def eval(model, save_path, test_path, device,threshold):
model.eval()
save_path = os.path.join(save_path, "submit")
os.makedirs(save_path, exist_ok=True)
img_path_root = os.path.join(test_path, "MSRA-TD500", "test")
# 预测所有测试图片
img_paths = [os.path.join(img_path_root, x) for x in os.listdir(img_path_root)]
for img_path in tqdm(img_paths, desc='test models'):
if not img_path.endswith('.JPG') and not img_path.endswith('.jpg'):
continue
img_name = os.path.basename(img_path).split('.')[0]
save_name = os.path.join(save_path, 'res_' + img_name + '.txt')
assert os.path.exists(img_path), 'file is not exists'
img = cv2.imread(img_path)
org_img = img.copy()
h, w = img.shape[:2]
img = scale_image(img)
# 将图片由(w,h)变为(1,img_channel,h,w)
tensor = transforms.ToTensor()(img)
tensor = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])(tensor)
tensor = tensor.unsqueeze_(0)
tensor = tensor.to(device)
with torch.no_grad():
preds = model(tensor)
preds, boxes_list = pse_decode(preds[0], config.scale, org_img)
np.savetxt(save_name, boxes_list.reshape(-1, 8), delimiter=',', fmt='%d')
# 开始计算 recall precision f1
f_score_new = get_msra_result(save_path, img_path_root)
return f_score_new
def predict(self, img: str, long_size: int = 2240):
'''
对传入的图像进行预测，支持图像地址,opecv 读取图片，偏慢
:param img: 图像地址
:param is_numpy:
:return:
'''
assert os.path.exists(img), 'file is not exists'
img = cv2.imread(img)
# img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
h, w = img.shape[:2]
scale = long_size / max(h, w)
img = cv2.resize(img, None, fx=scale, fy=scale)
# 将图片由(w,h)变为(1,img_channel,h,w)
tensor = transforms.ToTensor()(img)
tensor = tensor.unsqueeze_(0)
tensor = tensor.to(self.device)
with torch.no_grad():
torch.cuda.synchronize()
start = time.time()
preds = self.net(tensor)
preds, boxes_list = pse_decode(preds[0], self.scale)
scale = (preds.shape[1] / w, preds.shape[0] / h)
# print(scale)
# preds, boxes_list = decode(preds,num_pred=-1)
if len(boxes_list):
boxes_list = boxes_list / scale
torch.cuda.synchronize()
t = time.time() - start
return preds, boxes_list, t
def detect(self, image, img_show):
# get model output
preds = self.model.forward(image)
preds, boxes, contours = pse_decode(preds[0], self.scale,
self.threshold)
output = {'image': image, 'tr': preds, 'bbox': boxes}
return contours, output
def eval(model, workspace, test_path, device):
model.eval()
# torch.cuda.empty_cache() # speed up evaluating after training finished
img_path = os.path.join(test_path, 'Images/Test/')
save_path = os.path.join(workspace, 'output')
if os.path.exists(save_path):
shutil.rmtree(save_path, ignore_errors=True)
if not os.path.exists(save_path):
os.makedirs(save_path)
vis_ctw1500 = os.path.join(workspace, 'vis_ctw1500')
if os.path.exists(vis_ctw1500):
shutil.rmtree(vis_ctw1500, ignore_errors=True)
if not os.path.exists(vis_ctw1500):
os.makedirs(vis_ctw1500)
long_size = 1280
# 预测所有测试图片
img_paths = [os.path.join(img_path, x) for x in os.listdir(img_path)]
for idx, img_path in enumerate(tqdm(img_paths, desc='test models')):
img_name = os.path.basename(img_path).split('.')[0]
save_name = os.path.join(save_path, 'res_' + img_name + '.txt')
assert os.path.exists(img_path), 'file is not exists'
img = cv2.imread(img_path)
org_img = img.copy()
h, w = img.shape[:2]
img = scale_aligned_short(img)
tensor = transforms.ToTensor()(img)
tensor = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])(tensor)
tensor = tensor.unsqueeze_(0)
tensor = tensor.to(device)
with torch.no_grad():
preds = model(tensor)
preds, boxes_list = pse_decode(preds[0], config.scale, org_img)
if config.visualization:
for bbox in boxes_list:
bbox = np.array(bbox, np.int)
cv2.drawContours(org_img,
[bbox.reshape(int(bbox.shape[0] / 2), 2)], -1,
(0, 255, 0), 2)
org_img = cv2.resize(org_img, (640, 640))
debug(idx, img_path, [[org_img]], vis_ctw1500)
image_name = img_path.split('/')[-1].split('.')[0]
write_result_as_txt(image_name, boxes_list, save_path)
# recall precision f1
gt_path = os.path.join(test_path, 'gt/Test')
fid_path = os.path.join(workspace, 'res_tt.txt')
precision, recall, hmean = evl_totaltext(save_path, gt_path, fid_path)
# f_score_new = getresult(save_path,config.gt_name)
return precision, recall, hmean
def eval(model, save_path, test_path, device):
model.eval()
# torch.cuda.empty_cache() # speed up evaluating after training finished
img_path = os.path.join(test_path, 'img')
gt_path = os.path.join(test_path, 'gt')
if os.path.exists(save_path):
shutil.rmtree(save_path, ignore_errors=True)
if not os.path.exists(save_path):
os.makedirs(save_path)
# 预测所有测试图片
img_paths = [os.path.join(img_path, x) for x in os.listdir(img_path)]
for img_path in tqdm(img_paths, desc='test models'):
img_name = os.path.basename(img_path).split('.')[0]
save_name = os.path.join(save_path, 'res_' + img_name + '.txt')
assert os.path.exists(img_path), 'file is not exists'
img = cv2.imread(img_path)
h, w = img.shape[:2]
# if (w < h):
# if (h > MAX_LEN):
# scale = 1.0 * MAX_LEN / h
# w = w * scale
# h = MAX_LEN
# elif (h <= w):
# if (w > MAX_LEN):
# scale = 1.0 * MAX_LEN / w
# h = scale * h
# w = MAX_LEN
#
# w = int(w // 32 * 32)
# h = int(h // 32 * 32)
#
# img = cv2.resize(img, (h, w), interpolation=cv2.INTER_AREA)
# 将图片由(w,h)变为(1,img_channel,h,w)
tensor = transforms.ToTensor()(img)
tensor = tensor.unsqueeze_(0)
tensor = tensor.to(device)
with torch.no_grad():
preds = model(tensor)
preds, boxes_list = pse_decode(preds[0], config.scale)
scale = (preds.shape[1] * 1.0 / w, preds.shape[0] * 1.0 / h)
if len(boxes_list):
boxes_list = boxes_list / scale
np.savetxt(save_name,
boxes_list.reshape(-1, 8),
delimiter=',',
fmt='%d')
# 开始计算 recall precision f1
result_dict = cal_recall_precison_f1(gt_path, save_path)
return result_dict['recall'], result_dict['precision'], result_dict[
'hmean']
def eval(model, config, device, thre, long_size=1280):
model.eval()
img_path = os.path.join(config.testroot, 'test_img')
save_path = os.path.join(config.workspace, 'output_eval')
# if os.path.exists(save_path):
# shutil.rmtree(save_path, ignore_errors=True)
if not os.path.exists(save_path):
os.makedirs(save_path)
long_size = 2240
# 预测所有测试图片
img_paths = [os.path.join(img_path, x) for x in os.listdir(img_path)]
for img_path in tqdm(img_paths, desc='test models'):
img_name = os.path.basename(img_path).split('.')[0]
save_name = os.path.join(save_path, 'res_' + img_name + '.txt')
assert os.path.exists(img_path), 'file is not exists'
img = cv2.imread(img_path)
h, w = img.shape[:2]
scale = long_size / max(h, w)
img = cv2.resize(img, None, fx=scale, fy=scale)
# 将图片由(w,h)变为(1,img_channel,h,w)
tensor = transforms.ToTensor()(img)
tensor = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])(tensor)
tensor = tensor.unsqueeze_(0)
tensor = tensor.to(device)
with torch.no_grad():
preds = model(tensor)
preds, boxes_list = pse_decode(preds[0], config.scale)
scale = (preds.shape[1] * 1.0 / w, preds.shape[0] * 1.0 / h)
if len(boxes_list):
boxes_list = boxes_list / scale
np.savetxt(save_name,
boxes_list.reshape(-1, 8),
delimiter=',',
fmt='%d')
# 开始计算 recall precision f1
methodHmean, methodPrecision, methodRecall = getresult(
save_path, config.gt_name)
print("precision: {} , recall: {}, f1: {}".format(methodPrecision,
methodRecall,
methodHmean))
def valid_one_epoch(config, psenet, pseloss, optimizer, schedular, writer,
epoch):
# pbar = tqdm(total=100)
image_names, labels = load_imagename_labels(config.DATASET.LABEL_VAL_ROOT)
len_images = len(image_names)
long_edge = 2240
save_path = '/home/simple/mydemo/ocr_project/pytorch_psenet_project/pytorch_psenet/results'
if os.path.exists(save_path) == False:
os.mkdir(save_path)
gt_path = config.DATASET.LABEL_VAL_ROOT
for i in tqdm.tqdm(range(0, len_images)):
image_name = image_names[i] + '.jpg'
image = cv2.imread(
os.path.join(config.DATASET.IMAGE_VAL_ROOT, image_name))
image = cv2.cvtColor(image, code=cv2.COLOR_BGR2RGB)
image_show = image.copy()
# cv2.imshow('ttt',image_show)
h, w, c = image.shape
ratio = long_edge / max(h, w)
image = cv2.resize(image, dsize=None, fx=ratio, fy=ratio)
# image=torch.Tensor(image,dtype=torch.float32)
image = torchvision.transforms.ToTensor()(image).unsqueeze(0).to(
torch.device('cuda:' + config.CUDA.GPU))
save_name = os.path.join(save_path, 'res_' + image_names[i] + '.txt')
# put the batch dataset to GPU
# images = images.to(torch.device('cuda:' + config.CUDA.GPU))
# text_kernel_masks = text_kernel_masks.to(torch.device('cuda:' + config.CUDA.GPU))
# train_masks = train_masks.to(torch.device('cuda:' + config.CUDA.GPU))
# the model forward
with torch.no_grad():
output = psenet(image, True)
preds, boxes_list = pse_decode(output[0], config.VALID.SCALE)
scale = (preds.shape[1] * 1.0 / w, preds.shape[0] * 1.0 / h)
if len(boxes_list):
boxes_list = boxes_list / scale
np.savetxt(save_name,
boxes_list.reshape(-1, 8),
delimiter=',',
fmt='%d')
result_dict = cal_recall_precison_f1(gt_path, save_path)
print(result_dict)
return result_dict['recall'], result_dict['precision'], result_dict[
'hmean']
def eval(model, save_path, test_path, device):
model.eval()
img_path = os.path.join(test_path, 'val', 'image')
if os.path.exists(save_path):
shutil.rmtree(save_path, ignore_errors=True)
if not os.path.exists(save_path):
os.makedirs(save_path)
long_size = 2240
# 预测所有测试图片
img_paths = [os.path.join(img_path, x) for x in os.listdir(img_path)]
for img_path in tqdm(img_paths, desc='test models'):
img_name = os.path.basename(img_path).split('.')[0]
save_name = os.path.join(save_path, 'res_' + img_name + '.txt')
assert os.path.exists(img_path), 'file is not exists'
img = cv2.imread(img_path)
org_img = img.copy()
h, w = img.shape[:2]
#if max(h, w) > long_size:
# scale = long_size / max(h, w)
# img = cv2.resize(img, None, fx=scale, fy=scale)
img = scale_image(img)
# 将图片由(w,h)变为(1,img_channel,h,w)
tensor = transforms.ToTensor()(img)
tensor = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])(tensor)
tensor = tensor.unsqueeze_(0)
tensor = tensor.to(device)
with torch.no_grad():
preds = model(tensor)
preds, boxes_list = pse_decode(preds[0], config.scale, org_img,
config.is_test)
scale = (preds.shape[1] * 1.0 / w, preds.shape[0] * 1.0 / h)
if len(boxes_list):
boxes_list = boxes_list / scale
np.savetxt(save_name,
boxes_list.reshape(-1, 8),
delimiter=',',
fmt='%d')
# 开始计算 recall precision f1
f_score_new = getresult(save_path, config.gt_name)
return f_score_new
def eval(model, save_path, test_path, device):
model.eval()
# torch.cuda.empty_cache() # speed up evaluating after training finished
img_path = os.path.join(test_path, 'img')
gt_path = os.path.join(test_path, 'gt')
if os.path.exists(save_path):
shutil.rmtree(save_path, ignore_errors=True)
if not os.path.exists(save_path):
os.makedirs(save_path)
long_size = 2240
# 预测所有测试图片
img_paths = [os.path.join(img_path, x) for x in os.listdir(img_path)]
# img_paths = img_paths[:2]
for img_path in tqdm(img_paths, desc='test models'):
img_name = os.path.basename(img_path).split('.')[0]
save_name = os.path.join(save_path, 'res_' + img_name + '.txt')
assert os.path.exists(img_path), 'file is not exists'
img = cv2.imread(img_path)
h, w = img.shape[:2]
# if max(h, w) > long_size:
scale = long_size / max(h, w)
img = cv2.resize(img, None, fx=scale, fy=scale)
# 将图片由(w,h)变为(1,img_channel,h,w)
tensor = transforms.ToTensor()(img)
tensor = tensor.unsqueeze_(0)
tensor = tensor.to(device)
with torch.no_grad():
preds = model(tensor)
preds, boxes_list = pse_decode(preds[0], config.scale)
scale = (preds.shape[1] * 1.0 / w, preds.shape[0] * 1.0 / h)
if len(boxes_list):
boxes_list = boxes_list / scale
np.savetxt(save_name,
boxes_list.reshape(-1, 8),
delimiter=',',
fmt='%d')
# 开始计算 recall precision f
# print('222', save_path)
result_dict = cal_recall_precison_f1(gt_path, save_path)
# print('111', result_dict)
return result_dict['recall'], result_dict['precision'], result_dict[
'hmean']
def merge_eval(model,
save_path,
test_path,
device,
base_path=r'',
use_sub=True):
def get_filename_map(json_file):
merge_images_ids_to_corner = {
image_info["id"]: []
for image_info in json_file["images"]
}
merge_filename_to_info = {
image_info["filename"]: image_info
for image_info in json_file["images"]
}
for sub_image_info in json_file["sub_images"]:
merge_images_ids_to_corner[sub_image_info["id"]].append(
sub_image_info["corner"])
merge_filename_to_corner = {
image_info["filename"]:
merge_images_ids_to_corner[image_info["id"]]
for image_info in json_file["images"]
}
return merge_filename_to_corner, merge_filename_to_info
from utils.metric.metrics import GetScore
import json
model.eval()
metric_cls = GetScore()
get_polygon = GetPolygon()
raw_metrics = []
# torch.cuda.empty_cache() # speed up evaluating after training finished
with open(test_path) as f:
gt_file = json.load(f)
if os.path.exists(save_path):
shutil.rmtree(save_path, ignore_errors=True)
if not os.path.exists(save_path):
os.makedirs(save_path)
# 预测所有测试图片
imgs_name = [x['filename'] for x in gt_file["images"]]
filename_to_corner, filename_to_info = get_filename_map(gt_file)
for img_name in tqdm(imgs_name, desc="test models"):
img_path = os.path.join(base_path + 'img/', img_name)
assert os.path.exists(img_path), 'file is not exists'
img = cv2.imread(img_path)
h, w = img.shape[:2]
s = 0
if use_sub:
merge_pred = torch.zeros([config.n, h, w], requires_grad=False)
for corner in filename_to_corner[img_name]:
sub_im = img[corner[1]:corner[3], corner[0]:corner[2]]
# h, w = sub_im.shape[:2]
# scale = long_size / max(h, w)
# sub_im = cv2.resize(sub_im, None, fx=scale, fy=scale)
# 将图片由(w,h)变为(1,img_channel,h,w)
cv2.rectangle(img, (corner[0], corner[1]),
(corner[2], corner[3]), (255, s, 0), 5)
s += 20
tensor = transforms.ToTensor()(sub_im)
tensor = tensor.unsqueeze_(0)
tensor = tensor.to(device)
with torch.no_grad():
preds = model(tensor)
merge_pred[:, corner[1]:corner[3],
corner[0]:corner[2]] = preds[0]
preds = merge_pred
else:
# long_size = 3456
# scale = long_size / max(h, w)
# img = cv2.resize(img, None, fx=scale, fy=scale)
# 将图片由(w,h)变为(1,img_channel,h,w)
tensor = transforms.ToTensor()(img)
tensor = tensor.unsqueeze_(0)
tensor = tensor.to(device)
with torch.no_grad():
preds = model(tensor)
preds = preds[0]
preds = pse_decode(preds)
m = (preds * 100).astype(np.uint8)
m = cv2.resize(m, (1280, 720))
cv2.imshow("figure", img)
cv2.waitKey(0)
cv2.destroyWindow()
preds[preds > 0] = 1
pred_polygon = get_polygon(pred_map=preds, dest_shape=[h, w])
raw_metrics.append(
metric_cls(pred_polys=pred_polygon,
ignore_tags=np.zeros(len(
filename_to_info[img_name]["polygon"], ),
dtype=np.bool),
polys=filename_to_info[img_name]["polygon"]))
metrics = metric_cls.gather_measure(raw_metrics)
return metrics['recall'].avg, metrics['precision'].avg, metrics[
'fmeasure'].avg
def eval(model, workspace, test_path, kern_size_, device):
model.eval()
img_path = os.path.join(test_path, 'Images/Test/')
gt_path = os.path.join(test_path, 'gt/Test/')
save_path = os.path.join(workspace, 'output')
if os.path.exists(save_path):
shutil.rmtree(save_path, ignore_errors=True)
if not os.path.exists(save_path):
os.makedirs(save_path)
vis_ctw1500 = os.path.join(workspace, 'vis_ctw1500')
if os.path.exists(vis_ctw1500):
shutil.rmtree(vis_ctw1500, ignore_errors=True)
if not os.path.exists(vis_ctw1500):
os.makedirs(vis_ctw1500)
img_paths = [os.path.join(img_path, x) for x in os.listdir(img_path)]
gt_paths = [
os.path.join(gt_path, 'poly_gt_' + x.split('.')[0] + '.mat')
for x in os.listdir(img_path)
]
for idx, img_path in enumerate(tqdm(img_paths, desc='test models')):
img_name = os.path.basename(img_path).split('.')[0]
# 读取gt
gt_path_one = gt_paths[idx]
gt_box = get_bboxes(gt_path_one)
assert os.path.exists(img_path), 'file is not exists'
img = cv2.imread(img_path)
org_img = img.copy()
img = scale_aligned_short(img)
tensor = transforms.ToTensor()(img)
tensor = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])(tensor)
tensor = tensor.unsqueeze_(0)
tensor = tensor.to(device)
with torch.no_grad():
preds = model(tensor)
preds_1 = torch.sigmoid(preds[0])
preds_1 = preds_1.detach().cpu().numpy()
preds, boxes_list = pse_decode(preds[0], config.scale, org_img,
kern_size_)
if config.visualization:
for bbox in boxes_list:
bbox = np.array(bbox, np.int)
cv2.drawContours(org_img,
[bbox.reshape(int(bbox.shape[0] / 2), 2)], -1,
(0, 255, 0), 2)
for bbox in gt_box:
bbox = np.array(bbox, np.int)
cv2.drawContours(org_img,
[bbox.reshape(int(bbox.shape[0] / 2), 2)], -1,
(0, 0, 255), 2)
org_img = cv2.resize(org_img, (640, 640))
image_list = []
image_list.append(org_img)
for i in range(7):
score = (preds_1[i] * preds_1[-1]).copy().astype(np.float32)
score = cv2.resize(score, (640, 640))
score = np.expand_dims(score, -1)
score = np.concatenate((score, score, score), -1)
image_list.append(score * 255)
debug(idx, img_path, [image_list], vis_ctw1500)
image_name = img_path.split('/')[-1].split('.')[0]
write_result_as_txt(image_name, boxes_list, save_path)
# recall precision f1
gt_path = os.path.join(test_path, 'gt/Test')
fid_path = os.path.join(workspace, 'res_tt.txt')
shutil.rmtree(fid_path, ignore_errors=True)
precision, recall, hmean = evl_totaltext(save_path, gt_path, fid_path)
# f_score_new = getresult(save_path,config.gt_name)
return precision, recall, hmean
def eval(model, workspace, test_path, is_test, device):
model.eval()
# torch.cuda.empty_cache() # speed up evaluating after training finished
if is_test:
img_path = os.path.join(test_path, 'test')
save_path = os.path.join(workspace, 'output_test')
else:
img_path = os.path.join(test_path, 'val', 'image')
save_path = os.path.join(workspace, 'output')
gt_path = os.path.join(test_path, 'gt/Test/')
if os.path.exists(save_path):
shutil.rmtree(save_path, ignore_errors=True)
if not os.path.exists(save_path):
os.makedirs(save_path)
vis_icd17 = os.path.join(workspace, 'vis_icd17')
if os.path.exists(vis_icd17):
shutil.rmtree(vis_icd17, ignore_errors=True)
if not os.path.exists(vis_icd17):
os.makedirs(vis_icd17)
short_size = 1600
# 预测所有测试图片
img_paths = [os.path.join(img_path, x) for x in os.listdir(img_path)]
gt_paths = [
os.path.join(gt_path, 'poly_gt_' + x.split('.')[0] + '.mat')
for x in os.listdir(img_path)
]
for idx, img_path_one in enumerate(tqdm(img_paths, desc='test models')):
img_name = os.path.basename(img_path_one).split('.')[0]
if is_test:
save_name = os.path.join(
save_path, 'res_' + img_name.split('ts_')[-1] + '.txt')
else:
save_name = os.path.join(save_path, 'res_' + img_name + '.txt')
assert os.path.exists(img_path_one), 'file is not exists'
img = readImg(img_path_one)
org_img = img.copy()
h, w = img.shape[:2]
img = scale_image(img, short_size)
tensor = transforms.ToTensor()(img)
tensor = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])(tensor)
tensor = tensor.unsqueeze_(0)
tensor = tensor.to(device)
with torch.no_grad():
preds = model(tensor)
preds, boxes_list = pse_decode(preds[0], config.scale, org_img,
is_test)
# scale = (preds.shape[1] * 1.0 / w, preds.shape[0] * 1.0 / h)
# if len(boxes_list):
# boxes_list = boxes_list / scale
if is_test:
np.savetxt(save_name,
boxes_list.reshape(-1, 9),
delimiter=',',
fmt='%d')
else:
np.savetxt(save_name,
boxes_list.reshape(-1, 8),
delimiter=',',
fmt='%d')
# recall precision f1
if not is_test:
f_score_new = getresult(save_path, config.gt_name)
return f_score_new
