

# Data
print('==> Preparing data..')
transform_train = transforms.Compose([
    transforms.RandomCrop(32, padding=4),
    transforms.RandomHorizontalFlip(),
    transforms.ToTensor(),
    transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])

transform_test = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])

BatchSize = args.batch_size

trainset = torchvision.datasets.CIFAR100(root='./data', train=True, download=True, transform=transform_train)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=BatchSize, shuffle=True, num_workers=2)

testset = torchvision.datasets.CIFAR100(root='./data', train=False, download=True, transform=transform_test)
testloader = torch.utils.data.DataLoader(testset, batch_size=BatchSize, shuffle=False, num_workers=2)

norm = args.norm
print('==> Building model using {}..'.format(norm))
if norm == 'batchnorm':
    Norm = nn.BatchNorm2d
elif norm == 'zcanormsvdunstable':
    Norm = ZCANormSVDunstable
elif norm == 'zcanormpiunstable':
    Norm = ZCANormPIunstable
elif norm == 'zcanormsvdpi':
    Norm = ZCANormSVDPI
elif norm == 'pcanormsvdpi':
    Norm = myPCANormSVDPI


net = resnet18(Norm=Norm, num_classes=10)  # resnet18(Norm=Norm)

save_dir = 'runs/cifar100'  # 'runs'
model_name = net._get_name()
id = randint(0, 1000)
logdir = os.path.join(save_dir, model_name+'18'+'group4', '{}-bs{}'.format(norm, BatchSize), str(id))
if not os.path.isdir(logdir):
    os.makedirs(logdir)

writer = SummaryWriter(log_dir=logdir)
print('RUNDIR: {}'.format(logdir))

net = net.to(device)
if device == 'cuda':
    net = torch.nn.DataParallel(net)
    cudnn.benchmark = True

criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=args.lr, momentum=0.9, weight_decay=5e-4)
# optimizer = optim.Adam(net.parameters(), lr=args.lr, weight_decay=5e-4)

scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=100, gamma=0.1)

if args.resume:
    # Load checkpoint.
    print('==> Resuming from checkpoint..')
    assert os.path.isdir(logdir), 'Error: no checkpoint directory found!'
    checkpoint = torch.load('{}/best-{}-ckpt.t7'.format(logdir, model_name))
    net.load_state_dict(checkpoint['net'])
    best_acc = checkpoint['acc']
    start_epoch = checkpoint['epoch']

    scheduler.load_state_dict(checkpoint['scheduler'])
    optimizer.load_state_dict(checkpoint['optimizer'])


def train(epoch):
    print('\nEpoch: %d' % epoch)
    net.train()
    train_loss = 0
    correct = 0
    total = 0
    for batch_idx, (inputs, targets) in enumerate(trainloader):
        inputs, targets = inputs.to(device), targets.to(device)

        optimizer.zero_grad()
        outputs = net(inputs)
        loss = criterion(outputs, targets)

        if isnan(loss):
            print('nan found')
            print('paras from previous update:')
            for n, p in net.named_parameters():
                if p.requires_grad and ("layer1" not in n) and ("layer2" not in n) and ("layer3" not in n) and (
                        "layer4" not in n):
                    print('param/{}mean'.format(n), p.abs().mean().item(), epoch * len(trainloader) + batch_idx + 1)
                    print('param/{}max'.format(n), p.abs().max().item(), epoch * len(trainloader) + batch_idx + 1)
            sys.exit("Error message")

        loss.backward()

        optimizer.step()
        if batch_idx % 500 == 0:
            writer.add_scalar('loss/train_loss', loss.item(), epoch*len(trainloader)+batch_idx+1)
        train_loss += loss.item()
        _, predicted = outputs.max(1)
        total += targets.size(0)
        correct += predicted.eq(targets).sum().item()

        progress_bar(batch_idx, len(trainloader), 'Loss: %.3f | Acc: %.3f%% (%d/%d)'
                     % (train_loss/(batch_idx+1), 100.*correct/total, correct, total))

    acc = 100. * correct / total
    writer.add_scalar('loss/train_loss_avg', train_loss / len(trainloader), epoch)
    writer.add_scalar('train/accuracy', acc, epoch)
    writer.add_scalar('train/error', 100 - acc, epoch)


def test(epoch):
    global best_acc
    net.eval()
    test_loss = 0
    correct = 0
    total = 0
    with torch.no_grad():
        for batch_idx, (inputs, targets) in enumerate(testloader):
            inputs, targets = inputs.to(device), targets.to(device)
            outputs = net(inputs)
            loss = criterion(outputs, targets)

            test_loss += loss.item()
            _, predicted = outputs.max(1)
            total += targets.size(0)
            correct += predicted.eq(targets).sum().item()

            progress_bar(batch_idx, len(testloader), 'Loss: %.3f | Acc: %.3f%% (%d/%d)'
                % (test_loss/(batch_idx+1), 100.*correct/total, correct, total))

    # Save checkpoint.
    acc = 100.*correct/total
    writer.add_scalar('loss/test_loss_avg', test_loss / len(testloader), epoch)
    writer.add_scalar('test/accuracy', acc, epoch)
    writer.add_scalar('test/error', 100-acc, epoch)
    if acc > best_acc:
        print('Saving..')
        state = {
            'net': net.state_dict(),
            'acc': acc,
            'epoch': epoch,
            'optimizer': optimizer.state_dict(),
            'scheduler': scheduler.state_dict(),
        }
        # if not os.path.isdir('checkpoint'):
        #     os.mkdir('checkpoint')
        torch.save(state, os.path.join(logdir, 'best-{}-ckpt.t7'.format(model_name)))
        best_acc = acc


for epoch in range(start_epoch, int(100*3.5)):
    train(epoch)
    scheduler.step()
    test(epoch)