lanpa/tensorboardX

slider bar's values jumps when moving the bar's knob

ZhengRui opened this issue · 7 comments

I used add_image() to record some visualization results at each step/epoch, during training the image at each step/epoch is displayed well, but after the training is done, when i draw the knob left and right, it doesn't increase or decrease incrementally by step 1, instead step numbers can just goes like 1->2->4->6->12 ..., which is annoying, my college also has the same issue. I guess it is a front-end issue, can anyone post a fix? Thanks.

lanpa commented

The Tweaks section in readme may help.

I tried to change event_accumulator.TENSORS from 10 to 100, it seems has no effects, is it the right place that I should modify? As from search of my /usr/local/lib/python2.7/dist-packages/tensorboard/backend/application.py file, I only found event_accumulator in the event_accumulator.TENSORS setting.

lanpa commented

which version of TB are you using? I am on 0.1.7

I am on 0.1.8

lanpa commented

HI, I just tried 0.1.8 and found that by also modifying image_metadata in tensorboard/backend/application.py solves the problem.

DEFAULT_SIZE_GUIDANCE = {
    event_accumulator.TENSORS: 100,
}

# TODO(@wchargin): Once SQL mode is in play, replace this with an
# alternative that does not privilege first-party plugins.
DEFAULT_TENSOR_SIZE_GUIDANCE = {
    scalar_metadata.PLUGIN_NAME: 1000,
    image_metadata.PLUGIN_NAME: 100,
    audio_metadata.PLUGIN_NAME: 10,
    histogram_metadata.PLUGIN_NAME: 500,
}

awesome, that works, thanks, close this issue now.

I use the 'domo.py', but the PROJECTOR is not working.Why?
微信图片_20191120014746