train_autodeeplab.py

import argparse
import os
import numpy as np
from tqdm import tqdm
import sys
import torch
from collections import OrderedDict
from mypath import Path
from dataloaders import make_data_loader
from modeling.sync_batchnorm.replicate import patch_replication_callback
from modeling.deeplab import *
from utils.loss import SegmentationLosses
from utils.calculate_weights import calculate_weigths_labels
from utils.lr_scheduler import LR_Scheduler
from utils.saver import Saver
from utils.summaries import TensorboardSummary
from utils.metrics import Evaluator
from auto_deeplab import AutoDeeplab
from architect import Architect
import apex
try:
    from apex import amp
    APEX_AVAILABLE = True
except ModuleNotFoundError:
    APEX_AVAILABLE = False


print('working with pytorch version {}'.format(torch.__version__))
print('with cuda version {}'.format(torch.version.cuda))
print('cudnn enabled: {}'.format(torch.backends.cudnn.enabled))
print('cudnn version: {}'.format(torch.backends.cudnn.version()))

torch.backends.cudnn.benchmark = True

class Trainer(object):
    def __init__(self, args):
        self.args = args

        # Define Saver
        self.saver = Saver(args)
        self.saver.save_experiment_config()
        # Define Tensorboard Summary
        self.summary = TensorboardSummary(self.saver.experiment_dir)
        self.writer = self.summary.create_summary()
        self.use_amp = True if APEX_AVAILABLE else False
        self.opt_level = args.opt_level

        kwargs = {'num_workers': args.workers, 'pin_memory': True, 'drop_last':True}
        self.train_loaderA, self.train_loaderB, self.val_loader, self.test_loader, self.nclass = make_data_loader(args, **kwargs)

        if args.use_balanced_weights:
            classes_weights_path = os.path.join(Path.db_root_dir(args.dataset), args.dataset+'_classes_weights.npy')
            if os.path.isfile(classes_weights_path):
                weight = np.load(classes_weights_path)
            else:
                #if so, which trainloader to use?
                weight = calculate_weigths_labels(args.dataset, self.train_loader, self.nclass)
            weight = torch.from_numpy(weight.astype(np.float32))
        else:
            weight = None
        self.criterion = SegmentationLosses(weight=weight, cuda=args.cuda).build_loss(mode=args.loss_type)

        # Define network
        model = AutoDeeplab (num_classes=self.nclass, num_layers=12, criterion=self.criterion, filter_multiplier=self.args.filter_multiplier,
                             block_multiplier=self.args.block_multiplier, step=self.args.step)
        optimizer = torch.optim.SGD(
                model.weight_parameters(),
                args.lr,
                momentum=args.momentum,
                weight_decay=args.weight_decay
            )

        self.model, self.optimizer = model, optimizer

        self.architect_optimizer = torch.optim.Adam(self.model.arch_parameters(),
                                                    lr=args.arch_lr, betas=(0.9, 0.999),
                                                    weight_decay=args.arch_weight_decay)

        # Define Evaluator
        self.evaluator = Evaluator(self.nclass)
        # Define lr scheduler
        self.scheduler = LR_Scheduler(args.lr_scheduler, args.lr,
                                      args.epochs, len(self.train_loaderA), min_lr=args.min_lr)
        # TODO: Figure out if len(self.train_loader) should be devided by two ? in other module as well
        # Using cuda
        if args.cuda:
            self.model = self.model.cuda()


        # mixed precision
        if self.use_amp and args.cuda:
            keep_batchnorm_fp32 = True if (self.opt_level == 'O2' or self.opt_level == 'O3') else None

            # fix for current pytorch version with opt_level 'O1'
            if self.opt_level == 'O1' and torch.__version__ < '1.3':
                for module in self.model.modules():
                    if isinstance(module, torch.nn.modules.batchnorm._BatchNorm):
                        # Hack to fix BN fprop without affine transformation
                        if module.weight is None:
                            module.weight = torch.nn.Parameter(
                                torch.ones(module.running_var.shape, dtype=module.running_var.dtype,
                                           device=module.running_var.device), requires_grad=False)
                        if module.bias is None:
                            module.bias = torch.nn.Parameter(
                                torch.zeros(module.running_var.shape, dtype=module.running_var.dtype,
                                            device=module.running_var.device), requires_grad=False)

            # print(keep_batchnorm_fp32)
            self.model, [self.optimizer, self.architect_optimizer] = amp.initialize(
                self.model, [self.optimizer, self.architect_optimizer], opt_level=self.opt_level,
                keep_batchnorm_fp32=keep_batchnorm_fp32, loss_scale="dynamic")

            print('cuda finished')


        # Using data parallel
        if args.cuda and len(self.args.gpu_ids) >1:
            if self.opt_level == 'O2' or self.opt_level == 'O3':
                print('currently cannot run with nn.DataParallel and optimization level', self.opt_level)
            self.model = torch.nn.DataParallel(self.model, device_ids=self.args.gpu_ids)
            patch_replication_callback(self.model)
            print('training on multiple-GPUs')

        #checkpoint = torch.load(args.resume)
        #print('about to load state_dict')
        #self.model.load_state_dict(checkpoint['state_dict'])
        #print('model loaded')
        #sys.exit()

        # Resuming checkpoint
        self.best_pred = 0.0
        if args.resume is not None:
            if not os.path.isfile(args.resume):
                raise RuntimeError("=> no checkpoint found at '{}'" .format(args.resume))
            checkpoint = torch.load(args.resume)
            args.start_epoch = checkpoint['epoch']

            # if the weights are wrapped in module object we have to clean it
            if args.clean_module:
                self.model.load_state_dict(checkpoint['state_dict'])
                state_dict = checkpoint['state_dict']
                new_state_dict = OrderedDict()
                for k, v in state_dict.items():
                    name = k[7:]  # remove 'module.' of dataparallel
                    new_state_dict[name] = v
                self.model.load_state_dict(new_state_dict)

            else:
                if (torch.cuda.device_count() > 1 or args.load_parallel):
                    self.model.module.load_state_dict(checkpoint['state_dict'])
                else:
                    self.model.load_state_dict(checkpoint['state_dict'])


            if not args.ft:
                self.optimizer.load_state_dict(checkpoint['optimizer'])
            self.best_pred = checkpoint['best_pred']
            print("=> loaded checkpoint '{}' (epoch {})"
                  .format(args.resume, checkpoint['epoch']))

        # Clear start epoch if fine-tuning
        if args.ft:
            args.start_epoch = 0

    def training(self, epoch):
        train_loss = 0.0
        self.model.train()
        tbar = tqdm(self.train_loaderA)
        num_img_tr = len(self.train_loaderA)
        for i, sample in enumerate(tbar):
            image, target = sample['image'], sample['label']
            if self.args.cuda:
                image, target = image.cuda(), target.cuda()
            self.scheduler(self.optimizer, i, epoch, self.best_pred)
            self.optimizer.zero_grad()
            output = self.model(image)
            loss = self.criterion(output, target)
            if self.use_amp:
                with amp.scale_loss(loss, self.optimizer) as scaled_loss:
                    scaled_loss.backward()
            else:
                loss.backward()
            self.optimizer.step()

            if epoch >= self.args.alpha_epoch:
                search = next(iter(self.train_loaderB))
                image_search, target_search = search['image'], search['label']
                if self.args.cuda:
                    image_search, target_search = image_search.cuda (), target_search.cuda ()

                self.architect_optimizer.zero_grad()
                output_search = self.model(image_search)
                arch_loss = self.criterion(output_search, target_search)
                if self.use_amp:
                    with amp.scale_loss(arch_loss, self.architect_optimizer) as arch_scaled_loss:
                        arch_scaled_loss.backward()
                else:
                    arch_loss.backward()
                self.architect_optimizer.step()

            train_loss += loss.item()
            tbar.set_description('Train loss: %.3f' % (train_loss / (i + 1)))
            #self.writer.add_scalar('train/total_loss_iter', loss.item(), i + num_img_tr * epoch)

            # Show 10 * 3 inference results each epoch
            if i % (num_img_tr // 10) == 0:
                global_step = i + num_img_tr * epoch
                self.summary.visualize_image(self.writer, self.args.dataset, image, target, output, global_step)

            #torch.cuda.empty_cache()
        self.writer.add_scalar('train/total_loss_epoch', train_loss, epoch)
        print('[Epoch: %d, numImages: %5d]' % (epoch, i * self.args.batch_size + image.data.shape[0]))
        print('Loss: %.3f' % train_loss)

        if self.args.no_val:
            # save checkpoint every epoch
            is_best = False
            if torch.cuda.device_count() > 1:
                state_dict = self.model.module.state_dict()
            else:
                state_dict = self.model.state_dict()
            self.saver.save_checkpoint({
                'epoch': epoch + 1,
                'state_dict': state_dict,
                'optimizer': self.optimizer.state_dict(),
                'best_pred': self.best_pred,
            }, is_best)


    def validation(self, epoch):
        self.model.eval()
        self.evaluator.reset()
        tbar = tqdm(self.val_loader, desc='\r')
        test_loss = 0.0

        for i, sample in enumerate(tbar):
            image, target = sample['image'], sample['label']
            if self.args.cuda:
                image, target = image.cuda(), target.cuda()
            with torch.no_grad():
                output = self.model(image)
            loss = self.criterion(output, target)
            test_loss += loss.item()
            tbar.set_description('Test loss: %.3f' % (test_loss / (i + 1)))
            pred = output.data.cpu().numpy()
            target = target.cpu().numpy()
            pred = np.argmax(pred, axis=1)
            # Add batch sample into evaluator
            self.evaluator.add_batch(target, pred)

        # Fast test during the training
        Acc = self.evaluator.Pixel_Accuracy()
        Acc_class = self.evaluator.Pixel_Accuracy_Class()
        mIoU = self.evaluator.Mean_Intersection_over_Union()
        FWIoU = self.evaluator.Frequency_Weighted_Intersection_over_Union()
        self.writer.add_scalar('val/total_loss_epoch', test_loss, epoch)
        self.writer.add_scalar('val/mIoU', mIoU, epoch)
        self.writer.add_scalar('val/Acc', Acc, epoch)
        self.writer.add_scalar('val/Acc_class', Acc_class, epoch)
        self.writer.add_scalar('val/fwIoU', FWIoU, epoch)
        print('Validation:')
        print('[Epoch: %d, numImages: %5d]' % (epoch, i * self.args.batch_size + image.data.shape[0]))
        print("Acc:{}, Acc_class:{}, mIoU:{}, fwIoU: {}".format(Acc, Acc_class, mIoU, FWIoU))
        print('Loss: %.3f' % test_loss)
        new_pred = mIoU
        if new_pred > self.best_pred:
            is_best = True
            self.best_pred = new_pred
            if torch.cuda.device_count() > 1:
                state_dict = self.model.module.state_dict()
            else:
                state_dict = self.model.state_dict()
            self.saver.save_checkpoint({
                'epoch': epoch + 1,
                'state_dict': state_dict,
                'optimizer': self.optimizer.state_dict(),
                'best_pred': self.best_pred,
            }, is_best)

def main():
    parser = argparse.ArgumentParser(description="PyTorch DeeplabV3Plus Training")
    parser.add_argument('--backbone', type=str, default='resnet',
                        choices=['resnet', 'xception', 'drn', 'mobilenet'],
                        help='backbone name (default: resnet)')
    parser.add_argument('--opt_level', type=str, default='O0',
                        choices=['O0', 'O1', 'O2', 'O3'],
                        help='opt level for half percision training (default: O0)')
    parser.add_argument('--out-stride', type=int, default=16,
                        help='network output stride (default: 8)')
    parser.add_argument('--dataset', type=str, default='kd',
                        choices=['pascal', 'coco', 'cityscapes', 'kd'],
                        help='dataset name (default: pascal)')
    parser.add_argument('--autodeeplab', type=str, default='search',
                        choices=['search', 'train'])
    parser.add_argument('--use-sbd', action='store_true', default=False,
                        help='whether to use SBD dataset (default: True)')
    parser.add_argument('--load-parallel', type=int, default=0)
    parser.add_argument('--clean-module', type=int, default=0)
    parser.add_argument('--workers', type=int, default=0,
                        metavar='N', help='dataloader threads')
    parser.add_argument('--base_size', type=int, default=320,
                        help='base image size')
    parser.add_argument('--crop_size', type=int, default=320,
                        help='crop image size')
    parser.add_argument('--resize', type=int, default=512,
                        help='resize image size')
    parser.add_argument('--sync-bn', type=bool, default=None,
                        help='whether to use sync bn (default: auto)')
    parser.add_argument('--freeze-bn', type=bool, default=False,
                        help='whether to freeze bn parameters (default: False)')
    parser.add_argument('--loss-type', type=str, default='ce',
                        choices=['ce', 'focal'],
                        help='loss func type (default: ce)')
    # training hyper params
    parser.add_argument('--epochs', type=int, default=None, metavar='N',
                        help='number of epochs to train (default: auto)')
    parser.add_argument('--start_epoch', type=int, default=0,
                        metavar='N', help='start epochs (default:0)')
    parser.add_argument('--filter_multiplier', type=int, default=8)
    parser.add_argument('--block_multiplier', type=int, default=5)
    parser.add_argument('--step', type=int, default=5)
    parser.add_argument('--alpha_epoch', type=int, default=20,
                        metavar='N', help='epoch to start training alphas')
    parser.add_argument('--batch-size', type=int, default=2,
                        metavar='N', help='input batch size for \
                                training (default: auto)')
    parser.add_argument('--test-batch-size', type=int, default=None,
                        metavar='N', help='input batch size for \
                                testing (default: auto)')
    parser.add_argument('--use_balanced_weights', action='store_true', default=False,
                        help='whether to use balanced weights (default: False)')
    # optimizer params
    parser.add_argument('--lr', type=float, default=0.025, metavar='LR',
                        help='learning rate (default: auto)')
    parser.add_argument('--min_lr', type=float, default=0.001)
    parser.add_argument('--arch-lr', type=float, default=3e-3, metavar='LR',
                        help='learning rate for alpha and beta in architect searching process')

    parser.add_argument('--lr-scheduler', type=str, default='cos',
                        choices=['poly', 'step', 'cos'],
                        help='lr scheduler mode')
    parser.add_argument('--momentum', type=float, default=0.9,
                        metavar='M', help='momentum (default: 0.9)')
    parser.add_argument('--weight-decay', type=float, default=3e-4,
                        metavar='M', help='w-decay (default: 5e-4)')
    parser.add_argument('--arch-weight-decay', type=float, default=1e-3,
                        metavar='M', help='w-decay (default: 5e-4)')

    parser.add_argument('--nesterov', action='store_true', default=False,
                        help='whether use nesterov (default: False)')
    # cuda, seed and logging
    parser.add_argument('--no-cuda', action='store_true', default=
                        False, help='disables CUDA training')
    parser.add_argument('--gpu-ids', type=str, default='0',
                        help='use which gpu to train, must be a \
                        comma-separated list of integers only (default=0)')
    parser.add_argument('--seed', type=int, default=1, metavar='S',
                        help='random seed (default: 1)')
    # checking point
    parser.add_argument('--resume', type=str, default=None,
                        help='put the path to resuming file if needed')
    parser.add_argument('--checkname', type=str, default=None,
                        help='set the checkpoint name')
    # finetuning pre-trained models
    parser.add_argument('--ft', action='store_true', default=False,
                        help='finetuning on a different dataset')
    # evaluation option
    parser.add_argument('--eval-interval', type=int, default=1,
                        help='evaluuation interval (default: 1)')
    parser.add_argument('--no-val', action='store_true', default=False,
                        help='skip validation during training')

    args = parser.parse_args()
    args.cuda = not args.no_cuda and torch.cuda.is_available()
    if args.cuda:
        try:
            args.gpu_ids = [int(s) for s in args.gpu_ids.split(',')]
        except ValueError:
            raise ValueError('Argument --gpu_ids must be a comma-separated list of integers only')

    if args.sync_bn is None:
        if args.cuda and len(args.gpu_ids) > 1:
            args.sync_bn = True
        else:
            args.sync_bn = False

    # default settings for epochs, batch_size and lr
    if args.epochs is None:
        epoches = {
            'coco': 30,
            'cityscapes': 200,
            'pascal': 50,
            'kd':10
        }
        args.epochs = epoches[args.dataset.lower()]

    if args.batch_size is None:
        args.batch_size = 4 * len(args.gpu_ids)

    if args.test_batch_size is None:
        args.test_batch_size = args.batch_size

    #args.lr = args.lr / (4 * len(args.gpu_ids)) * args.batch_size


    if args.checkname is None:
        args.checkname = 'deeplab-'+str(args.backbone)
    print(args)
    torch.manual_seed(args.seed)
    trainer = Trainer(args)
    print('Starting Epoch:', trainer.args.start_epoch)
    print('Total Epoches:', trainer.args.epochs)
    for epoch in range(trainer.args.start_epoch, trainer.args.epochs):
        trainer.training(epoch)
        if not trainer.args.no_val and epoch % args.eval_interval == (args.eval_interval - 1):
            trainer.validation(epoch)

    trainer.writer.close()

if __name__ == "__main__":
   main()