def train(model, train_loader, criterion, epoch, vis):
model.train()
batch_loss = 0
for batch_idx, sample_batched in enumerate(train_loader):
data = sample_batched['image']
target = sample_batched['mask']
data, target = Variable(data.type(opt.dtype)), Variable(
target.type(opt.dtype))
optimizer.zero_grad()
output = model(data)
loss = criterion(output, target)
loss.backward()
optimizer.step()
batch_loss += loss.data[0]
if (batch_idx + 1) % opt.plot_every == 0:
ori_img_ = inverse_normalize(at.tonumpy(data[0]))
target_ = at.tonumpy(target[0])
pred_ = at.tonumpy(output[0])
vis.img('gt_img', ori_img_)
vis.img('gt_mask', target_)
vis.img('pred_mask', (pred_ >= 0.5).astype(np.float32))
batch_loss /= (batch_idx + 1)
print('epoch: ' + str(epoch) + ', train loss: ' + str(batch_loss))
with open('logs.txt', 'a') as file:
file.write('epoch: ' + str(epoch) + ', train loss: ' +
str(batch_loss) + '\n')
vis.plot('train loss', batch_loss)
def eval(dataloader, faster_rcnn, vis, test_num=10000):
pred_bboxes, pred_labels, pred_scores = list(), list(), list()
gt_bboxes, gt_labels, gt_difficults = list(), list(), list()
for ii, (imgs, sizes, gt_bboxes_, gt_labels_,
gt_difficults_) in tqdm(enumerate(dataloader)):
# plot groud truth bboxes
sizes = [sizes[0][0].item(), sizes[1][0].item()]
pred_bboxes_, pred_labels_, pred_scores_ = faster_rcnn.predict(
imgs, [sizes])
img = imgs.cuda().float()
ori_img_ = inverse_normalize(at.tonumpy(img[0]))
pred_img = visdom_bbox(ori_img_, at.tonumpy(pred_bboxes_[0]),
at.tonumpy(pred_labels_[0]).reshape(-1),
at.tonumpy(pred_scores_[0]))
vis.img('test_pred_img', pred_img)
gt_bboxes += list(gt_bboxes_.numpy())
gt_labels += list(gt_labels_.numpy())
gt_difficults += list(gt_difficults_.numpy())
pred_bboxes += pred_bboxes_
pred_labels += pred_labels_
pred_scores += pred_scores_
if ii == test_num:
break
result = eval_detection_voc(pred_bboxes,
pred_labels,
pred_scores,
gt_bboxes,
gt_labels,
gt_difficults,
use_07_metric=True)
return result
def train(**kwargs):
opt._parse(kwargs)
dataset = Dataset(opt)
# img, bbox, label, scale = dataset[0]
# 返回的img是被scale后的图像，可能已经被随机翻转了
# 返回的 bbox 按照 ymin xmin ymax xmax 排列
# H, W = size(im)
# 对于一张屏幕上显示的图片，a,b,c,d 代表 4 个顶点
# a ... b ymin
# . .
# c ... d ymax H高度 y的范围在 [0, H-1] 间
# xmin xmax
# W宽度 x的范围在 [0, W-1] 间
print('load data')
dataloader = torch.utils.data.DataLoader(dataset, batch_size=1, shuffle=True, \
# pin_memory=True,
num_workers=opt.num_workers)
faster_rcnn = FasterRCNNVGG16()
print('model construct completed')
trainer = FasterRCNNTrainer(faster_rcnn)
if opt.load_path:
trainer.load(opt.load_path)
print('load pretrained model from %s' % opt.load_path)
for epoch in range(opt.epoch):
for ii, (img, bbox_, label_, scale) in (enumerate(dataloader)):
print('step: ', ii)
scale = at.scalar(scale)
img, bbox, label = img.float(), bbox_, label_
img, bbox, label = Variable(img), Variable(bbox), Variable(label)
trainer.train_step(img, bbox, label, scale)
if ((ii + 1) % opt.plot_every == 0) and (epoch > 50):
# 运行多少步以后再predict一次，epoch跑的太少的话根本预测不准什么东西
# if os.path.exists(opt.debug_file):
# ipdb.set_trace()
# plot groud truth bboxes 画出原本的框
ori_img_ = inverse_normalize(at.tonumpy(img[0]))
gt_img = visdom_bbox(ori_img_, at.tonumpy(bbox_[0]),
at.tonumpy(label_[0]))
# gt_img np类型 范围是 [0, 1] 间 3 x H x W
# 这里要将 gt_img 这个带框，带标注的图像保存或者显示出来
# plot predicti bboxes
_bboxes, _labels, _scores = trainer.faster_rcnn.predict(
[ori_img_], visualize=True)
pred_img = visdom_bbox(ori_img_, at.tonumpy(_bboxes[0]),
at.tonumpy(_labels[0]).reshape(-1),
at.tonumpy(_scores[0]))
def imgflip(img, bbox, x_flip=True, y_flip=True):
imgs = at.tonumpy(img[0])
if y_flip:
imgs = imgs[:, ::-1, :]
if x_flip:
imgs = imgs[:, :, ::-1]
# print imgs
imgs = np.expand_dims(imgs, axis=0)
return inverse_normalize(imgs)
def train(**kwargs):
opt._parse(kwargs)
dataset = Dataset(opt)
print('load data')
dataloader = data_.DataLoader(dataset, \
batch_size=1, \
shuffle=True, \
# pin_memory=True,
num_workers=opt.num_workers)
testset = TestDataset(opt)
test_dataloader = data_.DataLoader(testset,
batch_size=1,
num_workers=opt.test_num_workers,
shuffle=False, \
pin_memory=True
)
faster_rcnn = FasterRCNNVGG16()
print('model construct completed')
trainer = FasterRCNNTrainer(faster_rcnn).cuda()
if opt.load_path:
trainer.load(opt.load_path)
print('load pretrained model from %s' % opt.load_path)
best_map = 0
lr_ = opt.lr
for epoch in range(opt.epoch):
trainer.reset_meters()
for ii, (img, bbox_, label_, scale) in tqdm(enumerate(dataloader)):
scale = at.scalar(scale)
img, bbox, label = img.cuda().float(), bbox_.cuda(), label_.cuda()
img, bbox, label = Variable(img), Variable(bbox), Variable(label)
trainer.train_step(img, bbox, label, scale)
if (ii + 1) % opt.plot_every == 0:
if os.path.exists(opt.debug_file):
ipdb.set_trace()
# plot groud truth bboxes
ori_img_ = inverse_normalize(at.tonumpy(img[0]))
# ori_img_ = (at.tonumpy(img[0]))
losses = trainer.get_meter_data()
print(losses)
write_image(ori_img_, at.tonumpy(bbox[0]), 'gt.png')
_bboxes = trainer.faster_rcnn.predict([ori_img_],
visualize=True)
_bboxes = at.tonumpy(_bboxes[0])
# plot predicted bboxes
write_image(ori_img_, _bboxes, 'pred.png')
print('saved an image')
if epoch == 13:
break
def train(individual, **kwargs):
opt._parse(kwargs)
dataset = Dataset(opt)
print('load data')
dataloader = data_.DataLoader(dataset,
batch_size=1,
shuffle=True,
num_workers=opt.num_workers)
testset = TestDataset(opt)
test_dataloader = data_.DataLoader(testset,
batch_size=1,
num_workers=opt.test_num_workers,
shuffle=False,
pin_memory=True)
faster_rcnn = FasterRCNN_mine(individual)
print('model construct completed')
trainer = FasterRCNNTrainer(faster_rcnn).cuda()
best_map = 0
lr_ = opt.lr
for epoch in range(opt.epoch):
trainer.reset_meters()
for ii, (img, bbox_, label_, scale) in tqdm(enumerate(dataloader)):
scale = at.scalar(scale)
img, bbox, label = img.cuda().float(), bbox_.cuda(), label_.cuda()
trainer.train_step(img, bbox, label, scale)
if (ii + 1) % opt.plot_every == 0:
if os.path.exists(opt.debug_file):
ipdb.set_trace()
ori_img_ = inverse_normalize(at.tonumpy(img[0]))
_bboxes, _labels, _scores = trainer.faster_rcnn.predict(
[ori_img_], visualize=True)
eval_result = eval(test_dataloader, faster_rcnn, test_num=opt.test_num)
lr_ = trainer.faster_rcnn.optimizer.param_groups[0]['lr']
best_path = None
if eval_result['map'] > best_map:
best_map = eval_result['map']
best_path = trainer.save(best_map=best_map)
if epoch == 9:
trainer.load(best_path)
trainer.faster_rcnn.scale_lr(opt.lr_decay)
lr_ = lr_ * opt.lr_decay
individual.accuracy = best_map
def eval_mAP(trainer, val_loader):
tqdm.monitor_interval = 0
mAP = []
for ii, sample in tqdm(enumerate(val_loader)):
if len(sample.keys()) == 5:
img_id, img, bbox, scale, label = sample['img_id'], sample['image'], sample['bbox'], sample['scale'], \
sample['label']
img, bbox, label = img.cuda().float(), bbox.cuda(), label.cuda()
img, bbox, label = Variable(img), Variable(bbox), Variable(label)
else:
img_id, img, scale = sample['img_id'], sample['image'], sample[
'scale']
bbox = np.zeros((1, 0, 4))
label = np.zeros((1, 0, 1))
img = img.cuda().float()
img = Variable(img)
# if bbox is None:
# continue
scale = at.scalar(scale)
ori_img_ = inverse_normalize(at.tonumpy(img[0]))
pred_boxes, pred_labels, pred_scores = trainer.faster_rcnn.predict(
[ori_img_], visualize=True)
pred_boxes = pred_boxes[0]
pred_labels = pred_labels[0]
pred_scores = pred_scores[0]
bbox = at.tonumpy(bbox[0])
# Rescale back
C, H, W = ori_img_.shape
ori_img_ = transform.resize(ori_img_,
(C, H * (1 / scale), W * (1 / scale)),
mode='reflect')
o_H, o_W = H * (1 / scale), W * (1 / scale)
pred_boxes = resize_bbox(pred_boxes, (H, W), (o_H, o_W))
bbox = resize_bbox(bbox, (H, W), (o_H, o_W))
mAP.append(map_iou(bbox, pred_boxes, pred_scores))
# if ii>=100:
# break
mAP = np.array(mAP)
mAP = mAP[mAP != np.array(None)].astype(np.float32)
return np.mean(mAP)
def run_test(model, test_loader):
pred_masks = []
img_ids = []
images = []
for batch_idx, sample_batched in tqdm(enumerate(test_loader)):
data, img_id = sample_batched['image'], sample_batched['img_id']
data = Variable(data.type(opt.dtype), volatile=True)
output = model.forward(data)
# output = (output > 0.5)
output = at.tonumpy(output)
for i in range(0, output.shape[0]):
pred_mask = np.squeeze(output[i])
id = img_id[i]
pred_mask = (pred_mask >= 0.5).astype(np.float32)
pred_masks.append(pred_mask)
img_ids.append(id)
ori_img_ = inverse_normalize(at.tonumpy(data[i]))
images.append(ori_img_)
return img_ids, images, pred_masks
def input_visual(imgs, boxes, labels):
imgs = inverse_normalize(at.tonumpy(imgs.squeeze())) / 255
plt.figure(figsize=(8, 8))
plt.imshow(imgs.transpose(1, 2, 0))
input_boxes = boxes.reshape(-1, 4)
w = input_boxes[:, 3] - input_boxes[:, 1]
h = input_boxes[:, 2] - input_boxes[:, 0]
for i in range(input_boxes.shape[0]):
plt.gca().add_patch(
Rectangle(input_boxes[i][[1, 0]],
w[i],
h[i],
fill=False,
edgecolor='r'))
plt.text(
input_boxes[i][1], input_boxes[i][0],
dv.VOC_BBOX_LABEL_NAMES[labels.reshape(-1,
len(input_boxes))[0][i]])
plt.axis("off")
plt.show()
def train(**kwargs):
opt._parse(kwargs) #获得config设置信息
dataset = Dataset(opt) #传入opt，利用设置的数据集参数来创建训练数据集
print('load data')
dataloader = data_.DataLoader(dataset, \ #用创建的训练数据集创建训练DataLoader,代码仅支持batch_size=1
batch_size=1, \
shuffle=True, \
# pin_memory=True,
num_workers=opt.num_workers)
testset = TestDataset(opt) #传入opt，利用设置的数据集参数来加载测试数据集
test_dataloader = data_.DataLoader(testset, #用创建的测试数据集创建训练DataLoader,代码仅支持batch_size=1
batch_size=1,
num_workers=opt.test_num_workers,
shuffle=False, \
pin_memory=True
)
faster_rcnn = FasterRCNNVGG16() #创建以vgg为backbone的FasterRCNN网络
print('model construct completed')
trainer = FasterRCNNTrainer(faster_rcnn).cuda() #把创建好的FasterRCNN网络放入训练器
if opt.load_path: #若有FasterRCNN网络的预训练加载，则加载load_path权重
trainer.load(opt.load_path) #训练器加载权重
print('load pretrained model from %s' % opt.load_path)
trainer.vis.text(dataset.db.label_names, win='labels')
best_map = 0 #初始化best_map，训练时用于判断是否需要保存模型，类似打擂台后面用
lr_ = opt.lr #得到预设的学习率
for epoch in range(opt.epoch): #开始训练，训练次数为opt.epoch
trainer.reset_meters()
for ii, (img, bbox_, label_, scale) in tqdm(enumerate(dataloader)):
scale = at.scalar(scale) #进行类别处理得到scale（待定）
#bbox是gt_box坐标(ymin, xmin, ymax, xmax)
#label是类别的下标VOC_BBOX_LABEL_NAMES
#img是图片，代码仅支持batch_size=1的训练
img, bbox, label = img.cuda().float(), bbox_.cuda(), label_.cuda() #使用gpu训练
trainer.train_step(img, bbox, label, scale) #预处理完毕，进入模型
if (ii + 1) % opt.plot_every == 0: #可视化内容，（跳过）
if os.path.exists(opt.debug_file):
ipdb.set_trace()
# plot loss
trainer.vis.plot_many(trainer.get_meter_data())
# plot groud truth bboxes
ori_img_ = inverse_normalize(at.tonumpy(img[0]))
gt_img = visdom_bbox(ori_img_,
at.tonumpy(bbox_[0]),
at.tonumpy(label_[0]))
trainer.vis.img('gt_img', gt_img)
# plot predicti bboxes
_bboxes, _labels, _scores = trainer.faster_rcnn.predict([ori_img_], visualize=True)
pred_img = visdom_bbox(ori_img_,
at.tonumpy(_bboxes[0]),
at.tonumpy(_labels[0]).reshape(-1),
at.tonumpy(_scores[0]))
trainer.vis.img('pred_img', pred_img)
# rpn confusion matrix(meter)
trainer.vis.text(str(trainer.rpn_cm.value().tolist()), win='rpn_cm')
# roi confusion matrix
trainer.vis.img('roi_cm', at.totensor(trainer.roi_cm.conf, False).float())
eval_result = eval(test_dataloader, faster_rcnn, test_num=opt.test_num) #训练一个epoch评估一次
trainer.vis.plot('test_map', eval_result['map']) #可视化内容，（跳过）
lr_ = trainer.faster_rcnn.optimizer.param_groups[0]['lr'] #获得当前的学习率
log_info = 'lr:{}, map:{},loss:{}'.format(str(lr_), #日志输出学习率，map，loss
str(eval_result['map']),
str(trainer.get_meter_data()))
trainer.vis.log(log_info) #可视化内容，（跳过）
if eval_result['map'] > best_map: #若这次评估的map大于之前最大的map则保存模型
best_map = eval_result['map'] #保存模型的map信息
best_path = trainer.save(best_map=best_map) #调用保存模型函数
if epoch == 9: #若训练到第9个epoch则加载之前最好的模型并且减低学习率继续训练
trainer.load(best_path) #加载模型
trainer.faster_rcnn.scale_lr(opt.lr_decay) #降低学习率
lr_ = lr_ * opt.lr_decay #获得当前学习率
if epoch == 13: #13个epoch停止训练
break
im_path_clone = str(img_id)
# save_path = _data.save_dir + 'f0rcnn' + im_path.split('/')[-1]
save_path = _data.save_dir + im_path_clone.split('/')[-1] + '.jpg'
save_path_adv = _data.save_dir_adv + im_path_clone.split(
'/')[-1] + '.jpg'
save_path_perturb = _data.save_dir_perturb + 'frcnn_perturb_' + im_path_clone.split(
'/')[-1] + '.jpg'
if not os.path.exists(_data.save_dir):
os.makedirs(_data.save_dir)
if not os.path.exists(_data.save_dir_adv):
os.makedirs(_data.save_dir_adv)
if not os.path.exists(_data.save_dir_perturb):
os.makedirs(_data.save_dir_perturb)
ori_img_ = inverse_normalize(at.tonumpy(img[0]))
# _bboxes, _labels, _scores = trainer.faster_rcnn.predict([ori_img_],\
# new_score=quality, visualize=True)
before = time.time()
adv_img, perturb, distance = trainer.attacker.perturb_mD_img(
img,
save_perturb=save_path_perturb,
rois=rois,
roi_scores=roi_scores)
after = time.time()
generate_time = after - before
total_time = total_time + generate_time
total_distance = total_distance + distance
adv_img_ = inverse_normalize(at.tonumpy(adv_img[0]))
perturb_ = inverse_normalize(at.tonumpy(perturb[0]))
del adv_img, perturb, img
def train_val():
print('load data')
train_loader, val_loader = get_train_val_loader(
opt.root_dir,
batch_size=opt.batch_size,
val_ratio=0.1,
shuffle=opt.shuffle,
num_workers=opt.num_workers,
pin_memory=opt.pin_memory)
faster_rcnn = FasterRCNNVGG16()
# faster_rcnn = FasterRCNNResNet50()
print('model construct completed')
trainer = FasterRCNNTrainer(faster_rcnn).cuda()
# if opt.load_path:
# trainer.load(opt.load_path)
# print('load pretrained model from %s' % opt.load_path)
# trainer.vis.text(dataset.db.label_names, win='labels')
best_map = 0
lr_ = opt.lr
for epoch in range(opt.epoch):
trainer.reset_meters()
tqdm.monitor_interval = 0
for ii, sample in tqdm(enumerate(train_loader)):
if len(sample.keys()) == 5:
img_id, img, bbox, scale, label = sample['img_id'], sample['image'], sample['bbox'], sample['scale'], \
sample['label']
img, bbox, label = img.cuda().float(), bbox.cuda(), label.cuda(
)
img, bbox, label = Variable(img), Variable(bbox), Variable(
label)
else:
img_id, img, bbox, scale, label = sample['img_id'], sample['image'], np.zeros((1, 0, 4)), \
sample['scale'], np.zeros((1, 0, 1))
img = img.cuda().float()
img = Variable(img)
if bbox.size == 0:
continue
scale = at.scalar(scale)
trainer.train_step(img_id, img, bbox, label, scale)
if (ii + 1) % opt.plot_every == 0:
# plot loss
trainer.vis.plot_many(trainer.get_meter_data())
# plot ground truth bboxes
ori_img_ = inverse_normalize(at.tonumpy(img[0]))
gt_img = visdom_bbox(ori_img_, img_id[0], at.tonumpy(bbox[0]),
at.tonumpy(label[0]))
trainer.vis.img('gt_img', gt_img)
# plot predicted bboxes
_bboxes, _labels, _scores = trainer.faster_rcnn.predict(
[ori_img_], visualize=True)
pred_img = visdom_bbox(ori_img_, img_id[0],
at.tonumpy(_bboxes[0]),
at.tonumpy(_labels[0]).reshape(-1),
at.tonumpy(_scores[0]))
trainer.vis.img('pred_img', pred_img)
# rpn confusion matrix(meter)
trainer.vis.text(str(trainer.rpn_cm.value().tolist()),
win='rpn_cm')
# roi confusion matrix
trainer.vis.img(
'roi_cm',
at.totensor(trainer.roi_cm.conf, False).float())
mAP = eval_mAP(trainer, val_loader)
trainer.vis.plot('val_mAP', mAP)
lr_ = trainer.faster_rcnn.optimizer.param_groups[0]['lr']
log_info = 'lr:{}, map:{},loss:{}'.format(
str(lr_), str(mAP), str(trainer.get_meter_data()))
trainer.vis.log(log_info)
if mAP > best_map:
best_map = mAP
best_path = trainer.save(best_map=best_map)
if epoch == opt.epoch - 1:
best_path = trainer.save()
if (epoch + 1) % 10 == 0:
trainer.load(best_path)
trainer.faster_rcnn.scale_lr(opt.lr_decay)
lr_ = lr_ * opt.lr_decay
def train(**kwargs):
# opt._parse(kwargs)
print('load data')
dataloader = get_train_loader(opt.root_dir,
batch_size=opt.batch_size,
shuffle=opt.shuffle,
num_workers=opt.num_workers,
pin_memory=opt.pin_memory)
faster_rcnn = FasterRCNNVGG16()
print('model construct completed')
trainer = FasterRCNNTrainer(faster_rcnn).cuda()
# if opt.load_path:
# trainer.load(opt.load_path)
# print('load pretrained model from %s' % opt.load_path)
# trainer.vis.text(dataset.db.label_names, win='labels')
best_map = 0
lr_ = opt.lr
for epoch in range(opt.epoch):
trainer.reset_meters()
for ii, sample in tqdm(enumerate(dataloader)):
if len(sample.keys()) == 5:
img_id, img, bbox_, scale, label_ = sample['img_id'], sample['image'], sample['bbox'], sample['scale'], \
sample['label']
img, bbox, label = img.cuda().float(), bbox_.cuda(
), label_.cuda()
img, bbox, label = Variable(img), Variable(bbox), Variable(
label)
else:
img_id, img, bbox, scale, label = sample['img_id'], sample['image'], np.zeros((1, 0, 4)), \
sample['scale'], np.zeros((1, 0, 1))
img = img.cuda().float()
img = Variable(img)
# if label.size == 0:
# continue
scale = at.scalar(scale)
trainer.train_step(img_id, img, bbox, label, scale)
if (ii + 1) % opt.plot_every == 0:
if os.path.exists(opt.debug_file):
ipdb.set_trace()
# plot loss
trainer.vis.plot_many(trainer.get_meter_data())
# plot ground truth bboxes
ori_img_ = inverse_normalize(at.tonumpy(img[0]))
gt_img = visdom_bbox(ori_img_, at.tonumpy(bbox_[0]),
at.tonumpy(label_[0]))
trainer.vis.img('gt_img', gt_img)
# plot predicted bboxes
_bboxes, _labels, _scores = trainer.faster_rcnn.predict(
[ori_img_], visualize=True)
pred_img = visdom_bbox(ori_img_, at.tonumpy(_bboxes[0]),
at.tonumpy(_labels[0]).reshape(-1),
at.tonumpy(_scores[0]))
trainer.vis.img('pred_img', pred_img)
# rpn confusion matrix(meter)
trainer.vis.text(str(trainer.rpn_cm.value().tolist()),
win='rpn_cm')
# roi confusion matrix
trainer.vis.img(
'roi_cm',
at.totensor(trainer.roi_cm.conf, False).float())
if epoch % 10 == 0:
best_path = trainer.save(best_map=best_map)
def cutmix_generate(img,
scale,
paste_scale,
paste_img,
attention_map,
bboxes,
labels,
paste_bboxes,
paste_labels,
info,
bbox_drop=True,
rescale=True,
device='cuda',
threshold=0.7,
overlap_threshold=0.7):
src_img = inverse_normalize(at.tonumpy(img.squeeze()))
src_img = torch.from_numpy(src_img).to(device)
target_img = inverse_normalize(at.tonumpy(paste_img.squeeze()))
target_img = torch.from_numpy(target_img.squeeze().copy()).to(device)
target_img_copy = target_img.clone()
src_size = src_img.shape
target_size = target_img.shape
img_size = [
target_size[1] if target_size[1] > src_size[1] else src_size[1],
target_size[2] if target_size[2] > src_size[2] else src_size[2]
]
if target_size[1] >= img_size[1] and target_size[2] >= src_size[2]:
new_scale = paste_scale
else:
new_scale = scale
attention_map = torch.nn.functional.interpolate(
attention_map.unsqueeze(0).unsqueeze(0), img_size).squeeze()
mask = torch.zeros_like(attention_map).to(device)
# print(attention_map.max(), attention_map.min(), attention_map.mean())
mean_threshold = attention_map.mean()
mask[attention_map > mean_threshold] = 1
mask[attention_map <= mean_threshold] = 0 # mask表示重要的区域
# print(torch.sum(mask==1)/mask.shape[0]/mask.shape[1])
# if torch.sum(mask)/mask.shape[0]/mask.shape[1] > 0.9:
# info["use_cutmix"] = 0
# res_img = pytorch_normalze(target_img_copy/255).unsqueeze(dim=0)
# return res_img, torch.from_numpy(paste_bboxes), torch.from_numpy(paste_labels), paste_scale, info
mask = mask.int()
x_scale = img_size[0] / src_size[1]
y_scale = img_size[1] / src_size[2]
x_scale_p = img_size[0] / target_size[1]
y_scale_p = img_size[1] / target_size[2]
h, w = img_size
src_img = torch.nn.functional.interpolate(src_img.unsqueeze(0),
img_size).squeeze()
target_img = torch.nn.functional.interpolate(target_img.unsqueeze(0),
img_size).squeeze()
if len(bboxes.shape) == 2 and bboxes.shape[1] == 4:
pass
else:
bboxes = bboxes.reshape([-1, 4])
bboxes = bboxes.int()
labels = labels.reshape(-1, )
ori_mask = torch.zeros(src_img.shape[1:]).to(device)
new_bboxes = []
new_labels = []
paste_bboxes_cp = paste_bboxes.copy()
bboxes_cp = bboxes.clone().detach()
res_img = target_img.clone()
b_h = (bboxes[:, 3] - bboxes[:, 1]).unsqueeze(dim=1)
b_w = (bboxes[:, 2] - bboxes[:, 0]).unsqueeze(dim=1)
b = torch.cat([b_h, b_w], axis=1)
areas = torch.prod(b, 1)
_, index = torch.sort(areas)
for i in index:
bboxes_cp[i, [0, 2]] = (bboxes[i, [0, 2]] * x_scale).int()
bboxes_cp[i, [1, 3]] = (bboxes[i, [1, 3]] * y_scale).int()
ymin, xmin, ymax, xmax = bboxes_cp[i].int()
x0 = ymin.item()
x1 = ymax.item()
y0 = xmin.item()
y1 = xmax.item()
ori_mask[x0:x1, y0:y1] = 1
ori_mask = ori_mask.int()
area = (x1 - x0) * (y1 - y0)
# print(area/w/h)
portion = torch.sum(mask[x0:x1, y0:y1] & ori_mask[x0:x1, y0:y1]) / area
if bbox_drop:
remain_flag = True if np.random.rand(1) > 0.2 else False
else:
remain_flag = True
if rescale:
rescale_flag = True if np.random.rand(1) > 0.0 else False
else:
rescale_flag = False
if rescale_flag:
if np.random.randn(1) >= 0.5:
rescale_conf = np.random.randint(low=10, high=1000,
size=1)[0] * 0.0001 + 1
else:
rescale_conf = np.random.rand(1)[0]
# print(rescale_conf)
if portion < threshold or not remain_flag or area / h / w > 0.8: # 不考虑占比大的bbox:
mask[x0:x1, y0:y1] = 0
for j in range(len(bboxes_cp)):
x = np.maximum(x0, bboxes_cp[j][0].detach().cpu().numpy())
y = np.maximum(y0, bboxes_cp[j][1].detach().cpu().numpy())
x = np.minimum(x1, bboxes_cp[j][2].detach().cpu().numpy())
y = np.minimum(y1, bboxes_cp[j][3].detach().cpu().numpy())
mask[x:x, y:y] = 1
else:
index_y = torch.where(
mask[ymin:ymax, xmin:xmax] == 1)[0].unique().sort()[0] + ymin
index_x = torch.where(mask[ymin:ymax, xmin:xmax] == 1)[1].unique(
).sort()[0] + xmin # bboxes尽量取得靠近图像边缘
mask_m = (mask[ymin:ymax, xmin:xmax]
& ori_mask[ymin:ymax, xmin:xmax]).bool()
mask_m = torch.stack([mask_m, mask_m, mask_m], dim=0)
sub_obj = src_img[:, ymin:ymax, xmin:xmax] # 取出来的目标
if len(index_y) < 2 or len(index_x) < 2:
print('!', index_y, type(index_y))
if rescale_flag and 0 < ymin * rescale_conf < h and 0 < ymax * rescale_conf < h and 0 < xmin * rescale_conf < w and 0 < xmax * rescale_conf < w and (
ymax - ymin) * rescale_conf * (xmax -
xmin) * rescale_conf > 400:
index_y = (index_y * rescale_conf).int()
index_x = (index_x * rescale_conf).int()
bbox_w = xmax.item() - xmin.item()
bbox_h = ymax.item() - ymin.item()
# print("before", ymin.item(), ymax.item(), xmin.item(), xmax.item())
ymin = int(ymin * rescale_conf)
ymax = int(ymax * rescale_conf)
xmin = int(xmin * rescale_conf)
xmax = int(xmax * rescale_conf)
# print("enlarge", ymin, ymax, xmin, xmax)
# print(bbox_h * rescale_conf, bbox_w *rescale_conf)
resize_f = tvtsf.Resize(
[int(bbox_h * rescale_conf),
int(bbox_w * rescale_conf)])
r_mask_m = resize_f(mask_m).bool()
# print(r_mask_m.shape, res_img[:, ymin:ymin+int(bbox_h*rescale_conf), xmin:xmin+int(bbox_w*rescale_conf)].shape)
res_img[:, ymin:ymin + int(bbox_h * rescale_conf), xmin:xmin +
int(bbox_w * rescale_conf)][r_mask_m] = resize_f(
sub_obj)[r_mask_m]
else:
# print(ymax*rescale_conf, ymin*rescale_conf, h)
# print(xmax*rescale_conf, xmin*rescale_conf, w)
res_img[:, ymin:ymax, xmin:xmax][mask_m] = sub_obj[mask_m]
new_bboxes.append([
index_y[0].item(), index_x[0].item(), index_y[-1].item(),
index_x[-1].item()
])
new_labels.append(labels[i])
new_bboxes_cp = np.concatenate([new_bboxes])
mask = mask & ori_mask
for i in range(paste_bboxes.shape[0]):
paste_bboxes_cp[i, [0, 2]] = (paste_bboxes[i, [0, 2]] *
x_scale_p).astype(np.int)
paste_bboxes_cp[i, [1, 3]] = (paste_bboxes[i, [1, 3]] *
y_scale_p).astype(np.int)
if len(new_bboxes_cp) > 0:
ymin, xmin, ymax, xmax = paste_bboxes_cp[i]
area = (ymax - ymin) * (xmax - xmin)
overlap = torch.sum(mask[int(ymin):int(ymax),
int(xmin):int(xmax)] == 1) / area
# print(overlap)
# lt = np.maximum(paste_bboxes_cp[i:i+1:,:2], new_bboxes_cp[:,:2])
# rb = np.minimum(paste_bboxes_cp[i:i+1:,2:], new_bboxes_cp[:,2:])
# overlap = np.prod(rb - lt, axis=1) / area
if (overlap > overlap_threshold).any():
continue
# if np.sum(mask[ymix:ymax, xmax:xmax] == 1) / area > 0.6 :
# continue
new_bboxes.append(paste_bboxes_cp[i])
new_labels.append(paste_labels[i])
if len(new_bboxes) == 0:
new_bboxes = None
else:
new_bboxes = torch.from_numpy(np.concatenate(
[new_bboxes])).float().to(device).unsqueeze(dim=0)
new_labels = torch.Tensor(new_labels).int().to(device).unsqueeze(dim=0)
mask = mask.int()
new_mask = mask
# res_img = src_img.mul(new_mask) + target_img.mul(1 - new_mask)
area = torch.sum(new_mask == 1).float().item() / (
torch.sum(ori_mask == 1) + 1e-8).float().item()
# compute iou
res_img = res_img / 255
res_img = res_img.float()
# if area < 0.6 or len(new_bboxes)==0:
# print(area)
if len(new_bboxes) == 0:
info["use_cutmix"] = 0
res_img = pytorch_normalze(target_img_copy / 255).unsqueeze(dim=0)
return res_img, torch.from_numpy(paste_bboxes), torch.from_numpy(
paste_labels), paste_scale, info
else:
info["use_cutmix"] = 1
res_img = pytorch_normalze(res_img).unsqueeze(dim=0)
return res_img, new_bboxes, new_labels, new_scale, info
def train(opt, faster_rcnn, dataloader, val_dataloader,
test_dataloader, trainer, lr_, best_map, start_epoch):
trainer.train()
for epoch in range(start_epoch, start_epoch+opt.epoch):
trainer.reset_meters()
pbar = tqdm(enumerate(dataloader), total=len(dataloader))
for ii, (img, bbox_, label_, scale) in pbar:
# Currently configured to predict (y_min, x_min, y_max, x_max)
# bbox_tmp = bbox_.clone()
# bbox_ = transform_bbox(bbox_)
scale = at.scalar(scale)
img, bbox, label = img.cuda().float(), bbox_.cuda(), label_.cuda()
losses = trainer.train_step(img, bbox, label, scale)
if ii % 100 == 0:
rpnloc = losses[0].cpu().data.numpy()
rpncls = losses[1].cpu().data.numpy()
roiloc = losses[2].cpu().data.numpy()
roicls = losses[3].cpu().data.numpy()
tot = losses[4].cpu().data.numpy()
pbar.set_description(f"Epoch: {epoch} | Batch: {ii} | RPNLoc Loss: {rpnloc:.4f} | RPNclc Loss: {rpncls:.4f} | ROIloc Loss: {roiloc:.4f} | ROIclc Loss: {roicls:.4f} | Total Loss: {tot:.4f}")
if (ii+1) % 1000 == 0:
eval_result = eval(val_dataloader, faster_rcnn, test_num=1000)
trainer.vis.plot('val_map', eval_result['map'])
lr_ = trainer.faster_rcnn.optimizer.param_groups[0]['lr']
val_log_info = 'lr:{}, map:{},loss:{}'.format(str(lr_),
str(eval_result['map']),
str(trainer.get_meter_data()))
trainer.vis.log(val_log_info)
print("Evaluation Results on Val Set ")
print(val_log_info)
print("\n\n")
if (ii + 1) % 100 == 0:
if os.path.exists(opt.debug_file):
ipdb.set_trace()
# plot loss
trainer.vis.plot_many(trainer.get_meter_data())
print(trainer.get_meter_data())
try:
ori_img_ = inverse_normalize(at.tonumpy(img[0]))
gt_img = visdom_bbox(ori_img_,
at.tonumpy(bbox_[0]),
at.tonumpy(label_[0]))
trainer.vis.img('gt_img', gt_img)
plt.show()
# plot predicti bboxes
_bboxes, _labels, _scores = trainer.faster_rcnn.predict([ori_img_], visualize=True)
pred_img = visdom_bbox(ori_img_,
at.tonumpy(_bboxes[0]),
at.tonumpy(_labels[0]).reshape(-1),
at.tonumpy(_scores[0]))
plt.show()
trainer.vis.img('pred_img', pred_img)
# rpn confusion matrix(meter)
trainer.vis.text(str(trainer.rpn_cm.value().tolist()),
win='rpn_cm')
# roi confusion matrix
trainer.vis.img('roi_cm', at.totensor(trainer.roi_cm.conf,
False).float())
except:
print("Cannot display images")
if (ii + 1) % 100 == 0:
eval_result = eval(val_dataloader, faster_rcnn, test_num=25)
trainer.vis.plot('val_map', eval_result['map'])
log_info = 'lr:{}, map:{},loss:{}'.format(str(lr_), str(
eval_result['map']), str(trainer.get_meter_data()))
trainer.vis.log(log_info)
# Save after every epoch
epoch_path = trainer.save(epoch, best_map=0)
eval_result = eval(test_dataloader, faster_rcnn, test_num=1000)
trainer.vis.plot('test_map', eval_result['map'])
lr_ = trainer.faster_rcnn.optimizer.param_groups[0]['lr']
test_log_info = 'lr:{}, map:{},loss:{}'.format(str(lr_),
str(eval_result['map']),
str(trainer.get_meter_data()))
trainer.vis.log(test_log_info)
print("Evaluation Results on Test Set ")
print(test_log_info)
print("\n\n")
if eval_result['map'] > best_map:
best_map = eval_result['map']
best_path = epoch_path
if epoch == 9:
trainer.load(best_path)
trainer.faster_rcnn.scale_lr(opt.lr_decay)
lr_ = lr_ * opt.lr_decay
if epoch == 13:
break
def train():
# Get the dataset
for phase in phases:
if phase == 'train':
if dataset_name == 'hollywood':
train_data_list_path = os.path.join(
opt.hollywood_dataset_root_path, 'hollywood_train.idl')
train_data_list = utils.get_phase_data_list(
train_data_list_path, dataset_name)
if dataset_name == 'brainwash':
train_data_list_path = os.path.join(
opt.brainwash_dataset_root_path, 'brainwash_train.idl')
train_data_list = utils.get_phase_data_list(
train_data_list_path, dataset_name)
elif phase == 'val':
if dataset_name == 'hollywood':
val_data_list_path = os.path.join(
opt.hollywood_dataset_root_path, 'hollywood_val.idl')
val_data_list = utils.get_phase_data_list(
val_data_list_path, dataset_name)
if dataset_name == 'brainwash':
val_data_list_path = os.path.join(
opt.brainwash_dataset_root_path, 'brainwash_val.idl')
val_data_list = utils.get_phase_data_list(
val_data_list_path, dataset_name)
elif phase == 'test':
if dataset_name == 'hollywood':
test_data_list_path = os.path.join(
opt.hollywood_dataset_root_path, 'hollywood_test.idl')
test_data_list = utils.get_phase_data_list(
test_data_list_path, dataset_name)
if dataset_name == 'brainwash':
test_data_list_path = os.path.join(
opt.brainwash_dataset_root_path, 'brainwash_test.idl')
test_data_list = utils.get_phase_data_list(
test_data_list_path, dataset_name)
print("Number of images for training: %s" % (len(train_data_list)))
print("Number of images for val: %s" % (len(val_data_list)))
print("Number of images for test: %s" % (len(test_data_list)))
if data_check_flag:
utils.check_loaded_data(train_data_list[random.randint(
1, len(train_data_list))])
utils.check_loaded_data(val_data_list[random.randint(
1, len(val_data_list))])
utils.check_loaded_data(test_data_list[random.randint(
1, len(test_data_list))])
# Load the train dataset
train_dataset = Dataset(train_data_list)
test_dataset = Dataset(val_data_list)
print("Load data.")
train_dataloader = data_.DataLoader(train_dataset,
batch_size=1,
shuffle=True,
num_workers=1)
test_dataloader = data_.DataLoader(test_dataset,
batch_size=1,
shuffle=True,
num_workers=1)
# Initialize the head detector.
head_detector_vgg16 = Head_Detector_VGG16(ratios=[1],
anchor_scales=[8, 16])
print("model construct completed")
trainer = Head_Detector_Trainer(head_detector_vgg16).cuda()
lr_ = opt.lr
for epoch in range(opt.epoch):
trainer.reset_meters()
for ii, (img, bbox_, scale) in enumerate(train_dataloader):
scale = at.scalar(scale)
img, bbox = img.cuda().float(), bbox_.cuda()
img, bbox = Variable(img), Variable(bbox)
_, _, _ = trainer.train_step(img, bbox, scale)
print("Forward and backward pass done.")
if (ii + 1) % opt.plot_every == 0:
trainer.vis.plot_many(trainer.get_meter_data())
ori_img_ = inverse_normalize(at.tonumpy(img[0]))
gt_img = visdom_bbox(ori_img_, at.tonumpy(bbox_[0]))
trainer.vis.img('gt_img', gt_img)
rois, _ = trainer.head_detector.predict(img,
scale=scale,
mode='visualize')
pred_img = visdom_bbox(ori_img_, at.tonumpy(rois))
trainer.vis.img('pred_img', pred_img)
trainer.vis.text(str(trainer.rpn_cm.value().tolist()),
win='rpn_cm')
avg_test_CorrLoc = eval(test_dataloader, head_detector_vgg16)
print("Epoch {} of {}.".format(epoch + 1, opt.epoch))
print(" test average corrLoc accuracy:\t\t{:.3f}".format(
avg_test_CorrLoc))
model_save_path = trainer.save(best_map=avg_test_CorrLoc)
if epoch == 8:
trainer.load(model_save_path)
trainer.head_detector.scale_lr(opt.lr_decay)
lr_ = lr_ * opt.lr_decay
def train(**kwargs):
opt._parse(
kwargs
) #将调用函数时候附加的参数用，config.py文件里面的opt._parse()进行解释，然后获取其数据存储的路径，之后放到Dataset里面！
dataset = Dataset(opt)
print('load data')
dataloader = data_.DataLoader(dataset,
batch_size=1,
shuffle=True,
num_workers=opt.num_workers)
# testset = TestDataset(opt)
# test_dataloader = data_.DataLoader(testset,
# batch_size=1,
# num_workers=opt.test_num_workers,
# shuffle=False,
# #pin_memory=True
# ) #pin_memory锁页内存,开启时使用显卡的内存，速度更快
faster_rcnn = FasterRCNN(7)
print('model construct completed')
pdb.set_trace()
trainer = Trainer(faster_rcnn).cuda()
#判断opt.load_path是否存在，如果存在，直接从opt.load_path读取预训练模型，然后将训练数据的label进行可视化操作
if opt.load_path:
trainer.load(opt.load_path)
print('load pretrained model from %s' % opt.load_path)
best_map = 0
lr_ = opt.lr
# 之后用一个for循环开始训练过程，而训练迭代的次数opt.epoch=14也在config.py文件中都预先定义好，属于超参数
for epoch in range(opt.epoch):
print('epoch {}/{}'.format(epoch, opt.epoch))
trainer.reset_meters() #首先在可视化界面重设所有数据
for ii, (img, bbox_, label_, scale) in tqdm(enumerate(dataloader)):
scale = scalar(scale)
img, bbox, label = img.cuda().float(), bbox_.cuda(), label_.cuda()
trainer.train_step(img, bbox, label, scale)
if (ii + 1) % opt.plot_every == 0:
if os.path.exists(opt.debug_file):
ipdb.set_trace()
#可视化画出loss
#可视化画出groudtruth bboxes
ori_img_ = inverse_normalize(array_tool.tonumpy(img[0]))
#可视化画出预测bboxes
# 调用faster_rcnn的predict函数进行预测，预测的结果保留在以_下划线开头的对象里面
_bboxes, _labels, _scores = trainer.faster_rcnn.predict(
[ori_img_], visualize=False)
#eval_result = eval(test_dataloader, faster_rcnn, test_num=opt.test_num)
lr_ = trainer.faster_rcnn.optimizer.param_groups[0]['lr']
log_info = 'lr:{}, map:{}, loss:{}'.format(
str(lr_), str(eval_result['map']), str(trainer.get_meter_data()))
if eval_result['map'] > best_map:
best_map = eval_result['map']
best_path = trainer.save(best_map=best_map)
if epoch == 9: #if判断语句如果学习的epoch达到了9就将学习率*0.1变成原来的十分之一
trainer.load(best_path)
trainer.faster_rcnn.scale_lr(opt.lr_decay)
lr_ = lr_ * opt.lr_decay
if epoch == 13:
break
def train(**kwargs):
opt._parse(kwargs)
dataset = Dataset(opt)
print('load data')
dataloader = data_.DataLoader(dataset, \
batch_size=1, \
shuffle=True, \
# pin_memory=True,
num_workers=opt.num_workers)
testset = TestDataset(opt)
test_dataloader = data_.DataLoader(testset,
batch_size=1,
num_workers=opt.test_num_workers,
shuffle=False, \
pin_memory=True
)
faster_rcnn = FasterRCNNVGG16()
print('model construct completed')
trainer = FasterRCNNTrainer(faster_rcnn).cuda()
if opt.load_path:
trainer.load(opt.load_path)
print('load pretrained model from %s' % opt.load_path)
trainer.vis.text(dataset.db.label_names, win='labels')
best_map = 0
lr_ = opt.lr
for epoch in range(opt.epoch):
trainer.reset_meters()
for ii, (img, bbox_, label_, scale) in tqdm(enumerate(dataloader)):
scale = at.scalar(scale)
img, bbox, label = img.cuda().float(), bbox_.cuda(), label_.cuda()
trainer.train_step(img, bbox, label, scale)
if (ii + 1) % opt.plot_every == 0:
if os.path.exists(opt.debug_file):
ipdb.set_trace()
# plot loss
trainer.vis.plot_many(trainer.get_meter_data())
# plot groud truth bboxes
ori_img_ = inverse_normalize(at.tonumpy(img[0]))
gt_img = visdom_bbox(ori_img_, at.tonumpy(bbox_[0]),
at.tonumpy(label_[0]))
trainer.vis.img('gt_img', gt_img)
# plot predicti bboxes
_bboxes, _labels, _scores = trainer.faster_rcnn.predict(
[ori_img_], visualize=True)
pred_img = visdom_bbox(ori_img_, at.tonumpy(_bboxes[0]),
at.tonumpy(_labels[0]).reshape(-1),
at.tonumpy(_scores[0]))
trainer.vis.img('pred_img', pred_img)
# rpn confusion matrix(meter)
trainer.vis.text(str(trainer.rpn_cm.value().tolist()),
win='rpn_cm')
# roi confusion matrix
trainer.vis.img(
'roi_cm',
at.totensor(trainer.roi_cm.conf, False).float())
eval_result = eval(test_dataloader, faster_rcnn, test_num=opt.test_num)
trainer.vis.plot('test_map', eval_result['map'])
lr_ = trainer.faster_rcnn.optimizer.param_groups[0]['lr']
log_info = 'lr:{}, map:{},loss:{}'.format(
str(lr_), str(eval_result['map']), str(trainer.get_meter_data()))
trainer.vis.log(log_info)
if eval_result['map'] > best_map:
best_map = eval_result['map']
best_path = trainer.save(best_map=best_map)
if epoch == 9:
trainer.load(best_path)
trainer.faster_rcnn.scale_lr(opt.lr_decay)
lr_ = lr_ * opt.lr_decay
if epoch == 13:
break
def train(**kwargs):
"""
训练
"""
#解析命令行参数，设置配置文件参数
opt._parse(kwargs)
#初始化Dataset参数
dataset = Dataset(opt)
print('load data')
#data_ 数据加载器（被重命名，pytorch方法）
dataloader = data_.DataLoader(dataset, \
batch_size=1, \
shuffle=True, \
# pin_memory=True,
num_workers=opt.num_workers)
#初始化TestDataset参数
testset = TestDataset(opt)
test_dataloader = data_.DataLoader(testset,
batch_size=1,
num_workers=opt.test_num_workers,
shuffle=False, \
pin_memory=True
)
#新建一个FasterRCNNVGG16
faster_rcnn = FasterRCNNVGG16()
print('model construct completed')
#新建一个trainer，并将网络模型转移到GPU上
#将FasterRCNNVGG16模型传入
trainer = FasterRCNNTrainer(faster_rcnn).cuda()
#如果存在，加载训练好的模型
if opt.load_path:
trainer.load(opt.load_path)
print('load pretrained model from %s' % opt.load_path)
#可视化类别 vis为visdom加载器
trainer.vis.text(dataset.db.label_names, win='labels')
#best_map存放的是 最优的mAP的网络参数
best_map = 0
lr_ = opt.lr
for epoch in range(opt.epoch):
#trainer方法 将平均精度的元组 和 混淆矩阵的值置0
trainer.reset_meters()
for ii, (img, bbox_, label_, scale) in tqdm(enumerate(dataloader)):
#调整数据的形状 scale：缩放倍数（输入图片尺寸 比上 输出数据的尺寸）
#1.6左右 供模型训练之前将模型规范化
scale = at.scalar(scale)
#将数据集转入到GPU上
#img 1x3x800x600 一张图片 三通道 大小800x600（不确定）
#bbox 1x1x4
#label 1x1
img, bbox, label = img.cuda().float(), bbox_.cuda(), label_.cuda()
#将数据转为V 变量，以便进行自动反向传播
img, bbox, label = Variable(img), Variable(bbox), Variable(label)
#训练并更新可学习参数（重点*****） 前向+反向，返回losses
trainer.train_step(img, bbox, label, scale)
#进行多个数据的可视化
if (ii + 1) % opt.plot_every == 0:
#进入调试模式
if os.path.exists(opt.debug_file):
ipdb.set_trace()
# plot loss 画五个损失
trainer.vis.plot_many(trainer.get_meter_data())
# plot groud truth bboxes img[0]，是压缩0位，形状变为[3x800x600]
#反向归一化，将img反向还原为原始图像，以便用于显示
ori_img_ = inverse_normalize(at.tonumpy(img[0]))
#通过原始图像，真实bbox，真实类别 进行显示
gt_img = visdom_bbox(ori_img_,
at.tonumpy(bbox_[0]),
at.tonumpy(label_[0]))
trainer.vis.img('gt_img', gt_img)
# plot predicti bboxes
#对原图进行预测，得到预测的bbox label scores
_bboxes, _labels, _scores = trainer.faster_rcnn.predict([ori_img_], visualize=True)
#通过原始图像、预测的bbox，预测的类别 以及概率 进行显示
pred_img = visdom_bbox(ori_img_,
at.tonumpy(_bboxes[0]),
at.tonumpy(_labels[0]).reshape(-1),
at.tonumpy(_scores[0]))
trainer.vis.img('pred_img', pred_img)
# rpn confusion matrix(meter)
#rpn混淆矩阵
trainer.vis.text(str(trainer.rpn_cm.value().tolist()), win='rpn_cm')
# roi confusion matrix
#roi混淆矩阵
trainer.vis.img('roi_cm', at.totensor(trainer.roi_cm.conf, False).float())
#使用验证集对当前的网络进行验证，返回一个字典，key值有AP,mAP
eval_result = eval(test_dataloader, faster_rcnn, test_num=opt.test_num)
#如果当前的map值优于best_map，则将当前值赋给best_map。将当前模型保留
if eval_result['map'] > best_map:
best_map = eval_result['map']
best_path = trainer.save(best_map=best_map)
#如果epoch到达9时，加载 当前的最优模型，并将学习率按lr_decay衰减调低
if epoch == 9:
trainer.load(best_path)
trainer.faster_rcnn.scale_lr(opt.lr_decay)
lr_ = lr_ * opt.lr_decay
#可视化验证集的test_map 和log信息
trainer.vis.plot('test_map', eval_result['map'])
log_info = 'lr:{}, map:{},loss:{}'.format(str(lr_),
str(eval_result['map']),
str(trainer.get_meter_data()))
trainer.vis.log(log_info)
if epoch == 13:
break
# This can be removed once PyTorch 0.4.x is out.
# See https://discuss.pytorch.org/t/question-about-rebuild-tensor-v2/14560
import torch._utils
try:
torch._utils._rebuild_tensor_v2
except AttributeError:
def _rebuild_tensor_v2(storage, storage_offset, size, stride, requires_grad, backward_hooks):
tensor = torch._utils._rebuild_tensor(storage, storage_offset, size, stride)
tensor.requires_grad = requires_grad
tensor._backward_hooks = backward_hooks
return tensor
torch._utils._rebuild_tensor_v2 = _rebuild_tensor_v2
#%%
faster_rcnn = FasterRCNNVGG16()
trainer = FasterRCNNTrainer(faster_rcnn).cuda()
trainer.load('./checkpoints/fasterrcnn_09031352_0')
opt.caffe_pretrain=True # this model was trained from caffe-pretrained model
# Plot examples on training set
dataset = RSNADataset(opt.root_dir)
for i in range(0, len(dataset)):
sample = dataset[i]
img = sample['image']
ori_img_ = inverse_normalize(at.tonumpy(img))
# plot predicti bboxes
_bboxes, _labels, _scores = trainer.faster_rcnn.predict([ori_img_], visualize=True)
pred_img = vis_bbox(ori_img_,
at.tonumpy(_bboxes[0]),
at.tonumpy(_labels[0]).reshape(-1),
at.tonumpy(_scores[0]))
def train(**kwargs):
opt._parse(kwargs)
dataset = Dataset(opt)
print('load data')
dataloader = data_.DataLoader(dataset,
batch_size=1,
shuffle=True, \
# pin_memory=True,
num_workers=opt.num_workers)
testset = TestDataset(opt)
test_dataloader = data_.DataLoader(testset,
batch_size=1,
num_workers=opt.test_num_workers,
shuffle=False,
pin_memory=True)
faster_rcnn = FasterRCNNVGG16()
print('model construct completed')
trainer = FasterRCNNTrainer(faster_rcnn).cuda()
if opt.load_path:
trainer.load(opt.load_path)
print('load pretrained model from %s' % opt.load_path)
trainer.vis.text(dataset.db.label_names, win='labels')
best_map = 0
best_ap = np.array([0.] * opt.label_number)
lr_ = opt.lr
vis = trainer.vis
starttime = datetime.datetime.now()
for epoch in range(opt.epoch):
trainer.reset_meters()
for ii, (img, bbox_, label_, scale) in tqdm(enumerate(dataloader)):
scale = at.scalar(scale)
img, bbox, label = img.cuda().float(), bbox_.cuda(), label_.cuda()
trainer.train_step(img, bbox, label, scale)
if (ii + 1) % opt.plot_every == 0:
if os.path.exists(opt.debug_file):
ipdb.set_trace()
# plot loss
trainer.vis.plot_many(trainer.get_meter_data())
# plot groud truth bboxes
ori_img_ = inverse_normalize(at.tonumpy(img[0]))
gt_img = visdom_bbox(ori_img_, at.tonumpy(bbox_[0]),
at.tonumpy(label_[0]))
trainer.vis.img('gt_img', gt_img)
# plot predicti bboxes
_bboxes, _labels, _scores = trainer.faster_rcnn.predict(
[ori_img_], visualize=True)
pred_img = visdom_bbox(ori_img_, at.tonumpy(_bboxes[0]),
at.tonumpy(_labels[0]).reshape(-1),
at.tonumpy(_scores[0]))
trainer.vis.img('pred_img', pred_img)
# rpn confusion matrix(meter)
trainer.vis.text(str(trainer.rpn_cm.value().tolist()),
win='rpn_cm')
# roi confusion matrix
roi_cm = at.totensor(trainer.roi_cm.conf, False).float()
trainer.vis.img('roi_cm', roi_cm)
eval_result = eval(test_dataloader,
faster_rcnn,
vis=vis,
test_num=opt.test_num)
best_ap = dict(zip(opt.VOC_BBOX_LABEL_NAMES, eval_result['ap']))
trainer.vis.plot('test_map', eval_result['map'])
lr_ = trainer.faster_rcnn.optimizer.param_groups[0]['lr']
log_info = 'lr:{}, map:{},loss:{}'.format(
str(lr_), str(eval_result['map']), str(trainer.get_meter_data()))
trainer.vis.log(log_info)
if eval_result['map'] > best_map:
print('roi_cm=\n', trainer.roi_cm.value())
plot_confusion_matrix(trainer.roi_cm.value(),
classes=('animal', 'plant', 'rock',
'background'),
normalize=False,
title='Normalized Confusion Matrix')
best_map = eval_result['map']
best_path = trainer.save(best_map=best_map, best_ap=best_ap)
if epoch == 9:
trainer.load(best_path)
trainer.faster_rcnn.scale_lr(opt.lr_decay)
lr_ = lr_ * opt.lr_decay
# if epoch == 13:
# break
endtime = datetime.datetime.now()
train_consum = (endtime - starttime).seconds
print("train_consum=", train_consum)
def train(**kwargs):
opt._parse(kwargs)
dataset = Dataset(opt)
# 300w_dataset = FaceLandmarksDataset()
print('load data')
dataloader = data_.DataLoader(dataset, \
batch_size=1, \
shuffle=True, \
pin_memory=True,\
num_workers=opt.num_workers)
testset = TestDataset(opt)
test_dataloader = data_.DataLoader(testset,
batch_size=1,
num_workers=opt.test_num_workers,
shuffle=False, \
pin_memory=True
)
faster_rcnn = FasterRCNNVGG16()
print('model construct completed')
attacker = attacks.DCGAN(train_adv=False)
if opt.load_attacker:
attacker.load(opt.load_attacker)
print('load attacker model from %s' % opt.load_attacker)
trainer = VictimFasterRCNNTrainer(faster_rcnn, attacker,
attack_mode=True).cuda()
# trainer = VictimFasterRCNNTrainer(faster_rcnn).cuda()
if opt.load_path:
trainer.load(opt.load_path)
print('load pretrained model from %s' % opt.load_path)
trainer.vis.text(dataset.db.label_names, win='labels')
# eval_result = eval(test_dataloader, faster_rcnn, test_num=2000)
best_map = 0
for epoch in range(opt.epoch):
trainer.reset_meters(adv=True)
for ii, (img, bbox_, label_, scale) in tqdm(enumerate(dataloader)):
ipdb.set_trace()
scale = at.scalar(scale)
img, bbox, label = img.cuda().float(), bbox_.cuda(), label_.cuda()
img, bbox, label = Variable(img), Variable(bbox), Variable(label)
trainer.train_step(img, bbox, label, scale)
if (ii) % opt.plot_every == 0:
if os.path.exists(opt.debug_file):
ipdb.set_trace()
# plot loss
trainer.vis.plot_many(trainer.get_meter_data())
trainer.vis.plot_many(trainer.get_meter_data(adv=True))
# plot groud truth bboxes
ori_img_ = inverse_normalize(at.tonumpy(img[0]))
gt_img = visdom_bbox(ori_img_, at.tonumpy(bbox_[0]),
at.tonumpy(label_[0]))
trainer.vis.img('gt_img', gt_img)
# plot predicted bboxes
_bboxes, _labels, _scores = trainer.faster_rcnn.predict(
[ori_img_], visualize=True)
pred_img = visdom_bbox(ori_img_, at.tonumpy(_bboxes[0]),
at.tonumpy(_labels[0]).reshape(-1),
at.tonumpy(_scores[0]))
trainer.vis.img('pred_img', pred_img)
if trainer.attacker is not None:
adv_img = trainer.attacker.perturb(img)
adv_img_ = inverse_normalize(at.tonumpy(adv_img[0]))
_bboxes, _labels, _scores = trainer.faster_rcnn.predict(
[adv_img_], visualize=True)
adv_pred_img = visdom_bbox(
adv_img_, at.tonumpy(_bboxes[0]),
at.tonumpy(_labels[0]).reshape(-1),
at.tonumpy(_scores[0]))
trainer.vis.img('adv_img', adv_pred_img)
# rpn confusion matrix(meter)
trainer.vis.text(str(trainer.rpn_cm.value().tolist()),
win='rpn_cm')
# roi confusion matrix
trainer.vis.img(
'roi_cm',
at.totensor(trainer.roi_cm.conf, False).float())
if (ii) % 500 == 0:
best_path = trainer.save(epochs=epoch, save_rcnn=True)
if epoch % 2 == 0:
best_path = trainer.save(epochs=epoch)
def train(**kwargs):
opt._parse(kwargs)
dataset = Dataset(opt)
print('load data')
dataloader = data_.DataLoader(dataset,
batch_size=1,
shuffle=True,
# pin_memory=True,
num_workers=opt.num_workers)
testset = TestDataset(opt)
test_dataloader = data_.DataLoader(testset,
batch_size=1,
num_workers=opt.test_num_workers,
shuffle=False,
pin_memory=True
)
testset_all = TestDataset_all(opt, 'test2')
test_all_dataloader = data_.DataLoader(testset_all,
batch_size=1,
num_workers=opt.test_num_workers,
shuffle=False,
pin_memory=True
)
tsf = Transform(opt.min_size, opt.max_size)
faster_rcnn = FasterRCNNVGG16()
trainer = FasterRCNNTrainer(faster_rcnn).cuda()
print('model construct completed')
# 加载训练过的模型，在config配置路径就可以了
if opt.load_path:
trainer.load(opt.load_path)
print('load pretrained model from %s' % opt.load_path)
#提取蒸馏知识所需要的软标签
if opt.is_distillation == True:
opt.predict_socre = 0.3
for ii, (imgs, sizes, gt_bboxes_, gt_labels_, scale, id_) in tqdm(enumerate(dataloader)):
if len(gt_bboxes_) == 0:
continue
sizes = [sizes[0][0].item(), sizes[1][0].item()]
pred_bboxes_, pred_labels_, pred_scores_, features_ = trainer.faster_rcnn.predict(imgs, [
sizes])
img_file = os.path.join(
opt.voc_data_dir, 'JPEGImages', id_[0] + '.jpg')
ori_img = read_image(img_file, color=True)
img, pred_bboxes_, pred_labels_, scale_ = tsf(
(ori_img, pred_bboxes_[0], pred_labels_[0]))
#去除软标签和真值标签重叠过多的部分，去除错误的软标签
pred_bboxes_, pred_labels_, pred_scores_ = py_cpu_nms(
gt_bboxes_[0], gt_labels_[0], pred_bboxes_, pred_labels_, pred_scores_[0])
#存储软标签，这样存储不会使得GPU占用过多
np.save('label/' + str(id_[0]) + '.npy', pred_labels_)
np.save('bbox/' + str(id_[0]) + '.npy', pred_bboxes_)
np.save('feature/' + str(id_[0]) + '.npy', features_)
np.save('score/' + str(id_[0]) + '.npy', pred_scores_)
opt.predict_socre = 0.05
t.cuda.empty_cache()
# visdom 显示所有类别标签名
trainer.vis.text(dataset.db.label_names, win='labels')
best_map = 0
lr_ = opt.lr
for epoch in range(opt.epoch):
print('epoch=%d' % epoch)
# 重置混淆矩阵
trainer.reset_meters()
# tqdm可以在长循环中添加一个进度提示信息，用户只需要封装任意的迭代器 tqdm(iterator)，
# 是一个快速、扩展性强
for ii, (img, sizes, bbox_, label_, scale, id_) in tqdm(enumerate(dataloader)):
if len(bbox_) == 0:
continue
scale = at.scalar(scale)
img, bbox, label = img.cuda().float(), bbox_.cuda(), label_.cuda()
# 训练的就这一步 下面的都是打印的信息
# 转化成pytorch能够计算的格式，转tensor格式
if opt.is_distillation == True:
#读取软标签
teacher_pred_labels = np.load(
'label/' + str(id_[0]) + '.npy')
teacher_pred_bboxes = np.load(
'bbox/' + str(id_[0]) + '.npy')
teacher_pred_features_ = np.load(
'feature/' + str(id_[0]) + '.npy')
teacher_pred_scores = np.load(
'score/' + str(id_[0]) + '.npy')
#格式转换
teacher_pred_bboxes = teacher_pred_bboxes.astype(np.float32)
teacher_pred_labels = teacher_pred_labels.astype(np.int32)
teacher_pred_scores = teacher_pred_scores.astype(np.float32)
#转成pytorch格式
teacher_pred_bboxes_ = at.totensor(teacher_pred_bboxes)
teacher_pred_labels_ = at.totensor(teacher_pred_labels)
teacher_pred_scores_ = at.totensor(teacher_pred_scores)
teacher_pred_features_ = at.totensor(teacher_pred_features_)
#使用GPU
teacher_pred_bboxes_ = teacher_pred_bboxes_.cuda()
teacher_pred_labels_ = teacher_pred_labels_.cuda()
teacher_pred_scores_ = teacher_pred_scores_.cuda()
teacher_pred_features_ = teacher_pred_features_.cuda()
# 如果dataset.py 中的Transform 设置了图像翻转,就要使用这个判读软标签是否一起翻转
if(teacher_pred_bboxes_[0][1] != bbox[0][0][1]):
_, o_C, o_H, o_W = img.shape
teacher_pred_bboxes_ = flip_bbox(
teacher_pred_bboxes_, (o_H, o_W), x_flip=True)
losses = trainer.train_step(img, bbox, label, scale, epoch,
teacher_pred_bboxes_, teacher_pred_labels_, teacher_pred_features_, teacher_pred_scores)
else:
trainer.train_step(img, bbox, label, scale, epoch)
# visdom显示的信息
if (ii + 1) % opt.plot_every == 0:
if os.path.exists(opt.debug_file):
ipdb.set_trace()
# plot loss
trainer.vis.plot_many(trainer.get_meter_data())
# plot groud truth bboxes
ori_img_ = inverse_normalize(at.tonumpy(img[0]))
gt_img = visdom_bbox(ori_img_,
at.tonumpy(bbox_[0]),
at.tonumpy(label_[0]))
trainer.vis.img('gt_img', gt_img)
gt_img = visdom_bbox(ori_img_,
at.tonumpy(teacher_pred_bboxes_),
at.tonumpy(teacher_pred_labels_),
at.tonumpy(teacher_pred_scores_))
trainer.vis.img('gt_img_all', gt_img)
# plot predicti bboxes
_bboxes, _labels, _scores, _ = trainer.faster_rcnn.predict(
[ori_img_], visualize=True)
pred_img = visdom_bbox(ori_img_,
at.tonumpy(_bboxes[0]),
at.tonumpy(_labels[0]).reshape(-1),
at.tonumpy(_scores[0]))
trainer.vis.img('pred_img', pred_img)
# 混淆矩阵
# rpn confusion matrix(meter)
trainer.vis.text(
str(trainer.rpn_cm.value().tolist()), win='rpn_cm')
# roi confusion matrix
trainer.vis.text(
str(trainer.roi_cm.value().tolist()), win='roi_cm')
# trainer.vis.img('roi_cm', at.totensor(
# trainer.roi_cm.value(), False).float())
eval_result = eval(test_dataloader, faster_rcnn, test_num=opt.test_num)
trainer.vis.plot('test_map', eval_result['map'])
lr_ = trainer.faster_rcnn.optimizer.param_groups[0]['lr']
log_info = 'lr:{},ap:{}, map:{},loss:{}'.format(str(lr_),
str(eval_result['ap']),
str(eval_result['map']),
str(trainer.get_meter_data()))
trainer.vis.log(log_info)
# 保存最好结果并记住路径
if eval_result['map'] > best_map:
best_map = eval_result['map']
best_path = trainer.save(best_map=best_map)
if epoch == 20:
trainer.save(best_map='20')
result = eval(test_all_dataloader,
trainer.faster_rcnn, test_num=5000)
print('20result={}'.format(str(result)))
# trainer.load(best_path)
# result=eval(test_all_dataloader,trainer.faster_rcnn,test_num=5000)
# print('bestmapresult={}'.format(str(result)))
break
# 每10轮加载前面最好权重，并且减少学习率
if epoch % 20 == 15:
trainer.load(best_path)
trainer.faster_rcnn.scale_lr(opt.lr_decay)
lr_ = lr_ * opt.lr_decay
def train(**kwargs):
opt.parse(kwargs)
print('loading data...')
trainset = TrainDataset(opt)
train_dataloader = torch.utils.data.DataLoader(trainset,
batch_size=1,
shuffle=True,
num_workers=opt.num_workers)
testset = TestDataset(opt)
test_dataloader = torch.utils.data.DataLoader(
testset,
batch_size=1,
num_workers=opt.test_num_workers,
shuffle=False,
pin_memory=True)
print('constructing model...')
if opt.model == 'vgg16':
faster_rcnn = FasterRCNNVGG16()
elif opt.model == 'resnet101':
faster_rcnn = FasterRCNNResNet101()
trainer = FasterRCNNTrainer(faster_rcnn).cuda()
print('loading model...')
if opt.load_path:
trainer.load(opt.load_path)
print('load pretrained model from %s' % opt.load_path)
else:
print('no pretrained model found')
trainer.vis.text('<br/>'.join(trainset.db.label_names), win='labels')
print('start training...')
best_map = 0.0
lr_ = opt.lr
for epoch in range(opt.epoch):
print("epoch : %d training ..." % epoch)
trainer.reset_meters()
for ii, (imgs_, bboxes_, labels_,
scales_) in tqdm(enumerate(train_dataloader)):
scales = at.scalar(scales_)
imgs, bboxes, labels = imgs_.cuda().float(), bboxes_.cuda(
), labels_.cuda()
trainer.train_step(imgs, bboxes, labels, scales)
if (ii + 1) % opt.plot_every == 0:
# plot loss
trainer.vis.plot_many(trainer.losses_data())
# generate plotted image
img = inverse_normalize(at.tonumpy(imgs_[0]))
# plot groud truth bboxes
bbox = at.tonumpy(bboxes_[0])
label = at.tonumpy(labels_[0])
img_gt = visdom_bbox(img, bbox, label)
trainer.vis.img('ground truth', img_gt)
bboxes__, labels__, scores__ = trainer.faster_rcnn.predict(
[img], visualize=True)
# plot prediction bboxes
bbox = at.tonumpy(bboxes__[0])
label = at.tonumpy(labels__[0]).reshape(-1)
score = at.tonumpy(scores__[0])
img_pred = visdom_bbox(img, bbox, label, score)
trainer.vis.img('prediction', img_pred)
# rpn confusion matrix(meter)
trainer.vis.text(str(trainer.rpn_cm.value().tolist()),
win='rpn_cm')
# roi confusion matrix
trainer.vis.img(
'roi_cm',
at.totensor(trainer.roi_cm.conf, False).float())
if ii + 1 == opt.train_num:
break
print("epoch : %d evaluating ..." % epoch)
eval_result = eval(test_dataloader, faster_rcnn, test_num=opt.test_num)
trainer.vis.plot('test_map', eval_result['map'])
lr_ = trainer.faster_rcnn.optimizer.param_groups[0]['lr']
log_info = vis_dict(
{
'epoch': '%s/%s' % (str(epoch), str(opt.epoch)),
'lr': lr_,
'map': float(eval_result['map']),
}, trainer.losses_data())
trainer.vis.log(log_info)
if eval_result['map'] > best_map:
best_map = eval_result['map']
best_path = trainer.save(best_map="%.4f" % best_map)
if epoch == 9:
trainer.load(best_path)
trainer.faster_rcnn.scale_lr(opt.lr_decay)
lr_ = lr_ * opt.lr_decay
def train(**kwargs): # *变量名, 表示任何多个无名参数, 它是一个tuple；**变量名, 表示关键字参数, 它是一个dict
opt._parse(kwargs) # 识别参数,传递过来的是一个字典,用parse来解析
dataset = Dataset(opt) # 作者自定义的Dataset类
print('读取数据中...')
# Dataloader 定义了一次获取批次数据的方法
dataloader = data_.DataLoader(dataset, \
batch_size=1, \
shuffle=True, \
# pin_memory=True,
num_workers=opt.num_workers) # PyTorch自带的DataLoader类,生成一个多线程迭代器来迭代dataset, 以供读取一个batch的数据
testset = TestDataset(opt, split='trainval')
# 测试集loader
test_dataloader = data_.DataLoader(testset,
batch_size=1,
num_workers=opt.test_num_workers,
shuffle=False, \
pin_memory=True
)
faster_rcnn = FasterRCNNVGG16() # 网络定义
print('模型构建完毕!')
trainer = FasterRCNNTrainer(
faster_rcnn).cuda() # 定义一个训练器,返回loss, .cuda()表示把返回的Tensor存入GPU
if opt.load_path: # 如果要加载预训练模型
trainer.load(opt.load_path)
print('已加载预训练参数 %s' % opt.load_path)
else:
print("未引入预训练参数, 随机初始化网络参数")
trainer.vis.text(dataset.db.label_names, win='labels') # 显示labels标题
best_map = 0 # 定义一个best_map
for epoch in range(opt.epoch): # 对于每一个epoch
trainer.reset_meters() # 重置测各种测量仪
# 对每一个数据
for ii, (img, bbox_, label_, scale) in tqdm(enumerate(dataloader)):
scale = at.scalar(scale) # 转化为标量
img, bbox, label = img.cuda().float(), bbox_.cuda(), label_.cuda(
) # 存入GPU
img, bbox, label = Variable(img), Variable(bbox), Variable(
label) # 转换成变量以供自动微分器使用
# TODO
trainer.train_step(img, bbox, label, scale) # 训练一步
if (ii + 1) % opt.plot_every == 0: # 如果到达"每多少次显示"
if os.path.exists(opt.debug_file):
ipdb.set_trace()
# plot loss
trainer.vis.plot_many(trainer.get_meter_data())
# plot groud truth bboxes
ori_img_ = inverse_normalize(at.tonumpy(img[0]))
gt_img = visdom_bbox(ori_img_, at.tonumpy(bbox_[0]),
at.tonumpy(label_[0]))
trainer.vis.img('gt_img', gt_img)
# plot predicti bboxes
_bboxes, _labels, _scores = trainer.faster_rcnn.predict(
[ori_img_], visualize=True)
pred_img = visdom_bbox(ori_img_, at.tonumpy(_bboxes[0]),
at.tonumpy(_labels[0]).reshape(-1),
at.tonumpy(_scores[0]))
trainer.vis.img('pred_img', pred_img)
# rpn confusion matrix(meter)
trainer.vis.text(str(trainer.rpn_cm.value().tolist()),
win='rpn_cm')
# roi confusion matrix
trainer.vis.img(
'roi_cm',
at.totensor(trainer.roi_cm.conf, False).float())
# 使用测试数据集来评价模型(此步里面包含预测信息)
eval_result = eval(test_dataloader, faster_rcnn, test_num=opt.test_num)
if eval_result['map'] > best_map:
best_map = eval_result['map']
best_path = trainer.save(
best_map=best_map) # 好到一定程度就存储模型, 存储在checkpoint文件夹内
if epoch == 9: # 到第9轮的时候读取模型, 并调整学习率
trainer.load(best_path)
trainer.faster_rcnn.scale_lr(opt.lr_decay)
trainer.vis.plot('test_map', eval_result['map'])
lr_ = trainer.faster_rcnn.optimizer.param_groups[0]['lr']
log_info = 'lr:{}, map:{},loss:{}'.format(
str(lr_), str(eval_result['map']), str(trainer.get_meter_data()))
trainer.vis.log(log_info)
# if epoch == 13: # 到第14轮的时候停止训练
# break
trainer.save(best_map=best_map)
def train(**kwargs):
opt._parse(
kwargs
) #将调用函数时候附加的参数用，config.py文件里面的opt._parse()进行解释，然后获取其数据存储的路径，之后放到Dataset里面！
dataset = Dataset(opt)
print('load data')
dataloader = data_.DataLoader(dataset,
batch_size=1,
shuffle=True,
num_workers=opt.num_workers)
testset = TestDataset(opt)
test_dataloader = data_.DataLoader(
testset,
batch_size=1,
num_workers=opt.test_num_workers,
shuffle=False,
#pin_memory=True
) #pin_memory锁页内存,开启时使用显卡的内存，速度更快
faster_rcnn = FasterRCNNVGG16()
print('model construct completed')
trainer = FasterRCNNTrainer(faster_rcnn).cuda()
#判断opt.load_path是否存在，如果存在，直接从opt.load_path读取预训练模型，然后将训练数据的label进行可视化操作
if opt.load_path:
trainer.load(opt.load_path)
print('load pretrained model from %s' % opt.load_path)
trainer.vis.text(dataset.dataset.label_names, win='labels')
best_map = 0
lr_ = opt.lr
# 之后用一个for循环开始训练过程，而训练迭代的次数opt.epoch=14也在config.py文件中都预先定义好，属于超参数
for epoch in range(opt.epoch):
print('epoch {}/{}'.format(epoch, opt.epoch))
trainer.reset_meters() #首先在可视化界面重设所有数据
for ii, (img, bbox_, label_, scale) in tqdm(enumerate(dataloader)):
scale = array_tool.scalar(scale)
img, bbox, label = img.cuda().float(), bbox_.cuda(), label_.cuda()
trainer.train_step(img, bbox, label, scale)
if (ii + 1) % opt.plot_every == 0:
if os.path.exists(opt.debug_file):
ipdb.set_trace()
#可视化画出loss
trainer.vis.plot_many(trainer.get_meter_data())
#可视化画出groudtruth bboxes
ori_img_ = inverse_normalize(array_tool.tonumpy(img[0]))
gt_img = visdom_bbox(ori_img_, array_tool.tonumpy(bbox_[0]),
array_tool.tonumpy(label_[0]))
trainer.vis.img('gt_img', gt_img)
#可视化画出预测bboxes
# 调用faster_rcnn的predict函数进行预测，预测的结果保留在以_下划线开头的对象里面
_bboxes, _labels, _scores = trainer.faster_rcnn.predict(
[ori_img_], visualize=True)
pred_img = visdom_bbox(
ori_img_, array_tool.tonumpy(_bboxes[0]),
array_tool.tonumpy(_labels[0]).reshape(-1),
array_tool.tonumpy(_scores[0]))
trainer.vis.img('pred_img', pred_img)
# 调用 trainer.vis.text将rpn_cm也就是RPN网络的混淆矩阵在可视化工具中显示出来
trainer.vis.text(str(trainer.rpn_cm.value().tolist()),
win='rpn_cm')
#将roi_cm也就是roihead网络的混淆矩阵在可视化工具中显示出来
trainer.vis.img(
'roi_cm',
array_tool.totensor(trainer.roi_cm.conf, False).float())
eval_result = eval(test_dataloader, faster_rcnn, test_num=opt.test_num)
trainer.vis.plot('test_map', eval_result['map'])
lr_ = trainer.faster_rcnn.optimizer.param_groups[0]['lr']
log_info = 'lr:{}, map:{}, loss{}'.format(
str(lr_), str(eval_result['map']), str(trainer.get_meter_data()))
trainer.vis.log(log_info) #将学习率以及map等信息及时显示更新
if eval_result['map'] > best_map:
best_map = eval_result['map']
best_path = trainer.save(best_map=best_map)
if epoch == 9: #if判断语句如果学习的epoch达到了9就将学习率*0.1变成原来的十分之一
trainer.load(best_path)
trainer.faster_rcnn.scale_lr(opt.lr_decay)
lr_ = lr_ * opt.lr_decay
if epoch == 13:
break
def train(**kwargs):
opt._parse(kwargs)
print('load data')
dataset = Dataset(opt)
dataloader = data_.DataLoader(dataset, \
batch_size=1, \
shuffle=True, \
# pin_memory=True,
num_workers=opt.num_workers)
testset = TestDataset(opt)
test_dataloader = data_.DataLoader(testset,
batch_size=1,
num_workers=opt.test_num_workers,
shuffle=False, \
pin_memory=True
)
faster_rcnn = FasterRCNNVGG16(n_fg_class=dataset.get_class_count(), anchor_scales=[1])
print('model construct completed')
trainer = FasterRCNNTrainer(faster_rcnn, n_fg_class=dataset.get_class_count())
if opt.use_cuda:
trainer = trainer.cuda()
if opt.load_path:
old_state = trainer.load(opt.load_path)
print('load pretrained model from %s' % opt.load_path)
if opt.validate_only:
num_eval_images = len(testset)
eval_result = eval(test_dataloader, faster_rcnn, test_num=num_eval_images)
print('Evaluation finished, obtained {} using {} out of {} images'.
format(eval_result, num_eval_images, len(testset)))
return
if opt.load_path and 'epoch' in old_state.keys():
starting_epoch = old_state['epoch'] + 1
print('Model was trained until epoch {}, continuing with epoch {}'.format(old_state['epoch'], starting_epoch))
else:
starting_epoch = 0
#trainer.vis.text(dataset.db.label_names, win='labels')
best_map = 0
lr_ = opt.lr
global_step = 0
for epoch in range(starting_epoch, opt.num_epochs):
lr_ = opt.lr * (opt.lr_decay ** (epoch // opt.epoch_decay))
trainer.faster_rcnn.set_lr(lr_)
print('Starting epoch {} with learning rate {}'.format(epoch, lr_))
trainer.reset_meters()
for ii, (img, bbox_, label_, scale) in tqdm(enumerate(dataloader), total=len(dataset)):
global_step = global_step + 1
scale = at.scalar(scale)
if opt.use_cuda:
img, bbox, label = img.cuda().float(), bbox_.float().cuda(), label_.float().cuda()
else:
img, bbox, label = img.float(), bbox_.float(), label_.float()
img, bbox, label = Variable(img), Variable(bbox), Variable(label)
losses = trainer.train_step(img, bbox, label, scale)
if (ii + 1) % opt.plot_every == 0:
if os.path.exists(opt.debug_file):
ipdb.set_trace()
# plot loss
#trainer.vis.plot_many(trainer.get_meter_data())
# plot groud truth bboxes
ori_img_ = inverse_normalize(at.tonumpy(img[0]))
gt_img = visdom_bbox(ori_img_,
at.tonumpy(bbox_[0]),
at.tonumpy(label_[0]),
label_names=dataset.get_class_names()+['BG'])
trainer.vis.img('gt_img', gt_img)
# plot predicti bboxes
_bboxes, _labels, _scores = trainer.faster_rcnn.predict([ori_img_], visualize=True)
pred_img = visdom_bbox(ori_img_,
at.tonumpy(_bboxes[0]),
at.tonumpy(_labels[0]).reshape(-1),
at.tonumpy(_scores[0]),
label_names=dataset.get_class_names()+['BG'])
trainer.vis.img('pred_img', pred_img)
# rpn confusion matrix(meter)
#trainer.vis.text(str(trainer.rpn_cm.value().tolist()), win='rpn_cm')
# roi confusion matrix
#trainer.vis.img('roi_cm', at.totensor(trainer.roi_cm.conf, False).float())
#print('Current total loss {}'.format(losses[-1].tolist()))
trainer.vis.plot('train_total_loss', losses[-1].tolist())
if (global_step) % opt.snapshot_every == 0:
snapshot_path = trainer.save(epoch=epoch)
print("Snapshotted to {}".format(snapshot_path))
#snapshot_path = trainer.save(epoch=epoch)
#print("After epoch {}: snapshotted to {}".format(epoch,snapshot_path))
eval_result = eval(test_dataloader, faster_rcnn, test_num=min(opt.test_num, len(testset)))
print(eval_result)
# TODO: this definitely is not good and will bias evaluation
if eval_result['map'] > best_map:
best_map = eval_result['map']
best_path = trainer.save(best_map=eval_result['map'],epoch=epoch)
print("After epoch {}: snapshotted to {}".format(epoch, best_path))
trainer.vis.plot('test_map', eval_result['map'])
def train(**kwargs):
opt._parse(kwargs)
dataset = Dataset(opt)
print('load data')
dataloader = data_.DataLoader(dataset, \
batch_size=1, \
shuffle=True, \
# pin_memory=True,
num_workers=opt.num_workers)
testset = TestDataset(opt)
test_dataloader = data_.DataLoader(testset,
batch_size=1,
num_workers=opt.test_num_workers,
shuffle=False, \
pin_memory=True
)
faster_rcnn = FasterRCNNVGG16()
print('model construct completed')
trainer = FasterRCNNTrainer(faster_rcnn).cuda()
if opt.load_path:
trainer.load(opt.load_path)
print('load pretrained model from {}'.format(opt.load_path))
# trainer.vis.text(dataset.db.label_names, win='labels')
adversary = None
if opt.flagadvtrain:
print("flagadvtrain turned: Adversarial training!")
atk = PGD.PGD(trainer, eps=16/255, alpha=3/255, steps=4)
# atk = torchattacks.PGD(trainer.faster_rcnn, eps=16, alpha=3, steps=4)
# adversary = PGDAttack(trainer.faster_rcnn, loss_fn=nn.CrossEntropyLoss(), eps=16, nb_iter=4, eps_iter=3,
# rand_init=True, clip_min=0.0, clip_max=1.0, targeted=False)
best_map = 0
lr_ = opt.lr
normal_total_loss = []
adv_total_loss = []
total_time = 0.0
total_imgs = 0
true_imgs = 0
for epoch in range(opt.epoch):
trainer.reset_meters()
once = True
for ii, (img, bbox_, label_, scale) in tqdm(enumerate(dataloader)):
scale = at.scalar(scale)
temp_img = copy.deepcopy(img).cuda()
img, bbox, label = img.cuda().float(), bbox_.cuda(), label_.cuda()
if opt.flagadvtrain:
before_time = time.time()
img = atk(img, bbox, label, scale)
after_time = time.time()
# with ctx_noparamgrad_and_eval(trainer.faster_rcnn):
# img = adversary.perturb(img, label)
# print("Adversarial training done!")
total_time += after_time - before_time
# print("Normal training starts\n")
# trainer.train_step(img, bbox, label, scale)
if (ii + 1) % opt.plot_every == 0:
# adv_total_loss.append(trainer.get_meter_data()["total_loss"])
if os.path.exists(opt.debug_file):
ipdb.set_trace()
# plot loss
# trainer.vis.plot_many(trainer.get_meter_data())
# plot groud truth bboxes
temp_ori_img_ = inverse_normalize(at.tonumpy(temp_img[0]))
# img2jpg(temp_ori_img_, "imgs/orig_images/", "gt_img{}".format(ii))
# temp_gt_img = visdom_bbox(temp_ori_img_,
# at.tonumpy(bbox_[0]),
# at.tonumpy(label_[0]))
# plt.figure()
# c, h, w = temp_gt_img.shape
# plt.imshow(np.reshape(temp_gt_img, (h, w, c)))
# plt.savefig("imgs/temp_orig_images/temp_gt_img{}".format(ii))
# plt.close()
ori_img_ = inverse_normalize(at.tonumpy(img[0]))
# print("GT Label is {} and pred_label is {}".format(label_[0],))
# img2jpg(ori_img_, "imgs/adv_images/", "adv_img{}".format(ii))
# gt_img = visdom_bbox(ori_img_,
# at.tonumpy(bbox_[0]),
# at.tonumpy(label_[0]))
# plt.figure()
# c, h, w = gt_img.shape
# plt.imshow(np.reshape(gt_img, (h, w, c)))
# plt.savefig("imgs/orig_images/gt_img{}".format(ii))
# plt.close()
# trainer.vis.img('gt_img', gt_img)
# plot predicti bboxes
_bboxes, _labels, _scores = trainer.faster_rcnn.predict([ori_img_], visualize=True)
fig1 = plt.figure()
ax1 = fig1.add_subplot(1,1,1)
# final1 = (at.tonumpy(img[0].cpu()).transpose(1,2,0).astype(np.uint8))
final1 = (ori_img_.transpose(1, 2, 0).astype(np.uint8))
ax1.imshow(final1)
gt_img = visdom_bbox(ax1,at.tonumpy(_bboxes[0]),at.tonumpy(_labels[0]))
fig1.savefig("imgs/adv_images/adv_img{}".format(ii))
plt.close()
_temp_bboxes, _temp_labels, _temp_scores = trainer.faster_rcnn.predict([temp_ori_img_], visualize=True)
fig2 = plt.figure()
ax2 = fig2.add_subplot(1, 1, 1)
final2 = (temp_ori_img_.transpose(1, 2, 0).astype(np.uint8))
# final2 = (at.tonumpy(temp_img[0].cpu()).transpose(1, 2, 0).astype(np.uint8))
ax2.imshow(final2)
gt_img = visdom_bbox(ax2, at.tonumpy(_temp_bboxes[0]), at.tonumpy(_temp_labels[0]))
fig2.savefig("imgs/orig_images/gt_img{}".format(ii))
plt.close()
# img2jpg(temp_gt_img, "imgs/orig_images/", "gt_img{}".format(ii))
# print("gt labels is {}, pred_orig_labels is {} and pred_adv_labels is {}".format(label_, _labels, _temp_labels))
total_imgs += 1
if len(_temp_labels) == 0:
continue
if _labels[0].shape[0] == _temp_labels[0].shape[0] and (_labels[0] == _temp_labels[0]).all() is True:
true_imgs += 1
# pred_img = visdom_bbox(ori_img_,
# at.tonumpy(_bboxes[0]),
# at.tonumpy(_labels[0]).reshape(-1),
# at.tonumpy(_scores[0]))
#
# print("Shape of temp_orig_img_ is {}".format(temp_ori_img_.shape))
# temp_pred_img = visdom_bbox(temp_ori_img_,
# at.tonumpy(_temp_bboxes[0]),
# at.tonumpy(_temp_labels[0]).reshape(-1),
# at.tonumpy(_temp_scores[0]))
#
# trainer.vis.img('pred_img', pred_img)
# rpn confusion matrix(meter)
# trainer.vis.text(str(trainer.rpn_cm.value().tolist()), win='rpn_cm')
# roi confusion matrix
# trainer.vis.img('roi_cm', at.totensor(trainer.roi_cm.conf, False).float())
# fig = plt.figure()
# ax1 = fig.add_subplot(2,1,1)
# ax1.plot(normal_total_loss)
# ax2 = fig.add_subplot(2,1,2)
# ax2.plot(adv_total_loss)
# fig.savefig("losses/both_loss{}".format(epoch))
# eval_result = eval(test_dataloader, faster_rcnn, test_num=opt.test_num,
# flagadvtrain=opt.flagadvtrain, adversary=atk)# adversary=adversary)
# trainer.vis.plot('test_map', eval_result['map'])
# lr_ = trainer.faster_rcnn.optimizer.param_groups[0]['lr']
# log_info = 'lr:{}, map:{},loss:{}'.format(str(lr_),
# str(eval_result['map']),
# str(trainer.get_meter_data()))
# print(log_info)
# # trainer.vis.log(log_info)
#
# if eval_result['map'] > best_map:
# best_map = eval_result['map']
# best_path = trainer.save(best_map=best_map)
# if epoch == 9:
# trainer.load(best_path)
# trainer.faster_rcnn.scale_lr(opt.lr_decay)
# lr_ = lr_ * opt.lr_decay
if epoch == 0:
break
if epoch == 13:
break
print("Total number of images is {}".format(total_imgs))
print("True images is {}".format(true_imgs))
print("Total time is {}".format(total_time))
print("Avg time is {}".format(total_time/total_imgs))
def train(**kwargs):
opt._parse(kwargs)
dataset = Dataset(opt)
print('load data')
dataloader = data_.DataLoader(dataset, \
batch_size=1, \
shuffle=True, \
# pin_memory=True,
num_workers=opt.num_workers)
testset = TestDataset(opt)
test_dataloader = data_.DataLoader(testset,
batch_size=1,
num_workers=opt.test_num_workers,
shuffle=False, \
pin_memory=True
)
faster_rcnn = FasterRCNNVGG16()
print('model construct completed')
trainer = FasterRCNNTrainer(faster_rcnn).cuda()
if opt.load_path:
trainer.load(opt.load_path)
print('load pretrained model from %s' % opt.load_path)
trainer.vis.text(dataset.db.label_names, win='labels')
best_map = 0
for epoch in range(opt.epoch):
trainer.reset_meters()
for ii, (img, bbox_, label_, scale) in tqdm(enumerate(dataloader)):
scale = at.scalar(scale)
img, bbox, label = img.cuda().float(), bbox_.cuda(), label_.cuda()
img, bbox, label = Variable(img), Variable(bbox), Variable(label)
trainer.train_step(img, bbox, label, scale)
if (ii + 1) % opt.plot_every == 0:
if os.path.exists(opt.debug_file):
ipdb.set_trace()
# plot loss
trainer.vis.plot_many(trainer.get_meter_data())
# plot groud truth bboxes
ori_img_ = inverse_normalize(at.tonumpy(img[0]))
gt_img = visdom_bbox(ori_img_,
at.tonumpy(bbox_[0]),
at.tonumpy(label_[0]))
trainer.vis.img('gt_img', gt_img)
# plot predicti bboxes
_bboxes, _labels, _scores = trainer.faster_rcnn.predict([ori_img_], visualize=True)
pred_img = visdom_bbox(ori_img_,
at.tonumpy(_bboxes[0]),
at.tonumpy(_labels[0]).reshape(-1),
at.tonumpy(_scores[0]))
trainer.vis.img('pred_img', pred_img)
# rpn confusion matrix(meter)
trainer.vis.text(str(trainer.rpn_cm.value().tolist()), win='rpn_cm')
# roi confusion matrix
trainer.vis.img('roi_cm', at.totensor(trainer.roi_cm.conf, False).float())
eval_result = eval(test_dataloader, faster_rcnn, test_num=opt.test_num)
if eval_result['map'] > best_map:
best_map = eval_result['map']
best_path = trainer.save(best_map=best_map)
if epoch == 9:
trainer.load(best_path)
trainer.faster_rcnn.scale_lr(opt.lr_decay)
trainer.vis.plot('test_map', eval_result['map'])
lr_ = trainer.faster_rcnn.optimizer.param_groups[0]['lr']
log_info = 'lr:{}, map:{},loss:{}'.format(str(lr_),
str(eval_result['map']),
str(trainer.get_meter_data()))
trainer.vis.log(log_info)
if epoch == 13:
break
def train(**kwargs):
opt._parse(kwargs)
log = SummaryWriter(log_dir=opt.log_dir)
dataset = Dataset(opt)
print('load data')
dataloader = data_.DataLoader(dataset,
batch_size=1,
shuffle=True,
pin_memory=True,
num_workers=opt.num_workers)
testset = TestDataset(opt)
test_dataloader = data_.DataLoader(testset,
batch_size=1,
num_workers=opt.test_num_workers,
shuffle=False,
pin_memory=True
)
# 配置文件
# cfg = VGConf()
# 训练数据集
# trainset = Dataset(cfg)
# valset = Dataset(cfg, valid=True)
# 加载数据
# print("load data2..")
# dataloader = DataLoader(dataloader, batch_size=1, shuffle=True,
# pin_memory=True, num_workers=opt.num_workers)
# valloader = DataLoader(test_dataloader, batch_size=1, shuffle=False,
# pin_memory=True, num_workers=opt.num_workers)
faster_rcnn = FasterRCNNVGG16()
print('model construct completed')
trainer = FasterRCNNTrainer(faster_rcnn).cuda()
if opt.load_path:
trainer.load(opt.load_path)
print('load pretrained model from %s' % opt.load_path)
# trainer.vis.text(dataset.db.label_names, win='labels')
best_map = 0
lr_ = opt.lr
idx = 0
for epoch in range(opt.epoch):
trainer.reset_meters()
for ii, (img, bbox_, label_, scale) in enumerate(dataloader):
scale = at.scalar(scale)
img, bbox, label = img.cuda().float(), bbox_.cuda(), label_.cuda()
trainer.train_step(img, bbox, label, scale)
# 获取损失值
losses = trainer.get_meter_data()
log.add_scalars(main_tag='Training(batch)',
tag_scalar_dict=losses,
global_step=idx)
idx = idx+1
if (ii + 1) % opt.plot_every == 0:
# if os.path.exists(opt.debug_file):
# ipdb.set_trace()
# plot loss
# trainer.vis.plot_many(trainer.get_meter_data())
print(trainer.get_meter_data())
# plot groud truth bboxes
ori_img_ = inverse_normalize(at.tonumpy(img[0]))
# gt_img = visdom_bbox(ori_img_,
# at.tonumpy(bbox_[0]),
# at.tonumpy(label_[0]))
# trainer.vis.img('gt_img', gt_img)
# plot predicti bboxes
# _bboxes, _labels, _scores = trainer.faster_rcnn.predict([ori_img_], visualize=True)
# pred_img = visdom_bbox(ori_img_,
# at.tonumpy(_bboxes[0]),
# at.tonumpy(_labels[0]).reshape(-1),
# at.tonumpy(_scores[0]))
# trainer.vis.img('pred_img', pred_img)
# rpn confusion matrix(meter)
# trainer.vis.text(str(trainer.rpn_cm.value().tolist()), win='rpn_cm')
# roi confusion matrix
# trainer.vis.img('roi_cm', at.totensor(trainer.roi_cm.conf, False).float())
eval_result = eval(test_dataloader, faster_rcnn, test_num=opt.test_num)
log.add_scalar(tag='mAP', scalar_value=eval_result['map'], global_step=epoch)
# trainer.vis.plot('test_map', eval_result['map'])
lr_ = trainer.faster_rcnn.optimizer.param_groups[0]['lr']
log_info = 'lr:{}, map:{},loss:{}'.format(str(lr_),
str(eval_result['map']),
str(trainer.get_meter_data()))
print(log_info)
# trainer.vis.log(log_info)
if eval_result['map'] > best_map:
best_map = eval_result['map']
best_path = trainer.save(best_map=best_map)
if epoch == 9:
trainer.load(best_path)
trainer.faster_rcnn.scale_lr(opt.lr_decay)
lr_ = lr_ * opt.lr_decay
if epoch == 13:
break
def eval(dataloader, faster_rcnn, trainer, dataset, test_num=10000):
with torch.no_grad():
print('Running validation')
# Each predicted box is organized as :`(y_{min}, x_{min}, y_{max}, x_{max}),
# Where y corresponds to the height and x to the width
pred_bboxes, pred_labels, pred_scores = list(), list(), list()
gt_bboxes, gt_labels, gt_difficults = list(), list(), list()
image_ids = list()
for ii, (imgs, sizes, gt_bboxes_, gt_labels_, gt_difficults_,
image_ids_) in tqdm(enumerate(dataloader), total=test_num):
sizes = [
sizes[0].detach().numpy().tolist()[0],
sizes[1].detach().numpy().tolist()[0]
]
pred_bboxes_, pred_labels_, pred_scores_ = faster_rcnn.predict(
imgs, [sizes])
# We have to add .copy() here to allow for the loaded image to be released after each iteration
gt_bboxes += list(gt_bboxes_.numpy().copy())
gt_labels += list(gt_labels_.numpy().copy())
gt_difficults += list(gt_difficults_.numpy().copy())
image_ids += list(image_ids_.numpy().copy())
pred_bboxes += [pp.copy() for pp in pred_bboxes_]
pred_labels += [pp.copy() for pp in pred_labels_]
pred_scores += [pp.copy() for pp in pred_scores_]
if ii == test_num: break
result = eval_detection_voc(pred_bboxes,
pred_labels,
pred_scores,
gt_bboxes,
gt_labels,
gt_difficults,
use_07_metric=True)
if opt.validate_only:
save_path = '{}_detections.npz'.format(opt.load_path)
np.savez(save_path,
pred_bboxes=pred_bboxes,
pred_labels=pred_labels,
pred_scores=pred_scores,
gt_bboxes=gt_bboxes,
gt_labels=gt_labels,
gt_difficults=gt_difficults,
image_ids=image_ids,
result=result)
else:
ori_img_ = inverse_normalize(at.tonumpy(imgs[0]))
gt_img = visdom_bbox(ori_img_,
at.tonumpy(gt_bboxes[-1]),
at.tonumpy(gt_labels[-1]),
label_names=dataset.get_class_names() +
['BG'])
trainer.vis.img('test_gt_img', gt_img)
# plot predicti bboxes
pred_img = visdom_bbox(ori_img_,
at.tonumpy(pred_bboxes[-1]),
at.tonumpy(pred_labels[-1]).reshape(-1),
at.tonumpy(pred_scores[-1]),
label_names=dataset.get_class_names() +
['BG'])
trainer.vis.img('test_pred_img', pred_img)
del imgs, gt_bboxes_, gt_labels_, gt_difficults_, image_ids_, pred_bboxes_, pred_labels_, pred_scores_
torch.cuda.empty_cache()
return result