我跑了ICDAR2015数据集，loss一般在05左右，降不下去了，训练集上测试效果几乎没效果，请问是还没有收敛吗？

In datasets.py line 351: "label = sample['label'].reshape([-1, 4, 2])", should be changed to "label = np.array(sample['label']['coor']).reshape([-1, 4, 2])".

Because sample['label'] is dict type, and it only need coordinates to draw the contours.

自己的数据集，修改好训练，报这个错，有大神遇到吗？
raise RuntimeError('each element in list of batch should be of equal size')
RuntimeError: each element in list of batch should be of equal size

where to download best accuaracy model? 哪里下载最佳模型？

hello,
Thanks for your sharing, I modify your code to complete my task based on pytorch. After fininshed my training process to test, it appears an error which is "IndexError: index 2 is out of bounds for dimension 0 with size 2"
Traceback (most recent call last):
File "main.py", line 184, in
test_model()
File "main.py", line 51, in test_model
all_boxes = postprocess.mask_to_box(out_1, out_2)
File "/home/liuchao/PixelLink-with-pytorch/postprocess.py", line 59, in mask_to_box
link_neighbors[:, i] = link_neighbors[:, i] & pixel_class[i]
IndexError: index 2 is out of bounds for dimension 0 with size 2
I'm grateful for u to help me solve the error above,Thanks~

File "main.py", line 138, in
main()
File "main.py", line 128, in main
train(config.epoch, iteration, dataloader, my_net, optimizer, optimizer2, device)
File "main.py", line 59, in train
for i_batch, sample in enumerate(dataloader):
File "/home/denny/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 317, in next
batch = self.collate_fn([self.dataset[i] for i in indices])
File "/home/denny/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 317, in
batch = self.collate_fn([self.dataset[i] for i in indices])
File "/home/denny/github_open_Sources/PixelLink-with-pytorch/datasets.py", line 84, in getitem
image, label = self.train_data_transform(index)
File "/home/denny/github_open_Sources/PixelLink-with-pytorch/datasets.py", line 103, in train_data_transform
img = self.read_image(self.images_dir, index)
File "/home/denny/github_open_Sources/PixelLink-with-pytorch/datasets.py", line 34, in read_image
image = ImgTransform.ReadImage(filename)
File "/home/denny/github_open_Sources/PixelLink-with-pytorch/ImgLib/ImgTransform.py", line 16, in ReadImage
with Image.open(filename) as img:
File "/home/denny/anaconda3/envs/pytorch/lib/python3.6/site-packages/PIL/Image.py", line 2609, in open
fp = builtins.open(filename, "rb")
FileNotFoundError: [Errno 2] No such file or directory: 'training_images/images/img_tensor(956).jpg'

hello, thanks for your codes and am interested in it..
As you said this version is still under developing, is there any new updates?
I didn't find the functions for testing and this version can only train IC15 dataset, right?
When will you upload new codes...really looking forward your reply.
Thanks again!!

我改了数据加载这里，看了下没啥问题，为啥会爆炸呢？请指教一下
reade_label()
with codecs.open(os.path.join(dir, self.labels_name[i-1]), encoding="utf-8_sig") as file:
data = file.readlines()
temp_str = ''
for l in data:
temp_str += l.strip('\r\n')
temp_json = json.loads(temp_str)
tmp = {}
tmp["coor"] = []
tmp["content"] = []
tmp["ignore"] = []
tmp["area"] = []
for line in temp_json["shapes"]:
tmp["content"].append(line["label"])
if line["label"] == "###":
tmp["ignore"].append(True)
else:
tmp["ignore"].append(False)
coor = []
if len(line["points"]) != 4:
print(os.path.join(dir, self.labels_name[i - 1]))
print(line["points"])
for n in line["points"]:
coor.append(int(n[0]))
coor.append(int(n[1]))
tmp["coor"].append(coor)
coor = np.array(coor).reshape([4, 2])
tmp["area"].append(cv2.contourArea(coor))

            # for line in data:
            #     content = line.split(",")
            #     coor = [int(n) for n in content[:8]]
            #
            #     content[8] = content[8].strip("\r\n")
            #     tmp["content"].append(content[8])
            #     if content[8] == "###":
            #         tmp["ignore"].append(True)
            #     else:
            #         tmp["ignore"].append(False)
            #     coor = np.array(coor).reshape([4,2])
            #     tmp["area"].append(cv2.contourArea(coor))
            res[i-1] = tmp
    return res

另外，请问加载图像时index为什么加1，我把加1去掉了？

index += 1

    filename = os.path.join(dir, self.images_name[index] + ".jpg")

postprocess.py line 89:128 should be changed to 256.

Sorry but where can I download the pretrained model, or which version do I need?

when I train this model, it happened this situation. And I have a RTX 1080 Ti with 12GB memory.

我用你的代码跑了下ICDAR 2015的数据集，
loss降到这么多时

在训练集上的测试结果如下，

这是因为我的还没收敛吗？我第一次训练文本检测的，不知道loss降到什么时候是收敛好了

File "/workspace/PixelLink-with-pytorch/main.py", line 144, in
main()
File "/workspace/PixelLink-with-pytorch/main.py", line 133, in main
train(config.epoch, iteration, dataloader, my_net, optimizer, optimizer2, device)
File "/workspace/PixelLink-with-pytorch/main.py", line 73, in train
pixel_pos_weights)
File "/workspace/PixelLink-with-pytorch/criterion.py", line 46, in pixel_loss
self.neg_pixel_weight[i] = self.neg_pixel_weight[i] & (neg_pixel_masks[i]==1)
RuntimeError: Expected object of scalar type Byte but got scalar type Bool for argument #2 'other' in call to _th_and