作者您好，我发现random_crop()函数得到的training mask中，很多裁剪结果没有包含任何valid text instances，即train_ic15.py/ohem_single()中的pos_num的值为0，我用您的训练代码跑了10个epoch，测试部分代码以及结果如下：

...
total_sample = 0
zero_sample = 0
def ohem_single(score, gt_text, training_mask):
    pos_num = (int)(np.sum(gt_text > 0.5)) - (int)(np.sum((gt_text > 0.5) & (training_mask <= 0.5)))
    print('pos_num: ', pos_num)
    global total_sample, zero_sample
    total_sample += 1
    if pos_num==0:
        zero_sample += 1
    print('black_ratio: {}/{} = {:.4f}%'.format(zero_sample, total_sample, zero_sample*100.0/total_sample))
...

部分打印结果如下:

...
('pos_num: ', 0)
black_ratio: 1899/9999 = 18.9919%
('pos_num: ', 17474)
black_ratio: 1899/10000 = 18.9900%

Epoch: [11 | 600] LR: 0.001000
('pos_num: ', 4287)
black_ratio: 1899/10001 = 18.9881%
('pos_num: ', 26144)
black_ratio: 1899/10002 = 18.9862%
...

也就是说，训练时，有接近1/5的裁剪是完全是无效的裁剪，请问您这样设置是故意为了做augmentation吗？ 期待您的解释@whai362

先是 include/pybind11/detail/common.h:113:10: fatal error: Python.h: No such file or directory #include <Python.h> ^~~~~~~~~~ compilation terminated. make: *** [adaptor.so] Error 1 这个错误 Makefile:10: recipe for target 'adaptor.so' failed

thanks for your great work and I found that as the number of text area increses, the progressive sacle expansion part consumes a lot of time. any advice to speed up the part?

To gain good performance, i have to set large long_size(>2200) which lower the efficiency. How could i balance the efficiency and performance ? Please give me some advise.

Hello sir @whai362 This repository is 5 months old and even without any source code inside it still get tons of stars. So when will you update the source code ?

作者，您好！ 首先感谢您的巨大工作，和贡献代码，我在阅读您的代码时发现icdar2015_loader.py文件中 rate = 1.0 - (1.0 - self.min_scale) / (self.kernel_num - 1) * i 这句代码和论文描述的公式不一样。

您能给一些解释吗？谢谢！为什么不一样？ri = 1 −(1 − m) × (n − i)/(n − 1) ;

Traceback (most recent call last): File "/home/yan/disk/All_python_Text/PSENet-master/test_ic15.py", line 220, in test(args) File "/home/yan/disk/All_python_Text/PSENet-master/test_ic15.py", line 154, in test pred = pse(kernels, args.min_kernel_area / (args.scale * args.scale)) File "/home/yan/disk/All_python_Text/PSENet-master/pse/init.py", line 14, in pse ret = np.array(cpse(polys, min_area), dtype='int32') TypeError: 'module' object is not callable

我在测试的时候就出现这个问题了,能帮帮我吗?

hello! I run tne test_ic15.py with C++ pse cversion: CUDA_VISIBLE_DEVICES='5' python test_ic15.py --scale 1 --resume /home/ltm/PSENet-master/ic15_res50.pth

then: make: Entering directory /home/ltm/PSENet-master/PSENet-master/pse' make:adaptor.so' is up to date. make: Leaving directory `/home/ltm/PSENet-master/PSENet-master/pse' Loading model and optimizer from checkpoint '/home/ltm/PSENet-master/ic15_res50.pth' Loaded checkpoint '/home/ltm/PSENet-master/ic15_res50.pth' (epoch 600) progress: 0 / 500 /usr/local/python3/lib/python3.6/site-packages/torch/nn/functional.py:2351: UserWarning: nn.functional.upsample is deprecated. Use nn.functional.interpolate instead. warnings.warn("nn.functional.upsample is deprecated. Use nn.functional.interpolate instead.") /usr/local/python3/lib/python3.6/site-packages/torch/nn/functional.py:2423: UserWarning: Default upsampling behavior when mode=bilinear is changed to align_corners=False since 0.4.0. Please specify align_corners=True if the old behavior is desired. See the documentation of nn.Upsample for details. "See the documentation of nn.Upsample for details.".format(mode)) Segmentation fault

besides，the python pse version is ok

please help, thanks

作者您好，感谢您的工作，这里有个问题想请问一下：

请问为什么在训练时用的是kernels = outputs[:, 1:, :, :]，具体体现在train_ic15.py 的122行，

而测试时用的是kernels = outputs[:, 0:args.kernel_num, :, :] * text，具体体现在test_ic15.py的132行，

那岂不是测试时outputs的最后一个channel完全没有用到？

配置好框架之后 train.py训练 报这个错 请问大家 我是哪里还没配好吗 File "train.py", line 213, in main(args) File "train.py", line 164, in main model = build_model(cfg.model) File "/home/liyanyan/PSE_origial/models/builder.py", line 11, in build_model model = models.dictcfg.type File "/home/liyanyan/PSE_origial/models/psenet.py", line 20, in init self.backbone = build_backbone(backbone) File "/home/liyanyan/PSE_origial/models/backbone/builder.py", line 11, in build_backbone backbone = models.backbone.dictcfg.type File "/home/liyanyan/PSE_origial/models/backbone/resnet.py", line 210, in resnet50 model.load_state_dict(load_url(model_urls['resnet50']), strict=False) File "/home/liyanyan/PSE_origial/models/backbone/resnet.py", line 234, in load_url return torch.load(cached_file, map_location=map_location) File "/home/liyanyan/.conda/envs/PSE-env/lib/python3.7/site-packages/torch/serialization.py", line 387, in load return _load(f, map_location, pickle_module, **pickle_load_args) File "/home/liyanyan/.conda/envs/PSE-env/lib/python3.7/site-packages/torch/serialization.py", line 564, in _load magic_number = pickle_module.load(f, **pickle_load_args) _pickle.UnpicklingError: invalid load key, '\xef'.

我按照readme文件中的指令训练完成后使用script.py文件进行精度验证，发现精度始终为0 Calculated!{"precision": 0.0, "recall": 0.0, "hmean": 0, "AP": 0} 查看网络输出后发现输出的res_img_xx.txt文件内，只有坐标，没有transcription 而计算精度时又需要用到transcription，所以精度一直为0.请问怎么解决这个问题？

score = torch.sigmoid(out[:, 0, :, :]) # out = (torch.sign(out - 1) + 1) / 2 # 0 1 # # text_mask = out[:, 0, :, :] # kernels = out[:, 1:cfg.test_cfg.kernel_num, :, :] * text_mask

    kernels = out[:, :cfg.test_cfg.kernel_num, :, :] > 0
    text_mask = kernels[:, :1, :, :]
    kernels[:, 1:, :, :] = kernels[:, 1:, :, :] * text_mask