train_imgreid_xent_vib.py

from __future__ import print_function
from __future__ import division

import os
import sys
import time
import datetime
import argparse
import os.path as osp
import numpy as np

import torch
import torch.nn as nn
import torch.backends.cudnn as cudnn
from torch.utils.data import DataLoader
from torch.optim import lr_scheduler

from torchreid import data_manager
from torchreid.dataset_loader_custom import ImageDataset
from torchreid import transforms as T
from torchreid import models
from torchreid.losses import CrossEntropyLabelSmooth, DeepSupervision,AngularLabelSmooth,AngleLoss,ConfidencePenalty,JSD_loss, InfoLoss, MultiHeadLossAutoTune
from torchreid.utils.iotools import save_checkpoint, check_isfile
from torchreid.utils.avgmeter import AverageMeter
from torchreid.utils.logger import Logger
from torchreid.utils.torchtools import set_bn_to_eval, count_num_param
from torchreid.utils.reidtools import visualize_ranked_results, plot_deltaTheta, drawTSNE
from torchreid.eval_metrics import evaluate
from torchreid.optimizers import init_optim

from tensorboardX import SummaryWriter
import random
import pdb

import math

from sklearn.metrics import pairwise_distances as pw

from torchreid.utils.re_ranking import re_ranking

## ECN
from torchreid.utils.ecn import ECN
parser = argparse.ArgumentParser(description='Train image model with cross entropy loss')
# Datasets
parser.add_argument('--root', type=str, default='data',
                    help="root path to data directory")
parser.add_argument('-d', '--dataset', type=str, default='market1501',
                    choices=data_manager.get_names())
parser.add_argument('-j', '--workers', default=4, type=int,
                    help="number of data loading workers (default: 4)")
parser.add_argument('--height', type=int, default=256,
                    help="height of an image (default: 256)")
parser.add_argument('--width', type=int, default=128,
                    help="width of an image (default: 128)")
parser.add_argument('--split-id', type=int, default=0,
                    help="split index (0-based)")
# CUHK03-specific setting
parser.add_argument('--cuhk03-labeled', action='store_true',
                    help="use labeled images, if false, detected images are used (default: False)")
parser.add_argument('--cuhk03-classic-split', action='store_true',
                    help="use classic split by Li et al. CVPR'14 (default: False)")
parser.add_argument('--use-metric-cuhk03', action='store_true',
                    help="use cuhk03-metric (default: False)")
# Optimization options
parser.add_argument('--optim', type=str, default='adam',
                    help="optimization algorithm (see optimizers.py)")
parser.add_argument('--max-epoch', default=60, type=int,
                    help="maximum epochs to run")
parser.add_argument('--start-epoch', default=0, type=int,
                    help="manual epoch number (useful on restarts)")
parser.add_argument('--train-batch', default=32, type=int,
                    help="train batch size")
parser.add_argument('--test-batch', default=100, type=int,
                    help="test batch size")
parser.add_argument('--lr', '--learning-rate', default=0.0003, type=float,
                    help="initial learning rate")
parser.add_argument('--stepsize', default=[20, 40], nargs='+', type=int,
                    help="stepsize to decay learning rate")
parser.add_argument('--gamma', default=0.1, type=float,
                    help="learning rate decay")
parser.add_argument('--weight-decay', default=5e-04, type=float,
                    help="weight decay (default: 5e-04)")
parser.add_argument('--fixbase-epoch', default=0, type=int,
                    help="epochs to fix base network (only train classifier, default: 0)")
parser.add_argument('--fixbase-lr', default=0.0003, type=float,
                    help="learning rate (when base network is frozen)")
parser.add_argument('--freeze-bn', action='store_true',
                    help="freeze running statistics in BatchNorm layers during training (default: False)")
parser.add_argument('--label-smooth', action='store_true',
                    help="use label smoothing regularizer in cross entropy loss")

parser.add_argument('--scheduler', action='store_true',
                    help="Enable learning rate schedular")

parser.add_argument('--test-rot', action='store_true',
                    help="Train only classifier to get rotation")
parser.add_argument('--plot-deltaTheta', action='store_true',
                    help="Plot DeltaTheta, only available in evaluation mode (default: False)")


# Architecture
parser.add_argument('-a', '--arch', type=str, default='resnet50', choices=models.get_names())
# Miscs
parser.add_argument('--print-freq', type=int, default=10,
                    help="print frequency")
parser.add_argument('--seed', type=int, default=1,
                    help="manual seed")
parser.add_argument('--resume', type=str, default='', metavar='PATH')
parser.add_argument('--load-weights', type=str, default='',
                    help="load pretrained weights but ignores layers that don't match in size")
parser.add_argument('--evaluate', action='store_true',
                    help="evaluation only")
parser.add_argument('--eval-step', type=int, default=-1,
                    help="run evaluation for every N epochs (set to -1 to test after training)")
parser.add_argument('--save-epoch', type=int, default=-1,
                    help="run evaluation for every N epochs (set to -1 to test after training)")
parser.add_argument('--start-eval', type=int, default=0,
                    help="start to evaluate after specific epoch")
parser.add_argument('--save-dir', type=str, default='log')
parser.add_argument('--use-cpu', action='store_true',
                    help="use cpu")
parser.add_argument('--gpu-devices', default='0', type=str,
                    help='gpu device ids for CUDA_VISIBLE_DEVICES')
parser.add_argument('--use-avai-gpus', action='store_true',
                    help="use available gpus instead of specified devices (this is useful when using managed clusters)")
parser.add_argument('--visualize-ranks', action='store_true',
                    help="visualize ranked results, only available in evaluation mode (default: False)")

parser.add_argument('--lambda-xent', type=float, default=1,
                    help="weight to balance cross entropy loss")

parser.add_argument('--beta', type=float, default=1,
                    help="weight to balance info loss")

parser.add_argument('--use-angular', action='store_true',
                    help="Use Angular Softmax (default: False)")
parser.add_argument('--draw-tsne', action='store_true',
                    help="Plot TSNE Clusters (default: False)")
parser.add_argument('--tsne-labels',type=int, default=3,
                    help="Number of TSNE Clusters (Default: 3)")
parser.add_argument('--mahalanobis', action='store_true',
                    help="Use mahalanobis (default: False)")
# global variables
#args = parser.parse_args()
#best_rank1 = -np.inf
parser.add_argument('--confidence-penalty', action='store_true',
                    help="use confidence penalty regularizer in cross entropy loss")
parser.add_argument('--confidence-beta',type=float, default=1,
                    help="set confidence penalty beta ")
parser.add_argument('--jsd', action='store_true',
                    help="use JSD in cross entropy loss")
parser.add_argument('--auto-tune-mtl', action='store_true',
                    help="Use Loss AutoTune")

parser.add_argument("--re-ranking", action='store_true',
                    help="Use k-reciprocal re-ranking (default: False)")
parser.add_argument("--use-ecn", action='store_true',
                    help="Use ECN re-ranking (default: False)")

parser.add_argument("--use-cosine", action='store_true',
                    help="Use cosine distance to rank (default: False)")
parser.add_argument('-sw', '--split-wild', type=str, default='large',
                    choices=["small", "medium", "large"])

def main(args):
    args = parser.parse_args(args)
    #global best_rank1
    best_rank1 = -np.inf
    torch.manual_seed(args.seed)
    # np.random.seed(args.seed)
    # random.seed(args.seed)
    if not args.use_avai_gpus: os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_devices
    use_gpu = torch.cuda.is_available()
    if args.use_cpu: use_gpu = False

    if not args.evaluate:
        sys.stdout = Logger(osp.join(args.save_dir, 'log_train.txt'))
    else:
        test_dir = args.save_dir
        if args.save_dir =='log':
            if args.resume:
                test_dir = os.path.dirname(args.resume)
            else:
                test_dir = os.path.dirname(args.load_weights)
        sys.stdout = Logger(osp.join(test_dir, 'log_test.txt'))
    print("==========\nArgs:{}\n==========".format(args))

    if use_gpu:
        print("Currently using GPU {}".format(args.gpu_devices))
        cudnn.benchmark = True
        torch.cuda.manual_seed_all(args.seed)
    else:
        print("Currently using CPU (GPU is highly recommended)")

    print("Initializing dataset {}".format(args.dataset))
    dataset = data_manager.init_imgreid_dataset(
        root=args.root, name=args.dataset, split_id=args.split_id,
        cuhk03_labeled=args.cuhk03_labeled, cuhk03_classic_split=args.cuhk03_classic_split, split_wild=args.split_wild
    )

    transform_train = T.Compose([
        T.Random2DTranslation(args.height, args.width),
        #T.Resize((args.height, args.width)),
        T.RandomSizedEarser(),
        T.RandomHorizontalFlip_custom(),
        #T.Resize((args.height, args.width)),
        T.ToTensor(),
        T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ])

    transform_test = T.Compose([
        T.Resize((args.height, args.width)),
        T.ToTensor(),
        T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ])

    pin_memory = True if use_gpu else False

    trainloader = DataLoader(
        ImageDataset(dataset.train, transform=transform_train),
        batch_size=args.train_batch, shuffle=True, num_workers=args.workers,
        pin_memory=pin_memory, drop_last=True,
    )

    queryloader = DataLoader(
        ImageDataset(dataset.query, transform=transform_test, return_path=args.draw_tsne),
        batch_size=args.test_batch, shuffle=False, num_workers=args.workers,
        pin_memory=pin_memory, drop_last=False,
    )

    galleryloader = DataLoader(
        ImageDataset(dataset.gallery, transform=transform_test, return_path=args.draw_tsne),
        batch_size=args.test_batch, shuffle=False, num_workers=args.workers,
        pin_memory=pin_memory, drop_last=False,
    )

    print("Initializing model: {}".format(args.arch))
    model = models.init_model(name=args.arch, num_classes=dataset.num_train_pids, loss={'xent','angular'} if args.use_angular else {'xent'}, use_gpu=use_gpu)
    print("Model size: {:.3f} M".format(count_num_param(model)))
    use_autoTune = False
    info_loss = InfoLoss()
    if not(args.use_angular):
        if args.label_smooth:
            print("Using Label Smoothing")
            criterion = CrossEntropyLabelSmooth(num_classes=dataset.num_train_pids, use_gpu=use_gpu)
        else:
            print("Using Normal Cross-Entropy")
            criterion = nn.CrossEntropyLoss()
        if args.jsd:
            print("Using JSD regularizer")
            criterion = (criterion,JSD_loss(dataset.num_train_pids),info_loss)
        else:
            if args.confidence_penalty:
                print("Using Confidence Penalty")
            criterion = (criterion,ConfidencePenalty(),info_loss)
    else:
        if args.label_smooth:
            print("Using Angular Label Smoothing")
            criterion = (AngularLabelSmooth(num_classes=dataset.num_train_pids, use_gpu=use_gpu),info_loss)
        else:
            print("Using Label Smoothing")
            criterion = (AngleLoss(),info_loss)
    if args.auto_tune_mtl:
        if args.confidence_penalty:
            print("Using AutoTune with CP")
            criterion = MultiHeadLossAutoTune(criterion, [args.lambda_xent, -args.confidence_beta, args.beta]).cuda()
        elif args.jsd:
            print("Using AutoTune with JSD")
            criterion = MultiHeadLossAutoTune(criterion, [args.lambda_xent, args.confidence_beta, args.beta]).cuda()
        else:
            print("Using AutoTune without CP or JSD")
            criterion = MultiHeadLossAutoTune((criterion[0], criterion[-1]), [args.lambda_xent , args.beta]).cuda()
        optimizer = init_optim(args.optim, list(model.parameters()) + list(criterion.parameters()), args.lr, args.weight_decay)
    else:
        optimizer = init_optim(args.optim, model.parameters(), args.lr, args.weight_decay)
    if args.scheduler:
        scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=args.stepsize, gamma=args.gamma)

    if args.fixbase_epoch > 0:
        if hasattr(model, 'classifier') and isinstance(model.classifier, nn.Module):
            if args.auto_tune_mtl:
                optimizer_tmp = init_optim(args.optim, list(model.classifier.parameters())+list(model.encoder.parameters())+ list(criterion.parameters()), args.fixbase_lr, args.weight_decay)
            else:
                optimizer_tmp = init_optim(args.optim, list(model.classifier.parameters())+list(model.encoder.parameters()), args.fixbase_lr, args.weight_decay)
        else:
            print("Warn: model has no attribute 'classifier' and fixbase_epoch is reset to 0")
            args.fixbase_epoch = 0

    if args.load_weights and check_isfile(args.load_weights):
        # load pretrained weights but ignore layers that don't match in size
        checkpoint = torch.load(args.load_weights)
        pretrain_dict = checkpoint['state_dict']
        model_dict = model.state_dict()
        pretrain_dict = {k: v for k, v in pretrain_dict.items() if k in model_dict and model_dict[k].size() == v.size()}
        model_dict.update(pretrain_dict)
        model.load_state_dict(model_dict)
        print("Loaded pretrained weights from '{}'".format(args.load_weights))

    if args.resume and check_isfile(args.resume):
        checkpoint = torch.load(args.resume)
        model.load_state_dict(checkpoint['state_dict'])
        args.start_epoch = checkpoint['epoch'] + 1
        best_rank1 = checkpoint['rank1']
        print("Loaded checkpoint from '{}'".format(args.resume))
        print("- start_epoch: {}\n- rank1: {}".format(args.start_epoch, best_rank1))

    if use_gpu:
        model = nn.DataParallel(model).cuda()

    if args.evaluate:
        print("Evaluate only")
        test_dir = args.save_dir
        if args.save_dir =='log':
            if args.resume:
                test_dir = os.path.dirname(args.resume)
            else:
                test_dir = os.path.dirname(args.load_weights)
        distmat = test(model, queryloader, galleryloader, use_gpu, args,writer=None,epoch=-1, return_distmat=True,draw_tsne=args.draw_tsne,tsne_clusters=args.tsne_labels, use_cosine = args.plot_deltaTheta)

        if args.visualize_ranks:
            visualize_ranked_results(
                distmat, dataset,
                save_dir=osp.join(test_dir, 'ranked_results'),
                topk=10,
            )
        if args.plot_deltaTheta:
            plot_deltaTheta(distmat, dataset,save_dir=osp.join(test_dir,'deltaTheta_results'), min_rank=1)
        return


    writer = SummaryWriter(log_dir=osp.join(args.save_dir, 'tensorboard'))
    start_time = time.time()
    train_time = 0
    best_epoch = args.start_epoch
    print("==> Start training")

    if args.fixbase_epoch > 0:
        print("Train classifier for {} epochs while keeping base network frozen".format(args.fixbase_epoch))

        for epoch in range(args.fixbase_epoch):
            start_train_time = time.time()
            train(epoch, model, criterion, optimizer_tmp, trainloader, use_gpu,writer, args, freeze_bn=True)
            train_time += round(time.time() - start_train_time)

        del optimizer_tmp
        print("Now open all layers for training")
    best_epoch = 0
    for epoch in range(args.start_epoch, args.max_epoch):
        start_train_time = time.time()
        train(epoch, model, criterion, optimizer, trainloader, use_gpu,writer, args)
        train_time += round(time.time() - start_train_time)

        if args.scheduler:
            scheduler.step()

        if (epoch + 1) > args.start_eval and ((args.save_epoch > 0 and (epoch + 1) % args.save_epoch == 0) or (args.eval_step > 0 and (epoch + 1) % args.eval_step == 0) or (epoch + 1) == args.max_epoch):
            if (epoch + 1) == args.max_epoch:
                if use_gpu:
                    state_dict = model.module.state_dict()
                else:
                    state_dict = model.state_dict()

                save_checkpoint({
                    'state_dict': state_dict,
                    'rank1': -1,
                    'epoch': epoch,
                }, False, osp.join(args.save_dir, 'beforeTesting_checkpoint_ep' + str(epoch + 1) + '.pth.tar'))
            is_best = False
            rank1 = -1
            if args.eval_step > 0:
                print("==> Test")

                rank1 = test(model, queryloader, galleryloader, use_gpu, args,writer=writer,epoch=epoch)

                is_best = rank1 > best_rank1

                if is_best:
                    best_rank1 = rank1
                    best_epoch = epoch + 1

            if use_gpu:
                state_dict = model.module.state_dict()
            else:
                state_dict = model.state_dict()


            save_checkpoint({
                'state_dict': state_dict,
                'rank1': rank1,
                'epoch': epoch,
            }, is_best, osp.join(args.save_dir, 'checkpoint_ep' + str(epoch + 1) + '.pth.tar'))

    print("==> Best Rank-1 {:.1%}, achieved at epoch {}".format(best_rank1, best_epoch))

    elapsed = round(time.time() - start_time)
    elapsed = str(datetime.timedelta(seconds=elapsed))
    train_time = str(datetime.timedelta(seconds=train_time))
    print("Finished. Total elapsed time (h:m:s): {}. Training time (h:m:s): {}.".format(elapsed, train_time))
    return best_rank1, best_epoch

def train(epoch, model, criterion, optimizer, trainloader, use_gpu,writer, args,freeze_bn=False):
    losses = AverageMeter()
    xent_losses = AverageMeter()
    info_losses = AverageMeter()
    batch_time = AverageMeter()
    data_time = AverageMeter()
    confidence_losses = AverageMeter()
    printed = False
    model.train()

    if freeze_bn or args.freeze_bn:
        model.apply(set_bn_to_eval)

    end = time.time()
    for batch_idx, (imgs, pids, _) in enumerate(trainloader):
        data_time.update(time.time() - end)

        if use_gpu:
            imgs, pids = imgs.cuda(), pids.cuda()

        (mu, std),outputs = model(imgs)
        text_dict = {}
        if not isinstance(criterion, MultiHeadLossAutoTune):
            if isinstance(outputs, tuple):
                xent_loss = DeepSupervision(criterion[0], outputs, pids)
                confidence_loss = DeepSupervision(criterion[1],outputs,pids)
            else:
                xent_loss = criterion[0](outputs, pids)
                confidence_loss = criterion[1](outputs,pids)

            info_loss = criterion[-1](mu.float(), std.float())

            if args.confidence_penalty:
                loss = args.lambda_xent *xent_loss + args.beta*info_loss - args.confidence_beta *confidence_loss
            elif args.jsd:
                loss = args.lambda_xent *xent_loss + args.beta*info_loss + args.confidence_beta *confidence_loss
            else:
                loss = args.lambda_xent *xent_loss + args.beta*info_loss
            confidence_losses.update(confidence_loss.item(), pids.size(0))
        else:

            if args.confidence_penalty or args.jsd:
                loss, individual_losses = criterion([outputs,outputs, mu], [pids, pids, std])
                confidence_loss = individual_losses[1]
            else:
                loss, individual_losses = criterion([outputs, mu], [pids, std])
                confidence_loss = 0
            xent_loss = individual_losses[0]
            info_loss = individual_losses[-1]
            text_dict = criterion.batch_meta()
            confidence_losses.update(0, pids.size(0))
        #info_loss = -0.5*(1-2*std.log()-(1+mu.pow(2))/(2*std.pow(2))).sum(1).mean().div(math.log(2))

        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

        batch_time.update(time.time() - end)

        losses.update(loss.item(), pids.size(0))

        xent_losses.update(xent_loss.item(), pids.size(0))
        info_losses.update(info_loss.item(), pids.size(0))
        if (batch_idx + 1) % args.print_freq == 0:
            if not printed:
              printed = True
            else:
              # Clean the current line
              sys.stdout.console.write("\033[F\033[K")
              #sys.stdout.console.write("\033[K")
            if args.jsd:
                print('Epoch: [{0}][{1}/{2}]\t'
                      'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
                      'Data {data_time.val:.4f} ({data_time.avg:.4f})\t'
                      'Xent_Loss {xent_loss.val:.4f} ({xent_loss.avg:.4f})\t'
                      'JSD_Loss {confidence_loss.val:.4f} ({confidence_loss.avg:.4f})\t'
                      'Info_Loss {info_loss.val:.4f} ({info_loss.avg:.4f})\t'
                      'Total_Loss {loss.val:.4f} ({loss.avg:.4f})\t'.format(
                       epoch + 1, batch_idx + 1, len(trainloader), batch_time=batch_time,
                       data_time=data_time,xent_loss=xent_losses,confidence_loss=confidence_losses, loss=losses), text_dict)
            else:
                print('Epoch: [{0}][{1}/{2}]\t'
                      'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
                      'Data {data_time.val:.4f} ({data_time.avg:.4f})\t'
                      'Xent_Loss {xent_loss.val:.4f} ({xent_loss.avg:.4f})\t'
                      'Confi_Loss {confidence_loss.val:.4f} ({confidence_loss.avg:.4f})\t'
                      'Info_Loss {info_loss.val:.4f} ({info_loss.avg:.4f})\t'
                      'Total_Loss {loss.val:.4f} ({loss.avg:.4f})\t'.format(
                       epoch + 1, batch_idx + 1, len(trainloader), batch_time=batch_time,
                       data_time=data_time,xent_loss=xent_losses,confidence_loss=confidence_losses,info_loss=info_losses, loss=losses), text_dict)

        end = time.time()

    writer.add_scalars(
      'loss',
      dict(loss=losses.avg,
            xent_loss=xent_losses.avg,
            info_loss=info_losses.avg,
            confidence_loss = confidence_losses.avg),
      epoch + 1)

def test(model, queryloader, galleryloader, use_gpu, args,writer,epoch, ranks=[1, 5, 10, 20], return_distmat=False,use_cosine = False,draw_tsne=False,tsne_clusters=3):

    batch_time = AverageMeter()

    model.eval()

    with torch.no_grad():
        qf, q_pids, q_camids = [], [], []
        qf_std = []
        q_imgPath = []
        for batch_idx, (input) in enumerate(queryloader):
            if not args.draw_tsne:
                imgs, pids, camids = input
            else:
                imgs, pids, camids,img_path = input
                q_imgPath.extend(img_path)
            if use_gpu: imgs = imgs.cuda()

            end = time.time()
            features,std = model(imgs)
            batch_time.update(time.time() - end)

            features = features.data.cpu()
            std = std.data.cpu()
            qf.append(features)
            qf_std.append(std)
            q_pids.extend(pids)
            q_camids.extend(camids)
        qf = torch.cat(qf, 0)
        qf_std = torch.cat(qf_std, 0)
        q_pids = np.asarray(q_pids)
        q_camids = np.asarray(q_camids)
        q_imgPath = np.asarray(q_imgPath)

        print("Extracted features for query set, obtained {}-by-{} matrix".format(qf.size(0), qf.size(1)))

        gf, g_pids, g_camids = [], [], []
        #gf_std = []
        g_imgPath = []
        end = time.time()
        for batch_idx, (input) in enumerate(galleryloader):
            if not args.draw_tsne:
                imgs, pids, camids = input
            else:
                imgs, pids, camids,img_path = input
                g_imgPath.extend(img_path)
            if use_gpu: imgs = imgs.cuda()

            end = time.time()
            features,_ = model(imgs)
            batch_time.update(time.time() - end)

            features = features.data.cpu()
            #std = std.data.cpu()
            gf.append(features)
            #gf_std.append(std)
            g_pids.extend(pids)
            g_camids.extend(camids)
        gf = torch.cat(gf, 0)
        #gf_std = torch.cat(gf_std, 0)
        g_pids = np.asarray(g_pids)
        g_camids = np.asarray(g_camids)
        g_imgPath = np.asarray(q_imgPath)

        print("Extracted features for gallery set, obtained {}-by-{} matrix".format(gf.size(0), gf.size(1)))

    print("==> BatchTime(s)/BatchSize(img): {:.3f}/{}".format(batch_time.avg, args.test_batch))
    m, n = qf.size(0), gf.size(0)

    if args.use_ecn:
        distmat= (ECN(qf.numpy(),gf.numpy(),k=25,t=3,q=8,method='rankdist')).transpose()
    elif args.mahalanobis:
        print("Using STD for Mahalanobis distance")
        distmat = torch.zeros((m,n))
        # #pdb.set_trace()
        # qf = qf.data.numpy()
        # gf= gf.data.numpy()
        # qf_std = qf_std.data.numpy()
        # for q_indx in range(int(m)):
        #     distmat[q_indx]= pw(np.expand_dims(qf[q_indx],axis=0),gf,metric='mahalanobis',n_jobs=8, VI=(np.eye(qf_std[q_indx].shape[0])*(1/qf_std[q_indx]).transpose()))
        #     print(q_indx)
        # pdb.set_trace()
        qf = qf / qf_std
        for q_indx in range(int(m)):
            gf_norm = gf * 1/qf_std[q_indx]
            distmat[q_indx] = torch.pow(qf[q_indx], 2).sum(dim=0, keepdim=True).expand(n) + \
                      torch.pow(gf_norm, 2).sum(dim=1, keepdim=True).squeeze()
            distmat[q_indx].unsqueeze(0).addmm_(1, -2, qf[q_indx].unsqueeze(0), gf_norm.t())
        distmat = distmat.numpy()
    elif not (use_cosine or args.use_cosine):

        distmat = torch.pow(qf, 2).sum(dim=1, keepdim=True).expand(m, n) + \
                  torch.pow(gf, 2).sum(dim=1, keepdim=True).expand(n, m).t()
        distmat.addmm_(1, -2, qf, gf.t())
        distmat = distmat.numpy()

        if args.re_ranking:
            distmat_q_q = torch.pow(qf, 2).sum(dim=1, keepdim=True).expand(m, m) + \
                      torch.pow(qf, 2).sum(dim=1, keepdim=True).expand(m, m).t()
            distmat_q_q.addmm_(1, -2, qf, qf.t())
            distmat_q_q = distmat_q_q.numpy()

            distmat_g_g = torch.pow(gf, 2).sum(dim=1, keepdim=True).expand(n, n) + \
                      torch.pow(gf, 2).sum(dim=1, keepdim=True).expand(n, n).t()
            distmat_g_g.addmm_(1, -2, gf, gf.t())
            distmat_g_g = distmat_g_g.numpy()

            print("Normal Re-Ranking")
            distmat = re_ranking(distmat, distmat_q_q, distmat_g_g, k1=20, k2=6, lambda_value=0.3)
    else:
        qf_norm = qf/qf.norm(dim=1)[:,None]
        gf_norm = gf/gf.norm(dim=1)[:,None]
        distmat = torch.addmm(1,torch.ones((m,n)),-1,qf_norm,gf_norm.transpose(0,1))
        distmat = distmat.numpy()

        if args.re_ranking:
            distmat_q_q = torch.addmm(1,torch.ones((m,m)),-1,qf_norm,qf_norm.transpose(0,1))
            distmat_q_q = distmat_q_q.numpy()

            distmat_g_g = torch.addmm(1,torch.ones((n,n)),-1,gf_norm,gf_norm.transpose(0,1))
            distmat_g_g = distmat_g_g.numpy()

            print("Re-Ranking with Cosine")
            distmat = re_ranking(distmat, distmat_q_q, distmat_g_g, k1=20, k2=6, lambda_value=0.3)

    print("Computing CMC and mAP")
    cmc, mAP = evaluate(distmat, q_pids, g_pids, q_camids, g_camids, use_metric_cuhk03=args.use_metric_cuhk03)

    print("Results ----------")
    print("mAP: {:.1%}".format(mAP))
    print("CMC curve")
    for r in ranks:
        print("Rank-{:<3}: {:.1%}".format(r, cmc[r-1]))
    print("------------------")

    if draw_tsne:
        drawTSNE(qf,gf,q_pids, g_pids, q_camids, g_camids,q_imgPath, g_imgPath,tsne_clusters,args.save_dir)
    if return_distmat:
        return distmat


    if writer != None:
        writer.add_scalars(
          'Testing',
          dict(rank_1=cmc[0],
               rank_5 = cmc[4],
               mAP=mAP),
          epoch + 1)
    return cmc[0]

def test_vib(model, queryloader, galleryloader, use_gpu, args,writer,epoch, ranks=[1, 5, 10, 20], return_distmat=False,use_cosine = False,draw_tsne=False,tsne_clusters=3):

    batch_time = AverageMeter()

    model.eval()

    with torch.no_grad():
        q_pids, q_camids = [], [], []
        for batch_idx, (imgs, pids, camids) in enumerate(queryloader):
            if use_gpu: imgs = imgs.cuda()

            end = time.time()
            features = model(imgs)
            batch_time.update(time.time() - end)

            features = features.data.cpu()
            qf_stat.append(features)
            q_pids.extend(pids)
            q_camids.extend(camids)
        qf_stat = torch.cat(qf_stat, 0)
        q_pids = np.asarray(q_pids)
        q_camids = np.asarray(q_camids)

        print("Extracted features for query set, obtained {}-by-{} matrix".format(qf_stat.size(0), qf_stat.size(1)))

        g_pids, g_camids = [], [], []
        end = time.time()
        for batch_idx, (imgs, pids, camids) in enumerate(galleryloader):
            if use_gpu: imgs = imgs.cuda()

            end = time.time()
            features = model(imgs)
            batch_time.update(time.time() - end)

            features = features.data.cpu()
            gf_stat.append(features)
            g_pids.extend(pids)
            g_camids.extend(camids)
        gf_stat = torch.cat(gf_stat, 0)
        g_pids = np.asarray(g_pids)
        g_camids = np.asarray(g_camids)

        print("Extracted features for gallery set, obtained {}-by-{} matrix".format(gf.size(0), gf.size(1)))

    print("==> BatchTime(s)/BatchSize(img): {:.3f}/{}".format(batch_time.avg, args.test_batch))

    m, n = qf_stat.size(0), gf_stat.size(0)
    score_board = torch.zeros((m,n,n),dtype=torch.int16)
    qf =torch.zeros(m,512)#,torch.zeros(m,512)
    for _ in range(args.sampling_count):
        qf_sample = model.reparametrize_n(qf_stat[:,0],qf_stat[:,1])
        qf = qf + qf_sample

    qf = qf / args.sampling_count

    for _ in range(args.sampling_count):

        #qf = model.reparametrize_n(qf_stat[:,0],qf_stat[:,1])
        gf = model.reparametrize_n(gf_stat[:,0],gf_stat[:,1])
        if not use_cosine:
            distmat = torch.pow(qf, 2).sum(dim=1, keepdim=True).expand(m, n) + \
                      torch.pow(gf, 2).sum(dim=1, keepdim=True).expand(n, m).t()
            distmat.addmm_(1, -2, qf, gf.t())
            distmat = distmat.numpy()
        else:
            qf_norm = qf/qf.norm(dim=1)[:,None]
            gf_norm = gf/gf.norm(dim=1)[:,None]
            distmat = torch.addmm(1,torch.ones((m,n)),-1,qf_norm,gf_norm.transpose(0,1))
            distmat = distmat.numpy()

        indices = np.argsort(distmat, axis=1)

        for indx in range(m):
            score_board[m, indices[indx],list(range(n))] += 1

    print("Computing CMC and mAP")
    cmc, mAP = evaluate(distmat, q_pids, g_pids, q_camids, g_camids, use_metric_cuhk03=args.use_metric_cuhk03)

    print("Results ----------")
    print("mAP: {:.1%}".format(mAP))
    print("CMC curve")
    for r in ranks:
        print("Rank-{:<3}: {:.1%}".format(r, cmc[r-1]))
    print("------------------")

    if draw_tsne:
        drawTSNE(qf,gf,q_pids, g_pids, q_camids, g_camids,tsne_clusters,args.save_dir)
    if return_distmat:
        return distmat


    if writer != None:
        writer.add_scalars(
          'Testing',
          dict(rank_1=cmc[0],
               rank_5 = cmc[4],
               mAP=mAP),
          epoch + 1)
    return cmc[0]

def accuracy(output, target, topk=(1,)):
    """Computes the precision@k for the specified values of k"""
    maxk = max(topk)
    batch_size = target.size(0)

    _, pred = output.topk(maxk, 1, True, True)
    pred = pred.t()
    correct = pred.eq(target.view(1, -1).expand_as(pred))

    res = []
    for k in topk:
        correct_k = correct[:k].view(-1).float().sum(0, keepdim=True)
        res.append(correct_k.mul_(100.0 / batch_size))
    return res

if __name__ == '__main__':
    main(sys.argv[1:])