def eval(dataloader, faster_rcnn, test_num=opt.test_num):
pred_bboxes, pred_labels, pred_scores, pred_masks = list(), list(), list(
), list()
gt_bboxes, gt_labels, gt_difficults, gt_masks = list(), list(), list(
), list()
for ii, (imgs, sizes, gt_bboxes_, gt_labels_, gt_difficults_,
gt_mask_) in tqdm(enumerate(dataloader)):
pred_bboxes_, pred_labels_, pred_scores_, pred_mask_ = faster_rcnn.predict(
imgs, visualize=True)
gt_bboxes += list(gt_bboxes_.numpy())
gt_labels += list(gt_labels_.numpy())
gt_masks += list(gt_mask_.numpy())
gt_difficults += list(gt_difficults_.numpy())
pred_bboxes += pred_bboxes_
pred_labels += pred_labels_
pred_scores += pred_scores_
pred_masks.append(pred_mask_)
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 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 eval(dataloader, faster_rcnn, test_num=10000):
"""
验证
"""
#空的list
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)):
sizes = [sizes[0][0], sizes[1][0]]
#预测目标框、类别、分数存入list
pred_bboxes_, pred_labels_, pred_scores_ = faster_rcnn.predict(imgs, [sizes])
#将真实的目标框、类别、difficults存入list
gt_bboxes += list(gt_bboxes_.numpy())
gt_labels += list(gt_labels_.numpy())
gt_difficults += list(gt_difficults_.numpy())
#将预测的目标框、类别、分数存入list
pred_bboxes += pred_bboxes_
pred_labels += pred_labels_
pred_scores += pred_scores_
if ii == test_num: break
#返回dictz字典，两个key值：AP、mAP
result = eval_detection_voc(
pred_bboxes, pred_labels, pred_scores,
gt_bboxes, gt_labels, gt_difficults,
use_07_metric=True)
return result
def eval(dataloader, faster_rcnn, test_num=2000):
# pred_bboxes, pred_labels, pred_scores, pred_depth, pred_y_rot = list(), list(), list(), list(), list()
pred_bboxes, pred_labels, pred_scores = list(), list(), list()
# gt_bboxes, gt_labels, gt_difficults, gt_depths, gt_y_rots = list(), list(), list(), list(), list()
gt_bboxes, gt_labels, gt_difficults = list(), list(), list()
# for ii, (imgs, sizes, gt_bboxes_, gt_labels_, gt_difficults_, gt_depths_, gt_y_rots_) in tqdm(enumerate(dataloader)):
for ii, (imgs, sizes, gt_bboxes_, gt_labels_,
gt_difficults_) in tqdm(enumerate(dataloader)):
sizes = [sizes[0][0].item(), sizes[1][0].item()]
pred_bboxes_, pred_labels_, pred_scores_, pred_depth_, pred_y_rot_ = faster_rcnn.predict(
imgs, [sizes])
gt_bboxes += list(gt_bboxes_.numpy())
gt_labels += list(gt_labels_.numpy())
gt_difficults += list(gt_difficults_.numpy())
# gt_depths += list(gt_depths_.numpy())
# gt_y_rots_ += list(gt_y_rots_.numpy())
pred_bboxes += pred_bboxes_
pred_labels += pred_labels_
pred_scores += pred_scores_
# pred_depth += pred_depth_
# pred_y_rot += pred_y_rot_
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)
# result = 0
# print("train.py:Func:eval:result: ", result)
# Here need to repalce the evaluation into a new one
return result
def eval(dataloader, faster_rcnn, test_num=10000):
# 预测框的位置，预测框的类别和分数
pred_bboxes, pred_labels, pred_scores = list(), list(), list()
# 真实框的位置，类别，是否为明显目标
gt_bboxes, gt_labels, gt_difficults = list(), list(), list()
# 一个for循环，从 enumerate(dataloader)里面依次读取数据，
# 读取的内容是: imgs图片，sizes尺寸，gt_boxes真实框的位置
# gt_labels真实框的类别以及gt_difficults
for ii, (imgs, sizes, gt_bboxes_, gt_labels_, gt_difficults_) in tqdm(enumerate(dataloader)):
sizes = [sizes[0][0].item(), sizes[1][0].item()]
# 用faster_rcnn.predict(imgs,[sizes]) 得出预测的pred_boxes_,
# pred_labels_,pred_scores_预测框位置，预测框标记以及预测框
# 的分数等等
pred_bboxes_, pred_labels_, pred_scores_ = faster_rcnn.predict(imgs, [sizes])
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
# 将pred_bbox,pred_label,pred_score ,gt_bbox,gt_label,gt_difficult
# 预测和真实的值全部依次添加到开始定义好的列表里面去，如果迭代次数等于测
# 试test_num，那么就跳出循环！调用 eval_detection_voc函数，接收上述的
# 六个列表参数，完成预测水平的评估！得到预测的结果
result = eval_detection_voc(
pred_bboxes, pred_labels, pred_scores,
gt_bboxes, gt_labels, gt_difficults,
use_07_metric=True)
return result
def eval(dataloader, faster_rcnn, test_num=10000, flagadvtrain=False, adversary=None):
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)):
sizes = [sizes[0][0].item(), sizes[1][0].item()]
# if flagadvtrain:
# scale = imgs[0].shape[2] / sizes[1]
# imgs, gt_bboxes_, gt_labels_ = imgs.cuda().float(), gt_bboxes_.cuda(), gt_labels_.cuda()
# imgs = adversary(imgs, gt_bboxes_, gt_labels_, scale)
# gt_bboxes_, gt_labels_ = gt_bboxes_.cpu(), gt_labels_.cpu()
pred_bboxes_, pred_labels_, pred_scores_ = faster_rcnn.predict(imgs, [sizes])
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 eval(dataloader, faster_rcnn, test_num=100):
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_,
id_) in enumerate(dataloader):
if len(gt_bboxes_) == 0:
continue
sizes = [sizes[0][0].item(), sizes[1][0].item()]
pred_bboxes_, pred_labels_, pred_scores_, _feature = faster_rcnn.predict(
imgs, [sizes])
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
# print len(pred_bboxes)
# 这个评价函数是返回ap 和map值 其中传入的pred_bboxes格式为3维的数组的list格式，
# 也就是说每个list都是一个3维数组(有batch的考量)
# 其他的同理
result = eval_detection_voc(pred_bboxes,
pred_labels,
pred_scores,
gt_bboxes,
gt_labels,
gt_difficults,
use_07_metric=True)
return result
def rfcn_md_eval(dataloader, r_fcn: RFCN, test_num=10000):
"""
eval model on the special dataset
:param dataloader: test dataset
:param r_fcn: model
:param test_num:
:return:
"""
pred_bboxes, pred_labels, pred_scores = list(), list(), list()
gt_bboxes, gt_labels, gt_difficults = list(), list(), list()
ii = -1
for (imgs, sizes, gt_bboxes_, gt_labels_,
gt_difficults_) in tqdm(dataloader):
ii += 1
sizes = [sizes[0][0].item(), sizes[1][0].item()]
imgs = imgs.cuda()
pred_bboxes_, pred_labels_, pred_scores_ = r_fcn.predict(imgs, [sizes])
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 eval(dataloader, faster_rcnn, test_num=10000):
pred_bboxes, pred_labels, pred_scores = list(), list(), list()
gt_bboxes, gt_labels, gt_difficults = list(), list(), list()
tk = tqdm(dataloader, total=int(len(dataloader)))
for (imgs, sizes, gt_bboxes_, gt_labels_, gt_difficults_) in tk:
sizes = [sizes[0][0].item(), sizes[1][0].item()]
pred_bboxes_, pred_labels_, pred_scores_ = faster_rcnn.predict(
imgs, [sizes])
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_
result = eval_detection_voc(pred_bboxes,
pred_labels,
pred_scores,
gt_bboxes,
gt_labels,
gt_difficults,
use_07_metric=True,
iou_thresh=0.2)
tk.set_postfix(score=result['map'])
return result
def eval(dataloader, faster_rcnn, 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)):
'''print("shape: ", imgs.shape)'''
sizes = [sizes[0][0].item(), sizes[1][0].item()]
pred_bboxes_, pred_labels_, pred_scores_ = faster_rcnn.predict(imgs, [sizes])
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)
i=0
for ap in result['ap']:
print(('AP for {} = {:.4f}'.format(VOC_BBOX_LABEL_NAMES[i], ap)))
i+=1
return result
def eval(dataloader, faster_rcnn, test_num=10000):
pred_bboxes, pred_labels, pred_scores = list(), list(), list()
gt_bboxes, gt_labels, gt_difficults = list(), list(), list()
gt_image_ids = list()
for ii, (imgs, sizes, gt_bboxes_, gt_labels_, gt_difficults_,
image_id) in tqdm(enumerate(dataloader)):
sizes = [sizes[0][0].item(), sizes[1][0].item()]
pred_bboxes_, pred_labels_, pred_scores_ = faster_rcnn.predict(
imgs, [sizes])
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_
gt_image_ids.append(image_id)
if ii == test_num: break
#if ii == 10: break
save_result_json(pred_bboxes, pred_labels, pred_scores, gt_bboxes,
gt_labels, gt_image_ids)
result = eval_detection_voc(pred_bboxes,
pred_labels,
pred_scores,
gt_bboxes,
gt_labels,
gt_difficults,
use_07_metric=True)
return result
def eval(dataloader, faster_rcnn, test_num=10000, prob_thre=0.7):
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)):
# imgs here are reshaped images
# sizes here are the original shape of images
sizes = [sizes[0][0], sizes[1][0]]
pred_bboxes_, pred_labels_, pred_scores_ = faster_rcnn.predict(
imgs, [sizes], prob_thre=prob_thre)
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 eval_model(dataloader, faster_rcnn, 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)):
sizes = [sizes[0][0].item(), sizes[1][0].item()]
#print('Eval start time - {}'.format(datetime.datetime.now().time()))
pred_bboxes_, pred_labels_, pred_scores_ = faster_rcnn.predict(
imgs, [sizes])
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 eval(dataloader, faster_rcnn, category=None, file=None, 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)):
sizes = [sizes[0][0].item(), sizes[1][0].item()]
pred_bboxes_, pred_labels_, pred_scores_ = faster_rcnn.predict(imgs, [sizes])
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)
i = 0
if file is not None:
file.write(str(category) + "," + str(result['map']) + ',')
for ap in result['ap']:
print(('AP for {} = {:.4f}'.format(VOC_BBOX_LABEL_NAMES[i], ap)))
if file is not None:
file.write(str(ap) + ",")
i += 1
if file is not None:
file.write("\n")
return result
def eval(self, dataloader, faster_rcnn, test_num=10000):
pred_bboxes, pred_labels, pred_scores = list(), list(), list()
gt_bboxes, gt_labels, gt_difficults = list(), list(), list()
total_losses = list()
for ii, (imgs, sizes, gt_bboxes_, gt_labels_, scale, gt_difficults_) \
in tqdm.tqdm(enumerate(dataloader)):
img = imgs.cuda().float()
bbox = gt_bboxes_.cuda()
label = gt_labels_.cuda()
sizes = [sizes[0][0].item(), sizes[1][0].item()]
pred_bboxes_, pred_labels_, pred_scores_ = \
faster_rcnn.predict(imgs, [sizes])
losses = self.trainer.forward(img, bbox, label, float(scale))
total_losses.append(losses.total_loss.item())
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)
total_loss = sum(total_losses) / len(total_losses)
return total_loss, result
def eval(dataloader, faster_rcnn, 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)):
sizes = [sizes[0][0], sizes[1][0]]
pred_bboxes_, pred_labels_, pred_scores_ = faster_rcnn.predict(
imgs, [sizes])
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 every_map(pred_bboxes,
pred_labels,
pred_scores,
gt_bboxes1,
gt_labels1,
gt_difficults1=None,
iou_thresh=0.5):
map_result = np.zeros((len(pred_labels)))
pred_bboxes = iter(pred_bboxes)
pred_labels = iter(pred_labels)
pred_scores = iter(pred_scores)
gt_bboxes1 = iter(gt_bboxes1)
gt_labels1 = iter(gt_labels1)
if gt_difficults1 is None:
gt_difficults1 = itertools.repeat(None)
else:
gt_difficults1 = iter(gt_difficults1)
n_pos = defaultdict(int)
score = defaultdict(list)
match = defaultdict(list)
# print map_result.shape
i = 0
for pred_bbox, pred_label, pred_score, gt_bbox, gt_label, gt_difficult in \
six.moves.zip(
pred_bboxes, pred_labels, pred_scores,
gt_bboxes1, gt_labels1, gt_difficults1):
pred_bboxes_, pred_labels_, pred_scores_ = list(), list(), list()
gt_bboxes_, gt_labels_, gt_difficults_ = list(), list(), list()
bbox1 = np.expand_dims(gt_bbox, axis=0)
label1 = np.expand_dims(pred_label, axis=0)
labels1 = np.expand_dims(gt_label, axis=0)
bounding1 = np.expand_dims(pred_bbox, axis=0)
confidence1 = np.expand_dims(pred_score, axis=0)
difficults1 = np.expand_dims(gt_difficult, axis=0)
gt_bboxes_ += list(bbox1)
gt_labels_ += list(labels1)
gt_difficults_ += list(difficults1)
pred_bboxes_ += list(bounding1)
pred_labels_ += list(label1)
pred_scores_ += list(confidence1)
# print pred_bboxes_[0].shape
result = eval_detection_voc(pred_bboxes_,
pred_labels_,
pred_scores_,
gt_bboxes_,
gt_labels_,
gt_difficults_,
use_07_metric=True)
map_result[i] = result['map']
i += 1
return map_result
def eval(dataloader, faster_rcnn, test_num, thresh=0.01):
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 enumerate(dataloader):
sizes = [sizes[0][0].item(), sizes[1][0].item()]
pred_bboxes_, pred_labels_, pred_scores_ = faster_rcnn.predict(
imgs, [sizes])
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
#print("gt_bboxes: ", len(gt_bboxes), gt_bboxes)
#print("gt_labels: ", len(gt_labels), gt_labels)
#print("gt_difficults: ", len(gt_difficults), gt_difficults)
#print("pred_bboxes: ", len(pred_bboxes), pred_bboxes)
#print("pred_labels: ", len(pred_labels), pred_labels)
#print("pred_scores: ", len(pred_scores), pred_scores)
#eval = Metric(visualize=False, visualization_root=None)
#filtered_score = []
#filtered_pred = []
#filtered_target = []
#for i in range(len(pred_bboxes)):
# if pred_scores[i] > thresh:
# filtered_score.append(pred_scores[i])
# filtered_pred.append(pred_bboxes[i])
# filtered_target.append(gt_bboxes[i])
#pred_list = pred_bboxes[i]
#target_list = gt_bboxes[i]
#pred_list = [p for p in pred_list if p[4] >= thresh]
#filtered_p = [p[:4] for p in pred_list]
#filtered_target = [p for p in target_list]
# image = None
# eval.eval_add_result(filtered_target, filtered_pred, image=image, image_name=i)
#precision, recall = evl.get_result()
#F1 = 2 * (precision * recall) / max((precision + recall), 1e-5)
#F2 = 5 * (precision * recall) / max((4 * precision + recall), 1e-5)
#print("detect time: ", time.time() - st)
#print("Prec: ", precision, "Rec: ", recall, "F1: ", F1, "F2: ", F2)
result = eval_detection_voc(pred_bboxes,
pred_labels,
pred_scores,
gt_bboxes,
gt_labels,
gt_difficults,
use_07_metric=True)
#print("detect time: ", time.time() - st)
print("precision and recall: ", result)
return result
def eval(seq_loader,
faster_rcnn,
signal_type,
scale=1.,
test_num=10000,
stop=False):
carrada = download('Carrada')
pred_bboxes, pred_labels, pred_scores = list(), list(), list()
gt_bboxes, gt_labels, gt_difficults = list(), list(), list()
# print('*** Evaluation ***')
for n_seq, sequence_data in tqdm(enumerate(seq_loader)):
seq_name, seq = sequence_data
# Overfit an image
# seq = [seq[155]] # large
seq = [seq[115]] # medium
# seq = [seq[28]] # small
path_to_frames = os.path.join(carrada, seq_name[0])
frame_set = TestCarradaDataset(opt, seq, 'box', signal_type,
path_to_frames)
frame_loader = data_.DataLoader(frame_set,
batch_size=1,
shuffle=False,
num_workers=opt.num_workers)
for ii, (imgs, sizes, gt_bboxes_, gt_labels_,
gt_difficults_) in tqdm(enumerate(frame_loader)):
sizes = [sizes[0][0].item(), sizes[1][0].item()]
imgs = normalize(imgs)
pred_bboxes_, pred_labels_, pred_scores_ = faster_rcnn.predict(
imgs, signal_type, [sizes])
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,
iou_thresh=0.5,
use_07_metric=True)
ious = bbox_iou(gt_bboxes[0], pred_bboxes[0])
try:
best_iou = ious.max()
except ValueError:
best_iou = 0
if stop:
import ipdb
ipdb.set_trace()
# print('Best IoU in validation: {}'.format(ious.max()))
return result, best_iou
def diversity_map(pred_bboxes, pred_labels, pred_scores, gt_bboxes1,
gt_labels1, gt_difficult1):
pred_bboxes_, pred_labels_, pred_scores_ = list(), list(), list()
gt_bboxes_, gt_labels_, gt_difficults_ = list(), list(), list()
result = eval_detection_voc(pred_bboxes,
pred_labels,
pred_scores,
gt_bboxes1,
gt_labels1,
use_07_metric=True)
return result['map']
def test(model, yolo_loss, dataloader, opt):
model.eval()
pred_bboxes, pred_labels, pred_scores = list(), list(), list()
gt_bboxes, gt_labels = list(), list()
for batch_i, (imgs, targets) in tqdm(enumerate(dataloader), total=len(dataloader)):
imgs = imgs.cuda()
targets = targets.numpy()
with torch.no_grad():
output = model(imgs)
all_output = []
for yolo, out in zip(yolo_loss, output):
all_output.append(yolo(out))
output = torch.cat(all_output, 1)
output = non_max_suppression(output, opt.num_classes, conf_thres=opt.conf_thres, nms_thres=opt.nms_thres)
# Compute average precision for each sample
for sample_i in range(targets.shape[0]):
# Get labels for sample where width is not zero (dummies)
annotations = targets[sample_i, targets[sample_i, :, 3] != 0]
# Extract detections
detections = output[sample_i]
if detections is None:
# If there are no detections but there are annotations mask as zero AP
detection_bboxes = np.zeros((0, 4))
detection_labels = np.zeros(0)
detection_scores = np.zeros(0)
else:
detections = detections.cpu().numpy()
detection_bboxes = detections[:, 0:4]
detection_scores = detections[:, 5]
detection_labels = detections[:, 6]
# Extract target boxes as (x1, y1, x2, y2)
target_boxes = np.zeros((len(annotations), 4))
target_boxes[:, 0] = (annotations[:, 1] - annotations[:, 3] / 2)
target_boxes[:, 1] = (annotations[:, 2] - annotations[:, 4] / 2)
target_boxes[:, 2] = (annotations[:, 1] + annotations[:, 3] / 2)
target_boxes[:, 3] = (annotations[:, 2] + annotations[:, 4] / 2)
target_boxes *= opt.img_size
target_labels = annotations[:, 0]
pred_bboxes.append(detection_bboxes)
pred_labels.append(detection_labels.astype(np.int))
pred_scores.append(detection_scores)
gt_bboxes.append(target_boxes)
gt_labels.append(target_labels.astype(np.int))
result = eval_detection_voc(
pred_bboxes, pred_labels, pred_scores,
gt_bboxes, gt_labels,
use_07_metric=True)
print("Mean Average Precision:" + str(result))
return result["map"]
def eval(dataloader, faster_rcnn, test_num=10000):
pred_bboxes, pred_labels, pred_scores = list(), list(), list()#预测结果List包含，框，类别，分数。
gt_bboxes, gt_labels, gt_difficults = list(), list(), list()#验证的gt框，类别，difficult疑问。
for ii, (imgs, sizes, gt_bboxes_, gt_labels_, gt_difficults_) in tqdm(enumerate(dataloader)):
sizes = [sizes[0][0].item(), sizes[1][0].item()] #尺寸的计算方式，疑问？
pred_bboxes_, pred_labels_, pred_scores_ = faster_rcnn.predict(imgs, [sizes]) #输入图片以及图片的尺寸进入模型得到预测结果
gt_bboxes += list(gt_bboxes_.numpy()) #真实框，存放N张图片的bboxes的list,[N,(X,4)]
gt_labels += list(gt_labels_.numpy()) #真实类别，存放N张图片的labels的list，[N,(X,)?]
gt_difficults += list(gt_difficults_.numpy()) #真实difficult,存放N张图片的bboxes的难易程度，[N,(X,)]
pred_bboxes += pred_bboxes_ #预测框,存放N张图片的预测bboxes的list,[N,(X,4)]
pred_labels += pred_labels_ #预测类别,存放N张图片的预测labels的list，[N,(X,)]
pred_scores += pred_scores_ #预测分数,存放N张图片的预测label的分数的list,[N,(X,)]
if ii == test_num: break #限制数量，预测test_num张图片
result = eval_detection_voc( #用voc评估方法评估
pred_bboxes, pred_labels, pred_scores,
gt_bboxes, gt_labels, gt_difficults,
use_07_metric=True)
return result
def eval(dataloader, faster_rcnn, test_num=981):
"""
验证
"""
#空的list
print('evaling')
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)):
sizes = [int(sizes[0][0]), int(sizes[1][0])]
#预测目标框、类别、分数存入list
pred_bboxes_, pred_labels_, pred_scores_ = faster_rcnn.predict(
imgs, [sizes])
#将真实的目标框、类别、difficults存入list
# print(pred_labels_)
print('第' + str(ii + 1) + '个图片预测结果个数:' + str(len(pred_labels_[0])))
gt_bboxes += list(gt_bboxes_.numpy())
gt_labels += list(gt_labels_.numpy())
gt_difficults += list(gt_difficults_.numpy())
#将预测的目标框、类别、分数存入list
pred_bboxes += pred_bboxes_
pred_labels += pred_labels_
pred_scores += pred_scores_
if ii == test_num: break
# if ii == 5: break
#返回dictz字典，两个key值：AP、mAP
# print
# for i in range(len(pred_labels)):
# print(len(pred_labels[i]))
# print('pre num is ', len(pred_labels))
results = []
for iou_thresh in [0.5, 0.55, 0.6, 0.65, 0.7, 0.75, 0.8, 0.85, 0.9, 0.95]:
result = eval_detection_voc(pred_bboxes,
pred_labels,
pred_scores,
gt_bboxes,
gt_labels,
gt_difficults,
use_07_metric=False,
iou_thresh=iou_thresh)
results.append(result)
return results
def eval(dataloader, faster_rcnn, 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)):
sizes = [sizes[0][0], sizes[1][0]]
pred_bboxes_, pred_labels_, pred_scores_ = faster_rcnn.predict(imgs, [sizes])
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 eval(seq_loader, faster_rcnn, signal_type, test_num=10000):
carrada = download('Carrada')
pred_bboxes, pred_labels, pred_scores = list(), list(), list()
gt_bboxes, gt_labels, gt_difficults = list(), list(), list()
# print('*** Evaluation ***')
for n_seq, sequence_data in tqdm(enumerate(seq_loader)):
seq_name, seq = sequence_data
path_to_frames = os.path.join(carrada, seq_name[0])
frame_set = TestCarradaDataset(opt, seq, 'box', signal_type,
path_to_frames)
frame_loader = data_.DataLoader(frame_set,
batch_size=1,
shuffle=False,
num_workers=opt.num_workers)
for ii, (imgs, sizes, gt_bboxes_, gt_labels_, gt_difficults_) in tqdm(enumerate(frame_loader)):
sizes = [sizes[0][0].item(), sizes[1][0].item()]
imgs = normalize(imgs)
pred_bboxes_, pred_labels_, pred_scores_ = faster_rcnn.predict(imgs, [sizes])
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)
ious = bbox_iou(gt_bboxes[0], pred_bboxes[0])
try:
best_iou = ious.max()
except ValueError:
best_iou = 0
return result, best_iou
def eval(dataloader, faster_rcnn, test_num=10000):
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)
return result
def eval(dataloader, faster_rcnn, test_num=10000):
pred_bboxes, pred_labels, pred_scores = list(), list(), list()
gt_bboxes, gt_labels, gt_difficults = list(), list(), list
for ii, (img, size, gt_bbox, gt_label,
gt_difficult) in tqdm(enumerate(dataloader)):
size = [size[0][0], size[1][0]] #why?
pred_bbox, pred_label, pred_score = faster_rcnn.predict(
img.numpy(), [size])
gt_bboxes += list(gt_bbox.numpy())
gt_labels += list(gt_label.numpy())
gt_difficults += list(gt_difficult.numpy())
pred_bboxes += pred_bbox
pred_labels += pred_label
pred_scores += pred_score
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 eval(dataloader, faster_rcnn, test_num=10000): # 评价函数
pred_bboxes, pred_labels, pred_scores = list(), list(), list() # 预测结果
gt_bboxes, gt_labels, gt_difficults = list(), list(), list() # 真实结果
for ii, (imgs, sizes, gt_bboxes_, gt_labels_,
gt_difficults_) in tqdm(enumerate(dataloader)): # ii 用来指示当前进度
sizes = [sizes[0][0], sizes[1][0]]
pred_bboxes_, pred_labels_, pred_scores_ = faster_rcnn.predict(
imgs, [sizes]) # 预测图片
gt_bboxes += list(gt_bboxes_.numpy())
gt_labels += list(gt_labels_.numpy())
gt_difficults += list(gt_difficults_.numpy()) # 这哥们英语一般
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) # 计算预测好坏情况
return result
def test(**kwargs):
opt._parse(kwargs)
device = opt.device
cnn = AlexNet()
in_features = cnn.classifier[6].in_features
cnn.classifier[6] = t.nn.Linear(in_features, 21)
cnn.load(opt.load_trained_path)
cnn.to(device)
svms = SVMs(method='load')
lrs = LRs(method='load')
dataset = NormalDataset(opt.voc_data_dir, split='test')
dataloader = t.utils.data.DataLoader(dataset,
batch_size=1,
num_workers=0,
shuffle=False)
ipdb.set_trace()
pred_bbox_last_, pred_label_last_, pred_score_last_, gt_bbox_, gt_label_ = [],[],[],[],[]
for ii, (img, bbox_imgs, bboxes, gt_bbox, gt_label,
_) in tqdm(enumerate(dataloader)):
# for ii in tqdm(range(len(dataset))):
# img,bbox_imgs,bboxes,gt_bbox,gt_label,_ = dataset.getitem(ii)
img = img[0].numpy()
bboxes = bboxes[0].numpy()
gt_bbox = gt_bbox[0].numpy()
gt_label = gt_label[0].numpy()
inputs = bbox_imgs.to(device)[0]
features = cnn.get_features(inputs).cpu().detach().numpy()
'''
imOut = cv2.UMat(img.numpy().transpose((1,2,0))*255).get()
for i, rect in enumerate(bboxes):
ymin, xmin, ymax, xmax = rect
cv2.rectangle(imOut, (xmin, ymin), (xmax, ymax), (0, 255, 0), 1, cv2.LINE_AA)
cv2.imwrite('./pic/1.jpg',imOut)
pred_label,pred_score,pred_bg = svms.predict(features)
mask = np.where(pred_bg == False)[0]
counts += 1
if mask.size == 0:
null_counts += 1
if ii == 10:
print("null per:{}".format(1.0*null_counts/counts))
break
pred_label = pred_label[mask]
pred_score = pred_score[mask]
pred_bboxes_unregress = bboxes[mask] # np.ndarray
bbox_features = features[mask]
'''
pred_label, pred_score = svms.predict(features)
pred_bboxes_unregress = bboxes
bbox_features = features
# mark
# 获取每个种类的bboxes和score，为了nms
pred_bboxes_2_nms = dict()
pred_score_2_nms = dict()
bbox_features_2_nms = dict()
for lab in np.unique(pred_label):
lab_mask = np.where(pred_label == lab)[0]
pred_bboxes_2_nms[opt.VOC_BBOX_LABEL_NAMES[
lab]] = pred_bboxes_unregress[lab_mask]
pred_score_2_nms[
opt.VOC_BBOX_LABEL_NAMES[lab]] = pred_score[lab_mask]
bbox_features_2_nms[
opt.VOC_BBOX_LABEL_NAMES[lab]] = bbox_features[lab_mask]
# nms & regression
pred_bbox_last = []
pred_label_last = []
pred_score_last = []
for cat, bbox_nms in pred_bboxes_2_nms.items():
mask = np.where(pred_score_2_nms[cat] > opt.nms_thresh)[0]
if mask.size == 0:
continue
else:
bbox_nms = bbox_nms[mask]
pre_score_2_nms[cat] = pre_score_2_nms[cat][mask]
features_2_nms = features_2_nms[mask]
keep_mask = nms(
t.from_numpy(bbox_nms[:, [1, 0, 3, 2]]).float(),
t.from_numpy(pred_score_2_nms[cat]).float(), opt.iou_thresh)
loc = lrs.predict(cat, bbox_features_2_nms[cat][keep_mask])
pred_bbox_cat = loc2bbox(bbox_nms[keep_mask], loc)
pred_score_last.append(pred_score_2_nms[cat][keep_mask])
pred_bbox_last.append(pred_bbox_cat)
pred_label_last.extend([cat] * pred_bbox_cat.shape[0])
if len(pred_label_last) > 0:
wrong_2_draw(img, np.vstack(pred_bbox_last),
np.array(pred_label_last), gt_bbox, gt_label)
pred_bbox_last_ += np.vstack(pred_bbox_last)
pred_label_last_ += np.array(pred_label_last)
pred_score_last_ += np.hstack(pred_Score_last)
gt_bbox_ += gt_bbox
gt_label_ += gt_label
# evaluation
res = eval_detection_voc(pred_bbox_last_, pred_label_last_,
pred_score_last_, gt_bbox_, gt_label_)
print(res)
pd.DataFrame(res).to_excel('./res.xlsx')
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
sizes = [sizes[0][0], sizes[1][0]]
<<<<<<< HEAD
pred_bboxes_, pred_labels_, pred_scores_ = faster_rcnn.predict(imgs,img_depth, [sizes])
=======
pred_bboxes_, pred_labels_, pred_scores_ = faster_rcnn.predict(imgs,[sizes])
>>>>>>> b43e1a358b5853ffb749ac931c9cd97a6dccf862
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)
print('load data')
dataloader = data_.DataLoader(dataset, \
batch_size=1, \
shuffle=True, \
#pin_memory=True,
num_workers=opt.num_workers)
testset = TestDataset(opt)