How to generate the train dataset?

[LearnerInGithub]

@zxytim @argman Hello, thanks for your sharing! Now I am curious about the format of the train dataset, so could provide a brief descriptions about it? Thanks very much!

[argman]

see in training samples

[LearnerInGithub]

@argman @zxytim Why the code always stuck at time.sleep(0.01) in the def get_batch(num_workers=10, **kwargs): after read through the the images?

[argman]

if you dont have enough cpu-cores, you should reduce the num_readers to a suitable number.

[LearnerInGithub]

@argman Even I change the readers number to 4(I am sure I have at least 28 free cores), the stuck problem still there, below is the part of running log: image_list.shape:(1000,) 1000 training images in ICDAR2015 enqueuer.queue isnot empty, sleep 0.01 enqueuer.queue isnot empty, sleep 0.01 enqueuer.queue isnot empty, sleep 0.01 . . . . enqueuer.queue isnot empty, sleep 0.01

[argman]

I dont encounter your problem, maybe it's the version of keras, I use keras to provide mulithread what's your keras version ?

[LearnerInGithub]

@argman My keras version is 2.0.0, and which version of keras you used?

[argman]

I've updated the code, no need to install keras now.

[LearnerInGithub]

@argman Now I faced new problem while training, the tips is warning, the warning info: icdar.py:210: RuntimeWarning: divide by zero encountered in float_scalars theta = np.arctan((poly[3][0] - poly[0][0]) / (poly[3][1] - poly[0][1])) icdar.py:216: RuntimeWarning: divide by zero encountered in float_scalars theta = np.arctan((poly[2][0] - poly[1][0]) / (poly[2][1] - poly[1][1]))

[argman]

its caused by some small polygons, you can ignore it.

[zxytim]

@LearnerInGithub The newest version is using np.arctan2, the problem should be fixed.

[rmmal]

@LearnerInGithub how did you solve the problem of being stucked in sleep(0.01) ??

[yangxuehang]

@LearnerInGithub @rmmal how did you solve the problem of being stucked in sleep(0.01) ??