Train on RCTW dataset

[Jacqueline720]

Has anybody trained pixel_link on the RCTW-17 from scratch? When I trained this dataset on pixel_link,the model can not converge:when I adjust learning rate as 1e-2,the loss goes to nan,while the loss goes to 0 when learning rate = 1e-3. Hope anybody have any idea about this problem~

[HoracceFeng]

Hi @Jacqueline720, have you solved this problem? In the source code, the image will be cropped to 512 while the images in RCTW dataset are dramatically larger, like 2000-3000 per edge. I have done some modification to make the whole image resize and put into training. But then I get the 'nan loss' once lr=1e-3.

Hi author @dengdan, would you give me some suggestions? Thx

[HoracceFeng]

Problem Solved. Many Questions are actually because of the TFrecords converter and mismatch of tf.summary.image(bounding_box). Will do a pipeline code for text segmentation later.

[yhpwsid]

Hi @Jacqueline720, have you solved this problem? In the source code, the image will be cropped to 512 while the images in RCTW dataset are dramatically larger, like 2000-3000 per edge. I have done some modification to make the whole image resize and put into training. But then I get the 'nan loss' once lr=1e-3.

Hi author @dengdan, would you give me some suggestions? Thx does the loss have anything to do with tf.summary.image? how did you solve the loss problem? thanks

[xiiiiiiii]

Has anybody trained pixel_link on the RCTW-17 from scratch? When I trained this dataset on pixel_link,the model can not converge:when I adjust learning rate as 1e-2,the loss goes to nan,while the loss goes to 0 when learning rate = 1e-3. Hope anybody have any idea about this problem~

你好，想问一下你在RCTW-17数据集上准确率召回率是多少，表现如何