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 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 test(img):
img = t.from_numpy(img)[None]
opt.caffe_pretrain=False # this model was trained from caffe-pretrained model
_bboxes, _labels, _scores = trainer.faster_rcnn.predict(img, visualize=True)
#output the 坐标
bboxes = at.tonumpy(_bboxes[0])
print(bboxes) #输出框的坐标，array格式
test_img = visdom_bbox(at.tonumpy(img[0]),
at.tonumpy(_bboxes[0]),
at.tonumpy(_labels[0]).reshape(-1),
at.tonumpy(_scores[0]).reshape(-1))
trainer.vis.img('test_img', test_img)
def predict(load_path, **kwargs):
"""
"""
"""parse parameters"""
opt.load_path = load_path
opt.parse(kwargs)
"""get images to be predicted"""
if not os.path.isdir(opt.predict_output_dir):
os.mkdir(opt.predict_output_dir)
img_files = os.listdir(opt.predict_input_dir)
img_files.sort()
img_paths = [
os.path.join(opt.predict_input_dir, name) for name in img_files
]
"""create model"""
rfcn_md = RFCN_ResNet101()
print('model construct completed')
rfcn_trainer = RFCN_Trainer(rfcn_md).cuda()
if opt.load_path:
rfcn_trainer.load(opt.load_path, load_viz_idx=opt.load_viz_x)
print('load pretrained model from %s' % opt.load_path)
"""predict"""
for img_path in tqdm(img_paths):
raw_img = read_image(img_path, color=True)
# plot predict bboxes
b_bboxes, b_labels, b_scores = rfcn_trainer.r_fcn.predict(
[raw_img], visualize=True)
pred_img = visdom_bbox(raw_img, tonumpy(b_bboxes[0]),
tonumpy(b_labels[0]).reshape(-1),
tonumpy(b_scores[0]))
file_name, file_ext = os.path.splitext(os.path.basename(img_path))
result = np.hstack([
b_labels[0][:, np.newaxis], b_scores[0][:, np.newaxis], b_bboxes[0]
])
# output to file
file_out_path = os.path.join(opt.predict_output_dir,
'res_' + file_name + '.txt')
np.savetxt(file_out_path, result, fmt='%.2f', delimiter=',')
img_out_path = os.path.join(opt.predict_output_dir,
file_name + '_res.jpg')
pred_img = np.flipud(pred_img).transpose((1, 2, 0)) * 255
cv2.imwrite(img_out_path, pred_img)
print('Done!')
def test(**kwargs):
opt._parse(kwargs)
faster_rcnn = FasterRCNNVGG16()
trainer = FasterRCNNTrainer(faster_rcnn).cuda()
trainer.load(
'C:/Users/86188/Desktop/simple-faster-rcnn-pytorch-master/checkpoints/fasterrcnn_08042317_0.9090909090909093'
)
print('load successs!')
img = read_image('test_img/test.jpg')
img = t.from_numpy(img)[None]
opt.caffe_pretrain = False # this model was trained from caffe-pretrained model
_bboxes, _labels, _scores = trainer.faster_rcnn.predict(img,
visualize=True)
test_img = visdom_bbox(at.tonumpy(img[0]), at.tonumpy(_bboxes[0]),
at.tonumpy(_labels[0]).reshape(-1),
at.tonumpy(_scores[0]).reshape(-1))
trainer.vis.img('test_img', test_img)
def detec_test_pic(pth, pic_test):
opt.load_path = opt.caffe_pretrain_path
opt.env = 'detec-tset-pic'
faster_rcnn = FasterRCNNVGG16()
trainer = FasterRCNNTrainer(faster_rcnn).cuda()
trainer.load(pth)
opt.caffe_pretrain = True # this model was trained from caffe-pretrained model
pic_index = 0
for pic in tqdm(os.listdir(pic_test)):
time.sleep(1)
img = read_image(os.path.join(pic_test, pic))
img = t.from_numpy(img)[None]
_bboxes, _labels, _scores = trainer.faster_rcnn.predict(img,
visualize=True)
pred_img = visdom_bbox(at.tonumpy(img[0]), at.tonumpy(_bboxes[0]),
at.tonumpy(_labels[0]).reshape(-1),
at.tonumpy(_scores[0]).reshape(-1))
trainer.vis.img('pred_img', pred_img)
pic_index += 1
if pic_index > 1000:
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
)
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): # *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)
faster_rcnn = FasterRCNNVGG16()
print('model construct completed')
trainer = FasterRCNNTrainer(faster_rcnn).cuda()
if opt.is_distilltion == False:
iteration_number = 10
path = opt.voc_data_dir + '/ImageSets/Main/trainval.txt'
datatxt = 0
f = open(path, "r")
for i in range(5000):
if i % 500 == 0:
datatxt = datatxt + 1
f2 = open(
opt.voc_data_dir + '/ImageSets/Main/' + str(datatxt) +
'.txt', "w")
f2.write(f.readline())
else:
iteration_number = 1
if opt.load_path:
trainer.load(opt.load_path)
print('load pretrained model from %s' % opt.load_path)
for jj in range(iteration_number):
t.cuda.empty_cache()
if jj > 0:
opt.datatxt = str(int(opt.datatxt) + 1)
opt.load_path = best_path
# 样本挖掘
print(opt.datatxt)
if opt.is_example_mining == True and opt.load_path != None:
if opt.example_type == 'mAP':
example_mining_map(trainer, opt.datatxt)
elif opt.example_type == 'loss':
example_mining_loss(opt.datatxt)
elif opt.example_type == 'diversity':
example_mining_diversity(trainer, opt.datatxt)
elif opt.example_type == 'mAP_diversity':
example_mining_map_diversity(trainer, opt.datatxt)
else:
example_mining_map_loss(trainer, opt.datatxt)
print('example mining completed')
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)
testset_all = TestDataset(opt, 'test')
test_all_dataloader = data_.DataLoader(
testset_all,
batch_size=1,
num_workers=opt.test_num_workers,
shuffle=False,
pin_memory=True)
# 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
# print(lr_)
t.cuda.empty_cache()
for epoch in range(opt.epoch):
t.cuda.empty_cache()
print('epoch=%d' % epoch)
if opt.example_type != 'mAP':
# 计算loss的数组初始化
loss = np.zeros(10000)
ID = list()
# 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 enumerate(dataloader):
if len(bbox_) == 0:
continue
t.cuda.empty_cache()
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_distilltion == True:
# inx = str(id_[0])
# inx = int(inx[-5:])
# teacher_pred_bboxes = pred_bboxes[int(index[inx])]
# teacher_pred_labels = pred_labels[int(index[inx])]
# teacher_pred_features_ = pred_features[int(index[inx])]
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_bboxes = teacher_pred_bboxes.astype(
np.float32)
teacher_pred_labels = teacher_pred_labels.astype(np.int32)
teacher_pred_bboxes_ = at.totensor(teacher_pred_bboxes)
teacher_pred_labels_ = at.totensor(teacher_pred_labels)
teacher_pred_bboxes_ = teacher_pred_bboxes_.cuda()
teacher_pred_labels_ = teacher_pred_labels_.cuda()
teacher_pred_features_ = teacher_pred_features_.cuda()
losses = trainer.train_step(img, bbox, label, scale, epoch,
teacher_pred_bboxes_,
teacher_pred_labels_,
teacher_pred_features_)
else:
losses = trainer.train_step(img, bbox, label, scale, epoch)
# 保存每一个样本的损失
if opt.example_type != 'mAP':
ID += list(id_)
loss[ii] = losses.total_loss
# 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)
# plot predicti bboxes
_bboxes, _labels, _scores, _ = trainer.faster_rcnn.predict(
[ori_img_], visualize=True)
print(at.tonumpy(_bboxes[0]).reshape(-1).shape)
print(at.tonumpy(_labels[0]).shape)
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 opt.example_type != 'mAP':
order = loss.argsort()[::-1]
f = open('loss.txt', "w")
for i in range(len(ID)):
f.write(ID[order[i]] + ' ' + str(loss[order[i]]) +
'\n')
f.close()
if epoch == 20:
#draw(test_dataloader, faster_rcnn, test_num=opt.test_num)
save_name = trainer.save(best_map='20')
f = open('result.txt', "a")
result = eval(test_all_dataloader,
trainer.faster_rcnn,
test_num=5000)
f.write(opt.datatxt + '\n')
f.write(save_name + '\n')
f.write(result + '\n')
f.close
print(result)
trainer.faster_rcnn.scale_lr(10)
lr_ = lr_ * 10
break
# 每10轮加载前面最好权重，并且减少学习率
if epoch % 20 == 15:
trainer.save(best_map='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):
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
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)
# opt.caffe_pretrain = True
TrainResume = False
dataset = Dataset(opt)
print('load dataset')
dataloader = data_.DataLoader(dataset,
batch_size=1,
shuffle=True,
pin_memory=True,
num_workers=opt.num_workers)
target_img_path = 'target_img.jpg'
target_img = read_image(target_img_path) / 255
target_img = torch.from_numpy(pytorch_normalze(target_img))
target_img = torch.unsqueeze(target_img, 0).numpy()
attacker = attacks.Blade_runner(train_BR=True)
if TrainResume:
attacker.load('checkpoints/attack_02152100_0.path')
# attacker = attacks_no_target.Blade_runner(train_BR=True)
faster_rcnn = FasterRCNNVGG16().eval()
faster_rcnn.cuda()
store(faster_rcnn)
trainer = BRFasterRcnnTrainer(faster_rcnn,
attacker,
layer_idx=layer_idies,
attack_mode=True).cuda()
if opt.load_path:
trainer.load(opt.load_path)
print('from %s Load model parameters' % opt.load_path)
trainer.vis.text(dataset.db.label_names, win='labels')
target_features_list = list()
img_feature = trainer.faster_rcnn(torch.from_numpy(target_img).cuda())
del img_feature
target_features = trainer.faster_rcnn.feature_maps
for target_feature_idx in target_features:
target_features_list.append(
target_features[target_feature_idx].cpu().detach().numpy())
del target_features
for epoch in range(opt.epoch):
trainer.reset_meters(BR=True)
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)
rois, roi_scores = faster_rcnn(img, flag=True)
if len(rois) != len(roi_scores):
print(
'The generated ROI and ROI score lengths are inconsistent')
trainer.train_step(img,
bbox,
label,
scale,
target_features_list,
rois=rois,
roi_scores=roi_scores)
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(BR=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,
rois=rois,
roi_scores=roi_scores)
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 and ii != 0:
best_path = trainer.save(epochs=ii, save_rcnn=False)
print('best path is %s' % best_path)
if epoch % 2 == 0:
best_path = trainer.save(epochs=epoch, save_rcnn=False)
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(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(**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'])
tensor_img = tensor_img.cuda()
# This preset filters bounding boxes with a score < 0.7
# and has to be set everytime before using predict()
self.faster_rcnn.use_preset('visualize')
pred_bboxes, pred_labels, pred_scores = self.faster_rcnn.predict(
tensor_img, [(img.shape[1], img.shape[2])])
box_filter = np.array(pred_scores[0]) > 0.7
return pred_bboxes[0][box_filter], pred_labels[0][
box_filter], pred_scores[0][box_filter]
if __name__ == '__main__':
det = PlasticDetector('checkpoints/fasterrcnn_07122125_0.5273599762268979',
True)
print('Loaded model.')
image_path = 'misc/demo.jpg'
test_image = PIL.Image.open(image_path)
print('Working on image {}'.format(image_path))
print(det.predict_image(test_image, 5))
pred_bboxes, pred_scores = det.predict_image(test_image, 1000)
pred_img = visdom_bbox(np.array(test_image.convert('RGB')).transpose(
(2, 0, 1)),
at.tonumpy(pred_bboxes[:, [1, 0, 3, 2]]),
at.tonumpy([1 for _ in pred_bboxes]),
at.tonumpy(pred_scores),
label_names=['Animal', 'BG'])
PIL.Image.fromarray((255 * pred_img).transpose(
(1, 2, 0)).astype(np.uint8)).save('output.jpg')
det.annotate_image(test_image, 5).save('output-annotate.jpg')
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
img, img_depth, bbox, label = img.cuda().float(), img_depth.cuda().float(), bbox_.cuda(), label_.cuda()
img, img_depth, bbox, label = Variable(img), Variable(img_depth), Variable(bbox), Variable(label)
trainer.train_step(img, img_depth, 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]))
ori_img_depth_ = inverse_normalize_depth(at.tonumpy(img_depth[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
<<<<<<< HEAD
_bboxes, _labels, _scores = trainer.faster_rcnn.predict(ori_img_,ori_img_depth_, visualize=True)
=======
_bboxes, _labels, _scores = trainer.faster_rcnn.predict(ori_img_, visualize=True)
>>>>>>> b43e1a358b5853ffb749ac931c9cd97a6dccf862
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)
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 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_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) # 全部的设置
dataset = Dataset(opt) # 数据集
print('load data')
dataloader = data_.DataLoader(dataset, \
batch_size=1, \
shuffle=True, \
# pin_memory=True,
num_workers=opt.num_workers)
# pin memory：锁页内存，内存为所欲为的时候为true，详情见：https://blog.csdn.net/yangwangnndd/article/details/95385628
# num worker：加载数据的线程数，默认为8。具体数值的选取由训练时间决定，当训练时间快于加载时间时则需要增加线程
# shuffle=True允许数据打乱排序
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: #接下来判断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
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): # 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
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 = at.scalar(scale)
img, bbox, label = img.cuda().float(), bbox_.cuda(), label_.cuda()
# 然后从训练数据中枚举dataloader,设置好缩放范围，将img,bbox,label,scale全部设置为可gpu加速
trainer.train_step(
img, bbox, label, scale
) # 调用trainer.py中的函数trainer.train_step(img,bbox,label,scale)进行一次参数迭代优化过程
# 判断数据读取次数是否能够整除plot_every(是否达到了画图次数)
if (ii + 1) % opt.plot_every == 0:
# 如果达到判断debug_file是否存在，用ipdb工具设置断点
if os.path.exists(opt.debug_file):
ipdb.set_trace()
# plot loss
# 调用trainer中的trainer.vis.plot_many(trainer.get_meter_data())将训练数据读取并上传完成可视化
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)
# 将每次迭代读取的图片用dataset文件里面的inverse_normalize()函数进行预处理，将处理后的图片调用Visdom_bbox
# 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)
# 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')
# 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'] # learning rate
log_info = 'lr:{}, map:{},loss:{}'.format(
str(lr_), str(eval_result['map']), str(trainer.get_meter_data()))
trainer.vis.log(log_info) # 将损失学习率以及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) # 用if判断语句永远保存效果最好的map
if epoch == 9:
trainer.load(best_path)
trainer.faster_rcnn.scale_lr(opt.lr_decay)
lr_ = lr_ * opt.lr_decay # 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
if epoch == 13:
break
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
)
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)
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 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
def train(**kwargs):
opt._parse(kwargs)
carrada = download('Carrada')
train_set = Carrada().get('Train')
val_set = Carrada().get('Validation')
test_set = Carrada().get('Test')
train_seqs = SequenceCarradaDataset(train_set)
val_seqs = SequenceCarradaDataset(val_set)
test_seqs = SequenceCarradaDataset(test_set)
train_seqs_loader = data_.DataLoader(train_seqs, \
batch_size=1, \
shuffle=True, \
# pin_memory=True,
num_workers=opt.num_workers)
val_seqs_loader = data_.DataLoader(val_seqs,
batch_size=1,
shuffle=False,
# pin_memory=True,
num_workers=opt.num_workers)
test_seqs_loader = data_.DataLoader(test_seqs,
batch_size=1,
shuffle=False,
# pin_memory=True,
num_workers=opt.num_workers)
# faster_rcnn = FasterRCNNVGG16(n_fg_class=3)
# faster_rcnn = FasterRCNNRESNET101(n_fg_class=3)
faster_rcnn = FasterRCNNRESNET18(n_fg_class=3)
print('model construct completed')
trainer = FasterRCNNTrainer(faster_rcnn).cuda()
scheduler = ExponentialLR(trainer.faster_rcnn.optimizer, gamma=0.9)
if opt.load_path:
trainer.load(opt.load_path)
print('load pretrained model from %s' % opt.load_path)
writer_path = os.path.join(opt.logs_path, opt.model_name)
os.makedirs(writer_path, exist_ok=True)
writer = SummaryWriter(writer_path)
iteration = 0
best_map = 0
lr_ = opt.lr
for epoch in range(opt.epoch):
print('Processing epoch: {}/{}'.format(epoch, opt.epoch))
trainer.reset_meters()
for n_seq, sequence_data in tqdm(enumerate(train_seqs_loader)):
seq_name, seq = sequence_data
path_to_frames = os.path.join(carrada, seq_name[0])
train_frame_set = CarradaDataset(opt, seq, 'box', opt.signal_type,
path_to_frames)
train_frame_loader = data_.DataLoader(train_frame_set,
batch_size=1,
shuffle=False,
num_workers=opt.num_workers)
for ii, (img, bbox_, label_, scale) in tqdm(enumerate(train_frame_loader)):
iteration += 1
scale = at.scalar(scale)
img, bbox, label = img.cuda().float(), bbox_.cuda(), label_.cuda()
img = normalize(img)
if opt.debug_step and (iteration+1) % opt.debug_step == 0:
trainer.train_step(img, bbox, label, scale, stop=True)
else:
trainer.train_step(img, bbox, label, scale)
if (iteration + 1) % opt.plot_every == 0:
if os.path.exists(opt.debug_file):
ipdb.set_trace()
train_results = trainer.get_meter_data()
writer.add_scalar('Losses/rpn_loc', train_results['rpn_loc_loss'],
iteration)
writer.add_scalar('Losses/rpn_cls', train_results['rpn_cls_loss'],
iteration)
writer.add_scalar('Losses/roi_loc', train_results['roi_loc_loss'],
iteration)
writer.add_scalar('Losses/roi_cls', train_results['roi_cls_loss'],
iteration)
writer.add_scalar('Losses/total', train_results['total_loss'],
iteration)
if (iteration + 1) % opt.img_every == 0:
ori_img_ = at.tonumpy(img[0])
gt_img = visdom_bbox(ori_img_,
at.tonumpy(bbox_[0]),
at.tonumpy(label_[0]))
gt_img_grid = make_grid(torch.from_numpy(gt_img))
writer.add_image('Ground_truth_img', gt_img_grid, iteration)
# plot predicti bboxes
_bboxes, _labels, _scores = trainer.faster_rcnn.predict([ori_img_], opt.signal_type,
visualize=True)
# FLAG: vis
pred_img = visdom_bbox(ori_img_,
at.tonumpy(_bboxes[0]),
at.tonumpy(_labels[0]).reshape(-1),
at.tonumpy(_scores[0]))
pred_img_grid = make_grid(torch.from_numpy(pred_img))
writer.add_image('Predicted_img', pred_img_grid, iteration)
if opt.train_eval and (iteration + 1) % opt.train_eval == 0:
train_eval_result, train_best_iou = eval(train_seqs_loader, faster_rcnn,
opt.signal_type)
writer.add_scalar('Train/mAP', train_eval_result['map'],
iteration)
writer.add_scalar('Train/Best_IoU', train_best_iou,
iteration)
eval_result, best_val_iou = eval(val_seqs_loader, faster_rcnn, opt.signal_type,
test_num=opt.test_num)
writer.add_scalar('Validation/mAP', eval_result['map'],
iteration)
writer.add_scalar('Validation/Best_IoU', best_val_iou,
iteration)
lr_ = scheduler.get_lr()[0]
writer.add_scalar('learning_rate', lr_, iteration)
log_info = 'lr:{}, map:{},loss:{}'.format(str(lr_),
str(eval_result['map']),
str(trainer.get_meter_data()))
print(log_info)
if eval_result['map'] > best_map:
test_result, test_best_iou = eval(test_seqs_loader, faster_rcnn, opt.signal_type,
test_num=opt.test_num)
writer.add_scalar('Test/mAP', test_result['map'],
iteration)
writer.add_scalar('Test/Best_IoU', test_best_iou,
iteration)
best_map = eval_result['map']
best_test_map = test_result['map']
best_path = trainer.save(best_val_map=best_map, best_test_map=best_test_map)
# best_path = trainer.save(best_map=best_map)
if (epoch + 1) % opt.lr_step == 0:
scheduler.step()
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！
opt._parse(kwargs)
dataset = Dataset(opt)
print('load data')
# #Dataset完成的任务见第二次推文数据预处理部分，
# 这里简单解释一下，就是用VOCBboxDataset作为数据
# 集，然后依次从样例数据库中读取图片出来，还调用了
# Transform(object)函数，完成图像的调整和随机翻转工作
dataloader = data_.DataLoader(dataset, \
batch_size=1, \
shuffle=True, \
# pin_memory=True,
num_workers=opt.num_workers)
testset = TestDataset(opt)
# 将数据装载到dataloader中，shuffle=True允许数据打乱排序，
# num_workers是设置数据分为几批处理，同样的将测试数据集也
# 进行同样的处理，然后装载到test_dataloader中
test_dataloader = data_.DataLoader(testset,
batch_size=1,
num_workers=opt.test_num_workers,
shuffle=False, \
pin_memory=True
)
# 定义faster_rcnn=FasterRCNNVGG16()训练模型
faster_rcnn = FasterRCNNVGG16()
print('model construct completed')
# 设置trainer = FasterRCNNTrainer(faster_rcnn).cuda()将
# FasterRCNNVGG16作为fasterrcnn的模型送入到FasterRCNNTrainer
# 中并设置好GPU加速
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
# 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文件中预先定义好，属于超参数
for epoch in range(opt.epoch):
# 首先在可视化界面重设所有数据
trainer.reset_meters()
for ii, (img, bbox_, label_, scale) in tqdm(enumerate(dataloader)):
scale = at.scalar(scale)
# 然后从训练数据中枚举dataloader,设置好缩放范围，
# 将img,bbox,label,scale全部设置为可gpu加速
img, bbox, label = img.cuda().float(), bbox_.cuda(), label_.cuda()
# 调用trainer.py中的函数trainer.train_step
# (img,bbox,label,scale)进行一次参数迭代优化过程
trainer.train_step(img, bbox, label, scale)
# 判断数据读取次数是否能够整除plot_every
# (是否达到了画图次数)，如果达到判断debug_file是否存在，
# 用ipdb工具设置断点，调用trainer中的trainer.vis.
# plot_many(trainer.get_meter_data())将训练数据读取并
# 上传完成可视化
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]))
# 将每次迭代读取的图片用dataset文件里面的inverse_normalize()
# 函数进行预处理，将处理后的图片调用Visdom_bbox可视化
trainer.vis.img('gt_img', gt_img)
# plot predicti 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_,
at.tonumpy(_bboxes[0]),
at.tonumpy(_labels[0]).reshape(-1),
at.tonumpy(_scores[0]))
# 利用同样的方法将原始图片以及边框类别的
# 预测结果同样在可视化工具中显示出来
trainer.vis.img('pred_img', pred_img)
# rpn confusion matrix(meter)
# 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')
# roi confusion matrix
# 可视化ROI head的混淆矩阵
trainer.vis.img('roi_cm', at.totensor(trainer.roi_cm.conf, False).float())
# 调用eval函数计算map等指标
eval_result = eval(test_dataloader, faster_rcnn, test_num=opt.test_num)
# 可视化map
trainer.vis.plot('test_map', eval_result['map'])
# 设置学习的learning rate
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()))
# 将损失学习率以及map等信息及时显示更新
trainer.vis.log(log_info)
# 用if判断语句永远保存效果最好的map
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
# 判断epoch==13结束训练验证过程
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=2,
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(7):
trainer.reset_meters()
for ii, (img, bbox_, label_, scale,
ori_img) in tqdm(enumerate(dataloader)):
scale = at.scalar(scale)
img, bbox, label = img.cuda().float(), bbox_.cuda(), label_.cuda()
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_ = (img * 0.225 + 0.45).clamp(min=0, max=1) * 255
gt_img = visdom_bbox(
at.tonumpy(ori_img_)[0],
at.tonumpy(bbox_)[0], label_[0].numpy())
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(
at.tonumpy(ori_img[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())
if epoch == 4:
trainer.faster_rcnn.scale_lr(opt.lr_decay)
eval_result = eval(test_dataloader, faster_rcnn, test_num=1e100)
print("eval_result")
trainer.save(mAP=eval_result["map"])
def RFCN_train(**kwargs):
"""
python train.py RFCN_train
"""
"""parse params"""
opt.parse(kwargs)
opt.batch_size = 1 # force set batch_size to 1
"""load train & test dataset"""
print('load data')
train_db = TrainDataset()
train_dataloader = DataLoader(train_db,
shuffle=True,
batch_size=opt.batch_size,
num_workers=opt.num_workers,
pin_memory=False)
test_db = TestDataset()
if opt.test_num < len(test_db):
test_db = torch.utils.data.Subset(test_db,
indices=torch.arange(opt.test_num))
test_dataloader = DataLoader(test_db,
shuffle=False,
batch_size=opt.test_batch_size,
num_workers=opt.test_num_workers,
pin_memory=False)
"""create model"""
rfcn_md = RFCN_ResNet101()
print('model construct completed')
rfcn_trainer = RFCN_Trainer(rfcn_md).cuda()
if opt.load_path:
rfcn_trainer.load(opt.load_path, load_viz_idx=opt.load_viz_x)
print('load pretrained model parameters from %s' % opt.load_path)
print("lr is:", rfcn_trainer.optimizer.param_groups[0]['lr'])
rfcn_trainer.train()
rfcn_trainer.viz.text(train_db.db.CLASS_NAME, win='labels')
best_map = 0
"""training"""
for epoch in range(opt.epoch_begin, opt.total_epoch):
rfcn_trainer.reset_meters()
step = -1
for (img, bbox_, label_, scale) in tqdm(train_dataloader):
step += 1
img, bbox, label = img.cuda().float(), bbox_.cuda(), label_.cuda()
scale = scale.item()
rfcn_trainer.train_step(img, bbox, label, scale)
if (step + 1) % opt.print_interval_steps == 0:
# plot loss
for k, v in rfcn_trainer.get_meter_data().items():
rfcn_trainer.viz.line(Y=np.array([v]),
X=np.array([rfcn_trainer.viz_index]),
win=k,
opts=dict(title=k,
xlabel='px',
ylable='loss'),
update=None if rfcn_trainer.viz_index
== 0 else 'append')
rfcn_trainer.viz_index += 1
# plot ground truth bboxes
ori_img_ = inverse_normalize(tonumpy(img[0]))
gt_img = visdom_bbox(ori_img_, tonumpy(bbox_[0]),
tonumpy(label_[0]))
rfcn_trainer.viz.image(gt_img,
win='gt_img',
opts={'title': 'gt_img'})
# plot predict bboxes
b_bboxes, b_labels, b_scores = rfcn_trainer.r_fcn.predict(
[ori_img_], visualize=True)
pred_img = visdom_bbox(ori_img_, tonumpy(b_bboxes[0]),
tonumpy(b_labels[0]).reshape(-1),
tonumpy(b_scores[0]))
rfcn_trainer.viz.image(pred_img,
win='pred_img',
opts={'title': 'predict image'})
# rpn confusion matrix(meter)
rfcn_trainer.viz.text(str(
rfcn_trainer.rpn_cm.value().tolist()),
win='rpn_cm')
# roi confusion matrix
rfcn_trainer.viz.image(rfcn_trainer.roi_cm.value().astype(
np.uint8),
win='roi_cm',
opts={'title': 'roi_cm'})
# get mAP
eval_result = rfcn_md_eval(test_dataloader,
rfcn_md,
test_num=opt.test_num)
lr_ = rfcn_trainer.optimizer.param_groups[0]['lr']
log_info = 'epoch:{}, lr:{}, map:{},loss:{}'.format(
str(epoch), str(lr_), str(eval_result['map']),
str(rfcn_trainer.get_meter_data()))
# plot mAP
rfcn_trainer.viz.line(Y=np.array([eval_result['map']]),
X=np.array([epoch]),
win='test_map',
opts=dict(title='test_map',
xlabel='px',
ylable='mAP'),
update=None if epoch == 0 else 'append')
# plot log text
rfcn_trainer.log(log_info)
print(log_info)
# if eval_result['map'].item() > best_map:
cur_map = eval_result['map']
cur_path = rfcn_trainer.save(best_map=cur_map)
if cur_map > best_map:
best_map = cur_map
best_path = cur_path
print("save model parameters to path: {}".format(cur_path))
# update learning rate
if (epoch + 1) in opt.LrMilestones:
rfcn_trainer.load(best_path)
print('update trainer weights from ', best_path, ' epoch is:',
epoch)
rfcn_trainer.scale_lr(epoch=epoch, gamma=opt.lr_gamma)
