mnist_train.py

import os
import argparse
import numpy as np
from tqdm import tqdm
import pandas as pd
import joblib
from collections import OrderedDict

import torch
import torch.backends.cudnn as cudnn
import torch.nn as nn
import torch.optim as optim
from torch.optim import lr_scheduler
import torchvision.transforms as transforms
import torchvision.datasets as datasets

from utils import *
from mnist import archs
import metrics


def parse_args():
    parser = argparse.ArgumentParser()

    parser.add_argument('--name', default=None,
                        help='model name: (default: arch+timestamp)')
    parser.add_argument('--arch', default='MNISTNet',
                        choices=archs.__all__,
                        help='model architecture')
    parser.add_argument('--metric', default='adacos',
                        choices=['adacos', 'arcface', 'sphereface', 'cosface', 'softmax'])
    parser.add_argument('--num-features', default=512, type=int,
                        help='dimention of embedded features')
    parser.add_argument('-b', '--batch-size', default=128, type=int,
                        metavar='N', help='mini-batch size (default: 128)')
    parser.add_argument('--epochs', default=100, type=int)
    parser.add_argument('--lr', '--learning-rate', default=1e-2, type=float)
    parser.add_argument('--min-lr', default=1e-3, type=float)
    parser.add_argument('--momentum', default=0.5, type=float)
    parser.add_argument('--weight-decay', default=1e-4, type=float)
    parser.add_argument('--nesterov', default=False, type=str2bool)
    parser.add_argument('--cpu', default=False, type=str2bool)

    args = parser.parse_args()

    return args


def train(args, train_loader, model, metric_fc, criterion, optimizer):
    losses = AverageMeter()
    acc1s = AverageMeter()

    model.train()
    metric_fc.train()

    for i, (input, target) in tqdm(enumerate(train_loader), total=len(train_loader)):
        if args.cpu:
            input = input.cpu()
            target = target.long().cpu()
        else:
            input = input.cuda()
            target = target.long().cuda() 

        feature = model(input)
        if args.metric == 'softmax':
            output = metric_fc(feature)
        else:
            output = metric_fc(feature, target)
        loss = criterion(output, target)

        acc1, = accuracy(output, target, topk=(1,))

        losses.update(loss.item(), input.size(0))
        acc1s.update(acc1.item(), input.size(0))

        # compute gradient and do optimizing step
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

    log = OrderedDict([
        ('loss', losses.avg),
        ('acc1', acc1s.avg),
    ])

    return log


def validate(args, val_loader, model, metric_fc, criterion):
    losses = AverageMeter()
    acc1s = AverageMeter()

    # switch to evaluate mode
    model.eval()
    metric_fc.eval()

    with torch.no_grad():
        for i, (input, target) in tqdm(enumerate(val_loader), total=len(val_loader)):
            if args.cpu:
                input = input.cpu()
                target = target.long().cpu()
            else:
                input = input.cuda()
                target = target.long().cuda()

            feature = model(input)
            if args.metric == 'softmax':
                output = metric_fc(feature)
            else:
                output = metric_fc(feature, target)
            loss = criterion(output, target)

            acc1, = accuracy(output, target, topk=(1,))

            losses.update(loss.item(), input.size(0))
            acc1s.update(acc1.item(), input.size(0))

    log = OrderedDict([
        ('loss', losses.avg),
        ('acc1', acc1s.avg),
    ])

    return log


def main():
    args = parse_args()

    if args.name is None:
        args.name = 'mnist_%s_%s_%dd' %(args.arch, args.metric, args.num_features)

    if not os.path.exists('models/%s' %args.name):
        os.makedirs('models/%s' %args.name)

    print('Config -----')
    for arg in vars(args):
        print('%s: %s' %(arg, getattr(args, arg)))
    print('------------')

    with open('models/%s/args.txt' %args.name, 'w') as f:
        for arg in vars(args):
            print('%s: %s' %(arg, getattr(args, arg)), file=f)

    joblib.dump(args, 'models/%s/args.pkl' %args.name)

    # criterion = nn.CrossEntropyLoss().cpu()

    if args.cpu:
        criterion = nn.CrossEntropyLoss().cpu()
    else:
        criterion = nn.CrossEntropyLoss().cuda()

    cudnn.benchmark = True

    transform_train = transforms.Compose([
        transforms.ToTensor(),
        transforms.Normalize((0.1307,), (0.3081,))
    ])
    transform_test = transforms.Compose([
        transforms.ToTensor(),
        transforms.Normalize((0.1307,), (0.3081,))
    ])

    train_set = datasets.MNIST(
        root='~/data',
        train=True,
        download=True,
        transform=transform_train)
    test_set = datasets.MNIST(
        root='~/data',
        train=False,
        download=True,
        transform=transform_train)

    train_loader = torch.utils.data.DataLoader(
        train_set,
        batch_size=args.batch_size,
        shuffle=True,
        num_workers=8)
    test_loader = torch.utils.data.DataLoader(
        test_set,
        batch_size=args.batch_size,
        shuffle=False,
        num_workers=8)

    # create model
    model = archs.__dict__[args.arch](args)

    if args.cpu:
        model = model.cpu()
    else:
        model = model.cuda()

    if args.metric == 'adacos':
        metric_fc = metrics.AdaCos(num_features=args.num_features, num_classes=10)
    elif args.metric == 'arcface':
        metric_fc = metrics.ArcFace(num_features=args.num_features, num_classes=10)
    elif args.metric == 'sphereface':
        metric_fc = metrics.SphereFace(num_features=args.num_features, num_classes=10)
    elif args.metric == 'cosface':
        metric_fc = metrics.CosFace(num_features=args.num_features, num_classes=10)
    else:
        metric_fc = nn.Linear(args.num_features, 10)
    if args.cpu:
        metric_fc = metric_fc.cpu()
    else:
        metric_fc = metric_fc.cuda()

    optimizer = optim.SGD(filter(lambda p: p.requires_grad, model.parameters()), lr=args.lr,
            momentum=args.momentum, weight_decay=args.weight_decay)

    scheduler = lr_scheduler.CosineAnnealingLR(optimizer,
            T_max=args.epochs, eta_min=args.min_lr)

    log = pd.DataFrame(index=[], columns=[
        'epoch', 'lr', 'loss', 'acc1', 'val_loss', 'val_acc1'
    ])

    best_loss = float('inf')
    for epoch in range(args.epochs):
        print('Epoch [%d/%d]' %(epoch+1, args.epochs))

        scheduler.step()

        # train for one epoch
        train_log = train(args, train_loader, model, metric_fc, criterion, optimizer)
        # evaluate on validation set
        val_log = validate(args, test_loader, model, metric_fc, criterion)

        print('loss %.4f - acc1 %.4f - val_loss %.4f - val_acc %.4f'
            %(train_log['loss'], train_log['acc1'], val_log['loss'], val_log['acc1']))

        tmp = pd.Series([
            epoch,
            scheduler.get_lr()[0],
            train_log['loss'],
            train_log['acc1'],
            val_log['loss'],
            val_log['acc1'],
        ], index=['epoch', 'lr', 'loss', 'acc1', 'val_loss', 'val_acc1'])

        log = log.append(tmp, ignore_index=True)
        log.to_csv('models/%s/log.csv' %args.name, index=False)

        if val_log['loss'] < best_loss:
            torch.save(model.state_dict(), 'models/%s/model.pth' %args.name)
            best_loss = val_log['loss']
            print("=> saved best model")


if __name__ == '__main__':
    main()