trainval_net_global_local.py

# coding:utf-8
# --------------------------------------------------------
# Pytorch multi-GPU Faster R-CNN
# Licensed under The MIT License [see LICENSE for details]
# Written by Jiasen Lu, Jianwei Yang, based on code from Ross Girshick
# --------------------------------------------------------
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import os
import numpy as np
import pprint
import pdb
import time
import _init_paths

from pdb import set_trace as breakpoint

import torch
from torch.autograd import Variable
import torch.nn as nn
from roi_data_layer.roidb import combined_roidb
from roi_data_layer.roibatchLoader import roibatchLoader

from model.utils.config import cfg, cfg_from_file, cfg_from_list, get_output_dir

from model.utils.net_utils import weights_normal_init, save_net, load_net, \
    adjust_learning_rate, save_checkpoint, clip_gradient, FocalLoss, sampler, calc_supp, EFocalLoss

from model.utils.parser_func import parse_args, set_dataset_args

from model.rpn.bbox_transform import clip_boxes
from model.nms.nms_wrapper import nms
from model.rpn.bbox_transform import bbox_transform_inv

import sys


from log_utils.utils import ReDirectSTD



def test_model_while_training(fasterRCNN, args):

  # args = parse_args()
  # args = set_dataset_args(args, test=True)
  # np.random.seed(cfg.RNG_SEED)

  if args.cfg_file is not None:
    cfg_from_file(args.cfg_file)
  if args.set_cfgs is not None:
    cfg_from_list(args.set_cfgs)

  cfg.TRAIN.USE_FLIPPED = False

  # args.imdbval_name = 'clipart_test' 

  imdb, roidb, ratio_list, ratio_index = combined_roidb(args.imdbval_name_target, False)

  # breakpoint()

  imdb.competition_mode(on=True)

  print('{:d} roidb entries'.format(len(roidb)))


  im_data = torch.FloatTensor(1)
  im_info = torch.FloatTensor(1)
  num_boxes = torch.LongTensor(1)
  gt_boxes = torch.FloatTensor(1)


  if args.cuda:
    im_data = im_data.cuda()
    im_info = im_info.cuda()
    num_boxes = num_boxes.cuda()
    gt_boxes = gt_boxes.cuda()

  # make variable
  im_data = Variable(im_data)
  im_info = Variable(im_info)
  num_boxes = Variable(num_boxes)
  gt_boxes = Variable(gt_boxes)

  if args.cuda:
    cfg.CUDA = True

  # if args.cuda:
  #   fasterRCNN.cuda()

  start = time.time()
  max_per_image = 100

  thresh = 0.0

  save_name = args.load_name.split('/')[-1]
  num_images = len(imdb.image_index)
  all_boxes = [[[] for _ in range(num_images)]
               for _ in range(imdb.num_classes)]

  output_dir = get_output_dir(imdb, save_name)
  dataset = roibatchLoader(roidb, ratio_list, ratio_index, 1, \
                        imdb.num_classes, training=False, normalize = False, path_return=True)
  dataloader = torch.utils.data.DataLoader(dataset, batch_size=1,
                            shuffle=False, num_workers=0, pin_memory=True)

  data_iter = iter(dataloader)

  _t = {'im_detect': time.time(), 'misc': time.time()}
  det_file = os.path.join(output_dir, 'detections.pkl')

  fasterRCNN.eval()
  empty_array = np.transpose(np.array([[],[],[],[],[]]), (1,0))
  for i in range(num_images):

      data = next(data_iter)
      im_data.data.resize_(data[0].size()).copy_(data[0])
      #print(data[0].size())
      im_info.data.resize_(data[1].size()).copy_(data[1])
      gt_boxes.data.resize_(data[2].size()).copy_(data[2])
      num_boxes.data.resize_(data[3].size()).copy_(data[3])

      det_tic = time.time()

      rois, cls_prob, bbox_pred, \
      rpn_loss_cls, rpn_loss_box, \
      RCNN_loss_cls, RCNN_loss_bbox, \
      rois_label, _, _ = fasterRCNN(im_data, im_info, gt_boxes, num_boxes)


      scores = cls_prob.data
      boxes = rois.data[:, :, 1:5]
      # d_pred = d_pred.data
      path = data[4]

      if cfg.TEST.BBOX_REG:
          # Apply bounding-box regression deltas
          box_deltas = bbox_pred.data
          if cfg.TRAIN.BBOX_NORMALIZE_TARGETS_PRECOMPUTED:
          # Optionally normalize targets by a precomputed mean and stdev
            if args.class_agnostic:
                box_deltas = box_deltas.view(-1, 4) * torch.FloatTensor(cfg.TRAIN.BBOX_NORMALIZE_STDS).cuda() \
                           + torch.FloatTensor(cfg.TRAIN.BBOX_NORMALIZE_MEANS).cuda()
                box_deltas = box_deltas.view(1, -1, 4)
            else:
                box_deltas = box_deltas.view(-1, 4) * torch.FloatTensor(cfg.TRAIN.BBOX_NORMALIZE_STDS).cuda() \
                           + torch.FloatTensor(cfg.TRAIN.BBOX_NORMALIZE_MEANS).cuda()
                box_deltas = box_deltas.view(1, -1, 4 * len(imdb.classes))

          pred_boxes = bbox_transform_inv(boxes, box_deltas, 1)
          pred_boxes = clip_boxes(pred_boxes, im_info.data, 1)
      else:
          # Simply repeat the boxes, once for each class
          pred_boxes = np.tile(boxes, (1, scores.shape[1]))

      pred_boxes /= data[1][0][2].item()

      scores = scores.squeeze()
      pred_boxes = pred_boxes.squeeze()
      det_toc = time.time()
      detect_time = det_toc - det_tic
      misc_tic = time.time()

      for j in range(1, imdb.num_classes):
          inds = torch.nonzero(scores[:,j]>thresh).view(-1)
          # if there is det
          if inds.numel() > 0:
            cls_scores = scores[:,j][inds]
            _, order = torch.sort(cls_scores, 0, True)
            if args.class_agnostic:
              cls_boxes = pred_boxes[inds, :]
            else:
              cls_boxes = pred_boxes[inds][:, j * 4:(j + 1) * 4]
            
            cls_dets = torch.cat((cls_boxes, cls_scores.unsqueeze(1)), 1)
            # cls_dets = torch.cat((cls_boxes, cls_scores), 1)
            cls_dets = cls_dets[order]
            keep = nms(cls_dets, cfg.TEST.NMS)
            cls_dets = cls_dets[keep.view(-1).long()]

            all_boxes[j][i] = cls_dets.cpu().numpy()
          else:
            all_boxes[j][i] = empty_array

      # Limit to max_per_image detections *over all classes*
      if max_per_image > 0:
          image_scores = np.hstack([all_boxes[j][i][:, -1]
                                    for j in range(1, imdb.num_classes)])
          if len(image_scores) > max_per_image:
              image_thresh = np.sort(image_scores)[-max_per_image]
              for j in range(1, imdb.num_classes):
                  keep = np.where(all_boxes[j][i][:, -1] >= image_thresh)[0]
                  all_boxes[j][i] = all_boxes[j][i][keep, :]

      # misc_toc = time.time()

      sys.stdout.write('im_detect: {:d}/{:d} {:.3f}s \r' \
          .format(i + 1, num_images, detect_time))
      sys.stdout.flush()

  imdb.evaluate_detections(all_boxes, output_dir)



if __name__ == '__main__':

    args = parse_args()

    print('Called with args:')
    print(args)
    args = set_dataset_args(args)
    if args.cfg_file is not None:
        cfg_from_file(args.cfg_file)
    if args.set_cfgs is not None:
        cfg_from_list(args.set_cfgs)


    log_file = './logs/' + args.imdb_name + '-' + args.imdb_name_target + '/' + args.stdout_file  
    ReDirectSTD(log_file, 'stdout', False)
    

    print('Using config:')
    pprint.pprint(cfg)
    np.random.seed(cfg.RNG_SEED)

    # torch.backends.cudnn.benchmark = True
    if torch.cuda.is_available() and not args.cuda:
        print("WARNING: You have a CUDA device, so you should probably run with --cuda")

    # train set
    # -- Note: Use validation set and disable the flipped to enable faster loading.
    cfg.TRAIN.USE_FLIPPED = True
    cfg.USE_GPU_NMS = args.cuda
    # source dataset
    imdb, roidb, ratio_list, ratio_index = combined_roidb(args.imdb_name)
    train_size = len(roidb)
    # target dataset
    imdb_t, roidb_t, ratio_list_t, ratio_index_t = combined_roidb(args.imdb_name_target)
    train_size_t = len(roidb_t)

    print('{:d} source roidb entries'.format(len(roidb)))
    print('{:d} target roidb entries'.format(len(roidb_t)))

    # breakpoint()

    output_dir = args.save_dir + "/" + args.net + "/" + args.dataset
    if not os.path.exists(output_dir):
        os.makedirs(output_dir)

    sampler_batch = sampler(train_size, args.batch_size)
    sampler_batch_t = sampler(train_size_t, args.batch_size)

    dataset_s = roibatchLoader(roidb, ratio_list, ratio_index, args.batch_size, \
                               imdb.num_classes, training=True)

    # breakpoint()
    
    dataloader_s = torch.utils.data.DataLoader(dataset_s, batch_size=args.batch_size,
                                               sampler=sampler_batch, num_workers=args.num_workers)
    dataset_t = roibatchLoader(roidb_t, ratio_list_t, ratio_index_t, args.batch_size, \
                               imdb.num_classes, training=True)
    dataloader_t = torch.utils.data.DataLoader(dataset_t, batch_size=args.batch_size,
                                               sampler=sampler_batch_t, num_workers=args.num_workers)
    # initilize the tensor holder here.
    im_data = torch.FloatTensor(1)
    im_info = torch.FloatTensor(1)
    num_boxes = torch.LongTensor(1)
    gt_boxes = torch.FloatTensor(1)
    # ship to cuda
    if args.cuda:
        im_data = im_data.cuda()
        im_info = im_info.cuda()
        num_boxes = num_boxes.cuda()
        gt_boxes = gt_boxes.cuda()

    # make variable
    im_data = Variable(im_data)
    im_info = Variable(im_info)
    num_boxes = Variable(num_boxes)
    gt_boxes = Variable(gt_boxes)
    if args.cuda:
        cfg.CUDA = True

    # initilize the network here.
    # from model.faster_rcnn.vgg16_global_local import vgg16
    
    from model.faster_rcnn.resnet_global_local import resnet

    # from model.faster_rcnn.resnet_global_local_pseudo_label import resnet

    if args.net == 'vgg16':
        fasterRCNN = vgg16(imdb.classes, pretrained=True, class_agnostic=args.class_agnostic, lc=args.lc,
                           gc=args.gc)
    elif args.net == 'res101':
        fasterRCNN = resnet(imdb.classes, 101, pretrained=True, class_agnostic=args.class_agnostic,
                            lc=args.lc, gc=args.gc)
    elif args.net == 'res50':
        fasterRCNN = resnet(imdb.classes, 50, pretrained=True, class_agnostic=args.class_agnostic, context=args.context)

    else:
        print("network is not defined")
        pdb.set_trace()

    fasterRCNN.create_architecture()

    # breakpoint()

    lr = cfg.TRAIN.LEARNING_RATE
    lr = args.lr
    # tr_momentum = cfg.TRAIN.MOMENTUM
    # tr_momentum = args.momentum

    params = []
    for key, value in dict(fasterRCNN.named_parameters()).items():
        if value.requires_grad:
            if 'bias' in key:
                params += [{'params': [value], 'lr': lr * (cfg.TRAIN.DOUBLE_BIAS + 1), \
                            'weight_decay': cfg.TRAIN.BIAS_DECAY and cfg.TRAIN.WEIGHT_DECAY or 0}]
            else:
                params += [{'params': [value], 'lr': lr, 'weight_decay': cfg.TRAIN.WEIGHT_DECAY}]

    if args.optimizer == "adam":
        lr = lr * 0.1
        optimizer = torch.optim.Adam(params)

    elif args.optimizer == "sgd":
        optimizer = torch.optim.SGD(params, momentum=cfg.TRAIN.MOMENTUM)

    if args.cuda:
        fasterRCNN.cuda()

    if args.resume:
        checkpoint = torch.load(args.load_name)
        args.session = checkpoint['session']
        args.start_epoch = checkpoint['epoch']
        fasterRCNN.load_state_dict(checkpoint['model'])
        optimizer.load_state_dict(checkpoint['optimizer'])
        lr = optimizer.param_groups[0]['lr']
        if 'pooling_mode' in checkpoint.keys():
            cfg.POOLING_MODE = checkpoint['pooling_mode']
        print("loaded checkpoint %s" % (args.load_name))

    if args.mGPUs:
        fasterRCNN = nn.DataParallel(fasterRCNN)
    iters_per_epoch = int(10000 / args.batch_size)
    if args.ef:
        FL = EFocalLoss(class_num=2, gamma=args.gamma)
    else:
        FL = FocalLoss(class_num=2, gamma=args.gamma)

    if args.use_tfboard:
        from tensorboardX import SummaryWriter

        logger = SummaryWriter("logs")

    # test_model_while_training(fasterRCNN, args)

    count_iter = 0
    for epoch in range(args.start_epoch, args.max_epochs + 1):
        # setting to train mode
        fasterRCNN.train()
        loss_temp = 0
        start = time.time()
        if epoch % (args.lr_decay_step + 1) == 0:
            adjust_learning_rate(optimizer, args.lr_decay_gamma)
            lr *= args.lr_decay_gamma

        data_iter_s = iter(dataloader_s)
        data_iter_t = iter(dataloader_t)
        for step in range(iters_per_epoch):
            try:
                data_s = next(data_iter_s)
            except:
                data_iter_s = iter(dataloader_s)
                data_s = next(data_iter_s)
            try:
                data_t = next(data_iter_t)
            except:
                data_iter_t = iter(dataloader_t)
                data_t = next(data_iter_t)
            #eta = 1.0
            count_iter += 1
            #put source data into variable
            im_data.data.resize_(data_s[0].size()).copy_(data_s[0])
            im_info.data.resize_(data_s[1].size()).copy_(data_s[1])
            gt_boxes.data.resize_(data_s[2].size()).copy_(data_s[2])
            num_boxes.data.resize_(data_s[3].size()).copy_(data_s[3])

            fasterRCNN.zero_grad()
            rois, cls_prob, bbox_pred, \
            rpn_loss_cls, rpn_loss_box, \
            RCNN_loss_cls, RCNN_loss_bbox, \
            rois_label, out_d_pixel, out_d = fasterRCNN(im_data, im_info, gt_boxes, num_boxes)
            loss = rpn_loss_cls.mean() + rpn_loss_box.mean() \
                   + RCNN_loss_cls.mean() + RCNN_loss_bbox.mean()
            loss_temp += loss.item()

            
            # domain label
            domain_s = Variable(torch.zeros(out_d.size(0)).long().cuda())
            # global alignment loss
            dloss_s = 0.5 * FL(out_d, domain_s)
            # local alignment loss
            dloss_s_p = 0.5 * torch.mean(out_d_pixel ** 2)

            #put target data into variable
            im_data.data.resize_(data_t[0].size()).copy_(data_t[0])
            im_info.data.resize_(data_t[1].size()).copy_(data_t[1])
            #gt is empty
            gt_boxes.data.resize_(1, 1, 5).zero_()
            num_boxes.data.resize_(1).zero_()
            out_d_pixel, out_d = fasterRCNN(im_data, im_info, gt_boxes, num_boxes, target=True)
            # domain label
            domain_t = Variable(torch.ones(out_d.size(0)).long().cuda())
            dloss_t = 0.5 * FL(out_d, domain_t)
            # local alignment loss
            dloss_t_p = 0.5 * torch.mean((1 - out_d_pixel) ** 2)
            if args.dataset == 'sim10k':
                loss += (dloss_s + dloss_t + dloss_s_p + dloss_t_p) * args.eta
            else:
                loss += (dloss_s + dloss_t + dloss_s_p + dloss_t_p)
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()

            if step % args.disp_interval == 0:
                end = time.time()
                if step > 0:
                    loss_temp /= (args.disp_interval + 1)

                if args.mGPUs:
                    loss_rpn_cls = rpn_loss_cls.mean().item()
                    loss_rpn_box = rpn_loss_box.mean().item()
                    loss_rcnn_cls = RCNN_loss_cls.mean().item()
                    loss_rcnn_box = RCNN_loss_bbox.mean().item()
                    fg_cnt = torch.sum(rois_label.data.ne(0))
                    bg_cnt = rois_label.data.numel() - fg_cnt
                else:
                    loss_rpn_cls = rpn_loss_cls.item()
                    loss_rpn_box = rpn_loss_box.item()
                    loss_rcnn_cls = RCNN_loss_cls.item()
                    loss_rcnn_box = RCNN_loss_bbox.item()
                    dloss_s = dloss_s.item()
                    dloss_t = dloss_t.item()
                    dloss_s_p = dloss_s_p.item()
                    dloss_t_p = dloss_t_p.item()
                    fg_cnt = torch.sum(rois_label.data.ne(0))
                    bg_cnt = rois_label.data.numel() - fg_cnt

                print("[session %d][epoch %2d][iter %4d/%4d] loss: %.4f, lr: %.2e" \
                      % (args.session, epoch, step, iters_per_epoch, loss_temp, lr))
                print("\t\t\tfg/bg=(%d/%d), time cost: %f" % (fg_cnt, bg_cnt, end - start))
                print(
                    "\t\t\trpn_cls: %.4f, rpn_box: %.4f, rcnn_cls: %.4f, rcnn_box %.4f dloss s: %.4f dloss t: %.4f dloss s pixel: %.4f dloss t pixel: %.4f eta: %.4f" \
                    % (loss_rpn_cls, loss_rpn_box, loss_rcnn_cls, loss_rcnn_box, dloss_s, dloss_t, dloss_s_p, dloss_t_p,
                       args.eta))
                if args.use_tfboard:
                    info = {
                        'loss': loss_temp,
                        'loss_rpn_cls': loss_rpn_cls,
                        'loss_rpn_box': loss_rpn_box,
                        'loss_rcnn_cls': loss_rcnn_cls,
                        'loss_rcnn_box': loss_rcnn_box
                    }
                    logger.add_scalars("logs_s_{}/losses".format(args.session), info,
                                       (epoch - 1) * iters_per_epoch + step)

                loss_temp = 0
                start = time.time()
        save_name = os.path.join(output_dir,
                                 'globallocal_target_{}_eta_{}_local_context_{}_global_context_{}_gamma_{}_session_{}_epoch_{}_step_{}.pth'.format(
                                     args.dataset_t, args.eta,
                                     args.lc, args.gc, args.gamma,
                                     args.session, epoch,
                                     step))

        test_model_while_training(fasterRCNN, args)

        save_checkpoint({
            'session': args.session,
            'epoch': epoch + 1,
            'model': fasterRCNN.module.state_dict() if args.mGPUs else fasterRCNN.state_dict(),
            'optimizer': optimizer.state_dict(),
            'pooling_mode': cfg.POOLING_MODE,
            'class_agnostic': args.class_agnostic,
        }, save_name)
        print('save model: {}'.format(save_name))

    if args.use_tfboard:
        logger.close()